freebsd-skq/sys/netinet/tcp_input.c

3828 lines
109 KiB
C
Raw Normal View History

/*-
* Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
1994-05-24 10:09:53 +00:00
* The Regents of the University of California. All rights reserved.
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
* Copyright (c) 2007-2008,2010
* Swinburne University of Technology, Melbourne, Australia.
* Copyright (c) 2009-2010 Lawrence Stewart <lstewart@freebsd.org>
* Copyright (c) 2010 The FreeBSD Foundation
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
* Copyright (c) 2010-2011 Juniper Networks, Inc.
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
* All rights reserved.
*
* Portions of this software were developed at the Centre for Advanced Internet
* Architectures, Swinburne University of Technology, by Lawrence Stewart,
* James Healy and David Hayes, made possible in part by a grant from the Cisco
* University Research Program Fund at Community Foundation Silicon Valley.
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
*
* Portions of this software were developed at the Centre for Advanced
* Internet Architectures, Swinburne University of Technology, Melbourne,
* Australia by David Hayes under sponsorship from the FreeBSD Foundation.
1994-05-24 10:09:53 +00:00
*
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
* Portions of this software were developed by Robert N. M. Watson under
* contract to Juniper Networks, Inc.
*
1994-05-24 10:09:53 +00:00
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 4. Neither the name of the University nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
* @(#)tcp_input.c 8.12 (Berkeley) 5/24/95
1994-05-24 10:09:53 +00:00
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_ipfw.h" /* for ipfw_fwd */
Initial import of RFC 2385 (TCP-MD5) digest support. This is the first of two commits; bringing in the kernel support first. This can be enabled by compiling a kernel with options TCP_SIGNATURE and FAST_IPSEC. For the uninitiated, this is a TCP option which provides for a means of authenticating TCP sessions which came into being before IPSEC. It is still relevant today, however, as it is used by many commercial router vendors, particularly with BGP, and as such has become a requirement for interconnect at many major Internet points of presence. Several parts of the TCP and IP headers, including the segment payload, are digested with MD5, including a shared secret. The PF_KEY interface is used to manage the secrets using security associations in the SADB. There is a limitation here in that as there is no way to map a TCP flow per-port back to an SPI without polluting tcpcb or using the SPD; the code to do the latter is unstable at this time. Therefore this code only supports per-host keying granularity. Whilst FAST_IPSEC is mutually exclusive with KAME IPSEC (and thus IPv6), TCP_SIGNATURE applies only to IPv4. For the vast majority of prospective users of this feature, this will not pose any problem. This implementation is output-only; that is, the option is honoured when responding to a host initiating a TCP session, but no effort is made [yet] to authenticate inbound traffic. This is, however, sufficient to interwork with Cisco equipment. Tested with a Cisco 2501 running IOS 12.0(27), and Quagga 0.96.4 with local patches. Patches for tcpdump to validate TCP-MD5 sessions are also available from me upon request. Sponsored by: sentex.net
2004-02-11 04:26:04 +00:00
#include "opt_inet.h"
#include "opt_inet6.h"
#include "opt_ipsec.h"
1997-09-16 18:36:06 +00:00
#include "opt_tcpdebug.h"
1994-05-24 10:09:53 +00:00
#include <sys/param.h>
#include <sys/kernel.h>
#include <sys/hhook.h>
1994-05-24 10:09:53 +00:00
#include <sys/malloc.h>
#include <sys/mbuf.h>
#include <sys/proc.h> /* for proc0 declaration */
1994-05-24 10:09:53 +00:00
#include <sys/protosw.h>
#include <sys/sdt.h>
#include <sys/signalvar.h>
1994-05-24 10:09:53 +00:00
#include <sys/socket.h>
#include <sys/socketvar.h>
#include <sys/sysctl.h>
#include <sys/syslog.h>
#include <sys/systm.h>
1994-05-24 10:09:53 +00:00
#include <machine/cpu.h> /* before tcp_seq.h, for tcp_random18() */
#include <vm/uma.h>
1994-05-24 10:09:53 +00:00
#include <net/if.h>
#include <net/if_var.h>
1994-05-24 10:09:53 +00:00
#include <net/route.h>
#include <net/vnet.h>
1994-05-24 10:09:53 +00:00
#define TCPSTATES /* for logging */
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
#include <netinet/cc.h>
1994-05-24 10:09:53 +00:00
#include <netinet/in.h>
#include <netinet/in_kdtrace.h>
#include <netinet/in_pcb.h>
1994-05-24 10:09:53 +00:00
#include <netinet/in_systm.h>
#include <netinet/in_var.h>
1994-05-24 10:09:53 +00:00
#include <netinet/ip.h>
#include <netinet/ip_icmp.h> /* required for icmp_var.h */
#include <netinet/icmp_var.h> /* for ICMP_BANDLIM */
1994-05-24 10:09:53 +00:00
#include <netinet/ip_var.h>
#include <netinet/ip_options.h>
#include <netinet/ip6.h>
#include <netinet/icmp6.h>
#include <netinet6/in6_pcb.h>
#include <netinet6/ip6_var.h>
#include <netinet6/nd6.h>
#ifdef TCP_RFC7413
#include <netinet/tcp_fastopen.h>
#endif
1994-05-24 10:09:53 +00:00
#include <netinet/tcp_fsm.h>
#include <netinet/tcp_seq.h>
#include <netinet/tcp_timer.h>
#include <netinet/tcp_var.h>
#include <netinet6/tcp6_var.h>
1994-05-24 10:09:53 +00:00
#include <netinet/tcpip.h>
There are times when it would be really nice to have a record of the last few packets and/or state transitions from each TCP socket. That would help with narrowing down certain problems we see in the field that are hard to reproduce without understanding the history of how we got into a certain state. This change provides just that. It saves copies of the last N packets in a list in the tcpcb. When the tcpcb is destroyed, the list is freed. I thought this was likely to be more performance-friendly than saving copies of the tcpcb. Plus, with the packets, you should be able to reverse-engineer what happened to the tcpcb. To enable the feature, you will need to compile a kernel with the TCPPCAP option. Even then, the feature defaults to being deactivated. You can activate it by setting a positive value for the number of captured packets. You can do that on either a global basis or on a per-socket basis (via a setsockopt call). There is no way to get the packets out of the kernel other than using kmem or getting a coredump. I thought that would help some of the legal/privacy concerns regarding such a feature. However, it should be possible to add a future effort to export them in PCAP format. I tested this at low scale, and found that there were no mbuf leaks and the peak mbuf usage appeared to be unchanged with and without the feature. The main performance concern I can envision is the number of mbufs that would be used on systems with a large number of sockets. If you save five packets per direction per socket and have 3,000 sockets, that will consume at least 30,000 mbufs just to keep these packets. I tried to reduce the concerns associated with this by limiting the number of clusters (not mbufs) that could be used for this feature. Again, in my testing, that appears to work correctly. Differential Revision: D3100 Submitted by: Jonathan Looney <jlooney at juniper dot net> Reviewed by: gnn, hiren
2015-10-14 00:35:37 +00:00
#ifdef TCPPCAP
#include <netinet/tcp_pcap.h>
#endif
#include <netinet/tcp_syncache.h>
#ifdef TCPDEBUG
1994-05-24 10:09:53 +00:00
#include <netinet/tcp_debug.h>
#endif /* TCPDEBUG */
#ifdef TCP_OFFLOAD
#include <netinet/tcp_offload.h>
#endif
#ifdef IPSEC
#include <netipsec/ipsec.h>
#include <netipsec/ipsec6.h>
#endif /*IPSEC*/
#include <machine/in_cksum.h>
#include <security/mac/mac_framework.h>
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
const int tcprexmtthresh = 3;
int tcp_log_in_vain = 0;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&tcp_log_in_vain, 0,
"Log all incoming TCP segments to closed ports");
VNET_DEFINE(int, blackhole) = 0;
#define V_blackhole VNET(blackhole)
SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(blackhole), 0,
"Do not send RST on segments to closed ports");
VNET_DEFINE(int, tcp_delack_enabled) = 1;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_delack_enabled), 0,
"Delay ACK to try and piggyback it onto a data packet");
VNET_DEFINE(int, drop_synfin) = 0;
#define V_drop_synfin VNET(drop_synfin)
SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(drop_synfin), 0,
"Drop TCP packets with SYN+FIN set");
Calculate the correct amount of bytes that are in-flight for a connection as suggested by RFC 6675. Currently differnt places in the stack tries to guess this in suboptimal ways. The main problem is that current calculations don't take sacked bytes into account. Sacked bytes are the bytes receiver acked via SACK option. This is suboptimal because it assumes that network has more outstanding (unacked) bytes than the actual value and thus sends less data by setting congestion window lower than what's possible which in turn may cause slower recovery from losses. As an example, one of the current calculations looks something like this: snd_nxt - snd_fack + sackhint.sack_bytes_rexmit New proposal from RFC 6675 is: snd_max - snd_una - sackhint.sacked_bytes + sackhint.sack_bytes_rexmit which takes sacked bytes into account which is a new addition to the sackhint struct. Only thing we are missing from RFC 6675 is isLost() i.e. segment being considered lost and thus adjusting pipe based on that which makes this calculation a bit on conservative side. The approach is very simple. We already process each ack with sack info in tcp_sack_doack() and extract sack blocks/holes out of it. We'd now also track this new variable sacked_bytes which keeps track of total sacked bytes reported. One downside to this approach is that we may get incorrect count of sacked_bytes if the other end decides to drop sack info in the ack because of memory pressure or some other reasons. But in this (not very likely) case also the pipe calculation would be conservative which is okay as opposed to being aggressive in sending packets into the network. Next step is to use this more accurate pipe estimation to drive congestion window adjustments. In collaboration with: rrs Reviewed by: jason_eggnet dot com, rrs MFC after: 2 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D3971
2015-10-28 22:57:51 +00:00
VNET_DEFINE(int, tcp_do_rfc6675_pipe) = 0;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc6675_pipe, CTLFLAG_VNET | CTLFLAG_RW,
Calculate the correct amount of bytes that are in-flight for a connection as suggested by RFC 6675. Currently differnt places in the stack tries to guess this in suboptimal ways. The main problem is that current calculations don't take sacked bytes into account. Sacked bytes are the bytes receiver acked via SACK option. This is suboptimal because it assumes that network has more outstanding (unacked) bytes than the actual value and thus sends less data by setting congestion window lower than what's possible which in turn may cause slower recovery from losses. As an example, one of the current calculations looks something like this: snd_nxt - snd_fack + sackhint.sack_bytes_rexmit New proposal from RFC 6675 is: snd_max - snd_una - sackhint.sacked_bytes + sackhint.sack_bytes_rexmit which takes sacked bytes into account which is a new addition to the sackhint struct. Only thing we are missing from RFC 6675 is isLost() i.e. segment being considered lost and thus adjusting pipe based on that which makes this calculation a bit on conservative side. The approach is very simple. We already process each ack with sack info in tcp_sack_doack() and extract sack blocks/holes out of it. We'd now also track this new variable sacked_bytes which keeps track of total sacked bytes reported. One downside to this approach is that we may get incorrect count of sacked_bytes if the other end decides to drop sack info in the ack because of memory pressure or some other reasons. But in this (not very likely) case also the pipe calculation would be conservative which is okay as opposed to being aggressive in sending packets into the network. Next step is to use this more accurate pipe estimation to drive congestion window adjustments. In collaboration with: rrs Reviewed by: jason_eggnet dot com, rrs MFC after: 2 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D3971
2015-10-28 22:57:51 +00:00
&VNET_NAME(tcp_do_rfc6675_pipe), 0,
"Use calculated pipe/in-flight bytes per RFC 6675");
VNET_DEFINE(int, tcp_do_rfc3042) = 1;
#define V_tcp_do_rfc3042 VNET(tcp_do_rfc3042)
SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3042, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_do_rfc3042), 0,
"Enable RFC 3042 (Limited Transmit)");
VNET_DEFINE(int, tcp_do_rfc3390) = 1;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_do_rfc3390), 0,
"Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
VNET_DEFINE(int, tcp_initcwnd_segments) = 10;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, initcwnd_segments,
CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_initcwnd_segments), 0,
"Slow-start flight size (initial congestion window) in number of segments");
VNET_DEFINE(int, tcp_do_rfc3465) = 1;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3465, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_do_rfc3465), 0,
"Enable RFC 3465 (Appropriate Byte Counting)");
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
VNET_DEFINE(int, tcp_abc_l_var) = 2;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc_l_var, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_abc_l_var), 2,
"Cap the max cwnd increment during slow-start to this number of segments");
static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, ecn, CTLFLAG_RW, 0, "TCP ECN");
VNET_DEFINE(int, tcp_do_ecn) = 0;
SYSCTL_INT(_net_inet_tcp_ecn, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_do_ecn), 0,
"TCP ECN support");
VNET_DEFINE(int, tcp_ecn_maxretries) = 1;
SYSCTL_INT(_net_inet_tcp_ecn, OID_AUTO, maxretries, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_ecn_maxretries), 0,
"Max retries before giving up on ECN");
FreeBSD-SA-14:19.tcp raised attention to the state of our stack towards blind SYN/RST spoofed attack. Originally our stack used in-window checks for incoming SYN/RST as proposed by RFC793. Later, circa 2003 the RST attack was mitigated using the technique described in P. Watson "Slipping in the window" paper [1]. After that, the checks were only relaxed for the sake of compatibility with some buggy TCP stacks. First, r192912 introduced the vulnerability, just fixed by aforementioned SA. Second, r167310 had slightly relaxed the default RST checks, instead of utilizing net.inet.tcp.insecure_rst sysctl. In 2010 a new technique for mitigation of these attacks was proposed in RFC5961 [2]. The idea is to send a "challenge ACK" packet to the peer, to verify that packet arrived isn't spoofed. If peer receives challenge ACK it should regenerate its RST or SYN with correct sequence number. This should not only protect against attacks, but also improve communication with broken stacks, so authors of reverted r167310 and r192912 won't be disappointed. [1] http://bandwidthco.com/whitepapers/netforensics/tcpip/TCP Reset Attacks.pdf [2] http://www.rfc-editor.org/rfc/rfc5961.txt Changes made: o Revert r167310. o Implement "challenge ACK" protection as specificed in RFC5961 against RST attack. On by default. - Carefully preserve r138098, which handles empty window edge case, not described by the RFC. - Update net.inet.tcp.insecure_rst description. o Implement "challenge ACK" protection as specificed in RFC5961 against SYN attack. On by default. - Provide net.inet.tcp.insecure_syn sysctl, to turn off RFC5961 protection. The changes were tested at Netflix. The tested box didn't show any anomalies compared to control box, except slightly increased number of TCP connection in LAST_ACK state. Reviewed by: rrs Sponsored by: Netflix Sponsored by: Nginx, Inc.
2014-09-16 11:07:25 +00:00
VNET_DEFINE(int, tcp_insecure_syn) = 0;
#define V_tcp_insecure_syn VNET(tcp_insecure_syn)
SYSCTL_INT(_net_inet_tcp, OID_AUTO, insecure_syn, CTLFLAG_VNET | CTLFLAG_RW,
FreeBSD-SA-14:19.tcp raised attention to the state of our stack towards blind SYN/RST spoofed attack. Originally our stack used in-window checks for incoming SYN/RST as proposed by RFC793. Later, circa 2003 the RST attack was mitigated using the technique described in P. Watson "Slipping in the window" paper [1]. After that, the checks were only relaxed for the sake of compatibility with some buggy TCP stacks. First, r192912 introduced the vulnerability, just fixed by aforementioned SA. Second, r167310 had slightly relaxed the default RST checks, instead of utilizing net.inet.tcp.insecure_rst sysctl. In 2010 a new technique for mitigation of these attacks was proposed in RFC5961 [2]. The idea is to send a "challenge ACK" packet to the peer, to verify that packet arrived isn't spoofed. If peer receives challenge ACK it should regenerate its RST or SYN with correct sequence number. This should not only protect against attacks, but also improve communication with broken stacks, so authors of reverted r167310 and r192912 won't be disappointed. [1] http://bandwidthco.com/whitepapers/netforensics/tcpip/TCP Reset Attacks.pdf [2] http://www.rfc-editor.org/rfc/rfc5961.txt Changes made: o Revert r167310. o Implement "challenge ACK" protection as specificed in RFC5961 against RST attack. On by default. - Carefully preserve r138098, which handles empty window edge case, not described by the RFC. - Update net.inet.tcp.insecure_rst description. o Implement "challenge ACK" protection as specificed in RFC5961 against SYN attack. On by default. - Provide net.inet.tcp.insecure_syn sysctl, to turn off RFC5961 protection. The changes were tested at Netflix. The tested box didn't show any anomalies compared to control box, except slightly increased number of TCP connection in LAST_ACK state. Reviewed by: rrs Sponsored by: Netflix Sponsored by: Nginx, Inc.
2014-09-16 11:07:25 +00:00
&VNET_NAME(tcp_insecure_syn), 0,
"Follow RFC793 instead of RFC5961 criteria for accepting SYN packets");
VNET_DEFINE(int, tcp_insecure_rst) = 0;
#define V_tcp_insecure_rst VNET(tcp_insecure_rst)
SYSCTL_INT(_net_inet_tcp, OID_AUTO, insecure_rst, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_insecure_rst), 0,
FreeBSD-SA-14:19.tcp raised attention to the state of our stack towards blind SYN/RST spoofed attack. Originally our stack used in-window checks for incoming SYN/RST as proposed by RFC793. Later, circa 2003 the RST attack was mitigated using the technique described in P. Watson "Slipping in the window" paper [1]. After that, the checks were only relaxed for the sake of compatibility with some buggy TCP stacks. First, r192912 introduced the vulnerability, just fixed by aforementioned SA. Second, r167310 had slightly relaxed the default RST checks, instead of utilizing net.inet.tcp.insecure_rst sysctl. In 2010 a new technique for mitigation of these attacks was proposed in RFC5961 [2]. The idea is to send a "challenge ACK" packet to the peer, to verify that packet arrived isn't spoofed. If peer receives challenge ACK it should regenerate its RST or SYN with correct sequence number. This should not only protect against attacks, but also improve communication with broken stacks, so authors of reverted r167310 and r192912 won't be disappointed. [1] http://bandwidthco.com/whitepapers/netforensics/tcpip/TCP Reset Attacks.pdf [2] http://www.rfc-editor.org/rfc/rfc5961.txt Changes made: o Revert r167310. o Implement "challenge ACK" protection as specificed in RFC5961 against RST attack. On by default. - Carefully preserve r138098, which handles empty window edge case, not described by the RFC. - Update net.inet.tcp.insecure_rst description. o Implement "challenge ACK" protection as specificed in RFC5961 against SYN attack. On by default. - Provide net.inet.tcp.insecure_syn sysctl, to turn off RFC5961 protection. The changes were tested at Netflix. The tested box didn't show any anomalies compared to control box, except slightly increased number of TCP connection in LAST_ACK state. Reviewed by: rrs Sponsored by: Netflix Sponsored by: Nginx, Inc.
2014-09-16 11:07:25 +00:00
"Follow RFC793 instead of RFC5961 criteria for accepting RST packets");
VNET_DEFINE(int, tcp_recvspace) = 1024*64;
#define V_tcp_recvspace VNET(tcp_recvspace)
SYSCTL_INT(_net_inet_tcp, TCPCTL_RECVSPACE, recvspace, CTLFLAG_VNET | CTLFLAG_RW,
&VNET_NAME(tcp_recvspace), 0, "Initial receive socket buffer size");
VNET_DEFINE(int, tcp_do_autorcvbuf) = 1;
#define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf)
SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_do_autorcvbuf), 0,
"Enable automatic receive buffer sizing");
VNET_DEFINE(int, tcp_autorcvbuf_inc) = 16*1024;
#define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc)
SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_autorcvbuf_inc), 0,
2007-03-19 19:00:51 +00:00
"Incrementor step size of automatic receive buffer");
VNET_DEFINE(int, tcp_autorcvbuf_max) = 2*1024*1024;
#define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max)
SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
&VNET_NAME(tcp_autorcvbuf_max), 0,
"Max size of automatic receive buffer");
Build on Jeff Roberson's linker-set based dynamic per-CPU allocator (DPCPU), as suggested by Peter Wemm, and implement a new per-virtual network stack memory allocator. Modify vnet to use the allocator instead of monolithic global container structures (vinet, ...). This change solves many binary compatibility problems associated with VIMAGE, and restores ELF symbols for virtualized global variables. Each virtualized global variable exists as a "reference copy", and also once per virtual network stack. Virtualized global variables are tagged at compile-time, placing the in a special linker set, which is loaded into a contiguous region of kernel memory. Virtualized global variables in the base kernel are linked as normal, but those in modules are copied and relocated to a reserved portion of the kernel's vnet region with the help of a the kernel linker. Virtualized global variables exist in per-vnet memory set up when the network stack instance is created, and are initialized statically from the reference copy. Run-time access occurs via an accessor macro, which converts from the current vnet and requested symbol to a per-vnet address. When "options VIMAGE" is not compiled into the kernel, normal global ELF symbols will be used instead and indirection is avoided. This change restores static initialization for network stack global variables, restores support for non-global symbols and types, eliminates the need for many subsystem constructors, eliminates large per-subsystem structures that caused many binary compatibility issues both for monitoring applications (netstat) and kernel modules, removes the per-function INIT_VNET_*() macros throughout the stack, eliminates the need for vnet_symmap ksym(2) munging, and eliminates duplicate definitions of virtualized globals under VIMAGE_GLOBALS. Bump __FreeBSD_version and update UPDATING. Portions submitted by: bz Reviewed by: bz, zec Discussed with: gnn, jamie, jeff, jhb, julian, sam Suggested by: peter Approved by: re (kensmith)
2009-07-14 22:48:30 +00:00
VNET_DEFINE(struct inpcbhead, tcb);
#define tcb6 tcb /* for KAME src sync over BSD*'s */
VNET_DEFINE(struct inpcbinfo, tcbinfo);
1994-05-24 10:09:53 +00:00
/*
* TCP statistics are stored in an "array" of counter(9)s.
*/
VNET_PCPUSTAT_DEFINE(struct tcpstat, tcpstat);
VNET_PCPUSTAT_SYSINIT(tcpstat);
SYSCTL_VNET_PCPUSTAT(_net_inet_tcp, TCPCTL_STATS, stats, struct tcpstat,
tcpstat, "TCP statistics (struct tcpstat, netinet/tcp_var.h)");
#ifdef VIMAGE
VNET_PCPUSTAT_SYSUNINIT(tcpstat);
#endif /* VIMAGE */
/*
* Kernel module interface for updating tcpstat. The argument is an index
* into tcpstat treated as an array.
*/
void
kmod_tcpstat_inc(int statnum)
{
counter_u64_add(VNET(tcpstat)[statnum], 1);
}
/*
* Wrapper for the TCP established input helper hook.
*/
void
hhook_run_tcp_est_in(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to)
{
struct tcp_hhook_data hhook_data;
if (V_tcp_hhh[HHOOK_TCP_EST_IN]->hhh_nhooks > 0) {
hhook_data.tp = tp;
hhook_data.th = th;
hhook_data.to = to;
hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_IN], &hhook_data,
tp->osd);
}
}
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
/*
* CC wrapper hook functions
*/
void
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
cc_ack_received(struct tcpcb *tp, struct tcphdr *th, uint16_t type)
{
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
INP_WLOCK_ASSERT(tp->t_inpcb);
tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th);
if (tp->snd_cwnd <= tp->snd_wnd)
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
tp->ccv->flags |= CCF_CWND_LIMITED;
else
tp->ccv->flags &= ~CCF_CWND_LIMITED;
if (type == CC_ACK) {
if (tp->snd_cwnd > tp->snd_ssthresh) {
tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
V_tcp_abc_l_var * tcp_maxseg(tp));
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
if (tp->t_bytes_acked >= tp->snd_cwnd) {
tp->t_bytes_acked -= tp->snd_cwnd;
tp->ccv->flags |= CCF_ABC_SENTAWND;
}
} else {
tp->ccv->flags &= ~CCF_ABC_SENTAWND;
tp->t_bytes_acked = 0;
}
}
if (CC_ALGO(tp)->ack_received != NULL) {
/* XXXLAS: Find a way to live without this */
tp->ccv->curack = th->th_ack;
CC_ALGO(tp)->ack_received(tp->ccv, type);
}
}
void
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
cc_conn_init(struct tcpcb *tp)
{
struct hc_metrics_lite metrics;
struct inpcb *inp = tp->t_inpcb;
u_int maxseg;
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
int rtt;
INP_WLOCK_ASSERT(tp->t_inpcb);
tcp_hc_get(&inp->inp_inc, &metrics);
maxseg = tcp_maxseg(tp);
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
if (tp->t_srtt == 0 && (rtt = metrics.rmx_rtt)) {
tp->t_srtt = rtt;
tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE;
TCPSTAT_INC(tcps_usedrtt);
if (metrics.rmx_rttvar) {
tp->t_rttvar = metrics.rmx_rttvar;
TCPSTAT_INC(tcps_usedrttvar);
} else {
/* default variation is +- 1 rtt */
tp->t_rttvar =
tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE;
}
TCPT_RANGESET(tp->t_rxtcur,
((tp->t_srtt >> 2) + tp->t_rttvar) >> 1,
tp->t_rttmin, TCPTV_REXMTMAX);
}
if (metrics.rmx_ssthresh) {
/*
* There's some sort of gateway or interface
* buffer limit on the path. Use this to set
* the slow start threshhold, but set the
* threshold to no less than 2*mss.
*/
tp->snd_ssthresh = max(2 * maxseg, metrics.rmx_ssthresh);
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
TCPSTAT_INC(tcps_usedssthresh);
}
/*
* Set the initial slow-start flight size.
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
*
* RFC5681 Section 3.1 specifies the default conservative values.
* RFC3390 specifies slightly more aggressive values.
* RFC6928 increases it to ten segments.
* Support for user specified value for initial flight size.
*
* If a SYN or SYN/ACK was lost and retransmitted, we have to
* reduce the initial CWND to one segment as congestion is likely
* requiring us to be cautious.
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
*/
if (tp->snd_cwnd == 1)
tp->snd_cwnd = maxseg; /* SYN(-ACK) lost */
else if (V_tcp_initcwnd_segments)
tp->snd_cwnd = min(V_tcp_initcwnd_segments * maxseg,
max(2 * maxseg, V_tcp_initcwnd_segments * 1460));
else if (V_tcp_do_rfc3390)
tp->snd_cwnd = min(4 * maxseg, max(2 * maxseg, 4380));
else {
/* Per RFC5681 Section 3.1 */
if (maxseg > 2190)
tp->snd_cwnd = 2 * maxseg;
else if (maxseg > 1095)
tp->snd_cwnd = 3 * maxseg;
else
tp->snd_cwnd = 4 * maxseg;
}
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
if (CC_ALGO(tp)->conn_init != NULL)
CC_ALGO(tp)->conn_init(tp->ccv);
}
void inline
cc_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type)
{
u_int maxseg;
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
INP_WLOCK_ASSERT(tp->t_inpcb);
switch(type) {
case CC_NDUPACK:
if (!IN_FASTRECOVERY(tp->t_flags)) {
tp->snd_recover = tp->snd_max;
if (tp->t_flags & TF_ECN_PERMIT)
tp->t_flags |= TF_ECN_SND_CWR;
}
break;
case CC_ECN:
if (!IN_CONGRECOVERY(tp->t_flags)) {
TCPSTAT_INC(tcps_ecn_rcwnd);
tp->snd_recover = tp->snd_max;
if (tp->t_flags & TF_ECN_PERMIT)
tp->t_flags |= TF_ECN_SND_CWR;
}
break;
case CC_RTO:
maxseg = tcp_maxseg(tp);
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
tp->t_dupacks = 0;
tp->t_bytes_acked = 0;
EXIT_RECOVERY(tp->t_flags);
tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 /
maxseg) * maxseg;
tp->snd_cwnd = maxseg;
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
break;
case CC_RTO_ERR:
TCPSTAT_INC(tcps_sndrexmitbad);
/* RTO was unnecessary, so reset everything. */
tp->snd_cwnd = tp->snd_cwnd_prev;
tp->snd_ssthresh = tp->snd_ssthresh_prev;
tp->snd_recover = tp->snd_recover_prev;
if (tp->t_flags & TF_WASFRECOVERY)
ENTER_FASTRECOVERY(tp->t_flags);
if (tp->t_flags & TF_WASCRECOVERY)
ENTER_CONGRECOVERY(tp->t_flags);
tp->snd_nxt = tp->snd_max;
TCP reuses t_rxtshift to determine the backoff timer used for both the persist state and the retransmit timer. However, the code that implements "bad retransmit recovery" only checks t_rxtshift to see if an ACK has been received in during the first retransmit timeout window. As a result, if ticks has wrapped over to a negative value and a socket is in the persist state, it can incorrectly treat an ACK from the remote peer as a "bad retransmit recovery" and restore saved values such as snd_ssthresh and snd_cwnd. However, if the socket has never had a retransmit timeout, then these saved values will be zero, so snd_ssthresh and snd_cwnd will be set to 0. If the socket is in fast recovery (this can be caused by excessive duplicate ACKs such as those fixed by 220794), then each ACK that arrives triggers either NewReno or SACK partial ACK handling which clamps snd_cwnd to be no larger than snd_ssthresh. In effect, the socket's send window is permamently stuck at 0 even though the remote peer is advertising a much larger window and pending data is only sent via TCP window probes (so one byte every few seconds). Fix this by adding a new TCP pcb flag (TF_PREVVALID) that indicates that the various snd_*_prev fields in the pcb are valid and only perform "bad retransmit recovery" if this flag is set in the pcb. The flag is set on the first retransmit timeout that occurs and is cleared on subsequent retransmit timeouts or when entering the persist state. Reviewed by: bz MFC after: 2 weeks
2011-04-29 15:40:12 +00:00
tp->t_flags &= ~TF_PREVVALID;
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
tp->t_badrxtwin = 0;
break;
}
if (CC_ALGO(tp)->cong_signal != NULL) {
if (th != NULL)
tp->ccv->curack = th->th_ack;
CC_ALGO(tp)->cong_signal(tp->ccv, type);
}
}
void inline
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
cc_post_recovery(struct tcpcb *tp, struct tcphdr *th)
{
INP_WLOCK_ASSERT(tp->t_inpcb);
/* XXXLAS: KASSERT that we're in recovery? */
if (CC_ALGO(tp)->post_recovery != NULL) {
tp->ccv->curack = th->th_ack;
CC_ALGO(tp)->post_recovery(tp->ccv);
}
/* XXXLAS: EXIT_RECOVERY ? */
tp->t_bytes_acked = 0;
}
#ifdef TCP_SIGNATURE
static inline int
tcp_signature_verify_input(struct mbuf *m, int off0, int tlen, int optlen,
struct tcpopt *to, struct tcphdr *th, u_int tcpbflag)
{
int ret;
tcp_fields_to_net(th);
ret = tcp_signature_verify(m, off0, tlen, optlen, to, th, tcpbflag);
tcp_fields_to_host(th);
return (ret);
}
#endif
/*
* Indicate whether this ack should be delayed. We can delay the ack if
* following conditions are met:
* - There is no delayed ack timer in progress.
* - Our last ack wasn't a 0-sized window. We never want to delay
* the ack that opens up a 0-sized window.
* - LRO wasn't used for this segment. We make sure by checking that the
* segment size is not larger than the MSS.
*/
#define DELAY_ACK(tp, tlen) \
((!tcp_timer_active(tp, TT_DELACK) && \
(tp->t_flags & TF_RXWIN0SENT) == 0) && \
(tlen <= tp->t_maxseg) && \
(V_tcp_delack_enabled || (tp->t_flags & TF_NEEDSYN)))
static void inline
cc_ecnpkt_handler(struct tcpcb *tp, struct tcphdr *th, uint8_t iptos)
{
INP_WLOCK_ASSERT(tp->t_inpcb);
if (CC_ALGO(tp)->ecnpkt_handler != NULL) {
switch (iptos & IPTOS_ECN_MASK) {
case IPTOS_ECN_CE:
tp->ccv->flags |= CCF_IPHDR_CE;
break;
case IPTOS_ECN_ECT0:
tp->ccv->flags &= ~CCF_IPHDR_CE;
break;
case IPTOS_ECN_ECT1:
tp->ccv->flags &= ~CCF_IPHDR_CE;
break;
}
if (th->th_flags & TH_CWR)
tp->ccv->flags |= CCF_TCPHDR_CWR;
else
tp->ccv->flags &= ~CCF_TCPHDR_CWR;
if (tp->t_flags & TF_DELACK)
tp->ccv->flags |= CCF_DELACK;
else
tp->ccv->flags &= ~CCF_DELACK;
CC_ALGO(tp)->ecnpkt_handler(tp->ccv);
if (tp->ccv->flags & CCF_ACKNOW)
tcp_timer_activate(tp, TT_DELACK, tcp_delacktime);
}
}
1994-05-24 10:09:53 +00:00
/*
* TCP input handling is split into multiple parts:
* tcp6_input is a thin wrapper around tcp_input for the extended
* ip6_protox[] call format in ip6_input
* tcp_input handles primary segment validation, inpcb lookup and
* SYN processing on listen sockets
* tcp_do_segment processes the ACK and text of the segment for
* establishing, established and closing connections
1994-05-24 10:09:53 +00:00
*/
#ifdef INET6
int
tcp6_input(struct mbuf **mp, int *offp, int proto)
{
struct mbuf *m = *mp;
struct in6_ifaddr *ia6;
struct ip6_hdr *ip6;
IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), IPPROTO_DONE);
/*
* draft-itojun-ipv6-tcp-to-anycast
* better place to put this in?
*/
ip6 = mtod(m, struct ip6_hdr *);
ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */);
if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) {
struct ip6_hdr *ip6;
ifa_free(&ia6->ia_ifa);
ip6 = mtod(m, struct ip6_hdr *);
icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR,
(caddr_t)&ip6->ip6_dst - (caddr_t)ip6);
return (IPPROTO_DONE);
}
if (ia6)
ifa_free(&ia6->ia_ifa);
return (tcp_input(mp, offp, proto));
}
#endif /* INET6 */
int
tcp_input(struct mbuf **mp, int *offp, int proto)
1994-05-24 10:09:53 +00:00
{
struct mbuf *m = *mp;
struct tcphdr *th = NULL;
struct ip *ip = NULL;
struct inpcb *inp = NULL;
struct tcpcb *tp = NULL;
struct socket *so = NULL;
u_char *optp = NULL;
int off0;
int optlen = 0;
#ifdef INET
int len;
#endif
int tlen = 0, off;
int drop_hdrlen;
int thflags;
int rstreason = 0; /* For badport_bandlim accounting purposes */
#ifdef TCP_SIGNATURE
uint8_t sig_checked = 0;
#endif
uint8_t iptos = 0;
struct m_tag *fwd_tag = NULL;
#ifdef INET6
struct ip6_hdr *ip6 = NULL;
int isipv6;
#else
const void *ip6 = NULL;
#endif /* INET6 */
struct tcpopt to; /* options in this segment */
char *s = NULL; /* address and port logging */
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
int ti_locked;
#ifdef TCPDEBUG
/*
* The size of tcp_saveipgen must be the size of the max ip header,
* now IPv6.
*/
u_char tcp_saveipgen[IP6_HDR_LEN];
struct tcphdr tcp_savetcp;
short ostate = 0;
#endif
#ifdef INET6
isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? 1 : 0;
#endif
off0 = *offp;
m = *mp;
*mp = NULL;
to.to_flags = 0;
TCPSTAT_INC(tcps_rcvtotal);
#ifdef INET6
if (isipv6) {
/* IP6_EXTHDR_CHECK() is already done at tcp6_input(). */
if (m->m_len < (sizeof(*ip6) + sizeof(*th))) {
m = m_pullup(m, sizeof(*ip6) + sizeof(*th));
if (m == NULL) {
TCPSTAT_INC(tcps_rcvshort);
return (IPPROTO_DONE);
}
}
ip6 = mtod(m, struct ip6_hdr *);
th = (struct tcphdr *)((caddr_t)ip6 + off0);
tlen = sizeof(*ip6) + ntohs(ip6->ip6_plen) - off0;
if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) {
if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
th->th_sum = m->m_pkthdr.csum_data;
else
th->th_sum = in6_cksum_pseudo(ip6, tlen,
IPPROTO_TCP, m->m_pkthdr.csum_data);
th->th_sum ^= 0xffff;
} else
th->th_sum = in6_cksum(m, IPPROTO_TCP, off0, tlen);
if (th->th_sum) {
TCPSTAT_INC(tcps_rcvbadsum);
goto drop;
}
/*
* Be proactive about unspecified IPv6 address in source.
* As we use all-zero to indicate unbounded/unconnected pcb,
* unspecified IPv6 address can be used to confuse us.
*
* Note that packets with unspecified IPv6 destination is
* already dropped in ip6_input.
*/
if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) {
/* XXX stat */
goto drop;
}
}
#endif
#if defined(INET) && defined(INET6)
else
#endif
#ifdef INET
{
/*
* Get IP and TCP header together in first mbuf.
* Note: IP leaves IP header in first mbuf.
*/
if (off0 > sizeof (struct ip)) {
ip_stripoptions(m);
off0 = sizeof(struct ip);
}
if (m->m_len < sizeof (struct tcpiphdr)) {
if ((m = m_pullup(m, sizeof (struct tcpiphdr)))
== NULL) {
TCPSTAT_INC(tcps_rcvshort);
return (IPPROTO_DONE);
}
}
ip = mtod(m, struct ip *);
th = (struct tcphdr *)((caddr_t)ip + off0);
tlen = ntohs(ip->ip_len) - off0;
if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
th->th_sum = m->m_pkthdr.csum_data;
else
th->th_sum = in_pseudo(ip->ip_src.s_addr,
ip->ip_dst.s_addr,
htonl(m->m_pkthdr.csum_data + tlen +
IPPROTO_TCP));
th->th_sum ^= 0xffff;
} else {
struct ipovly *ipov = (struct ipovly *)ip;
/*
* Checksum extended TCP header and data.
*/
len = off0 + tlen;
bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
ipov->ih_len = htons(tlen);
th->th_sum = in_cksum(m, len);
/* Reset length for SDT probes. */
ip->ip_len = htons(tlen + off0);
}
if (th->th_sum) {
TCPSTAT_INC(tcps_rcvbadsum);
goto drop;
}
/* Re-initialization for later version check */
ip->ip_v = IPVERSION;
}
#endif /* INET */
1994-05-24 10:09:53 +00:00
#ifdef INET6
if (isipv6)
iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff;
#endif
#if defined(INET) && defined(INET6)
else
#endif
#ifdef INET
iptos = ip->ip_tos;
#endif
1994-05-24 10:09:53 +00:00
/*
* Check that TCP offset makes sense,
* pull out TCP options and adjust length. XXX
*/
off = th->th_off << 2;
1994-05-24 10:09:53 +00:00
if (off < sizeof (struct tcphdr) || off > tlen) {
TCPSTAT_INC(tcps_rcvbadoff);
1994-05-24 10:09:53 +00:00
goto drop;
}
tlen -= off; /* tlen is used instead of ti->ti_len */
1994-05-24 10:09:53 +00:00
if (off > sizeof (struct tcphdr)) {
#ifdef INET6
if (isipv6) {
IP6_EXTHDR_CHECK(m, off0, off, IPPROTO_DONE);
ip6 = mtod(m, struct ip6_hdr *);
th = (struct tcphdr *)((caddr_t)ip6 + off0);
}
#endif
#if defined(INET) && defined(INET6)
else
#endif
#ifdef INET
{
if (m->m_len < sizeof(struct ip) + off) {
if ((m = m_pullup(m, sizeof (struct ip) + off))
== NULL) {
TCPSTAT_INC(tcps_rcvshort);
return (IPPROTO_DONE);
}
ip = mtod(m, struct ip *);
th = (struct tcphdr *)((caddr_t)ip + off0);
1994-05-24 10:09:53 +00:00
}
}
#endif
1994-05-24 10:09:53 +00:00
optlen = off - sizeof (struct tcphdr);
optp = (u_char *)(th + 1);
1994-05-24 10:09:53 +00:00
}
thflags = th->th_flags;
1994-05-24 10:09:53 +00:00
/*
* Convert TCP protocol specific fields to host format.
*/
tcp_fields_to_host(th);
1994-05-24 10:09:53 +00:00
/*
* Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options.
*/
drop_hdrlen = off0 + off;
1994-05-24 10:09:53 +00:00
/*
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
* Locate pcb for segment; if we're likely to add or remove a
* connection then first acquire pcbinfo lock. There are three cases
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
* where we might discover later we need a write lock despite the
* flags: ACKs moving a connection out of the syncache, ACKs for a
* connection in TIMEWAIT and SYNs not targeting a listening socket.
1994-05-24 10:09:53 +00:00
*/
if ((thflags & (TH_FIN | TH_RST)) != 0) {
INP_INFO_RLOCK(&V_tcbinfo);
ti_locked = TI_RLOCKED;
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
} else
ti_locked = TI_UNLOCKED;
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
/*
* Grab info from PACKET_TAG_IPFORWARD tag prepended to the chain.
*/
2012-12-18 08:14:16 +00:00
if (
#ifdef INET6
(isipv6 && (m->m_flags & M_IP6_NEXTHOP))
#ifdef INET
|| (!isipv6 && (m->m_flags & M_IP_NEXTHOP))
#endif
#endif
#if defined(INET) && !defined(INET6)
(m->m_flags & M_IP_NEXTHOP)
#endif
)
fwd_tag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL);
Convert ipfw to use PFIL_HOOKS. This is change is transparent to userland and preserves the ipfw ABI. The ipfw core packet inspection and filtering functions have not been changed, only how ipfw is invoked is different. However there are many changes how ipfw is and its add-on's are handled: In general ipfw is now called through the PFIL_HOOKS and most associated magic, that was in ip_input() or ip_output() previously, is now done in ipfw_check_[in|out]() in the ipfw PFIL handler. IPDIVERT is entirely handled within the ipfw PFIL handlers. A packet to be diverted is checked if it is fragmented, if yes, ip_reass() gets in for reassembly. If not, or all fragments arrived and the packet is complete, divert_packet is called directly. For 'tee' no reassembly attempt is made and a copy of the packet is sent to the divert socket unmodified. The original packet continues its way through ip_input/output(). ipfw 'forward' is done via m_tag's. The ipfw PFIL handlers tag the packet with the new destination sockaddr_in. A check if the new destination is a local IP address is made and the m_flags are set appropriately. ip_input() and ip_output() have some more work to do here. For ip_input() the m_flags are checked and a packet for us is directly sent to the 'ours' section for further processing. Destination changes on the input path are only tagged and the 'srcrt' flag to ip_forward() is set to disable destination checks and ICMP replies at this stage. The tag is going to be handled on output. ip_output() again checks for m_flags and the 'ours' tag. If found, the packet will be dropped back to the IP netisr where it is going to be picked up by ip_input() again and the directly sent to the 'ours' section. When only the destination changes, the route's 'dst' is overwritten with the new destination from the forward m_tag. Then it jumps back at the route lookup again and skips the firewall check because it has been marked with M_SKIP_FIREWALL. ipfw 'forward' has to be compiled into the kernel with 'option IPFIREWALL_FORWARD' to enable it. DUMMYNET is entirely handled within the ipfw PFIL handlers. A packet for a dummynet pipe or queue is directly sent to dummynet_io(). Dummynet will then inject it back into ip_input/ip_output() after it has served its time. Dummynet packets are tagged and will continue from the next rule when they hit the ipfw PFIL handlers again after re-injection. BRIDGING and IPFW_ETHER are not changed yet and use ipfw_chk() directly as they did before. Later this will be changed to dedicated ETHER PFIL_HOOKS. More detailed changes to the code: conf/files Add netinet/ip_fw_pfil.c. conf/options Add IPFIREWALL_FORWARD option. modules/ipfw/Makefile Add ip_fw_pfil.c. net/bridge.c Disable PFIL_HOOKS if ipfw for bridging is active. Bridging ipfw is still directly invoked to handle layer2 headers and packets would get a double ipfw when run through PFIL_HOOKS as well. netinet/ip_divert.c Removed divert_clone() function. It is no longer used. netinet/ip_dummynet.[ch] Neither the route 'ro' nor the destination 'dst' need to be stored while in dummynet transit. Structure members and associated macros are removed. netinet/ip_fastfwd.c Removed all direct ipfw handling code and replace it with the new 'ipfw forward' handling code. netinet/ip_fw.h Removed 'ro' and 'dst' from struct ip_fw_args. netinet/ip_fw2.c (Re)moved some global variables and the module handling. netinet/ip_fw_pfil.c New file containing the ipfw PFIL handlers and module initialization. netinet/ip_input.c Removed all direct ipfw handling code and replace it with the new 'ipfw forward' handling code. ip_forward() does not longer require the 'next_hop' struct sockaddr_in argument. Disable early checks if 'srcrt' is set. netinet/ip_output.c Removed all direct ipfw handling code and replace it with the new 'ipfw forward' handling code. netinet/ip_var.h Add ip_reass() as general function. (Used from ipfw PFIL handlers for IPDIVERT.) netinet/raw_ip.c Directly check if ipfw and dummynet control pointers are active. netinet/tcp_input.c Rework the 'ipfw forward' to local code to work with the new way of forward tags. netinet/tcp_sack.c Remove include 'opt_ipfw.h' which is not needed here. sys/mbuf.h Remove m_claim_next() macro which was exclusively for ipfw 'forward' and is no longer needed. Approved by: re (scottl)
2004-08-17 22:05:54 +00:00
findpcb:
#ifdef INVARIANTS
if (ti_locked == TI_RLOCKED) {
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
} else {
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
}
#endif
#ifdef INET6
if (isipv6 && fwd_tag != NULL) {
struct sockaddr_in6 *next_hop6;
next_hop6 = (struct sockaddr_in6 *)(fwd_tag + 1);
/*
* Transparently forwarded. Pretend to be the destination.
* Already got one like this?
*/
inp = in6_pcblookup_mbuf(&V_tcbinfo,
&ip6->ip6_src, th->th_sport, &ip6->ip6_dst, th->th_dport,
INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif, m);
if (!inp) {
/*
* It's new. Try to find the ambushing socket.
* Because we've rewritten the destination address,
* any hardware-generated hash is ignored.
*/
inp = in6_pcblookup(&V_tcbinfo, &ip6->ip6_src,
th->th_sport, &next_hop6->sin6_addr,
next_hop6->sin6_port ? ntohs(next_hop6->sin6_port) :
th->th_dport, INPLOOKUP_WILDCARD |
INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif);
}
} else if (isipv6) {
inp = in6_pcblookup_mbuf(&V_tcbinfo, &ip6->ip6_src,
th->th_sport, &ip6->ip6_dst, th->th_dport,
INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB,
m->m_pkthdr.rcvif, m);
}
#endif /* INET6 */
#if defined(INET6) && defined(INET)
else
#endif
#ifdef INET
if (fwd_tag != NULL) {
Convert ipfw to use PFIL_HOOKS. This is change is transparent to userland and preserves the ipfw ABI. The ipfw core packet inspection and filtering functions have not been changed, only how ipfw is invoked is different. However there are many changes how ipfw is and its add-on's are handled: In general ipfw is now called through the PFIL_HOOKS and most associated magic, that was in ip_input() or ip_output() previously, is now done in ipfw_check_[in|out]() in the ipfw PFIL handler. IPDIVERT is entirely handled within the ipfw PFIL handlers. A packet to be diverted is checked if it is fragmented, if yes, ip_reass() gets in for reassembly. If not, or all fragments arrived and the packet is complete, divert_packet is called directly. For 'tee' no reassembly attempt is made and a copy of the packet is sent to the divert socket unmodified. The original packet continues its way through ip_input/output(). ipfw 'forward' is done via m_tag's. The ipfw PFIL handlers tag the packet with the new destination sockaddr_in. A check if the new destination is a local IP address is made and the m_flags are set appropriately. ip_input() and ip_output() have some more work to do here. For ip_input() the m_flags are checked and a packet for us is directly sent to the 'ours' section for further processing. Destination changes on the input path are only tagged and the 'srcrt' flag to ip_forward() is set to disable destination checks and ICMP replies at this stage. The tag is going to be handled on output. ip_output() again checks for m_flags and the 'ours' tag. If found, the packet will be dropped back to the IP netisr where it is going to be picked up by ip_input() again and the directly sent to the 'ours' section. When only the destination changes, the route's 'dst' is overwritten with the new destination from the forward m_tag. Then it jumps back at the route lookup again and skips the firewall check because it has been marked with M_SKIP_FIREWALL. ipfw 'forward' has to be compiled into the kernel with 'option IPFIREWALL_FORWARD' to enable it. DUMMYNET is entirely handled within the ipfw PFIL handlers. A packet for a dummynet pipe or queue is directly sent to dummynet_io(). Dummynet will then inject it back into ip_input/ip_output() after it has served its time. Dummynet packets are tagged and will continue from the next rule when they hit the ipfw PFIL handlers again after re-injection. BRIDGING and IPFW_ETHER are not changed yet and use ipfw_chk() directly as they did before. Later this will be changed to dedicated ETHER PFIL_HOOKS. More detailed changes to the code: conf/files Add netinet/ip_fw_pfil.c. conf/options Add IPFIREWALL_FORWARD option. modules/ipfw/Makefile Add ip_fw_pfil.c. net/bridge.c Disable PFIL_HOOKS if ipfw for bridging is active. Bridging ipfw is still directly invoked to handle layer2 headers and packets would get a double ipfw when run through PFIL_HOOKS as well. netinet/ip_divert.c Removed divert_clone() function. It is no longer used. netinet/ip_dummynet.[ch] Neither the route 'ro' nor the destination 'dst' need to be stored while in dummynet transit. Structure members and associated macros are removed. netinet/ip_fastfwd.c Removed all direct ipfw handling code and replace it with the new 'ipfw forward' handling code. netinet/ip_fw.h Removed 'ro' and 'dst' from struct ip_fw_args. netinet/ip_fw2.c (Re)moved some global variables and the module handling. netinet/ip_fw_pfil.c New file containing the ipfw PFIL handlers and module initialization. netinet/ip_input.c Removed all direct ipfw handling code and replace it with the new 'ipfw forward' handling code. ip_forward() does not longer require the 'next_hop' struct sockaddr_in argument. Disable early checks if 'srcrt' is set. netinet/ip_output.c Removed all direct ipfw handling code and replace it with the new 'ipfw forward' handling code. netinet/ip_var.h Add ip_reass() as general function. (Used from ipfw PFIL handlers for IPDIVERT.) netinet/raw_ip.c Directly check if ipfw and dummynet control pointers are active. netinet/tcp_input.c Rework the 'ipfw forward' to local code to work with the new way of forward tags. netinet/tcp_sack.c Remove include 'opt_ipfw.h' which is not needed here. sys/mbuf.h Remove m_claim_next() macro which was exclusively for ipfw 'forward' and is no longer needed. Approved by: re (scottl)
2004-08-17 22:05:54 +00:00
struct sockaddr_in *next_hop;
next_hop = (struct sockaddr_in *)(fwd_tag+1);
/*
Remove (almost all) global variables that were used to hold packet forwarding state ("annotations") during ip processing. The code is considerably cleaner now. The variables removed by this change are: ip_divert_cookie used by divert sockets ip_fw_fwd_addr used for transparent ip redirection last_pkt used by dynamic pipes in dummynet Removal of the first two has been done by carrying the annotations into volatile structs prepended to the mbuf chains, and adding appropriate code to add/remove annotations in the routines which make use of them, i.e. ip_input(), ip_output(), tcp_input(), bdg_forward(), ether_demux(), ether_output_frame(), div_output(). On passing, remove a bug in divert handling of fragmented packet. Now it is the fragment at offset 0 which sets the divert status of the whole packet, whereas formerly it was the last incoming fragment to decide. Removal of last_pkt required a change in the interface of ip_fw_chk() and dummynet_io(). On passing, use the same mechanism for dummynet annotations and for divert/forward annotations. option IPFIREWALL_FORWARD is effectively useless, the code to implement it is very small and is now in by default to avoid the obfuscation of conditionally compiled code. NOTES: * there is at least one global variable left, sro_fwd, in ip_output(). I am not sure if/how this can be removed. * I have deliberately avoided gratuitous style changes in this commit to avoid cluttering the diffs. Minor stule cleanup will likely be necessary * this commit only focused on the IP layer. I am sure there is a number of global variables used in the TCP and maybe UDP stack. * despite the number of files touched, there are absolutely no API's or data structures changed by this commit (except the interfaces of ip_fw_chk() and dummynet_io(), which are internal anyways), so an MFC is quite safe and unintrusive (and desirable, given the improved readability of the code). MFC after: 10 days
2002-06-22 11:51:02 +00:00
* Transparently forwarded. Pretend to be the destination.
* already got one like this?
*/
inp = in_pcblookup_mbuf(&V_tcbinfo, ip->ip_src, th->th_sport,
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
ip->ip_dst, th->th_dport, INPLOOKUP_WLOCKPCB,
m->m_pkthdr.rcvif, m);
if (!inp) {
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
/*
* It's new. Try to find the ambushing socket.
* Because we've rewritten the destination address,
* any hardware-generated hash is ignored.
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
*/
inp = in_pcblookup(&V_tcbinfo, ip->ip_src,
th->th_sport, next_hop->sin_addr,
next_hop->sin_port ? ntohs(next_hop->sin_port) :
th->th_dport, INPLOOKUP_WILDCARD |
INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif);
}
} else
inp = in_pcblookup_mbuf(&V_tcbinfo, ip->ip_src,
th->th_sport, ip->ip_dst, th->th_dport,
INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB,
m->m_pkthdr.rcvif, m);
#endif /* INET */
1994-05-24 10:09:53 +00:00
/*
2007-03-21 18:36:49 +00:00
* If the INPCB does not exist then all data in the incoming
* segment is discarded and an appropriate RST is sent back.
Add code to allow the system to handle multiple routing tables. This particular implementation is designed to be fully backwards compatible and to be MFC-able to 7.x (and 6.x) Currently the only protocol that can make use of the multiple tables is IPv4 Similar functionality exists in OpenBSD and Linux. From my notes: ----- One thing where FreeBSD has been falling behind, and which by chance I have some time to work on is "policy based routing", which allows different packet streams to be routed by more than just the destination address. Constraints: ------------ I want to make some form of this available in the 6.x tree (and by extension 7.x) , but FreeBSD in general needs it so I might as well do it in -current and back port the portions I need. One of the ways that this can be done is to have the ability to instantiate multiple kernel routing tables (which I will now refer to as "Forwarding Information Bases" or "FIBs" for political correctness reasons). Which FIB a particular packet uses to make the next hop decision can be decided by a number of mechanisms. The policies these mechanisms implement are the "Policies" referred to in "Policy based routing". One of the constraints I have if I try to back port this work to 6.x is that it must be implemented as a EXTENSION to the existing ABIs in 6.x so that third party applications do not need to be recompiled in timespan of the branch. This first version will not have some of the bells and whistles that will come with later versions. It will, for example, be limited to 16 tables in the first commit. Implementation method, Compatible version. (part 1) ------------------------------- For this reason I have implemented a "sufficient subset" of a multiple routing table solution in Perforce, and back-ported it to 6.x. (also in Perforce though not always caught up with what I have done in -current/P4). The subset allows a number of FIBs to be defined at compile time (8 is sufficient for my purposes in 6.x) and implements the changes needed to allow IPV4 to use them. I have not done the changes for ipv6 simply because I do not need it, and I do not have enough knowledge of ipv6 (e.g. neighbor discovery) needed to do it. Other protocol families are left untouched and should there be users with proprietary protocol families, they should continue to work and be oblivious to the existence of the extra FIBs. To understand how this is done, one must know that the current FIB code starts everything off with a single dimensional array of pointers to FIB head structures (One per protocol family), each of which in turn points to the trie of routes available to that family. The basic change in the ABI compatible version of the change is to extent that array to be a 2 dimensional array, so that instead of protocol family X looking at rt_tables[X] for the table it needs, it looks at rt_tables[Y][X] when for all protocol families except ipv4 Y is always 0. Code that is unaware of the change always just sees the first row of the table, which of course looks just like the one dimensional array that existed before. The entry points rtrequest(), rtalloc(), rtalloc1(), rtalloc_ign() are all maintained, but refer only to the first row of the array, so that existing callers in proprietary protocols can continue to do the "right thing". Some new entry points are added, for the exclusive use of ipv4 code called in_rtrequest(), in_rtalloc(), in_rtalloc1() and in_rtalloc_ign(), which have an extra argument which refers the code to the correct row. In addition, there are some new entry points (currently called rtalloc_fib() and friends) that check the Address family being looked up and call either rtalloc() (and friends) if the protocol is not IPv4 forcing the action to row 0 or to the appropriate row if it IS IPv4 (and that info is available). These are for calling from code that is not specific to any particular protocol. The way these are implemented would change in the non ABI preserving code to be added later. One feature of the first version of the code is that for ipv4, the interface routes show up automatically on all the FIBs, so that no matter what FIB you select you always have the basic direct attached hosts available to you. (rtinit() does this automatically). You CAN delete an interface route from one FIB should you want to but by default it's there. ARP information is also available in each FIB. It's assumed that the same machine would have the same MAC address, regardless of which FIB you are using to get to it. This brings us as to how the correct FIB is selected for an outgoing IPV4 packet. Firstly, all packets have a FIB associated with them. if nothing has been done to change it, it will be FIB 0. The FIB is changed in the following ways. Packets fall into one of a number of classes. 1/ locally generated packets, coming from a socket/PCB. Such packets select a FIB from a number associated with the socket/PCB. This in turn is inherited from the process, but can be changed by a socket option. The process in turn inherits it on fork. I have written a utility call setfib that acts a bit like nice.. setfib -3 ping target.example.com # will use fib 3 for ping. It is an obvious extension to make it a property of a jail but I have not done so. It can be achieved by combining the setfib and jail commands. 2/ packets received on an interface for forwarding. By default these packets would use table 0, (or possibly a number settable in a sysctl(not yet)). but prior to routing the firewall can inspect them (see below). (possibly in the future you may be able to associate a FIB with packets received on an interface.. An ifconfig arg, but not yet.) 3/ packets inspected by a packet classifier, which can arbitrarily associate a fib with it on a packet by packet basis. A fib assigned to a packet by a packet classifier (such as ipfw) would over-ride a fib associated by a more default source. (such as cases 1 or 2). 4/ a tcp listen socket associated with a fib will generate accept sockets that are associated with that same fib. 5/ Packets generated in response to some other packet (e.g. reset or icmp packets). These should use the FIB associated with the packet being reponded to. 6/ Packets generated during encapsulation. gif, tun and other tunnel interfaces will encapsulate using the FIB that was in effect withthe proces that set up the tunnel. thus setfib 1 ifconfig gif0 [tunnel instructions] will set the fib for the tunnel to use to be fib 1. Routing messages would be associated with their process, and thus select one FIB or another. messages from the kernel would be associated with the fib they refer to and would only be received by a routing socket associated with that fib. (not yet implemented) In addition Netstat has been edited to be able to cope with the fact that the array is now 2 dimensional. (It looks in system memory using libkvm (!)). Old versions of netstat see only the first FIB. In addition two sysctls are added to give: a) the number of FIBs compiled in (active) b) the default FIB of the calling process. Early testing experience: ------------------------- Basically our (IronPort's) appliance does this functionality already using ipfw fwd but that method has some drawbacks. For example, It can't fully simulate a routing table because it can't influence the socket's choice of local address when a connect() is done. Testing during the generating of these changes has been remarkably smooth so far. Multiple tables have co-existed with no notable side effects, and packets have been routes accordingly. ipfw has grown 2 new keywords: setfib N ip from anay to any count ip from any to any fib N In pf there seems to be a requirement to be able to give symbolic names to the fibs but I do not have that capacity. I am not sure if it is required. SCTP has interestingly enough built in support for this, called VRFs in Cisco parlance. it will be interesting to see how that handles it when it suddenly actually does something. Where to next: -------------------- After committing the ABI compatible version and MFCing it, I'd like to proceed in a forward direction in -current. this will result in some roto-tilling in the routing code. Firstly: the current code's idea of having a separate tree per protocol family, all of the same format, and pointed to by the 1 dimensional array is a bit silly. Especially when one considers that there is code that makes assumptions about every protocol having the same internal structures there. Some protocols don't WANT that sort of structure. (for example the whole idea of a netmask is foreign to appletalk). This needs to be made opaque to the external code. My suggested first change is to add routing method pointers to the 'domain' structure, along with information pointing the data. instead of having an array of pointers to uniform structures, there would be an array pointing to the 'domain' structures for each protocol address domain (protocol family), and the methods this reached would be called. The methods would have an argument that gives FIB number, but the protocol would be free to ignore it. When the ABI can be changed it raises the possibilty of the addition of a fib entry into the "struct route". Currently, the structure contains the sockaddr of the desination, and the resulting fib entry. To make this work fully, one could add a fib number so that given an address and a fib, one can find the third element, the fib entry. Interaction with the ARP layer/ LL layer would need to be revisited as well. Qing Li has been working on this already. This work was sponsored by Ironport Systems/Cisco Reviewed by: several including rwatson, bz and mlair (parts each) Obtained from: Ironport systems/Cisco
2008-05-09 23:03:00 +00:00
* XXX MRT Send RST using which routing table?
1994-05-24 10:09:53 +00:00
*/
if (inp == NULL) {
2007-03-21 18:36:49 +00:00
/*
* Log communication attempts to ports that are not
* in use.
*/
if ((tcp_log_in_vain == 1 && (thflags & TH_SYN)) ||
tcp_log_in_vain == 2) {
if ((s = tcp_log_vain(NULL, th, (void *)ip, ip6)))
log(LOG_INFO, "%s; %s: Connection attempt "
"to closed port\n", s, __func__);
}
2007-03-21 18:36:49 +00:00
/*
* When blackholing do not respond with a RST but
* completely ignore the segment and drop it.
*/
if ((V_blackhole == 1 && (thflags & TH_SYN)) ||
V_blackhole == 2)
goto dropunlock;
2007-03-21 18:36:49 +00:00
rstreason = BANDLIM_RST_CLOSEDPORT;
goto dropwithreset;
}
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
INP_WLOCK_ASSERT(inp);
if ((inp->inp_flowtype == M_HASHTYPE_NONE) &&
(M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) &&
((inp->inp_socket == NULL) ||
(inp->inp_socket->so_options & SO_ACCEPTCONN) == 0)) {
inp->inp_flowid = m->m_pkthdr.flowid;
inp->inp_flowtype = M_HASHTYPE_GET(m);
}
#ifdef IPSEC
#ifdef INET6
if (isipv6 && ipsec6_in_reject(m, inp)) {
goto dropunlock;
} else
#endif /* INET6 */
if (ipsec4_in_reject(m, inp) != 0) {
goto dropunlock;
}
#endif /* IPSEC */
/*
* Check the minimum TTL for socket.
*/
if (inp->inp_ip_minttl != 0) {
#ifdef INET6
if (isipv6) {
if (inp->inp_ip_minttl > ip6->ip6_hlim)
goto dropunlock;
} else
#endif
if (inp->inp_ip_minttl > ip->ip_ttl)
goto dropunlock;
}
/*
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
* A previous connection in TIMEWAIT state is supposed to catch stray
* or duplicate segments arriving late. If this segment was a
* legitimate new connection attempt, the old INPCB gets removed and
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
* we can try again to find a listening socket.
*
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
* At this point, due to earlier optimism, we may hold only an inpcb
* lock, and not the inpcbinfo write lock. If so, we need to try to
* acquire it, or if that fails, acquire a reference on the inpcb,
* drop all locks, acquire a global write lock, and then re-acquire
* the inpcb lock. We may at that point discover that another thread
* has tried to free the inpcb, in which case we need to loop back
* and try to find a new inpcb to deliver to.
*
* XXXRW: It may be time to rethink timewait locking.
*/
relocked:
if (inp->inp_flags & INP_TIMEWAIT) {
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
if (ti_locked == TI_UNLOCKED) {
if (INP_INFO_TRY_RLOCK(&V_tcbinfo) == 0) {
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
in_pcbref(inp);
INP_WUNLOCK(inp);
INP_INFO_RLOCK(&V_tcbinfo);
ti_locked = TI_RLOCKED;
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
INP_WLOCK(inp);
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
if (in_pcbrele_wlocked(inp)) {
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
inp = NULL;
goto findpcb;
}
} else
ti_locked = TI_RLOCKED;
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
}
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
if (thflags & TH_SYN)
tcp_dooptions(&to, optp, optlen, TO_SYN);
/*
* NB: tcp_twcheck unlocks the INP and frees the mbuf.
*/
if (tcp_twcheck(inp, &to, th, m, tlen))
goto findpcb;
INP_INFO_RUNLOCK(&V_tcbinfo);
return (IPPROTO_DONE);
}
/*
* The TCPCB may no longer exist if the connection is winding
* down or it is in the CLOSED state. Either way we drop the
* segment and send an appropriate response.
*/
1994-05-24 10:09:53 +00:00
tp = intotcpcb(inp);
if (tp == NULL || tp->t_state == TCPS_CLOSED) {
rstreason = BANDLIM_RST_CLOSEDPORT;
goto dropwithreset;
}
1995-05-30 08:16:23 +00:00
#ifdef TCP_OFFLOAD
if (tp->t_flags & TF_TOE) {
tcp_offload_input(tp, m);
m = NULL; /* consumed by the TOE driver */
goto dropunlock;
}
#endif
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
/*
* We've identified a valid inpcb, but it could be that we need an
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
* inpcbinfo write lock but don't hold it. In this case, attempt to
* acquire using the same strategy as the TIMEWAIT case above. If we
* relock, we have to jump back to 'relocked' as the connection might
* now be in TIMEWAIT.
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
*/
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
#ifdef INVARIANTS
if ((thflags & (TH_FIN | TH_RST)) != 0)
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
#endif
if (!((tp->t_state == TCPS_ESTABLISHED && (thflags & TH_SYN) == 0) ||
(tp->t_state == TCPS_LISTEN && (thflags & TH_SYN) &&
!(tp->t_flags & TF_FASTOPEN)))) {
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
if (ti_locked == TI_UNLOCKED) {
if (INP_INFO_TRY_RLOCK(&V_tcbinfo) == 0) {
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
in_pcbref(inp);
INP_WUNLOCK(inp);
INP_INFO_RLOCK(&V_tcbinfo);
ti_locked = TI_RLOCKED;
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
INP_WLOCK(inp);
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
if (in_pcbrele_wlocked(inp)) {
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
inp = NULL;
goto findpcb;
}
goto relocked;
} else
ti_locked = TI_RLOCKED;
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
}
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
}
#ifdef MAC
INP_WLOCK_ASSERT(inp);
if (mac_inpcb_check_deliver(inp, m))
goto dropunlock;
#endif
so = inp->inp_socket;
KASSERT(so != NULL, ("%s: so == NULL", __func__));
#ifdef TCPDEBUG
if (so->so_options & SO_DEBUG) {
ostate = tp->t_state;
#ifdef INET6
if (isipv6) {
bcopy((char *)ip6, (char *)tcp_saveipgen, sizeof(*ip6));
} else
#endif
bcopy((char *)ip, (char *)tcp_saveipgen, sizeof(*ip));
tcp_savetcp = *th;
}
#endif /* TCPDEBUG */
/*
* When the socket is accepting connections (the INPCB is in LISTEN
* state) we look into the SYN cache if this is a new connection
* attempt or the completion of a previous one.
*/
if (so->so_options & SO_ACCEPTCONN) {
struct in_conninfo inc;
2007-04-20 15:21:29 +00:00
KASSERT(tp->t_state == TCPS_LISTEN, ("%s: so accepting but "
"tp not listening", __func__));
bzero(&inc, sizeof(inc));
#ifdef INET6
if (isipv6) {
inc.inc_flags |= INC_ISIPV6;
inc.inc6_faddr = ip6->ip6_src;
inc.inc6_laddr = ip6->ip6_dst;
} else
#endif
{
inc.inc_faddr = ip->ip_src;
inc.inc_laddr = ip->ip_dst;
}
inc.inc_fport = th->th_sport;
inc.inc_lport = th->th_dport;
inc.inc_fibnum = so->so_fibnum;
/*
* Check for an existing connection attempt in syncache if
* the flag is only ACK. A successful lookup creates a new
* socket appended to the listen queue in SYN_RECEIVED state.
*/
if ((thflags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK) {
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
/*
* Parse the TCP options here because
* syncookies need access to the reflected
* timestamp.
*/
tcp_dooptions(&to, optp, optlen, 0);
/*
* NB: syncache_expand() doesn't unlock
* inp and tcpinfo locks.
*/
if (!syncache_expand(&inc, &to, th, &so, m)) {
Rewrite of TCP syncookies to remove locking requirements and to enhance functionality: - Remove a rwlock aquisition/release per generated syncookie. Locking is now integrated with the bucket row locking of syncache itself and syncookies no longer add any additional lock overhead. - Syncookie secrets are different for and stored per syncache buck row. Secrets expire after 16 seconds and are reseeded on-demand. - The computational overhead for syncookie generation and verification is one MD5 hash computation as before. - Syncache can be turned off and run with syncookies only by setting the sysctl net.inet.tcp.syncookies_only=1. This implementation extends the orginal idea and first implementation of FreeBSD by using not only the initial sequence number field to store information but also the timestamp field if present. This way we can keep track of the entire state we need to know to recreate the session in its original form. Almost all TCP speakers implement RFC1323 timestamps these days. For those that do not we still have to live with the known shortcomings of the ISN only SYN cookies. The use of the timestamp field causes the timestamps to be randomized if syncookies are enabled. The idea of SYN cookies is to encode and include all necessary information about the connection setup state within the SYN-ACK we send back and thus to get along without keeping any local state until the ACK to the SYN-ACK arrives (if ever). Everything we need to know should be available from the information we encoded in the SYN-ACK. A detailed description of the inner working of the syncookies mechanism is included in the comments in tcp_syncache.c. Reviewed by: silby (slightly earlier version) Sponsored by: TCP/IP Optimization Fundraise 2005
2006-09-13 13:08:27 +00:00
/*
* No syncache entry or ACK was not
* for our SYN/ACK. Send a RST.
* NB: syncache did its own logging
* of the failure cause.
Rewrite of TCP syncookies to remove locking requirements and to enhance functionality: - Remove a rwlock aquisition/release per generated syncookie. Locking is now integrated with the bucket row locking of syncache itself and syncookies no longer add any additional lock overhead. - Syncookie secrets are different for and stored per syncache buck row. Secrets expire after 16 seconds and are reseeded on-demand. - The computational overhead for syncookie generation and verification is one MD5 hash computation as before. - Syncache can be turned off and run with syncookies only by setting the sysctl net.inet.tcp.syncookies_only=1. This implementation extends the orginal idea and first implementation of FreeBSD by using not only the initial sequence number field to store information but also the timestamp field if present. This way we can keep track of the entire state we need to know to recreate the session in its original form. Almost all TCP speakers implement RFC1323 timestamps these days. For those that do not we still have to live with the known shortcomings of the ISN only SYN cookies. The use of the timestamp field causes the timestamps to be randomized if syncookies are enabled. The idea of SYN cookies is to encode and include all necessary information about the connection setup state within the SYN-ACK we send back and thus to get along without keeping any local state until the ACK to the SYN-ACK arrives (if ever). Everything we need to know should be available from the information we encoded in the SYN-ACK. A detailed description of the inner working of the syncookies mechanism is included in the comments in tcp_syncache.c. Reviewed by: silby (slightly earlier version) Sponsored by: TCP/IP Optimization Fundraise 2005
2006-09-13 13:08:27 +00:00
*/
rstreason = BANDLIM_RST_OPENPORT;
goto dropwithreset;
}
#ifdef TCP_RFC7413
new_tfo_socket:
#endif
if (so == NULL) {
/*
* We completed the 3-way handshake
* but could not allocate a socket
* either due to memory shortage,
* listen queue length limits or
* global socket limits. Send RST
* or wait and have the remote end
* retransmit the ACK for another
* try.
*/
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"Socket allocation failed due to "
"limits or memory shortage, %s\n",
s, __func__,
V_tcp_sc_rst_sock_fail ?
"sending RST" : "try again");
if (V_tcp_sc_rst_sock_fail) {
rstreason = BANDLIM_UNLIMITED;
goto dropwithreset;
} else
goto dropunlock;
}
/*
* Socket is created in state SYN_RECEIVED.
* Unlock the listen socket, lock the newly
* created socket and update the tp variable.
*/
INP_WUNLOCK(inp); /* listen socket */
inp = sotoinpcb(so);
/*
* New connection inpcb is already locked by
* syncache_expand().
*/
INP_WLOCK_ASSERT(inp);
tp = intotcpcb(inp);
KASSERT(tp->t_state == TCPS_SYN_RECEIVED,
("%s: ", __func__));
#ifdef TCP_SIGNATURE
if (sig_checked == 0) {
tcp_dooptions(&to, optp, optlen,
(thflags & TH_SYN) ? TO_SYN : 0);
if (!tcp_signature_verify_input(m, off0, tlen,
optlen, &to, th, tp->t_flags)) {
/*
* In SYN_SENT state if it receives an
* RST, it is allowed for further
* processing.
*/
if ((thflags & TH_RST) == 0 ||
(tp->t_state == TCPS_SYN_SENT) == 0)
goto dropunlock;
}
sig_checked = 1;
}
#endif
/*
* Process the segment and the data it
* contains. tcp_do_segment() consumes
* the mbuf chain and unlocks the inpcb.
*/
tp->t_fb->tfb_tcp_do_segment(m, th, so, tp, drop_hdrlen, tlen,
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
iptos, ti_locked);
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
return (IPPROTO_DONE);
}
/*
* Segment flag validation for new connection attempts:
*
* Our (SYN|ACK) response was rejected.
* Check with syncache and remove entry to prevent
* retransmits.
*
* NB: syncache_chkrst does its own logging of failure
* causes.
*/
if (thflags & TH_RST) {
syncache_chkrst(&inc, th);
goto dropunlock;
}
/*
* We can't do anything without SYN.
*/
if ((thflags & TH_SYN) == 0) {
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"SYN is missing, segment ignored\n",
s, __func__);
TCPSTAT_INC(tcps_badsyn);
goto dropunlock;
}
/*
* (SYN|ACK) is bogus on a listen socket.
*/
if (thflags & TH_ACK) {
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"SYN|ACK invalid, segment rejected\n",
s, __func__);
syncache_badack(&inc); /* XXX: Not needed! */
TCPSTAT_INC(tcps_badsyn);
rstreason = BANDLIM_RST_OPENPORT;
goto dropwithreset;
}
/*
* If the drop_synfin option is enabled, drop all
* segments with both the SYN and FIN bits set.
* This prevents e.g. nmap from identifying the
* TCP/IP stack.
* XXX: Poor reasoning. nmap has other methods
* and is constantly refining its stack detection
* strategies.
* XXX: This is a violation of the TCP specification
* and was used by RFC1644.
*/
if ((thflags & TH_FIN) && V_drop_synfin) {
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"SYN|FIN segment ignored (based on "
"sysctl setting)\n", s, __func__);
TCPSTAT_INC(tcps_badsyn);
2011-01-07 21:40:34 +00:00
goto dropunlock;
}
/*
* Segment's flags are (SYN) or (SYN|FIN).
*
* TH_PUSH, TH_URG, TH_ECE, TH_CWR are ignored
* as they do not affect the state of the TCP FSM.
* The data pointed to by TH_URG and th_urp is ignored.
*/
KASSERT((thflags & (TH_RST|TH_ACK)) == 0,
("%s: Listen socket: TH_RST or TH_ACK set", __func__));
KASSERT(thflags & (TH_SYN),
("%s: Listen socket: TH_SYN not set", __func__));
#ifdef INET6
/*
* If deprecated address is forbidden,
* we do not accept SYN to deprecated interface
* address to prevent any new inbound connection from
* getting established.
* When we do not accept SYN, we send a TCP RST,
* with deprecated source address (instead of dropping
* it). We compromise it as it is much better for peer
* to send a RST, and RST will be the final packet
* for the exchange.
*
* If we do not forbid deprecated addresses, we accept
* the SYN packet. RFC2462 does not suggest dropping
* SYN in this case.
* If we decipher RFC2462 5.5.4, it says like this:
* 1. use of deprecated addr with existing
* communication is okay - "SHOULD continue to be
* used"
* 2. use of it with new communication:
* (2a) "SHOULD NOT be used if alternate address
* with sufficient scope is available"
* (2b) nothing mentioned otherwise.
* Here we fall into (2b) case as we have no choice in
* our source address selection - we must obey the peer.
*
* The wording in RFC2462 is confusing, and there are
* multiple description text for deprecated address
* handling - worse, they are not exactly the same.
* I believe 5.5.4 is the best one, so we follow 5.5.4.
*/
if (isipv6 && !V_ip6_use_deprecated) {
struct in6_ifaddr *ia6;
ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */);
if (ia6 != NULL &&
(ia6->ia6_flags & IN6_IFF_DEPRECATED)) {
ifa_free(&ia6->ia_ifa);
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"Connection attempt to deprecated "
"IPv6 address rejected\n",
s, __func__);
rstreason = BANDLIM_RST_OPENPORT;
goto dropwithreset;
}
if (ia6)
ifa_free(&ia6->ia_ifa);
}
#endif /* INET6 */
/*
* Basic sanity checks on incoming SYN requests:
* Don't respond if the destination is a link layer
* broadcast according to RFC1122 4.2.3.10, p. 104.
* If it is from this socket it must be forged.
* Don't respond if the source or destination is a
* global or subnet broad- or multicast address.
* Note that it is quite possible to receive unicast
* link-layer packets with a broadcast IP address. Use
* in_broadcast() to find them.
*/
if (m->m_flags & (M_BCAST|M_MCAST)) {
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"Connection attempt from broad- or multicast "
"link layer address ignored\n", s, __func__);
goto dropunlock;
}
#ifdef INET6
if (isipv6) {
if (th->th_dport == th->th_sport &&
IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, &ip6->ip6_src)) {
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"Connection attempt to/from self "
"ignored\n", s, __func__);
goto dropunlock;
}
if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) {
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"Connection attempt from/to multicast "
"address ignored\n", s, __func__);
goto dropunlock;
}
}
#endif
#if defined(INET) && defined(INET6)
else
#endif
#ifdef INET
{
if (th->th_dport == th->th_sport &&
ip->ip_dst.s_addr == ip->ip_src.s_addr) {
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"Connection attempt from/to self "
"ignored\n", s, __func__);
goto dropunlock;
}
if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif)) {
if ((s = tcp_log_addrs(&inc, th, NULL, NULL)))
log(LOG_DEBUG, "%s; %s: Listen socket: "
"Connection attempt from/to broad- "
"or multicast address ignored\n",
s, __func__);
goto dropunlock;
}
}
#endif
/*
* SYN appears to be valid. Create compressed TCP state
* for syncache.
*/
#ifdef TCPDEBUG
if (so->so_options & SO_DEBUG)
tcp_trace(TA_INPUT, ostate, tp,
(void *)tcp_saveipgen, &tcp_savetcp, 0);
#endif
TCP_PROBE3(debug__input, tp, th, mtod(m, const char *));
tcp_dooptions(&to, optp, optlen, TO_SYN);
#ifdef TCP_RFC7413
if (syncache_add(&inc, &to, th, inp, &so, m, NULL, NULL))
goto new_tfo_socket;
#else
syncache_add(&inc, &to, th, inp, &so, m, NULL, NULL);
#endif
/*
* Entry added to syncache and mbuf consumed.
* Only the listen socket is unlocked by syncache_add().
*/
if (ti_locked == TI_RLOCKED) {
INP_INFO_RUNLOCK(&V_tcbinfo);
ti_locked = TI_UNLOCKED;
}
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
return (IPPROTO_DONE);
} else if (tp->t_state == TCPS_LISTEN) {
/*
* When a listen socket is torn down the SO_ACCEPTCONN
* flag is removed first while connections are drained
* from the accept queue in a unlock/lock cycle of the
* ACCEPT_LOCK, opening a race condition allowing a SYN
* attempt go through unhandled.
*/
goto dropunlock;
}
#ifdef TCP_SIGNATURE
if (sig_checked == 0) {
tcp_dooptions(&to, optp, optlen,
(thflags & TH_SYN) ? TO_SYN : 0);
if (!tcp_signature_verify_input(m, off0, tlen, optlen, &to,
th, tp->t_flags)) {
/*
* In SYN_SENT state if it receives an RST, it is
* allowed for further processing.
*/
if ((thflags & TH_RST) == 0 ||
(tp->t_state == TCPS_SYN_SENT) == 0)
goto dropunlock;
}
sig_checked = 1;
}
#endif
TCP_PROBE5(receive, NULL, tp, mtod(m, const char *), tp, th);
/*
* Segment belongs to a connection in SYN_SENT, ESTABLISHED or later
* state. tcp_do_segment() always consumes the mbuf chain, unlocks
* the inpcb, and unlocks pcbinfo.
*/
tp->t_fb->tfb_tcp_do_segment(m, th, so, tp, drop_hdrlen, tlen, iptos, ti_locked);
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
return (IPPROTO_DONE);
dropwithreset:
TCP_PROBE5(receive, NULL, tp, mtod(m, const char *), tp, th);
if (ti_locked == TI_RLOCKED) {
INP_INFO_RUNLOCK(&V_tcbinfo);
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
ti_locked = TI_UNLOCKED;
}
#ifdef INVARIANTS
else {
KASSERT(ti_locked == TI_UNLOCKED, ("%s: dropwithreset "
"ti_locked: %d", __func__, ti_locked));
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
}
#endif
if (inp != NULL) {
tcp_dropwithreset(m, th, tp, tlen, rstreason);
INP_WUNLOCK(inp);
} else
tcp_dropwithreset(m, th, NULL, tlen, rstreason);
m = NULL; /* mbuf chain got consumed. */
goto drop;
dropunlock:
if (m != NULL)
TCP_PROBE5(receive, NULL, tp, mtod(m, const char *), tp, th);
if (ti_locked == TI_RLOCKED) {
INP_INFO_RUNLOCK(&V_tcbinfo);
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
ti_locked = TI_UNLOCKED;
}
#ifdef INVARIANTS
else {
KASSERT(ti_locked == TI_UNLOCKED, ("%s: dropunlock "
"ti_locked: %d", __func__, ti_locked));
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
}
#endif
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
if (inp != NULL)
INP_WUNLOCK(inp);
drop:
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
if (s != NULL)
free(s, M_TCPLOG);
if (m != NULL)
m_freem(m);
return (IPPROTO_DONE);
}
void
tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
struct tcpcb *tp, int drop_hdrlen, int tlen, uint8_t iptos,
int ti_locked)
{
One of the ways to detect loss is to count duplicate acks coming back from the other end till it reaches predetermined threshold which is 3 for us right now. Once that happens, we trigger fast-retransmit to do loss recovery. Main problem with the current implementation is that we don't honor SACK information well to detect whether an incoming ack is a dupack or not. RFC6675 has latest recommendations for that. According to it, dupack is a segment that arrives carrying a SACK block that identifies previously unknown information between snd_una and snd_max even if it carries new data, changes the advertised window, or moves the cumulative acknowledgment point. With the prevalence of Selective ACK (SACK) these days, improper handling can lead to delayed loss recovery. With the fix, new behavior looks like following: 0) th_ack < snd_una --> ignore Old acks are ignored. 1) th_ack == snd_una, !sack_changed --> ignore Acks with SACK enabled but without any new SACK info in them are ignored. 2) th_ack == snd_una, window == old_window --> increment Increment on a good dupack. 3) th_ack == snd_una, window != old_window, sack_changed --> increment When SACK enabled, it's okay to have advertized window changed if the ack has new SACK info. 4) th_ack > snd_una --> reset to 0 Reset to 0 when left edge moves. 5) th_ack > snd_una, sack_changed --> increment Increment if left edge moves but there is new SACK info. Here, sack_changed is the indicator that incoming ack has previously unknown SACK info in it. Note: This fix is not fully compliant to RFC6675. That may require a few changes to current implementation in order to keep per-sackhole dupack counter and change to the way we mark/handle sack holes. PR: 203663 Reviewed by: jtl MFC after: 3 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D4225
2015-12-08 21:21:48 +00:00
int thflags, acked, ourfinisacked, needoutput = 0, sack_changed;
int rstreason, todrop, win;
u_long tiwin;
char *s;
struct in_conninfo *inc;
struct mbuf *mfree;
struct tcpopt to;
int tfo_syn;
#ifdef TCPDEBUG
/*
* The size of tcp_saveipgen must be the size of the max ip header,
* now IPv6.
*/
u_char tcp_saveipgen[IP6_HDR_LEN];
struct tcphdr tcp_savetcp;
short ostate = 0;
#endif
thflags = th->th_flags;
inc = &tp->t_inpcb->inp_inc;
tp->sackhint.last_sack_ack = 0;
One of the ways to detect loss is to count duplicate acks coming back from the other end till it reaches predetermined threshold which is 3 for us right now. Once that happens, we trigger fast-retransmit to do loss recovery. Main problem with the current implementation is that we don't honor SACK information well to detect whether an incoming ack is a dupack or not. RFC6675 has latest recommendations for that. According to it, dupack is a segment that arrives carrying a SACK block that identifies previously unknown information between snd_una and snd_max even if it carries new data, changes the advertised window, or moves the cumulative acknowledgment point. With the prevalence of Selective ACK (SACK) these days, improper handling can lead to delayed loss recovery. With the fix, new behavior looks like following: 0) th_ack < snd_una --> ignore Old acks are ignored. 1) th_ack == snd_una, !sack_changed --> ignore Acks with SACK enabled but without any new SACK info in them are ignored. 2) th_ack == snd_una, window == old_window --> increment Increment on a good dupack. 3) th_ack == snd_una, window != old_window, sack_changed --> increment When SACK enabled, it's okay to have advertized window changed if the ack has new SACK info. 4) th_ack > snd_una --> reset to 0 Reset to 0 when left edge moves. 5) th_ack > snd_una, sack_changed --> increment Increment if left edge moves but there is new SACK info. Here, sack_changed is the indicator that incoming ack has previously unknown SACK info in it. Note: This fix is not fully compliant to RFC6675. That may require a few changes to current implementation in order to keep per-sackhole dupack counter and change to the way we mark/handle sack holes. PR: 203663 Reviewed by: jtl MFC after: 3 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D4225
2015-12-08 21:21:48 +00:00
sack_changed = 0;
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
/*
* If this is either a state-changing packet or current state isn't
* established, we require a write lock on tcbinfo. Otherwise, we
* allow the tcbinfo to be in either alocked or unlocked, as the
* caller may have unnecessarily acquired a write lock due to a race.
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
*/
if ((thflags & (TH_SYN | TH_FIN | TH_RST)) != 0 ||
tp->t_state != TCPS_ESTABLISHED) {
KASSERT(ti_locked == TI_RLOCKED, ("%s ti_locked %d for "
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
"SYN/FIN/RST/!EST", __func__, ti_locked));
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
} else {
#ifdef INVARIANTS
if (ti_locked == TI_RLOCKED)
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
else {
KASSERT(ti_locked == TI_UNLOCKED, ("%s: EST "
"ti_locked: %d", __func__, ti_locked));
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
}
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
#endif
}
INP_WLOCK_ASSERT(tp->t_inpcb);
KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
__func__));
KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
__func__));
1994-05-24 10:09:53 +00:00
There are times when it would be really nice to have a record of the last few packets and/or state transitions from each TCP socket. That would help with narrowing down certain problems we see in the field that are hard to reproduce without understanding the history of how we got into a certain state. This change provides just that. It saves copies of the last N packets in a list in the tcpcb. When the tcpcb is destroyed, the list is freed. I thought this was likely to be more performance-friendly than saving copies of the tcpcb. Plus, with the packets, you should be able to reverse-engineer what happened to the tcpcb. To enable the feature, you will need to compile a kernel with the TCPPCAP option. Even then, the feature defaults to being deactivated. You can activate it by setting a positive value for the number of captured packets. You can do that on either a global basis or on a per-socket basis (via a setsockopt call). There is no way to get the packets out of the kernel other than using kmem or getting a coredump. I thought that would help some of the legal/privacy concerns regarding such a feature. However, it should be possible to add a future effort to export them in PCAP format. I tested this at low scale, and found that there were no mbuf leaks and the peak mbuf usage appeared to be unchanged with and without the feature. The main performance concern I can envision is the number of mbufs that would be used on systems with a large number of sockets. If you save five packets per direction per socket and have 3,000 sockets, that will consume at least 30,000 mbufs just to keep these packets. I tried to reduce the concerns associated with this by limiting the number of clusters (not mbufs) that could be used for this feature. Again, in my testing, that appears to work correctly. Differential Revision: D3100 Submitted by: Jonathan Looney <jlooney at juniper dot net> Reviewed by: gnn, hiren
2015-10-14 00:35:37 +00:00
#ifdef TCPPCAP
/* Save segment, if requested. */
tcp_pcap_add(th, m, &(tp->t_inpkts));
#endif
1994-05-24 10:09:53 +00:00
/*
* Segment received on connection.
* Reset idle time and keep-alive timer.
* XXX: This should be done after segment
* validation to ignore broken/spoofed segs.
1994-05-24 10:09:53 +00:00
*/
tp->t_rcvtime = ticks;
if (TCPS_HAVEESTABLISHED(tp->t_state))
tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp));
1994-05-24 10:09:53 +00:00
/*
* Scale up the window into a 32-bit value.
* For the SYN_SENT state the scale is zero.
*/
tiwin = th->th_win << tp->snd_scale;
/*
* TCP ECN processing.
*/
if (tp->t_flags & TF_ECN_PERMIT) {
if (thflags & TH_CWR)
tp->t_flags &= ~TF_ECN_SND_ECE;
switch (iptos & IPTOS_ECN_MASK) {
case IPTOS_ECN_CE:
tp->t_flags |= TF_ECN_SND_ECE;
TCPSTAT_INC(tcps_ecn_ce);
break;
case IPTOS_ECN_ECT0:
TCPSTAT_INC(tcps_ecn_ect0);
break;
case IPTOS_ECN_ECT1:
TCPSTAT_INC(tcps_ecn_ect1);
break;
}
/* Process a packet differently from RFC3168. */
cc_ecnpkt_handler(tp, th, iptos);
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
/* Congestion experienced. */
if (thflags & TH_ECE) {
cc_cong_signal(tp, th, CC_ECN);
}
}
/*
* Parse options on any incoming segment.
*/
tcp_dooptions(&to, (u_char *)(th + 1),
(th->th_off << 2) - sizeof(struct tcphdr),
(thflags & TH_SYN) ? TO_SYN : 0);
/*
* If echoed timestamp is later than the current time,
Rewrite of TCP syncookies to remove locking requirements and to enhance functionality: - Remove a rwlock aquisition/release per generated syncookie. Locking is now integrated with the bucket row locking of syncache itself and syncookies no longer add any additional lock overhead. - Syncookie secrets are different for and stored per syncache buck row. Secrets expire after 16 seconds and are reseeded on-demand. - The computational overhead for syncookie generation and verification is one MD5 hash computation as before. - Syncache can be turned off and run with syncookies only by setting the sysctl net.inet.tcp.syncookies_only=1. This implementation extends the orginal idea and first implementation of FreeBSD by using not only the initial sequence number field to store information but also the timestamp field if present. This way we can keep track of the entire state we need to know to recreate the session in its original form. Almost all TCP speakers implement RFC1323 timestamps these days. For those that do not we still have to live with the known shortcomings of the ISN only SYN cookies. The use of the timestamp field causes the timestamps to be randomized if syncookies are enabled. The idea of SYN cookies is to encode and include all necessary information about the connection setup state within the SYN-ACK we send back and thus to get along without keeping any local state until the ACK to the SYN-ACK arrives (if ever). Everything we need to know should be available from the information we encoded in the SYN-ACK. A detailed description of the inner working of the syncookies mechanism is included in the comments in tcp_syncache.c. Reviewed by: silby (slightly earlier version) Sponsored by: TCP/IP Optimization Fundraise 2005
2006-09-13 13:08:27 +00:00
* fall back to non RFC1323 RTT calculation. Normalize
* timestamp if syncookies were used when this connection
* was established.
*/
Rewrite of TCP syncookies to remove locking requirements and to enhance functionality: - Remove a rwlock aquisition/release per generated syncookie. Locking is now integrated with the bucket row locking of syncache itself and syncookies no longer add any additional lock overhead. - Syncookie secrets are different for and stored per syncache buck row. Secrets expire after 16 seconds and are reseeded on-demand. - The computational overhead for syncookie generation and verification is one MD5 hash computation as before. - Syncache can be turned off and run with syncookies only by setting the sysctl net.inet.tcp.syncookies_only=1. This implementation extends the orginal idea and first implementation of FreeBSD by using not only the initial sequence number field to store information but also the timestamp field if present. This way we can keep track of the entire state we need to know to recreate the session in its original form. Almost all TCP speakers implement RFC1323 timestamps these days. For those that do not we still have to live with the known shortcomings of the ISN only SYN cookies. The use of the timestamp field causes the timestamps to be randomized if syncookies are enabled. The idea of SYN cookies is to encode and include all necessary information about the connection setup state within the SYN-ACK we send back and thus to get along without keeping any local state until the ACK to the SYN-ACK arrives (if ever). Everything we need to know should be available from the information we encoded in the SYN-ACK. A detailed description of the inner working of the syncookies mechanism is included in the comments in tcp_syncache.c. Reviewed by: silby (slightly earlier version) Sponsored by: TCP/IP Optimization Fundraise 2005
2006-09-13 13:08:27 +00:00
if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
to.to_tsecr -= tp->ts_offset;
if (TSTMP_GT(to.to_tsecr, tcp_ts_getticks()))
Rewrite of TCP syncookies to remove locking requirements and to enhance functionality: - Remove a rwlock aquisition/release per generated syncookie. Locking is now integrated with the bucket row locking of syncache itself and syncookies no longer add any additional lock overhead. - Syncookie secrets are different for and stored per syncache buck row. Secrets expire after 16 seconds and are reseeded on-demand. - The computational overhead for syncookie generation and verification is one MD5 hash computation as before. - Syncache can be turned off and run with syncookies only by setting the sysctl net.inet.tcp.syncookies_only=1. This implementation extends the orginal idea and first implementation of FreeBSD by using not only the initial sequence number field to store information but also the timestamp field if present. This way we can keep track of the entire state we need to know to recreate the session in its original form. Almost all TCP speakers implement RFC1323 timestamps these days. For those that do not we still have to live with the known shortcomings of the ISN only SYN cookies. The use of the timestamp field causes the timestamps to be randomized if syncookies are enabled. The idea of SYN cookies is to encode and include all necessary information about the connection setup state within the SYN-ACK we send back and thus to get along without keeping any local state until the ACK to the SYN-ACK arrives (if ever). Everything we need to know should be available from the information we encoded in the SYN-ACK. A detailed description of the inner working of the syncookies mechanism is included in the comments in tcp_syncache.c. Reviewed by: silby (slightly earlier version) Sponsored by: TCP/IP Optimization Fundraise 2005
2006-09-13 13:08:27 +00:00
to.to_tsecr = 0;
}
/*
* If timestamps were negotiated during SYN/ACK they should
* appear on every segment during this session and vice versa.
*/
if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS)) {
if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
log(LOG_DEBUG, "%s; %s: Timestamp missing, "
"no action\n", s, __func__);
free(s, M_TCPLOG);
}
}
if (!(tp->t_flags & TF_RCVD_TSTMP) && (to.to_flags & TOF_TS)) {
if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
"no action\n", s, __func__);
free(s, M_TCPLOG);
}
}
1994-05-24 10:09:53 +00:00
/*
* Process options only when we get SYN/ACK back. The SYN case
* for incoming connections is handled in tcp_syncache.
* According to RFC1323 the window field in a SYN (i.e., a <SYN>
* or <SYN,ACK>) segment itself is never scaled.
* XXX this is traditional behavior, may need to be cleaned up.
1994-05-24 10:09:53 +00:00
*/
if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
if ((to.to_flags & TOF_SCALE) &&
(tp->t_flags & TF_REQ_SCALE)) {
tp->t_flags |= TF_RCVD_SCALE;
tp->snd_scale = to.to_wscale;
}
/*
* Initial send window. It will be updated with
* the next incoming segment to the scaled value.
*/
tp->snd_wnd = th->th_win;
if (to.to_flags & TOF_TS) {
tp->t_flags |= TF_RCVD_TSTMP;
tp->ts_recent = to.to_tsval;
tp->ts_recent_age = tcp_ts_getticks();
}
if (to.to_flags & TOF_MSS)
tcp_mss(tp, to.to_mss);
if ((tp->t_flags & TF_SACK_PERMIT) &&
(to.to_flags & TOF_SACKPERM) == 0)
tp->t_flags &= ~TF_SACK_PERMIT;
}
1995-05-30 08:16:23 +00:00
/*
1994-05-24 10:09:53 +00:00
* Header prediction: check for the two common cases
* of a uni-directional data xfer. If the packet has
* no control flags, is in-sequence, the window didn't
* change and we're not retransmitting, it's a
* candidate. If the length is zero and the ack moved
* forward, we're the sender side of the xfer. Just
* free the data acked & wake any higher level process
* that was blocked waiting for space. If the length
* is non-zero and the ack didn't move, we're the
* receiver side. If we're getting packets in-order
* (the reassembly queue is empty), add the data to
* the socket buffer and note that we need a delayed ack.
* Make sure that the hidden state-flags are also off.
* Since we check for TCPS_ESTABLISHED first, it can only
* be TH_NEEDSYN.
1994-05-24 10:09:53 +00:00
*/
if (tp->t_state == TCPS_ESTABLISHED &&
th->th_seq == tp->rcv_nxt &&
(thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK &&
tp->snd_nxt == tp->snd_max &&
tiwin && tiwin == tp->snd_wnd &&
((tp->t_flags & (TF_NEEDSYN|TF_NEEDFIN)) == 0) &&
LIST_EMPTY(&tp->t_segq) &&
((to.to_flags & TOF_TS) == 0 ||
TSTMP_GEQ(to.to_tsval, tp->ts_recent)) ) {
1994-05-24 10:09:53 +00:00
1995-05-30 08:16:23 +00:00
/*
1994-05-24 10:09:53 +00:00
* If last ACK falls within this segment's sequence numbers,
* record the timestamp.
* NOTE that the test is modified according to the latest
* proposal of the tcplw@cray.com list (Braden 1993/04/26).
1994-05-24 10:09:53 +00:00
*/
if ((to.to_flags & TOF_TS) != 0 &&
SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
tp->ts_recent_age = tcp_ts_getticks();
tp->ts_recent = to.to_tsval;
1994-05-24 10:09:53 +00:00
}
if (tlen == 0) {
if (SEQ_GT(th->th_ack, tp->snd_una) &&
SEQ_LEQ(th->th_ack, tp->snd_max) &&
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
!IN_RECOVERY(tp->t_flags) &&
(to.to_flags & TOF_SACK) == 0 &&
TAILQ_EMPTY(&tp->snd_holes)) {
1994-05-24 10:09:53 +00:00
/*
* This is a pure ack for outstanding data.
1994-05-24 10:09:53 +00:00
*/
if (ti_locked == TI_RLOCKED)
INP_INFO_RUNLOCK(&V_tcbinfo);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
ti_locked = TI_UNLOCKED;
TCPSTAT_INC(tcps_predack);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
/*
* "bad retransmit" recovery.
*/
if (tp->t_rxtshift == 1 &&
TCP reuses t_rxtshift to determine the backoff timer used for both the persist state and the retransmit timer. However, the code that implements "bad retransmit recovery" only checks t_rxtshift to see if an ACK has been received in during the first retransmit timeout window. As a result, if ticks has wrapped over to a negative value and a socket is in the persist state, it can incorrectly treat an ACK from the remote peer as a "bad retransmit recovery" and restore saved values such as snd_ssthresh and snd_cwnd. However, if the socket has never had a retransmit timeout, then these saved values will be zero, so snd_ssthresh and snd_cwnd will be set to 0. If the socket is in fast recovery (this can be caused by excessive duplicate ACKs such as those fixed by 220794), then each ACK that arrives triggers either NewReno or SACK partial ACK handling which clamps snd_cwnd to be no larger than snd_ssthresh. In effect, the socket's send window is permamently stuck at 0 even though the remote peer is advertising a much larger window and pending data is only sent via TCP window probes (so one byte every few seconds). Fix this by adding a new TCP pcb flag (TF_PREVVALID) that indicates that the various snd_*_prev fields in the pcb are valid and only perform "bad retransmit recovery" if this flag is set in the pcb. The flag is set on the first retransmit timeout that occurs and is cleared on subsequent retransmit timeouts or when entering the persist state. Reviewed by: bz MFC after: 2 weeks
2011-04-29 15:40:12 +00:00
tp->t_flags & TF_PREVVALID &&
(int)(ticks - tp->t_badrxtwin) < 0) {
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
cc_cong_signal(tp, th, CC_RTO_ERR);
}
/*
* Recalculate the transmit timer / rtt.
*
* Some boxes send broken timestamp replies
* during the SYN+ACK phase, ignore
* timestamps of 0 or we could calculate a
* huge RTT and blow up the retransmit timer.
*/
if ((to.to_flags & TOF_TS) != 0 &&
to.to_tsecr) {
u_int t;
t = tcp_ts_getticks() - to.to_tsecr;
if (!tp->t_rttlow || tp->t_rttlow > t)
tp->t_rttlow = t;
tcp_xmit_timer(tp,
TCP_TS_TO_TICKS(t) + 1);
} else if (tp->t_rtttime &&
SEQ_GT(th->th_ack, tp->t_rtseq)) {
if (!tp->t_rttlow ||
tp->t_rttlow > ticks - tp->t_rtttime)
tp->t_rttlow = ticks - tp->t_rtttime;
tcp_xmit_timer(tp,
ticks - tp->t_rtttime);
}
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
acked = BYTES_THIS_ACK(tp, th);
/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
hhook_run_tcp_est_in(tp, th, &to);
TCPSTAT_INC(tcps_rcvackpack);
TCPSTAT_ADD(tcps_rcvackbyte, acked);
1994-05-24 10:09:53 +00:00
sbdrop(&so->so_snd, acked);
if (SEQ_GT(tp->snd_una, tp->snd_recover) &&
SEQ_LEQ(th->th_ack, tp->snd_recover))
tp->snd_recover = th->th_ack - 1;
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
/*
* Let the congestion control algorithm update
* congestion control related information. This
* typically means increasing the congestion
* window.
*/
cc_ack_received(tp, th, CC_ACK);
tp->snd_una = th->th_ack;
/*
* Pull snd_wl2 up to prevent seq wrap relative
* to th_ack.
*/
tp->snd_wl2 = th->th_ack;
tp->t_dupacks = 0;
1994-05-24 10:09:53 +00:00
m_freem(m);
/*
* If all outstanding data are acked, stop
* retransmit timer, otherwise restart timer
* using current (possibly backed-off) value.
* If process is waiting for space,
* wakeup/selwakeup/signal. If data
* are ready to send, let tcp_output
* decide between more output or persist.
*/
#ifdef TCPDEBUG
if (so->so_options & SO_DEBUG)
tcp_trace(TA_INPUT, ostate, tp,
(void *)tcp_saveipgen,
&tcp_savetcp, 0);
#endif
TCP_PROBE3(debug__input, tp, th,
mtod(m, const char *));
1994-05-24 10:09:53 +00:00
if (tp->snd_una == tp->snd_max)
tcp_timer_activate(tp, TT_REXMT, 0);
else if (!tcp_timer_active(tp, TT_PERSIST))
tcp_timer_activate(tp, TT_REXMT,
tp->t_rxtcur);
sowwakeup(so);
if (sbavail(&so->so_snd))
(void) tp->t_fb->tfb_tcp_output(tp);
goto check_delack;
1994-05-24 10:09:53 +00:00
}
} else if (th->th_ack == tp->snd_una &&
tlen <= sbspace(&so->so_rcv)) {
int newsize = 0; /* automatic sockbuf scaling */
1994-05-24 10:09:53 +00:00
/*
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
* This is a pure, in-sequence data packet with
* nothing on the reassembly queue and we have enough
* buffer space to take it.
1994-05-24 10:09:53 +00:00
*/
if (ti_locked == TI_RLOCKED)
INP_INFO_RUNLOCK(&V_tcbinfo);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
ti_locked = TI_UNLOCKED;
/* Clean receiver SACK report if present */
if ((tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks)
tcp_clean_sackreport(tp);
TCPSTAT_INC(tcps_preddat);
tp->rcv_nxt += tlen;
/*
* Pull snd_wl1 up to prevent seq wrap relative to
* th_seq.
*/
tp->snd_wl1 = th->th_seq;
/*
* Pull rcv_up up to prevent seq wrap relative to
* rcv_nxt.
*/
tp->rcv_up = tp->rcv_nxt;
TCPSTAT_INC(tcps_rcvpack);
TCPSTAT_ADD(tcps_rcvbyte, tlen);
#ifdef TCPDEBUG
if (so->so_options & SO_DEBUG)
tcp_trace(TA_INPUT, ostate, tp,
(void *)tcp_saveipgen, &tcp_savetcp, 0);
#endif
TCP_PROBE3(debug__input, tp, th, mtod(m, const char *));
/*
* Automatic sizing of receive socket buffer. Often the send
* buffer size is not optimally adjusted to the actual network
* conditions at hand (delay bandwidth product). Setting the
* buffer size too small limits throughput on links with high
* bandwidth and high delay (eg. trans-continental/oceanic links).
*
* On the receive side the socket buffer memory is only rarely
* used to any significant extent. This allows us to be much
* more aggressive in scaling the receive socket buffer. For
* the case that the buffer space is actually used to a large
* extent and we run out of kernel memory we can simply drop
* the new segments; TCP on the sender will just retransmit it
* later. Setting the buffer size too big may only consume too
* much kernel memory if the application doesn't read() from
* the socket or packet loss or reordering makes use of the
* reassembly queue.
*
* The criteria to step up the receive buffer one notch are:
* 1. Application has not set receive buffer size with
* SO_RCVBUF. Setting SO_RCVBUF clears SB_AUTOSIZE.
* 2. the number of bytes received during the time it takes
* one timestamp to be reflected back to us (the RTT);
* 3. received bytes per RTT is within seven eighth of the
* current socket buffer size;
* 4. receive buffer size has not hit maximal automatic size;
*
* This algorithm does one step per RTT at most and only if
* we receive a bulk stream w/o packet losses or reorderings.
* Shrinking the buffer during idle times is not necessary as
* it doesn't consume any memory when idle.
*
* TODO: Only step up if the application is actually serving
* the buffer to better manage the socket buffer resources.
*/
if (V_tcp_do_autorcvbuf &&
(to.to_flags & TOF_TS) &&
to.to_tsecr &&
(so->so_rcv.sb_flags & SB_AUTOSIZE)) {
if (TSTMP_GT(to.to_tsecr, tp->rfbuf_ts) &&
to.to_tsecr - tp->rfbuf_ts < hz) {
if (tp->rfbuf_cnt >
(so->so_rcv.sb_hiwat / 8 * 7) &&
so->so_rcv.sb_hiwat <
V_tcp_autorcvbuf_max) {
newsize =
min(so->so_rcv.sb_hiwat +
V_tcp_autorcvbuf_inc,
V_tcp_autorcvbuf_max);
}
/* Start over with next RTT. */
tp->rfbuf_ts = 0;
tp->rfbuf_cnt = 0;
} else
tp->rfbuf_cnt += tlen; /* add up */
}
/* Add data to socket buffer. */
SOCKBUF_LOCK(&so->so_rcv);
if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
m_freem(m);
} else {
/*
* Set new socket buffer size.
* Give up when limit is reached.
*/
if (newsize)
if (!sbreserve_locked(&so->so_rcv,
newsize, so, NULL))
so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
m_adj(m, drop_hdrlen); /* delayed header drop */
sbappendstream_locked(&so->so_rcv, m, 0);
}
/* NB: sorwakeup_locked() does an implicit unlock. */
sorwakeup_locked(so);
if (DELAY_ACK(tp, tlen)) {
tp->t_flags |= TF_DELACK;
} else {
tp->t_flags |= TF_ACKNOW;
tp->t_fb->tfb_tcp_output(tp);
}
goto check_delack;
1994-05-24 10:09:53 +00:00
}
}
/*
* Calculate amount of space in receive window,
* and then do TCP input processing.
* Receive window is amount of space in rcv queue,
* but not less than advertised window.
*/
win = sbspace(&so->so_rcv);
if (win < 0)
win = 0;
tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt));
1994-05-24 10:09:53 +00:00
/* Reset receive buffer auto scaling when not in bulk receive mode. */
tp->rfbuf_ts = 0;
tp->rfbuf_cnt = 0;
1994-05-24 10:09:53 +00:00
switch (tp->t_state) {
/*
* If the state is SYN_RECEIVED:
* if seg contains an ACK, but not for our SYN/ACK, send a RST.
*/
case TCPS_SYN_RECEIVED:
if ((thflags & TH_ACK) &&
(SEQ_LEQ(th->th_ack, tp->snd_una) ||
SEQ_GT(th->th_ack, tp->snd_max))) {
rstreason = BANDLIM_RST_OPENPORT;
goto dropwithreset;
}
#ifdef TCP_RFC7413
if (tp->t_flags & TF_FASTOPEN) {
/*
* When a TFO connection is in SYN_RECEIVED, the
* only valid packets are the initial SYN, a
* retransmit/copy of the initial SYN (possibly with
* a subset of the original data), a valid ACK, a
* FIN, or a RST.
*/
if ((thflags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK)) {
rstreason = BANDLIM_RST_OPENPORT;
goto dropwithreset;
} else if (thflags & TH_SYN) {
/* non-initial SYN is ignored */
if ((tcp_timer_active(tp, TT_DELACK) ||
tcp_timer_active(tp, TT_REXMT)))
goto drop;
} else if (!(thflags & (TH_ACK|TH_FIN|TH_RST))) {
goto drop;
}
}
#endif
break;
1994-05-24 10:09:53 +00:00
/*
* If the state is SYN_SENT:
* if seg contains an ACK, but not for our SYN, drop the input.
* if seg contains a RST, then drop the connection.
* if seg does not contain SYN, then drop it.
* Otherwise this is an acceptable SYN segment
* initialize tp->rcv_nxt and tp->irs
* if seg contains ack then advance tp->snd_una
* if seg contains an ECE and ECN support is enabled, the stream
* is ECN capable.
1994-05-24 10:09:53 +00:00
* if SYN has been acked change to ESTABLISHED else SYN_RCVD state
* arrange for segment to be acked (eventually)
* continue processing rest of data/controls, beginning with URG
*/
case TCPS_SYN_SENT:
if ((thflags & TH_ACK) &&
(SEQ_LEQ(th->th_ack, tp->iss) ||
SEQ_GT(th->th_ack, tp->snd_max))) {
rstreason = BANDLIM_UNLIMITED;
goto dropwithreset;
}
if ((thflags & (TH_ACK|TH_RST)) == (TH_ACK|TH_RST)) {
TCP_PROBE5(connect__refused, NULL, tp,
mtod(m, const char *), tp, th);
tp = tcp_drop(tp, ECONNREFUSED);
}
if (thflags & TH_RST)
1994-05-24 10:09:53 +00:00
goto drop;
if (!(thflags & TH_SYN))
1994-05-24 10:09:53 +00:00
goto drop;
tp->irs = th->th_seq;
1994-05-24 10:09:53 +00:00
tcp_rcvseqinit(tp);
if (thflags & TH_ACK) {
TCPSTAT_INC(tcps_connects);
soisconnected(so);
#ifdef MAC
mac_socketpeer_set_from_mbuf(m, so);
#endif
/* Do window scaling on this connection? */
if ((tp->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) ==
(TF_RCVD_SCALE|TF_REQ_SCALE)) {
tp->rcv_scale = tp->request_r_scale;
}
tp->rcv_adv += imin(tp->rcv_wnd,
TCP_MAXWIN << tp->rcv_scale);
tp->snd_una++; /* SYN is acked */
/*
* If there's data, delay ACK; if there's also a FIN
* ACKNOW will be turned on later.
*/
if (DELAY_ACK(tp, tlen) && tlen != 0)
tcp_timer_activate(tp, TT_DELACK,
tcp_delacktime);
else
tp->t_flags |= TF_ACKNOW;
if ((thflags & TH_ECE) && V_tcp_do_ecn) {
tp->t_flags |= TF_ECN_PERMIT;
TCPSTAT_INC(tcps_ecn_shs);
}
/*
* Received <SYN,ACK> in SYN_SENT[*] state.
* Transitions:
* SYN_SENT --> ESTABLISHED
* SYN_SENT* --> FIN_WAIT_1
*/
tp->t_starttime = ticks;
if (tp->t_flags & TF_NEEDFIN) {
tcp_state_change(tp, TCPS_FIN_WAIT_1);
tp->t_flags &= ~TF_NEEDFIN;
thflags &= ~TH_SYN;
} else {
tcp_state_change(tp, TCPS_ESTABLISHED);
TCP_PROBE5(connect__established, NULL, tp,
mtod(m, const char *), tp, th);
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
cc_conn_init(tp);
tcp_timer_activate(tp, TT_KEEP,
TP_KEEPIDLE(tp));
}
} else {
/*
* Received initial SYN in SYN-SENT[*] state =>
* simultaneous open.
* If it succeeds, connection is * half-synchronized.
* Otherwise, do 3-way handshake:
* SYN-SENT -> SYN-RECEIVED
* SYN-SENT* -> SYN-RECEIVED*
*/
tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
tcp_timer_activate(tp, TT_REXMT, 0);
tcp_state_change(tp, TCPS_SYN_RECEIVED);
1995-05-30 08:16:23 +00:00
}
1994-05-24 10:09:53 +00:00
KASSERT(ti_locked == TI_RLOCKED, ("%s: trimthenstep6: "
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
"ti_locked %d", __func__, ti_locked));
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
INP_WLOCK_ASSERT(tp->t_inpcb);
1994-05-24 10:09:53 +00:00
/*
* Advance th->th_seq to correspond to first data byte.
1994-05-24 10:09:53 +00:00
* If data, trim to stay within window,
* dropping FIN if necessary.
*/
th->th_seq++;
if (tlen > tp->rcv_wnd) {
todrop = tlen - tp->rcv_wnd;
1994-05-24 10:09:53 +00:00
m_adj(m, -todrop);
tlen = tp->rcv_wnd;
thflags &= ~TH_FIN;
TCPSTAT_INC(tcps_rcvpackafterwin);
TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
1994-05-24 10:09:53 +00:00
}
tp->snd_wl1 = th->th_seq - 1;
tp->rcv_up = th->th_seq;
/*
* Client side of transaction: already sent SYN and data.
* If the remote host used T/TCP to validate the SYN,
* our data will be ACK'd; if so, enter normal data segment
* processing in the middle of step 5, ack processing.
* Otherwise, goto step 6.
1995-05-30 08:16:23 +00:00
*/
if (thflags & TH_ACK)
goto process_ACK;
1994-05-24 10:09:53 +00:00
goto step6;
/*
* If the state is LAST_ACK or CLOSING or TIME_WAIT:
* do normal processing.
*
* NB: Leftover from RFC1644 T/TCP. Cases to be reused later.
1995-05-30 08:16:23 +00:00
*/
case TCPS_LAST_ACK:
case TCPS_CLOSING:
break; /* continue normal processing */
1994-05-24 10:09:53 +00:00
}
/*
* States other than LISTEN or SYN_SENT.
* First check the RST flag and sequence number since reset segments
* are exempt from the timestamp and connection count tests. This
* fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix
* below which allowed reset segments in half the sequence space
* to fall though and be processed (which gives forged reset
* segments with a random sequence number a 50 percent chance of
* killing a connection).
* Then check timestamp, if present.
* Then check the connection count, if present.
1995-05-30 08:16:23 +00:00
* Then check that at least some bytes of segment are within
1994-05-24 10:09:53 +00:00
* receive window. If segment begins before rcv_nxt,
* drop leading data (and SYN); if nothing left, just ack.
*/
if (thflags & TH_RST) {
FreeBSD-SA-14:19.tcp raised attention to the state of our stack towards blind SYN/RST spoofed attack. Originally our stack used in-window checks for incoming SYN/RST as proposed by RFC793. Later, circa 2003 the RST attack was mitigated using the technique described in P. Watson "Slipping in the window" paper [1]. After that, the checks were only relaxed for the sake of compatibility with some buggy TCP stacks. First, r192912 introduced the vulnerability, just fixed by aforementioned SA. Second, r167310 had slightly relaxed the default RST checks, instead of utilizing net.inet.tcp.insecure_rst sysctl. In 2010 a new technique for mitigation of these attacks was proposed in RFC5961 [2]. The idea is to send a "challenge ACK" packet to the peer, to verify that packet arrived isn't spoofed. If peer receives challenge ACK it should regenerate its RST or SYN with correct sequence number. This should not only protect against attacks, but also improve communication with broken stacks, so authors of reverted r167310 and r192912 won't be disappointed. [1] http://bandwidthco.com/whitepapers/netforensics/tcpip/TCP Reset Attacks.pdf [2] http://www.rfc-editor.org/rfc/rfc5961.txt Changes made: o Revert r167310. o Implement "challenge ACK" protection as specificed in RFC5961 against RST attack. On by default. - Carefully preserve r138098, which handles empty window edge case, not described by the RFC. - Update net.inet.tcp.insecure_rst description. o Implement "challenge ACK" protection as specificed in RFC5961 against SYN attack. On by default. - Provide net.inet.tcp.insecure_syn sysctl, to turn off RFC5961 protection. The changes were tested at Netflix. The tested box didn't show any anomalies compared to control box, except slightly increased number of TCP connection in LAST_ACK state. Reviewed by: rrs Sponsored by: Netflix Sponsored by: Nginx, Inc.
2014-09-16 11:07:25 +00:00
/*
* RFC5961 Section 3.2
*
* - RST drops connection only if SEG.SEQ == RCV.NXT.
* - If RST is in window, we send challenge ACK.
*
* Note: to take into account delayed ACKs, we should
* test against last_ack_sent instead of rcv_nxt.
* Note 2: we handle special case of closed window, not
* covered by the RFC.
*/
if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
(tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) {
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
KASSERT(ti_locked == TI_RLOCKED,
FreeBSD-SA-14:19.tcp raised attention to the state of our stack towards blind SYN/RST spoofed attack. Originally our stack used in-window checks for incoming SYN/RST as proposed by RFC793. Later, circa 2003 the RST attack was mitigated using the technique described in P. Watson "Slipping in the window" paper [1]. After that, the checks were only relaxed for the sake of compatibility with some buggy TCP stacks. First, r192912 introduced the vulnerability, just fixed by aforementioned SA. Second, r167310 had slightly relaxed the default RST checks, instead of utilizing net.inet.tcp.insecure_rst sysctl. In 2010 a new technique for mitigation of these attacks was proposed in RFC5961 [2]. The idea is to send a "challenge ACK" packet to the peer, to verify that packet arrived isn't spoofed. If peer receives challenge ACK it should regenerate its RST or SYN with correct sequence number. This should not only protect against attacks, but also improve communication with broken stacks, so authors of reverted r167310 and r192912 won't be disappointed. [1] http://bandwidthco.com/whitepapers/netforensics/tcpip/TCP Reset Attacks.pdf [2] http://www.rfc-editor.org/rfc/rfc5961.txt Changes made: o Revert r167310. o Implement "challenge ACK" protection as specificed in RFC5961 against RST attack. On by default. - Carefully preserve r138098, which handles empty window edge case, not described by the RFC. - Update net.inet.tcp.insecure_rst description. o Implement "challenge ACK" protection as specificed in RFC5961 against SYN attack. On by default. - Provide net.inet.tcp.insecure_syn sysctl, to turn off RFC5961 protection. The changes were tested at Netflix. The tested box didn't show any anomalies compared to control box, except slightly increased number of TCP connection in LAST_ACK state. Reviewed by: rrs Sponsored by: Netflix Sponsored by: Nginx, Inc.
2014-09-16 11:07:25 +00:00
("%s: TH_RST ti_locked %d, th %p tp %p",
__func__, ti_locked, th, tp));
KASSERT(tp->t_state != TCPS_SYN_SENT,
("%s: TH_RST for TCPS_SYN_SENT th %p tp %p",
__func__, th, tp));
if (V_tcp_insecure_rst ||
tp->last_ack_sent == th->th_seq) {
TCPSTAT_INC(tcps_drops);
FreeBSD-SA-14:19.tcp raised attention to the state of our stack towards blind SYN/RST spoofed attack. Originally our stack used in-window checks for incoming SYN/RST as proposed by RFC793. Later, circa 2003 the RST attack was mitigated using the technique described in P. Watson "Slipping in the window" paper [1]. After that, the checks were only relaxed for the sake of compatibility with some buggy TCP stacks. First, r192912 introduced the vulnerability, just fixed by aforementioned SA. Second, r167310 had slightly relaxed the default RST checks, instead of utilizing net.inet.tcp.insecure_rst sysctl. In 2010 a new technique for mitigation of these attacks was proposed in RFC5961 [2]. The idea is to send a "challenge ACK" packet to the peer, to verify that packet arrived isn't spoofed. If peer receives challenge ACK it should regenerate its RST or SYN with correct sequence number. This should not only protect against attacks, but also improve communication with broken stacks, so authors of reverted r167310 and r192912 won't be disappointed. [1] http://bandwidthco.com/whitepapers/netforensics/tcpip/TCP Reset Attacks.pdf [2] http://www.rfc-editor.org/rfc/rfc5961.txt Changes made: o Revert r167310. o Implement "challenge ACK" protection as specificed in RFC5961 against RST attack. On by default. - Carefully preserve r138098, which handles empty window edge case, not described by the RFC. - Update net.inet.tcp.insecure_rst description. o Implement "challenge ACK" protection as specificed in RFC5961 against SYN attack. On by default. - Provide net.inet.tcp.insecure_syn sysctl, to turn off RFC5961 protection. The changes were tested at Netflix. The tested box didn't show any anomalies compared to control box, except slightly increased number of TCP connection in LAST_ACK state. Reviewed by: rrs Sponsored by: Netflix Sponsored by: Nginx, Inc.
2014-09-16 11:07:25 +00:00
/* Drop the connection. */
switch (tp->t_state) {
case TCPS_SYN_RECEIVED:
so->so_error = ECONNREFUSED;
goto close;
case TCPS_ESTABLISHED:
case TCPS_FIN_WAIT_1:
case TCPS_FIN_WAIT_2:
case TCPS_CLOSE_WAIT:
so->so_error = ECONNRESET;
close:
tcp_state_change(tp, TCPS_CLOSED);
/* FALLTHROUGH */
default:
tp = tcp_close(tp);
}
} else {
TCPSTAT_INC(tcps_badrst);
/* Send challenge ACK. */
tcp_respond(tp, mtod(m, void *), th, m,
tp->rcv_nxt, tp->snd_nxt, TH_ACK);
tp->last_ack_sent = tp->rcv_nxt;
m = NULL;
}
}
goto drop;
}
FreeBSD-SA-14:19.tcp raised attention to the state of our stack towards blind SYN/RST spoofed attack. Originally our stack used in-window checks for incoming SYN/RST as proposed by RFC793. Later, circa 2003 the RST attack was mitigated using the technique described in P. Watson "Slipping in the window" paper [1]. After that, the checks were only relaxed for the sake of compatibility with some buggy TCP stacks. First, r192912 introduced the vulnerability, just fixed by aforementioned SA. Second, r167310 had slightly relaxed the default RST checks, instead of utilizing net.inet.tcp.insecure_rst sysctl. In 2010 a new technique for mitigation of these attacks was proposed in RFC5961 [2]. The idea is to send a "challenge ACK" packet to the peer, to verify that packet arrived isn't spoofed. If peer receives challenge ACK it should regenerate its RST or SYN with correct sequence number. This should not only protect against attacks, but also improve communication with broken stacks, so authors of reverted r167310 and r192912 won't be disappointed. [1] http://bandwidthco.com/whitepapers/netforensics/tcpip/TCP Reset Attacks.pdf [2] http://www.rfc-editor.org/rfc/rfc5961.txt Changes made: o Revert r167310. o Implement "challenge ACK" protection as specificed in RFC5961 against RST attack. On by default. - Carefully preserve r138098, which handles empty window edge case, not described by the RFC. - Update net.inet.tcp.insecure_rst description. o Implement "challenge ACK" protection as specificed in RFC5961 against SYN attack. On by default. - Provide net.inet.tcp.insecure_syn sysctl, to turn off RFC5961 protection. The changes were tested at Netflix. The tested box didn't show any anomalies compared to control box, except slightly increased number of TCP connection in LAST_ACK state. Reviewed by: rrs Sponsored by: Netflix Sponsored by: Nginx, Inc.
2014-09-16 11:07:25 +00:00
/*
* RFC5961 Section 4.2
* Send challenge ACK for any SYN in synchronized state.
*/
if ((thflags & TH_SYN) && tp->t_state != TCPS_SYN_SENT &&
tp->t_state != TCPS_SYN_RECEIVED) {
KASSERT(ti_locked == TI_RLOCKED,
FreeBSD-SA-14:19.tcp raised attention to the state of our stack towards blind SYN/RST spoofed attack. Originally our stack used in-window checks for incoming SYN/RST as proposed by RFC793. Later, circa 2003 the RST attack was mitigated using the technique described in P. Watson "Slipping in the window" paper [1]. After that, the checks were only relaxed for the sake of compatibility with some buggy TCP stacks. First, r192912 introduced the vulnerability, just fixed by aforementioned SA. Second, r167310 had slightly relaxed the default RST checks, instead of utilizing net.inet.tcp.insecure_rst sysctl. In 2010 a new technique for mitigation of these attacks was proposed in RFC5961 [2]. The idea is to send a "challenge ACK" packet to the peer, to verify that packet arrived isn't spoofed. If peer receives challenge ACK it should regenerate its RST or SYN with correct sequence number. This should not only protect against attacks, but also improve communication with broken stacks, so authors of reverted r167310 and r192912 won't be disappointed. [1] http://bandwidthco.com/whitepapers/netforensics/tcpip/TCP Reset Attacks.pdf [2] http://www.rfc-editor.org/rfc/rfc5961.txt Changes made: o Revert r167310. o Implement "challenge ACK" protection as specificed in RFC5961 against RST attack. On by default. - Carefully preserve r138098, which handles empty window edge case, not described by the RFC. - Update net.inet.tcp.insecure_rst description. o Implement "challenge ACK" protection as specificed in RFC5961 against SYN attack. On by default. - Provide net.inet.tcp.insecure_syn sysctl, to turn off RFC5961 protection. The changes were tested at Netflix. The tested box didn't show any anomalies compared to control box, except slightly increased number of TCP connection in LAST_ACK state. Reviewed by: rrs Sponsored by: Netflix Sponsored by: Nginx, Inc.
2014-09-16 11:07:25 +00:00
("tcp_do_segment: TH_SYN ti_locked %d", ti_locked));
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
FreeBSD-SA-14:19.tcp raised attention to the state of our stack towards blind SYN/RST spoofed attack. Originally our stack used in-window checks for incoming SYN/RST as proposed by RFC793. Later, circa 2003 the RST attack was mitigated using the technique described in P. Watson "Slipping in the window" paper [1]. After that, the checks were only relaxed for the sake of compatibility with some buggy TCP stacks. First, r192912 introduced the vulnerability, just fixed by aforementioned SA. Second, r167310 had slightly relaxed the default RST checks, instead of utilizing net.inet.tcp.insecure_rst sysctl. In 2010 a new technique for mitigation of these attacks was proposed in RFC5961 [2]. The idea is to send a "challenge ACK" packet to the peer, to verify that packet arrived isn't spoofed. If peer receives challenge ACK it should regenerate its RST or SYN with correct sequence number. This should not only protect against attacks, but also improve communication with broken stacks, so authors of reverted r167310 and r192912 won't be disappointed. [1] http://bandwidthco.com/whitepapers/netforensics/tcpip/TCP Reset Attacks.pdf [2] http://www.rfc-editor.org/rfc/rfc5961.txt Changes made: o Revert r167310. o Implement "challenge ACK" protection as specificed in RFC5961 against RST attack. On by default. - Carefully preserve r138098, which handles empty window edge case, not described by the RFC. - Update net.inet.tcp.insecure_rst description. o Implement "challenge ACK" protection as specificed in RFC5961 against SYN attack. On by default. - Provide net.inet.tcp.insecure_syn sysctl, to turn off RFC5961 protection. The changes were tested at Netflix. The tested box didn't show any anomalies compared to control box, except slightly increased number of TCP connection in LAST_ACK state. Reviewed by: rrs Sponsored by: Netflix Sponsored by: Nginx, Inc.
2014-09-16 11:07:25 +00:00
TCPSTAT_INC(tcps_badsyn);
if (V_tcp_insecure_syn &&
SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
tp = tcp_drop(tp, ECONNRESET);
rstreason = BANDLIM_UNLIMITED;
} else {
/* Send challenge ACK. */
tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt,
tp->snd_nxt, TH_ACK);
tp->last_ack_sent = tp->rcv_nxt;
m = NULL;
}
goto drop;
}
/*
1994-05-24 10:09:53 +00:00
* RFC 1323 PAWS: If we have a timestamp reply on this segment
* and it's less than ts_recent, drop it.
*/
if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent &&
TSTMP_LT(to.to_tsval, tp->ts_recent)) {
1994-05-24 10:09:53 +00:00
/* Check to see if ts_recent is over 24 days old. */
if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) {
1994-05-24 10:09:53 +00:00
/*
* Invalidate ts_recent. If this segment updates
* ts_recent, the age will be reset later and ts_recent
* will get a valid value. If it does not, setting
* ts_recent to zero will at least satisfy the
* requirement that zero be placed in the timestamp
* echo reply when ts_recent isn't valid. The
* age isn't reset until we get a valid ts_recent
* because we don't want out-of-order segments to be
* dropped when ts_recent is old.
*/
tp->ts_recent = 0;
} else {
TCPSTAT_INC(tcps_rcvduppack);
TCPSTAT_ADD(tcps_rcvdupbyte, tlen);
TCPSTAT_INC(tcps_pawsdrop);
if (tlen)
2002-12-17 00:24:48 +00:00
goto dropafterack;
goto drop;
1994-05-24 10:09:53 +00:00
}
}
/*
* In the SYN-RECEIVED state, validate that the packet belongs to
* this connection before trimming the data to fit the receive
* window. Check the sequence number versus IRS since we know
* the sequence numbers haven't wrapped. This is a partial fix
* for the "LAND" DoS attack.
*/
if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) {
rstreason = BANDLIM_RST_OPENPORT;
goto dropwithreset;
}
todrop = tp->rcv_nxt - th->th_seq;
1994-05-24 10:09:53 +00:00
if (todrop > 0) {
if (thflags & TH_SYN) {
thflags &= ~TH_SYN;
th->th_seq++;
if (th->th_urp > 1)
th->th_urp--;
1994-05-24 10:09:53 +00:00
else
thflags &= ~TH_URG;
1994-05-24 10:09:53 +00:00
todrop--;
}
/*
* Following if statement from Stevens, vol. 2, p. 960.
*/
if (todrop > tlen
|| (todrop == tlen && (thflags & TH_FIN) == 0)) {
1994-05-24 10:09:53 +00:00
/*
* Any valid FIN must be to the left of the window.
* At this point the FIN must be a duplicate or out
* of sequence; drop it.
1994-05-24 10:09:53 +00:00
*/
thflags &= ~TH_FIN;
/*
* Send an ACK to resynchronize and drop any data.
* But keep on processing for RST or ACK.
*/
tp->t_flags |= TF_ACKNOW;
todrop = tlen;
TCPSTAT_INC(tcps_rcvduppack);
TCPSTAT_ADD(tcps_rcvdupbyte, todrop);
1994-05-24 10:09:53 +00:00
} else {
TCPSTAT_INC(tcps_rcvpartduppack);
TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop);
1994-05-24 10:09:53 +00:00
}
drop_hdrlen += todrop; /* drop from the top afterwards */
th->th_seq += todrop;
tlen -= todrop;
if (th->th_urp > todrop)
th->th_urp -= todrop;
1994-05-24 10:09:53 +00:00
else {
thflags &= ~TH_URG;
th->th_urp = 0;
1994-05-24 10:09:53 +00:00
}
}
/*
* If new data are received on a connection after the
* user processes are gone, then RST the other end.
*/
if ((so->so_state & SS_NOFDREF) &&
tp->t_state > TCPS_CLOSE_WAIT && tlen) {
KASSERT(ti_locked == TI_RLOCKED, ("%s: SS_NOFDEREF && "
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
"CLOSE_WAIT && tlen ti_locked %d", __func__, ti_locked));
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
log(LOG_DEBUG, "%s; %s: %s: Received %d bytes of data "
"after socket was closed, "
"sending RST and removing tcpcb\n",
s, __func__, tcpstates[tp->t_state], tlen);
free(s, M_TCPLOG);
}
1994-05-24 10:09:53 +00:00
tp = tcp_close(tp);
TCPSTAT_INC(tcps_rcvafterclose);
rstreason = BANDLIM_UNLIMITED;
1994-05-24 10:09:53 +00:00
goto dropwithreset;
}
/*
* If segment ends after window, drop trailing data
* (and PUSH and FIN); if nothing left, just ACK.
*/
todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
1994-05-24 10:09:53 +00:00
if (todrop > 0) {
TCPSTAT_INC(tcps_rcvpackafterwin);
if (todrop >= tlen) {
TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen);
1994-05-24 10:09:53 +00:00
/*
* If window is closed can only take segments at
* window edge, and have to drop data and PUSH from
* incoming segments. Continue processing, but
* remember to ack. Otherwise, drop segment
* and ack.
*/
if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
1994-05-24 10:09:53 +00:00
tp->t_flags |= TF_ACKNOW;
TCPSTAT_INC(tcps_rcvwinprobe);
1994-05-24 10:09:53 +00:00
} else
goto dropafterack;
} else
TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
1994-05-24 10:09:53 +00:00
m_adj(m, -todrop);
tlen -= todrop;
thflags &= ~(TH_PUSH|TH_FIN);
1994-05-24 10:09:53 +00:00
}
/*
* If last ACK falls within this segment's sequence numbers,
* record its timestamp.
* NOTE:
* 1) That the test incorporates suggestions from the latest
* proposal of the tcplw@cray.com list (Braden 1993/04/26).
* 2) That updating only on newer timestamps interferes with
* our earlier PAWS tests, so this check should be solely
* predicated on the sequence space of this segment.
* 3) That we modify the segment boundary check to be
* Last.ACK.Sent <= SEG.SEQ + SEG.Len
* instead of RFC1323's
* Last.ACK.Sent < SEG.SEQ + SEG.Len,
* This modified check allows us to overcome RFC1323's
* limitations as described in Stevens TCP/IP Illustrated
* Vol. 2 p.869. In such cases, we can still calculate the
* RTT correctly when RCV.NXT == Last.ACK.Sent.
1994-05-24 10:09:53 +00:00
*/
if ((to.to_flags & TOF_TS) != 0 &&
SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
((thflags & (TH_SYN|TH_FIN)) != 0))) {
tp->ts_recent_age = tcp_ts_getticks();
tp->ts_recent = to.to_tsval;
1994-05-24 10:09:53 +00:00
}
/*
* If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN
* flag is on (half-synchronized state), then queue data for
* later processing; else drop segment and return.
*/
if ((thflags & TH_ACK) == 0) {
if (tp->t_state == TCPS_SYN_RECEIVED ||
(tp->t_flags & TF_NEEDSYN)) {
#ifdef TCP_RFC7413
if (tp->t_state == TCPS_SYN_RECEIVED &&
tp->t_flags & TF_FASTOPEN) {
tp->snd_wnd = tiwin;
cc_conn_init(tp);
}
#endif
goto step6;
} else if (tp->t_flags & TF_ACKNOW)
goto dropafterack;
else
goto drop;
}
1995-05-30 08:16:23 +00:00
1994-05-24 10:09:53 +00:00
/*
* Ack processing.
*/
switch (tp->t_state) {
/*
* In SYN_RECEIVED state, the ack ACKs our SYN, so enter
* ESTABLISHED state and continue processing.
* The ACK was checked above.
1994-05-24 10:09:53 +00:00
*/
case TCPS_SYN_RECEIVED:
TCPSTAT_INC(tcps_connects);
1994-05-24 10:09:53 +00:00
soisconnected(so);
/* Do window scaling? */
if ((tp->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) ==
(TF_RCVD_SCALE|TF_REQ_SCALE)) {
tp->rcv_scale = tp->request_r_scale;
tp->snd_wnd = tiwin;
1994-05-24 10:09:53 +00:00
}
/*
1995-05-30 08:16:23 +00:00
* Make transitions:
* SYN-RECEIVED -> ESTABLISHED
* SYN-RECEIVED* -> FIN-WAIT-1
*/
tp->t_starttime = ticks;
if (tp->t_flags & TF_NEEDFIN) {
tcp_state_change(tp, TCPS_FIN_WAIT_1);
tp->t_flags &= ~TF_NEEDFIN;
} else {
tcp_state_change(tp, TCPS_ESTABLISHED);
TCP_PROBE5(accept__established, NULL, tp,
mtod(m, const char *), tp, th);
#ifdef TCP_RFC7413
if (tp->t_tfo_pending) {
tcp_fastopen_decrement_counter(tp->t_tfo_pending);
tp->t_tfo_pending = NULL;
/*
* Account for the ACK of our SYN prior to
* regular ACK processing below.
*/
tp->snd_una++;
}
/*
* TFO connections call cc_conn_init() during SYN
* processing. Calling it again here for such
* connections is not harmless as it would undo the
* snd_cwnd reduction that occurs when a TFO SYN|ACK
* is retransmitted.
*/
if (!(tp->t_flags & TF_FASTOPEN))
#endif
cc_conn_init(tp);
tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp));
}
1995-05-30 08:16:23 +00:00
/*
* If segment contains data or ACK, will call tcp_reass()
* later; if not, do so now to pass queued data to user.
*/
if (tlen == 0 && (thflags & TH_FIN) == 0)
(void) tcp_reass(tp, (struct tcphdr *)0, 0,
(struct mbuf *)0);
tp->snd_wl1 = th->th_seq - 1;
/* FALLTHROUGH */
1994-05-24 10:09:53 +00:00
/*
* In ESTABLISHED state: drop duplicate ACKs; ACK out of range
* ACKs. If the ack is in the range
* tp->snd_una < th->th_ack <= tp->snd_max
* then advance tp->snd_una to th->th_ack and drop
1994-05-24 10:09:53 +00:00
* data from the retransmission queue. If this ACK reflects
* more up to date window information we update our window information.
*/
case TCPS_ESTABLISHED:
case TCPS_FIN_WAIT_1:
case TCPS_FIN_WAIT_2:
case TCPS_CLOSE_WAIT:
case TCPS_CLOSING:
case TCPS_LAST_ACK:
if (SEQ_GT(th->th_ack, tp->snd_max)) {
TCPSTAT_INC(tcps_rcvacktoomuch);
goto dropafterack;
}
if ((tp->t_flags & TF_SACK_PERMIT) &&
((to.to_flags & TOF_SACK) ||
!TAILQ_EMPTY(&tp->snd_holes)))
One of the ways to detect loss is to count duplicate acks coming back from the other end till it reaches predetermined threshold which is 3 for us right now. Once that happens, we trigger fast-retransmit to do loss recovery. Main problem with the current implementation is that we don't honor SACK information well to detect whether an incoming ack is a dupack or not. RFC6675 has latest recommendations for that. According to it, dupack is a segment that arrives carrying a SACK block that identifies previously unknown information between snd_una and snd_max even if it carries new data, changes the advertised window, or moves the cumulative acknowledgment point. With the prevalence of Selective ACK (SACK) these days, improper handling can lead to delayed loss recovery. With the fix, new behavior looks like following: 0) th_ack < snd_una --> ignore Old acks are ignored. 1) th_ack == snd_una, !sack_changed --> ignore Acks with SACK enabled but without any new SACK info in them are ignored. 2) th_ack == snd_una, window == old_window --> increment Increment on a good dupack. 3) th_ack == snd_una, window != old_window, sack_changed --> increment When SACK enabled, it's okay to have advertized window changed if the ack has new SACK info. 4) th_ack > snd_una --> reset to 0 Reset to 0 when left edge moves. 5) th_ack > snd_una, sack_changed --> increment Increment if left edge moves but there is new SACK info. Here, sack_changed is the indicator that incoming ack has previously unknown SACK info in it. Note: This fix is not fully compliant to RFC6675. That may require a few changes to current implementation in order to keep per-sackhole dupack counter and change to the way we mark/handle sack holes. PR: 203663 Reviewed by: jtl MFC after: 3 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D4225
2015-12-08 21:21:48 +00:00
sack_changed = tcp_sack_doack(tp, &to, th->th_ack);
Calculate the correct amount of bytes that are in-flight for a connection as suggested by RFC 6675. Currently differnt places in the stack tries to guess this in suboptimal ways. The main problem is that current calculations don't take sacked bytes into account. Sacked bytes are the bytes receiver acked via SACK option. This is suboptimal because it assumes that network has more outstanding (unacked) bytes than the actual value and thus sends less data by setting congestion window lower than what's possible which in turn may cause slower recovery from losses. As an example, one of the current calculations looks something like this: snd_nxt - snd_fack + sackhint.sack_bytes_rexmit New proposal from RFC 6675 is: snd_max - snd_una - sackhint.sacked_bytes + sackhint.sack_bytes_rexmit which takes sacked bytes into account which is a new addition to the sackhint struct. Only thing we are missing from RFC 6675 is isLost() i.e. segment being considered lost and thus adjusting pipe based on that which makes this calculation a bit on conservative side. The approach is very simple. We already process each ack with sack info in tcp_sack_doack() and extract sack blocks/holes out of it. We'd now also track this new variable sacked_bytes which keeps track of total sacked bytes reported. One downside to this approach is that we may get incorrect count of sacked_bytes if the other end decides to drop sack info in the ack because of memory pressure or some other reasons. But in this (not very likely) case also the pipe calculation would be conservative which is okay as opposed to being aggressive in sending packets into the network. Next step is to use this more accurate pipe estimation to drive congestion window adjustments. In collaboration with: rrs Reviewed by: jason_eggnet dot com, rrs MFC after: 2 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D3971
2015-10-28 22:57:51 +00:00
else
/*
* Reset the value so that previous (valid) value
* from the last ack with SACK doesn't get used.
*/
tp->sackhint.sacked_bytes = 0;
/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
hhook_run_tcp_est_in(tp, th, &to);
if (SEQ_LEQ(th->th_ack, tp->snd_una)) {
u_int maxseg;
maxseg = tcp_maxseg(tp);
One of the ways to detect loss is to count duplicate acks coming back from the other end till it reaches predetermined threshold which is 3 for us right now. Once that happens, we trigger fast-retransmit to do loss recovery. Main problem with the current implementation is that we don't honor SACK information well to detect whether an incoming ack is a dupack or not. RFC6675 has latest recommendations for that. According to it, dupack is a segment that arrives carrying a SACK block that identifies previously unknown information between snd_una and snd_max even if it carries new data, changes the advertised window, or moves the cumulative acknowledgment point. With the prevalence of Selective ACK (SACK) these days, improper handling can lead to delayed loss recovery. With the fix, new behavior looks like following: 0) th_ack < snd_una --> ignore Old acks are ignored. 1) th_ack == snd_una, !sack_changed --> ignore Acks with SACK enabled but without any new SACK info in them are ignored. 2) th_ack == snd_una, window == old_window --> increment Increment on a good dupack. 3) th_ack == snd_una, window != old_window, sack_changed --> increment When SACK enabled, it's okay to have advertized window changed if the ack has new SACK info. 4) th_ack > snd_una --> reset to 0 Reset to 0 when left edge moves. 5) th_ack > snd_una, sack_changed --> increment Increment if left edge moves but there is new SACK info. Here, sack_changed is the indicator that incoming ack has previously unknown SACK info in it. Note: This fix is not fully compliant to RFC6675. That may require a few changes to current implementation in order to keep per-sackhole dupack counter and change to the way we mark/handle sack holes. PR: 203663 Reviewed by: jtl MFC after: 3 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D4225
2015-12-08 21:21:48 +00:00
if (tlen == 0 &&
(tiwin == tp->snd_wnd ||
(tp->t_flags & TF_SACK_PERMIT))) {
/*
* If this is the first time we've seen a
* FIN from the remote, this is not a
* duplicate and it needs to be processed
* normally. This happens during a
* simultaneous close.
*/
if ((thflags & TH_FIN) &&
(TCPS_HAVERCVDFIN(tp->t_state) == 0)) {
tp->t_dupacks = 0;
break;
}
TCPSTAT_INC(tcps_rcvdupack);
1994-05-24 10:09:53 +00:00
/*
* If we have outstanding data (other than
* a window probe), this is a completely
* duplicate ack (ie, window info didn't
* change and FIN isn't set),
* the ack is the biggest we've
1994-05-24 10:09:53 +00:00
* seen and we've seen exactly our rexmt
* threshhold of them, assume a packet
* has been dropped and retransmit it.
* Kludge snd_nxt & the congestion
* window so we send only this one
* packet.
*
* We know we're losing at the current
* window size so do congestion avoidance
* (set ssthresh to half the current window
* and pull our congestion window back to
* the new ssthresh).
*
* Dup acks mean that packets have left the
1995-05-30 08:16:23 +00:00
* network (they're now cached at the receiver)
1994-05-24 10:09:53 +00:00
* so bump cwnd by the amount in the receiver
* to keep a constant cwnd packets in the
* network.
*
* When using TCP ECN, notify the peer that
* we reduced the cwnd.
1994-05-24 10:09:53 +00:00
*/
One of the ways to detect loss is to count duplicate acks coming back from the other end till it reaches predetermined threshold which is 3 for us right now. Once that happens, we trigger fast-retransmit to do loss recovery. Main problem with the current implementation is that we don't honor SACK information well to detect whether an incoming ack is a dupack or not. RFC6675 has latest recommendations for that. According to it, dupack is a segment that arrives carrying a SACK block that identifies previously unknown information between snd_una and snd_max even if it carries new data, changes the advertised window, or moves the cumulative acknowledgment point. With the prevalence of Selective ACK (SACK) these days, improper handling can lead to delayed loss recovery. With the fix, new behavior looks like following: 0) th_ack < snd_una --> ignore Old acks are ignored. 1) th_ack == snd_una, !sack_changed --> ignore Acks with SACK enabled but without any new SACK info in them are ignored. 2) th_ack == snd_una, window == old_window --> increment Increment on a good dupack. 3) th_ack == snd_una, window != old_window, sack_changed --> increment When SACK enabled, it's okay to have advertized window changed if the ack has new SACK info. 4) th_ack > snd_una --> reset to 0 Reset to 0 when left edge moves. 5) th_ack > snd_una, sack_changed --> increment Increment if left edge moves but there is new SACK info. Here, sack_changed is the indicator that incoming ack has previously unknown SACK info in it. Note: This fix is not fully compliant to RFC6675. That may require a few changes to current implementation in order to keep per-sackhole dupack counter and change to the way we mark/handle sack holes. PR: 203663 Reviewed by: jtl MFC after: 3 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D4225
2015-12-08 21:21:48 +00:00
/*
* Following 2 kinds of acks should not affect
* dupack counting:
* 1) Old acks
* 2) Acks with SACK but without any new SACK
* information in them. These could result from
* any anomaly in the network like a switch
* duplicating packets or a possible DoS attack.
*/
if (th->th_ack != tp->snd_una ||
((tp->t_flags & TF_SACK_PERMIT) &&
!sack_changed))
break;
else if (!tcp_timer_active(tp, TT_REXMT))
1994-05-24 10:09:53 +00:00
tp->t_dupacks = 0;
else if (++tp->t_dupacks > tcprexmtthresh ||
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
IN_FASTRECOVERY(tp->t_flags)) {
cc_ack_received(tp, th, CC_DUPACK);
if ((tp->t_flags & TF_SACK_PERMIT) &&
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
IN_FASTRECOVERY(tp->t_flags)) {
int awnd;
/*
* Compute the amount of data in flight first.
* We can inject new data into the pipe iff
2011-01-07 21:40:34 +00:00
* we have less than 1/2 the original window's
* worth of data in flight.
*/
Calculate the correct amount of bytes that are in-flight for a connection as suggested by RFC 6675. Currently differnt places in the stack tries to guess this in suboptimal ways. The main problem is that current calculations don't take sacked bytes into account. Sacked bytes are the bytes receiver acked via SACK option. This is suboptimal because it assumes that network has more outstanding (unacked) bytes than the actual value and thus sends less data by setting congestion window lower than what's possible which in turn may cause slower recovery from losses. As an example, one of the current calculations looks something like this: snd_nxt - snd_fack + sackhint.sack_bytes_rexmit New proposal from RFC 6675 is: snd_max - snd_una - sackhint.sacked_bytes + sackhint.sack_bytes_rexmit which takes sacked bytes into account which is a new addition to the sackhint struct. Only thing we are missing from RFC 6675 is isLost() i.e. segment being considered lost and thus adjusting pipe based on that which makes this calculation a bit on conservative side. The approach is very simple. We already process each ack with sack info in tcp_sack_doack() and extract sack blocks/holes out of it. We'd now also track this new variable sacked_bytes which keeps track of total sacked bytes reported. One downside to this approach is that we may get incorrect count of sacked_bytes if the other end decides to drop sack info in the ack because of memory pressure or some other reasons. But in this (not very likely) case also the pipe calculation would be conservative which is okay as opposed to being aggressive in sending packets into the network. Next step is to use this more accurate pipe estimation to drive congestion window adjustments. In collaboration with: rrs Reviewed by: jason_eggnet dot com, rrs MFC after: 2 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D3971
2015-10-28 22:57:51 +00:00
if (V_tcp_do_rfc6675_pipe)
awnd = tcp_compute_pipe(tp);
else
awnd = (tp->snd_nxt - tp->snd_fack) +
tp->sackhint.sack_bytes_rexmit;
if (awnd < tp->snd_ssthresh) {
tp->snd_cwnd += maxseg;
if (tp->snd_cwnd > tp->snd_ssthresh)
tp->snd_cwnd = tp->snd_ssthresh;
}
} else
tp->snd_cwnd += maxseg;
(void) tp->t_fb->tfb_tcp_output(tp);
goto drop;
} else if (tp->t_dupacks == tcprexmtthresh) {
1994-05-24 10:09:53 +00:00
tcp_seq onxt = tp->snd_nxt;
/*
* If we're doing sack, check to
* see if we're already in sack
* recovery. If we're not doing sack,
* check to see if we're in newreno
* recovery.
*/
if (tp->t_flags & TF_SACK_PERMIT) {
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
if (IN_FASTRECOVERY(tp->t_flags)) {
tp->t_dupacks = 0;
break;
}
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
} else {
if (SEQ_LEQ(th->th_ack,
tp->snd_recover)) {
tp->t_dupacks = 0;
break;
}
}
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
/* Congestion signal before ack. */
cc_cong_signal(tp, th, CC_NDUPACK);
cc_ack_received(tp, th, CC_DUPACK);
tcp_timer_activate(tp, TT_REXMT, 0);
tp->t_rtttime = 0;
if (tp->t_flags & TF_SACK_PERMIT) {
TCPSTAT_INC(
tcps_sack_recovery_episode);
tp->sack_newdata = tp->snd_nxt;
tp->snd_cwnd = maxseg;
(void) tp->t_fb->tfb_tcp_output(tp);
goto drop;
}
tp->snd_nxt = th->th_ack;
tp->snd_cwnd = maxseg;
(void) tp->t_fb->tfb_tcp_output(tp);
KASSERT(tp->snd_limited <= 2,
("%s: tp->snd_limited too big",
__func__));
1994-05-24 10:09:53 +00:00
tp->snd_cwnd = tp->snd_ssthresh +
maxseg *
(tp->t_dupacks - tp->snd_limited);
1994-05-24 10:09:53 +00:00
if (SEQ_GT(onxt, tp->snd_nxt))
tp->snd_nxt = onxt;
goto drop;
} else if (V_tcp_do_rfc3042) {
/*
* Process first and second duplicate
* ACKs. Each indicates a segment
* leaving the network, creating room
* for more. Make sure we can send a
* packet on reception of each duplicate
* ACK by increasing snd_cwnd by one
* segment. Restore the original
* snd_cwnd after packet transmission.
*/
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
cc_ack_received(tp, th, CC_DUPACK);
u_long oldcwnd = tp->snd_cwnd;
tcp_seq oldsndmax = tp->snd_max;
u_int sent;
int avail;
KASSERT(tp->t_dupacks == 1 ||
tp->t_dupacks == 2,
("%s: dupacks not 1 or 2",
__func__));
if (tp->t_dupacks == 1)
tp->snd_limited = 0;
tp->snd_cwnd =
(tp->snd_nxt - tp->snd_una) +
(tp->t_dupacks - tp->snd_limited) *
maxseg;
/*
* Only call tcp_output when there
* is new data available to be sent.
* Otherwise we would send pure ACKs.
*/
SOCKBUF_LOCK(&so->so_snd);
avail = sbavail(&so->so_snd) -
(tp->snd_nxt - tp->snd_una);
SOCKBUF_UNLOCK(&so->so_snd);
if (avail > 0)
(void) tp->t_fb->tfb_tcp_output(tp);
sent = tp->snd_max - oldsndmax;
if (sent > maxseg) {
KASSERT((tp->t_dupacks == 2 &&
tp->snd_limited == 0) ||
(sent == maxseg + 1 &&
tp->t_flags & TF_SENTFIN),
("%s: sent too much",
__func__));
tp->snd_limited = 2;
} else if (sent > 0)
++tp->snd_limited;
tp->snd_cwnd = oldcwnd;
goto drop;
1994-05-24 10:09:53 +00:00
}
One of the ways to detect loss is to count duplicate acks coming back from the other end till it reaches predetermined threshold which is 3 for us right now. Once that happens, we trigger fast-retransmit to do loss recovery. Main problem with the current implementation is that we don't honor SACK information well to detect whether an incoming ack is a dupack or not. RFC6675 has latest recommendations for that. According to it, dupack is a segment that arrives carrying a SACK block that identifies previously unknown information between snd_una and snd_max even if it carries new data, changes the advertised window, or moves the cumulative acknowledgment point. With the prevalence of Selective ACK (SACK) these days, improper handling can lead to delayed loss recovery. With the fix, new behavior looks like following: 0) th_ack < snd_una --> ignore Old acks are ignored. 1) th_ack == snd_una, !sack_changed --> ignore Acks with SACK enabled but without any new SACK info in them are ignored. 2) th_ack == snd_una, window == old_window --> increment Increment on a good dupack. 3) th_ack == snd_una, window != old_window, sack_changed --> increment When SACK enabled, it's okay to have advertized window changed if the ack has new SACK info. 4) th_ack > snd_una --> reset to 0 Reset to 0 when left edge moves. 5) th_ack > snd_una, sack_changed --> increment Increment if left edge moves but there is new SACK info. Here, sack_changed is the indicator that incoming ack has previously unknown SACK info in it. Note: This fix is not fully compliant to RFC6675. That may require a few changes to current implementation in order to keep per-sackhole dupack counter and change to the way we mark/handle sack holes. PR: 203663 Reviewed by: jtl MFC after: 3 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D4225
2015-12-08 21:21:48 +00:00
}
1994-05-24 10:09:53 +00:00
break;
One of the ways to detect loss is to count duplicate acks coming back from the other end till it reaches predetermined threshold which is 3 for us right now. Once that happens, we trigger fast-retransmit to do loss recovery. Main problem with the current implementation is that we don't honor SACK information well to detect whether an incoming ack is a dupack or not. RFC6675 has latest recommendations for that. According to it, dupack is a segment that arrives carrying a SACK block that identifies previously unknown information between snd_una and snd_max even if it carries new data, changes the advertised window, or moves the cumulative acknowledgment point. With the prevalence of Selective ACK (SACK) these days, improper handling can lead to delayed loss recovery. With the fix, new behavior looks like following: 0) th_ack < snd_una --> ignore Old acks are ignored. 1) th_ack == snd_una, !sack_changed --> ignore Acks with SACK enabled but without any new SACK info in them are ignored. 2) th_ack == snd_una, window == old_window --> increment Increment on a good dupack. 3) th_ack == snd_una, window != old_window, sack_changed --> increment When SACK enabled, it's okay to have advertized window changed if the ack has new SACK info. 4) th_ack > snd_una --> reset to 0 Reset to 0 when left edge moves. 5) th_ack > snd_una, sack_changed --> increment Increment if left edge moves but there is new SACK info. Here, sack_changed is the indicator that incoming ack has previously unknown SACK info in it. Note: This fix is not fully compliant to RFC6675. That may require a few changes to current implementation in order to keep per-sackhole dupack counter and change to the way we mark/handle sack holes. PR: 203663 Reviewed by: jtl MFC after: 3 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D4225
2015-12-08 21:21:48 +00:00
} else {
/*
* This ack is advancing the left edge, reset the
* counter.
*/
tp->t_dupacks = 0;
/*
* If this ack also has new SACK info, increment the
* counter as per rfc6675.
*/
if ((tp->t_flags & TF_SACK_PERMIT) && sack_changed)
tp->t_dupacks++;
1994-05-24 10:09:53 +00:00
}
KASSERT(SEQ_GT(th->th_ack, tp->snd_una),
("%s: th_ack <= snd_una", __func__));
1994-05-24 10:09:53 +00:00
/*
* If the congestion window was inflated to account
* for the other side's cached packets, retract it.
*/
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
if (IN_FASTRECOVERY(tp->t_flags)) {
if (SEQ_LT(th->th_ack, tp->snd_recover)) {
if (tp->t_flags & TF_SACK_PERMIT)
tcp_sack_partialack(tp, th);
else
tcp_newreno_partial_ack(tp, th);
} else
cc_post_recovery(tp, th);
}
/*
* If we reach this point, ACK is not a duplicate,
* i.e., it ACKs something we sent.
*/
if (tp->t_flags & TF_NEEDSYN) {
1995-05-30 08:16:23 +00:00
/*
* T/TCP: Connection was half-synchronized, and our
* SYN has been ACK'd (so connection is now fully
* synchronized). Go to non-starred state,
* increment snd_una for ACK of SYN, and check if
* we can do window scaling.
*/
tp->t_flags &= ~TF_NEEDSYN;
tp->snd_una++;
/* Do window scaling? */
if ((tp->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) ==
(TF_RCVD_SCALE|TF_REQ_SCALE)) {
tp->rcv_scale = tp->request_r_scale;
/* Send window already scaled. */
}
}
process_ACK:
INP_WLOCK_ASSERT(tp->t_inpcb);
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
acked = BYTES_THIS_ACK(tp, th);
TCPSTAT_INC(tcps_rcvackpack);
TCPSTAT_ADD(tcps_rcvackbyte, acked);
1994-05-24 10:09:53 +00:00
/*
* If we just performed our first retransmit, and the ACK
* arrives within our recovery window, then it was a mistake
* to do the retransmit in the first place. Recover our
* original cwnd and ssthresh, and proceed to transmit where
* we left off.
*/
TCP reuses t_rxtshift to determine the backoff timer used for both the persist state and the retransmit timer. However, the code that implements "bad retransmit recovery" only checks t_rxtshift to see if an ACK has been received in during the first retransmit timeout window. As a result, if ticks has wrapped over to a negative value and a socket is in the persist state, it can incorrectly treat an ACK from the remote peer as a "bad retransmit recovery" and restore saved values such as snd_ssthresh and snd_cwnd. However, if the socket has never had a retransmit timeout, then these saved values will be zero, so snd_ssthresh and snd_cwnd will be set to 0. If the socket is in fast recovery (this can be caused by excessive duplicate ACKs such as those fixed by 220794), then each ACK that arrives triggers either NewReno or SACK partial ACK handling which clamps snd_cwnd to be no larger than snd_ssthresh. In effect, the socket's send window is permamently stuck at 0 even though the remote peer is advertising a much larger window and pending data is only sent via TCP window probes (so one byte every few seconds). Fix this by adding a new TCP pcb flag (TF_PREVVALID) that indicates that the various snd_*_prev fields in the pcb are valid and only perform "bad retransmit recovery" if this flag is set in the pcb. The flag is set on the first retransmit timeout that occurs and is cleared on subsequent retransmit timeouts or when entering the persist state. Reviewed by: bz MFC after: 2 weeks
2011-04-29 15:40:12 +00:00
if (tp->t_rxtshift == 1 && tp->t_flags & TF_PREVVALID &&
(int)(ticks - tp->t_badrxtwin) < 0)
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
cc_cong_signal(tp, th, CC_RTO_ERR);
1994-05-24 10:09:53 +00:00
/*
* If we have a timestamp reply, update smoothed
* round trip time. If no timestamp is present but
* transmit timer is running and timed sequence
* number was acked, update smoothed round trip time.
* Since we now have an rtt measurement, cancel the
* timer backoff (cf., Phil Karn's retransmit alg.).
* Recompute the initial retransmit timer.
*
* Some boxes send broken timestamp replies
* during the SYN+ACK phase, ignore
* timestamps of 0 or we could calculate a
* huge RTT and blow up the retransmit timer.
1994-05-24 10:09:53 +00:00
*/
if ((to.to_flags & TOF_TS) != 0 && to.to_tsecr) {
u_int t;
t = tcp_ts_getticks() - to.to_tsecr;
if (!tp->t_rttlow || tp->t_rttlow > t)
tp->t_rttlow = t;
tcp_xmit_timer(tp, TCP_TS_TO_TICKS(t) + 1);
} else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq)) {
if (!tp->t_rttlow || tp->t_rttlow > ticks - tp->t_rtttime)
tp->t_rttlow = ticks - tp->t_rtttime;
tcp_xmit_timer(tp, ticks - tp->t_rtttime);
}
1994-05-24 10:09:53 +00:00
/*
* If all outstanding data is acked, stop retransmit
* timer and remember to restart (more output or persist).
* If there is more data to be acked, restart retransmit
* timer, using current (possibly backed-off) value.
*/
if (th->th_ack == tp->snd_max) {
tcp_timer_activate(tp, TT_REXMT, 0);
1994-05-24 10:09:53 +00:00
needoutput = 1;
} else if (!tcp_timer_active(tp, TT_PERSIST))
tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
/*
* If no data (only SYN) was ACK'd,
* skip rest of ACK processing.
*/
if (acked == 0)
goto step6;
1994-05-24 10:09:53 +00:00
/*
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
* Let the congestion control algorithm update congestion
* control related information. This typically means increasing
* the congestion window.
1994-05-24 10:09:53 +00:00
*/
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
cc_ack_received(tp, th, CC_ACK);
SOCKBUF_LOCK(&so->so_snd);
if (acked > sbavail(&so->so_snd)) {
tp->snd_wnd -= sbavail(&so->so_snd);
mfree = sbcut_locked(&so->so_snd,
(int)sbavail(&so->so_snd));
1994-05-24 10:09:53 +00:00
ourfinisacked = 1;
} else {
mfree = sbcut_locked(&so->so_snd, acked);
tp->snd_wnd -= acked;
1994-05-24 10:09:53 +00:00
ourfinisacked = 0;
}
/* NB: sowwakeup_locked() does an implicit unlock. */
sowwakeup_locked(so);
m_freem(mfree);
/* Detect una wraparound. */
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
if (!IN_RECOVERY(tp->t_flags) &&
SEQ_GT(tp->snd_una, tp->snd_recover) &&
SEQ_LEQ(th->th_ack, tp->snd_recover))
tp->snd_recover = th->th_ack - 1;
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
/* XXXLAS: Can this be moved up into cc_post_recovery? */
if (IN_RECOVERY(tp->t_flags) &&
SEQ_GEQ(th->th_ack, tp->snd_recover)) {
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
EXIT_RECOVERY(tp->t_flags);
}
tp->snd_una = th->th_ack;
if (tp->t_flags & TF_SACK_PERMIT) {
if (SEQ_GT(tp->snd_una, tp->snd_recover))
tp->snd_recover = tp->snd_una;
}
1994-05-24 10:09:53 +00:00
if (SEQ_LT(tp->snd_nxt, tp->snd_una))
tp->snd_nxt = tp->snd_una;
switch (tp->t_state) {
/*
* In FIN_WAIT_1 STATE in addition to the processing
* for the ESTABLISHED state if our FIN is now acknowledged
* then enter FIN_WAIT_2.
*/
case TCPS_FIN_WAIT_1:
if (ourfinisacked) {
/*
* If we can't receive any more
* data, then closing user can proceed.
* Starting the timer is contrary to the
* specification, but if we don't get a FIN
* we'll hang forever.
*
* XXXjl:
* we should release the tp also, and use a
* compressed state.
1994-05-24 10:09:53 +00:00
*/
if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
soisdisconnected(so);
tcp_timer_activate(tp, TT_2MSL,
(tcp_fast_finwait2_recycle ?
tcp_finwait2_timeout :
TP_MAXIDLE(tp)));
}
tcp_state_change(tp, TCPS_FIN_WAIT_2);
1994-05-24 10:09:53 +00:00
}
break;
/*
1994-05-24 10:09:53 +00:00
* In CLOSING STATE in addition to the processing for
* the ESTABLISHED state if the ACK acknowledges our FIN
* then enter the TIME-WAIT state, otherwise ignore
* the segment.
*/
case TCPS_CLOSING:
if (ourfinisacked) {
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
tcp_twstart(tp);
INP_INFO_RUNLOCK(&V_tcbinfo);
m_freem(m);
return;
1994-05-24 10:09:53 +00:00
}
break;
/*
* In LAST_ACK, we may still be waiting for data to drain
* and/or to be acked, as well as for the ack of our FIN.
* If our FIN is now acknowledged, delete the TCB,
* enter the closed state and return.
*/
case TCPS_LAST_ACK:
if (ourfinisacked) {
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1994-05-24 10:09:53 +00:00
tp = tcp_close(tp);
goto drop;
}
break;
}
}
step6:
INP_WLOCK_ASSERT(tp->t_inpcb);
1994-05-24 10:09:53 +00:00
/*
* Update window information.
* Don't look at window if no ACK: TAC's send garbage on first SYN.
1994-05-24 10:09:53 +00:00
*/
if ((thflags & TH_ACK) &&
(SEQ_LT(tp->snd_wl1, th->th_seq) ||
(tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
(tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
/* keep track of pure window updates */
if (tlen == 0 &&
tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
TCPSTAT_INC(tcps_rcvwinupd);
1994-05-24 10:09:53 +00:00
tp->snd_wnd = tiwin;
tp->snd_wl1 = th->th_seq;
tp->snd_wl2 = th->th_ack;
1994-05-24 10:09:53 +00:00
if (tp->snd_wnd > tp->max_sndwnd)
tp->max_sndwnd = tp->snd_wnd;
needoutput = 1;
1994-05-24 10:09:53 +00:00
}
/*
* Process segments with URG.
*/
if ((thflags & TH_URG) && th->th_urp &&
1994-05-24 10:09:53 +00:00
TCPS_HAVERCVDFIN(tp->t_state) == 0) {
/*
* This is a kludge, but if we receive and accept
* random urgent pointers, we'll crash in
* soreceive. It's hard to imagine someone
* actually wanting to send this much urgent data.
*/
SOCKBUF_LOCK(&so->so_rcv);
if (th->th_urp + sbavail(&so->so_rcv) > sb_max) {
th->th_urp = 0; /* XXX */
thflags &= ~TH_URG; /* XXX */
SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */
1994-05-24 10:09:53 +00:00
goto dodata; /* XXX */
}
/*
* If this segment advances the known urgent pointer,
* then mark the data stream. This should not happen
* in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since
1995-05-30 08:16:23 +00:00
* a FIN has been received from the remote side.
1994-05-24 10:09:53 +00:00
* In these states we ignore the URG.
*
* According to RFC961 (Assigned Protocols),
* the urgent pointer points to the last octet
* of urgent data. We continue, however,
* to consider it to indicate the first octet
1995-05-30 08:16:23 +00:00
* of data past the urgent section as the original
1994-05-24 10:09:53 +00:00
* spec states (in one of two places).
*/
if (SEQ_GT(th->th_seq+th->th_urp, tp->rcv_up)) {
tp->rcv_up = th->th_seq + th->th_urp;
so->so_oobmark = sbavail(&so->so_rcv) +
1994-05-24 10:09:53 +00:00
(tp->rcv_up - tp->rcv_nxt) - 1;
if (so->so_oobmark == 0)
so->so_rcv.sb_state |= SBS_RCVATMARK;
1994-05-24 10:09:53 +00:00
sohasoutofband(so);
tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA);
}
SOCKBUF_UNLOCK(&so->so_rcv);
1994-05-24 10:09:53 +00:00
/*
* Remove out of band data so doesn't get presented to user.
* This can happen independent of advancing the URG pointer,
* but if two URG's are pending at once, some out-of-band
* data may creep in... ick.
*/
if (th->th_urp <= (u_long)tlen &&
!(so->so_options & SO_OOBINLINE)) {
/* hdr drop is delayed */
tcp_pulloutofband(so, th, m, drop_hdrlen);
}
} else {
1994-05-24 10:09:53 +00:00
/*
* If no out of band data is expected,
* pull receive urgent pointer along
* with the receive window.
*/
if (SEQ_GT(tp->rcv_nxt, tp->rcv_up))
tp->rcv_up = tp->rcv_nxt;
}
1994-05-24 10:09:53 +00:00
dodata: /* XXX */
INP_WLOCK_ASSERT(tp->t_inpcb);
1994-05-24 10:09:53 +00:00
/*
* Process the segment text, merging it into the TCP sequencing queue,
* and arranging for acknowledgment of receipt if necessary.
* This process logically involves adjusting tp->rcv_wnd as data
* is presented to the user (this happens in tcp_usrreq.c,
* case PRU_RCVD). If a FIN has already been received on this
* connection then we just ignore the text.
*/
tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
(tp->t_flags & TF_FASTOPEN));
if ((tlen || (thflags & TH_FIN) || tfo_syn) &&
1994-05-24 10:09:53 +00:00
TCPS_HAVERCVDFIN(tp->t_state) == 0) {
tcp_seq save_start = th->th_seq;
m_adj(m, drop_hdrlen); /* delayed header drop */
/*
* Insert segment which includes th into TCP reassembly queue
* with control block tp. Set thflags to whether reassembly now
* includes a segment with FIN. This handles the common case
* inline (segment is the next to be received on an established
* connection, and the queue is empty), avoiding linkage into
* and removal from the queue and repetition of various
* conversions.
* Set DELACK for segments received in order, but ack
* immediately when segments are out of order (so
* fast retransmit can work).
*/
if (th->th_seq == tp->rcv_nxt &&
LIST_EMPTY(&tp->t_segq) &&
(TCPS_HAVEESTABLISHED(tp->t_state) ||
tfo_syn)) {
if (DELAY_ACK(tp, tlen) || tfo_syn)
tp->t_flags |= TF_DELACK;
else
tp->t_flags |= TF_ACKNOW;
tp->rcv_nxt += tlen;
thflags = th->th_flags & TH_FIN;
TCPSTAT_INC(tcps_rcvpack);
TCPSTAT_ADD(tcps_rcvbyte, tlen);
SOCKBUF_LOCK(&so->so_rcv);
if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
m_freem(m);
else
sbappendstream_locked(&so->so_rcv, m, 0);
/* NB: sorwakeup_locked() does an implicit unlock. */
sorwakeup_locked(so);
} else {
/*
* XXX: Due to the header drop above "th" is
* theoretically invalid by now. Fortunately
* m_adj() doesn't actually frees any mbufs
* when trimming from the head.
*/
thflags = tcp_reass(tp, th, &tlen, m);
tp->t_flags |= TF_ACKNOW;
}
if (tlen > 0 && (tp->t_flags & TF_SACK_PERMIT))
tcp_update_sack_list(tp, save_start, save_start + tlen);
#if 0
1994-05-24 10:09:53 +00:00
/*
* Note the amount of data that peer has sent into
* our window, in order to estimate the sender's
* buffer size.
* XXX: Unused.
1994-05-24 10:09:53 +00:00
*/
if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt))
len = so->so_rcv.sb_hiwat - (tp->rcv_adv - tp->rcv_nxt);
else
len = so->so_rcv.sb_hiwat;
#endif
1994-05-24 10:09:53 +00:00
} else {
m_freem(m);
thflags &= ~TH_FIN;
1994-05-24 10:09:53 +00:00
}
/*
* If FIN is received ACK the FIN and let the user know
* that the connection is closing.
*/
if (thflags & TH_FIN) {
1994-05-24 10:09:53 +00:00
if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
socantrcvmore(so);
/*
* If connection is half-synchronized
* (ie NEEDSYN flag on) then delay ACK,
* so it may be piggybacked when SYN is sent.
* Otherwise, since we received a FIN then no
* more input can be expected, send ACK now.
*/
if (tp->t_flags & TF_NEEDSYN)
tp->t_flags |= TF_DELACK;
1995-05-30 08:16:23 +00:00
else
tp->t_flags |= TF_ACKNOW;
1994-05-24 10:09:53 +00:00
tp->rcv_nxt++;
}
switch (tp->t_state) {
/*
1994-05-24 10:09:53 +00:00
* In SYN_RECEIVED and ESTABLISHED STATES
* enter the CLOSE_WAIT state.
*/
case TCPS_SYN_RECEIVED:
tp->t_starttime = ticks;
/* FALLTHROUGH */
1994-05-24 10:09:53 +00:00
case TCPS_ESTABLISHED:
tcp_state_change(tp, TCPS_CLOSE_WAIT);
1994-05-24 10:09:53 +00:00
break;
/*
1994-05-24 10:09:53 +00:00
* If still in FIN_WAIT_1 STATE FIN has not been acked so
* enter the CLOSING state.
*/
case TCPS_FIN_WAIT_1:
tcp_state_change(tp, TCPS_CLOSING);
1994-05-24 10:09:53 +00:00
break;
/*
1994-05-24 10:09:53 +00:00
* In FIN_WAIT_2 state enter the TIME_WAIT state,
1995-05-30 08:16:23 +00:00
* starting the time-wait timer, turning off the other
1994-05-24 10:09:53 +00:00
* standard timers.
*/
case TCPS_FIN_WAIT_2:
INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
KASSERT(ti_locked == TI_RLOCKED, ("%s: dodata "
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
"TCP_FIN_WAIT_2 ti_locked: %d", __func__,
ti_locked));
tcp_twstart(tp);
INP_INFO_RUNLOCK(&V_tcbinfo);
return;
1994-05-24 10:09:53 +00:00
}
}
if (ti_locked == TI_RLOCKED)
INP_INFO_RUNLOCK(&V_tcbinfo);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
ti_locked = TI_UNLOCKED;
#ifdef TCPDEBUG
if (so->so_options & SO_DEBUG)
tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen,
&tcp_savetcp, 0);
#endif
TCP_PROBE3(debug__input, tp, th, mtod(m, const char *));
1994-05-24 10:09:53 +00:00
/*
* Return any desired output.
*/
if (needoutput || (tp->t_flags & TF_ACKNOW))
(void) tp->t_fb->tfb_tcp_output(tp);
check_delack:
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
KASSERT(ti_locked == TI_UNLOCKED, ("%s: check_delack ti_locked %d",
__func__, ti_locked));
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
INP_WLOCK_ASSERT(tp->t_inpcb);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
if (tp->t_flags & TF_DELACK) {
tp->t_flags &= ~TF_DELACK;
tcp_timer_activate(tp, TT_DELACK, tcp_delacktime);
}
INP_WUNLOCK(tp->t_inpcb);
return;
1994-05-24 10:09:53 +00:00
dropafterack:
/*
* Generate an ACK dropping incoming segment if it occupies
* sequence space, where the ACK reflects our state.
*
* We can now skip the test for the RST flag since all
* paths to this code happen after packets containing
* RST have been dropped.
*
* In the SYN-RECEIVED state, don't send an ACK unless the
* segment we received passes the SYN-RECEIVED ACK test.
* If it fails send a RST. This breaks the loop in the
* "LAND" DoS attack, and also prevents an ACK storm
* between two listening ports that have been sent forged
* SYN segments, each with the source address of the other.
1994-05-24 10:09:53 +00:00
*/
if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
(SEQ_GT(tp->snd_una, th->th_ack) ||
SEQ_GT(th->th_ack, tp->snd_max)) ) {
rstreason = BANDLIM_RST_OPENPORT;
goto dropwithreset;
}
#ifdef TCPDEBUG
if (so->so_options & SO_DEBUG)
tcp_trace(TA_DROP, ostate, tp, (void *)tcp_saveipgen,
&tcp_savetcp, 0);
#endif
TCP_PROBE3(debug__input, tp, th, mtod(m, const char *));
if (ti_locked == TI_RLOCKED)
INP_INFO_RUNLOCK(&V_tcbinfo);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
ti_locked = TI_UNLOCKED;
1994-05-24 10:09:53 +00:00
tp->t_flags |= TF_ACKNOW;
(void) tp->t_fb->tfb_tcp_output(tp);
INP_WUNLOCK(tp->t_inpcb);
m_freem(m);
return;
1994-05-24 10:09:53 +00:00
dropwithreset:
if (ti_locked == TI_RLOCKED)
INP_INFO_RUNLOCK(&V_tcbinfo);
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
ti_locked = TI_UNLOCKED;
if (tp != NULL) {
tcp_dropwithreset(m, th, tp, tlen, rstreason);
INP_WUNLOCK(tp->t_inpcb);
} else
tcp_dropwithreset(m, th, NULL, tlen, rstreason);
return;
drop:
if (ti_locked == TI_RLOCKED) {
INP_INFO_RUNLOCK(&V_tcbinfo);
Decompose the current single inpcbinfo lock into two locks: - The existing ipi_lock continues to protect the global inpcb list and inpcb counter. This lock is now relegated to a small number of allocation and free operations, and occasional operations that walk all connections (including, awkwardly, certain UDP multicast receive operations -- something to revisit). - A new ipi_hash_lock protects the two inpcbinfo hash tables for looking up connections and bound sockets, manipulated using new INP_HASH_*() macros. This lock, combined with inpcb locks, protects the 4-tuple address space. Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb connection locks, so may be acquired while manipulating a connection on which a lock is already held, avoiding the need to acquire the inpcbinfo lock preemptively when a binding change might later be required. As a result, however, lookup operations necessarily go through a reference acquire while holding the lookup lock, later acquiring an inpcb lock -- if required. A new function in_pcblookup() looks up connections, and accepts flags indicating how to return the inpcb. Due to lock order changes, callers no longer need acquire locks before performing a lookup: the lookup routine will acquire the ipi_hash_lock as needed. In the future, it will also be able to use alternative lookup and locking strategies transparently to callers, such as pcbgroup lookup. New lookup flags are, supplementing the existing INPLOOKUP_WILDCARD flag: INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb Callers must pass exactly one of these flags (for the time being). Some notes: - All protocols are updated to work within the new regime; especially, TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely eliminated, and global hash lock hold times are dramatically reduced compared to previous locking. - The TCP syncache still relies on the pcbinfo lock, something that we may want to revisit. - Support for reverting to the FreeBSD 7.x locking strategy in TCP input is no longer available -- hash lookup locks are now held only very briefly during inpcb lookup, rather than for potentially extended periods. However, the pcbinfo ipi_lock will still be acquired if a connection state might change such that a connection is added or removed. - Raw IP sockets continue to use the pcbinfo ipi_lock for protection, due to maintaining their own hash tables. - The interface in6_pcblookup_hash_locked() is maintained, which allows callers to acquire hash locks and perform one or more lookups atomically with 4-tuple allocation: this is required only for TCPv6, as there is no in6_pcbconnect_setup(), which there should be. - UDPv6 locking remains significantly more conservative than UDPv4 locking, which relates to source address selection. This needs attention, as it likely significantly reduces parallelism in this code for multithreaded socket use (such as in BIND). - In the UDPv4 and UDPv6 multicast cases, we need to revisit locking somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which is no longer sufficient. A second check once the inpcb lock is held should do the trick, keeping the general case from requiring the inpcb lock for every inpcb visited. - This work reminds us that we need to revisit locking of the v4/v6 flags, which may be accessed lock-free both before and after this change. - Right now, a single lock name is used for the pcbhash lock -- this is undesirable, and probably another argument is required to take care of this (or a char array name field in the pcbinfo?). This is not an MFC candidate for 8.x due to its impact on lookup and locking semantics. It's possible some of these issues could be worked around with compatibility wrappers, if necessary. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-05-30 09:43:55 +00:00
ti_locked = TI_UNLOCKED;
}
Move from solely write-locking the global tcbinfo in tcp_input() to read-locking in the TCP input path, allowing greater TCP input parallelism where multiple ithreads or ithread and netisr are able to run in parallel. Previously, most TCP input paths held a write lock on the global tcbinfo lock, effectively serializing TCP input. Before looking up the connection, acquire a write lock if a potentially state-changing flag is set on the TCP segment header (FIN, RST, SYN), and otherwise a read lock. We may later have to upgrade to a write lock in certain cases (ACKs received by the syncache or during TIMEWAIT) in order to support global state transitions, but this is never required for steady-state packets. Upgrading from a write lock to a read lock must be done as a trylock operation to avoid deadlocks, and actually violates the lock order as the tcbinfo lock preceeds the inpcb lock held at the time of upgrade. If the trylock fails, we bump the refcount on the inpcb, drop both locks, and re-acquire in-order. If another thread has freed the connection while the locks are dropped, we free the inpcb and repeat the lookup (this should hardly ever or never happen in practice). For now, maintain a number of new counters measuring how many times various cases execute, and in particular whether various optimistic assumptions about when read locks can be used, whether upgrades are done using the fast path, and whether connections close in practice in the above-described race, actually occur. MFC after: 6 weeks Discussed with: kmacy Reviewed by: bz, gnn, kmacy Tested by: kmacy
2008-12-08 20:27:00 +00:00
#ifdef INVARIANTS
else
INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
#endif
/*
* Drop space held by incoming segment and return.
*/
#ifdef TCPDEBUG
if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
tcp_trace(TA_DROP, ostate, tp, (void *)tcp_saveipgen,
&tcp_savetcp, 0);
#endif
TCP_PROBE3(debug__input, tp, th, mtod(m, const char *));
if (tp != NULL)
INP_WUNLOCK(tp->t_inpcb);
m_freem(m);
}
/*
* Issue RST and make ACK acceptable to originator of segment.
* The mbuf must still include the original packet header.
* tp may be NULL.
*/
void
tcp_dropwithreset(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp,
int tlen, int rstreason)
{
#ifdef INET
struct ip *ip;
#endif
#ifdef INET6
struct ip6_hdr *ip6;
#endif
if (tp != NULL) {
INP_WLOCK_ASSERT(tp->t_inpcb);
}
/* Don't bother if destination was broadcast/multicast. */
if ((th->th_flags & TH_RST) || m->m_flags & (M_BCAST|M_MCAST))
1994-05-24 10:09:53 +00:00
goto drop;
#ifdef INET6
if (mtod(m, struct ip *)->ip_v == 6) {
ip6 = mtod(m, struct ip6_hdr *);
if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
goto drop;
/* IPv6 anycast check is done at tcp6_input() */
}
#endif
#if defined(INET) && defined(INET6)
else
#endif
#ifdef INET
{
ip = mtod(m, struct ip *);
if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
goto drop;
}
#endif
/* Perform bandwidth limiting. */
if (badport_bandlim(rstreason) < 0)
goto drop;
/* tcp_respond consumes the mbuf chain. */
if (th->th_flags & TH_ACK) {
tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0,
th->th_ack, TH_RST);
} else {
if (th->th_flags & TH_SYN)
tlen++;
tcp_respond(tp, mtod(m, void *), th, m, th->th_seq+tlen,
(tcp_seq)0, TH_RST|TH_ACK);
1994-05-24 10:09:53 +00:00
}
return;
drop:
m_freem(m);
1994-05-24 10:09:53 +00:00
}
/*
* Parse TCP options and place in tcpopt.
*/
void
tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, int flags)
1994-05-24 10:09:53 +00:00
{
int opt, optlen;
to->to_flags = 0;
1994-05-24 10:09:53 +00:00
for (; cnt > 0; cnt -= optlen, cp += optlen) {
opt = cp[0];
if (opt == TCPOPT_EOL)
break;
if (opt == TCPOPT_NOP)
optlen = 1;
else {
if (cnt < 2)
break;
1994-05-24 10:09:53 +00:00
optlen = cp[1];
if (optlen < 2 || optlen > cnt)
1994-05-24 10:09:53 +00:00
break;
}
switch (opt) {
case TCPOPT_MAXSEG:
if (optlen != TCPOLEN_MAXSEG)
continue;
if (!(flags & TO_SYN))
1994-05-24 10:09:53 +00:00
continue;
to->to_flags |= TOF_MSS;
bcopy((char *)cp + 2,
(char *)&to->to_mss, sizeof(to->to_mss));
to->to_mss = ntohs(to->to_mss);
1994-05-24 10:09:53 +00:00
break;
case TCPOPT_WINDOW:
if (optlen != TCPOLEN_WINDOW)
continue;
if (!(flags & TO_SYN))
1994-05-24 10:09:53 +00:00
continue;
to->to_flags |= TOF_SCALE;
to->to_wscale = min(cp[2], TCP_MAX_WINSHIFT);
1994-05-24 10:09:53 +00:00
break;
case TCPOPT_TIMESTAMP:
if (optlen != TCPOLEN_TIMESTAMP)
continue;
to->to_flags |= TOF_TS;
bcopy((char *)cp + 2,
(char *)&to->to_tsval, sizeof(to->to_tsval));
to->to_tsval = ntohl(to->to_tsval);
bcopy((char *)cp + 6,
(char *)&to->to_tsecr, sizeof(to->to_tsecr));
to->to_tsecr = ntohl(to->to_tsecr);
1994-05-24 10:09:53 +00:00
break;
Initial import of RFC 2385 (TCP-MD5) digest support. This is the first of two commits; bringing in the kernel support first. This can be enabled by compiling a kernel with options TCP_SIGNATURE and FAST_IPSEC. For the uninitiated, this is a TCP option which provides for a means of authenticating TCP sessions which came into being before IPSEC. It is still relevant today, however, as it is used by many commercial router vendors, particularly with BGP, and as such has become a requirement for interconnect at many major Internet points of presence. Several parts of the TCP and IP headers, including the segment payload, are digested with MD5, including a shared secret. The PF_KEY interface is used to manage the secrets using security associations in the SADB. There is a limitation here in that as there is no way to map a TCP flow per-port back to an SPI without polluting tcpcb or using the SPD; the code to do the latter is unstable at this time. Therefore this code only supports per-host keying granularity. Whilst FAST_IPSEC is mutually exclusive with KAME IPSEC (and thus IPv6), TCP_SIGNATURE applies only to IPv4. For the vast majority of prospective users of this feature, this will not pose any problem. This implementation is output-only; that is, the option is honoured when responding to a host initiating a TCP session, but no effort is made [yet] to authenticate inbound traffic. This is, however, sufficient to interwork with Cisco equipment. Tested with a Cisco 2501 running IOS 12.0(27), and Quagga 0.96.4 with local patches. Patches for tcpdump to validate TCP-MD5 sessions are also available from me upon request. Sponsored by: sentex.net
2004-02-11 04:26:04 +00:00
#ifdef TCP_SIGNATURE
/*
* XXX In order to reply to a host which has set the
* TCP_SIGNATURE option in its initial SYN, we have to
* record the fact that the option was observed here
* for the syncache code to perform the correct response.
*/
case TCPOPT_SIGNATURE:
if (optlen != TCPOLEN_SIGNATURE)
continue;
to->to_flags |= TOF_SIGNATURE;
to->to_signature = cp + 2;
Initial import of RFC 2385 (TCP-MD5) digest support. This is the first of two commits; bringing in the kernel support first. This can be enabled by compiling a kernel with options TCP_SIGNATURE and FAST_IPSEC. For the uninitiated, this is a TCP option which provides for a means of authenticating TCP sessions which came into being before IPSEC. It is still relevant today, however, as it is used by many commercial router vendors, particularly with BGP, and as such has become a requirement for interconnect at many major Internet points of presence. Several parts of the TCP and IP headers, including the segment payload, are digested with MD5, including a shared secret. The PF_KEY interface is used to manage the secrets using security associations in the SADB. There is a limitation here in that as there is no way to map a TCP flow per-port back to an SPI without polluting tcpcb or using the SPD; the code to do the latter is unstable at this time. Therefore this code only supports per-host keying granularity. Whilst FAST_IPSEC is mutually exclusive with KAME IPSEC (and thus IPv6), TCP_SIGNATURE applies only to IPv4. For the vast majority of prospective users of this feature, this will not pose any problem. This implementation is output-only; that is, the option is honoured when responding to a host initiating a TCP session, but no effort is made [yet] to authenticate inbound traffic. This is, however, sufficient to interwork with Cisco equipment. Tested with a Cisco 2501 running IOS 12.0(27), and Quagga 0.96.4 with local patches. Patches for tcpdump to validate TCP-MD5 sessions are also available from me upon request. Sponsored by: sentex.net
2004-02-11 04:26:04 +00:00
break;
2004-02-13 18:21:45 +00:00
#endif
case TCPOPT_SACK_PERMITTED:
if (optlen != TCPOLEN_SACK_PERMITTED)
continue;
if (!(flags & TO_SYN))
continue;
if (!V_tcp_do_sack)
continue;
to->to_flags |= TOF_SACKPERM;
break;
case TCPOPT_SACK:
if (optlen <= 2 || (optlen - 2) % TCPOLEN_SACK != 0)
continue;
if (flags & TO_SYN)
continue;
to->to_flags |= TOF_SACK;
to->to_nsacks = (optlen - 2) / TCPOLEN_SACK;
to->to_sacks = cp + 2;
TCPSTAT_INC(tcps_sack_rcv_blocks);
break;
#ifdef TCP_RFC7413
case TCPOPT_FAST_OPEN:
if ((optlen != TCPOLEN_FAST_OPEN_EMPTY) &&
(optlen < TCPOLEN_FAST_OPEN_MIN) &&
(optlen > TCPOLEN_FAST_OPEN_MAX))
continue;
if (!(flags & TO_SYN))
continue;
if (!V_tcp_fastopen_enabled)
continue;
to->to_flags |= TOF_FASTOPEN;
to->to_tfo_len = optlen - 2;
to->to_tfo_cookie = to->to_tfo_len ? cp + 2 : NULL;
break;
#endif
default:
continue;
1994-05-24 10:09:53 +00:00
}
}
}
/*
* Pull out of band byte out of a segment so
* it doesn't appear in the user's data queue.
* It is still reflected in the segment length for
* sequencing purposes.
*/
void
tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m,
int off)
1994-05-24 10:09:53 +00:00
{
int cnt = off + th->th_urp - 1;
1995-05-30 08:16:23 +00:00
1994-05-24 10:09:53 +00:00
while (cnt >= 0) {
if (m->m_len > cnt) {
char *cp = mtod(m, caddr_t) + cnt;
struct tcpcb *tp = sototcpcb(so);
INP_WLOCK_ASSERT(tp->t_inpcb);
1994-05-24 10:09:53 +00:00
tp->t_iobc = *cp;
tp->t_oobflags |= TCPOOB_HAVEDATA;
bcopy(cp+1, cp, (unsigned)(m->m_len - cnt - 1));
m->m_len--;
if (m->m_flags & M_PKTHDR)
m->m_pkthdr.len--;
1994-05-24 10:09:53 +00:00
return;
}
cnt -= m->m_len;
m = m->m_next;
if (m == NULL)
1994-05-24 10:09:53 +00:00
break;
}
panic("tcp_pulloutofband");
}
/*
* Collect new round-trip time estimate
* and update averages and current timeout.
*/
void
tcp_xmit_timer(struct tcpcb *tp, int rtt)
1994-05-24 10:09:53 +00:00
{
int delta;
INP_WLOCK_ASSERT(tp->t_inpcb);
TCPSTAT_INC(tcps_rttupdated);
tp->t_rttupdated++;
if (tp->t_srtt != 0) {
/*
* srtt is stored as fixed point with 5 bits after the
* binary point (i.e., scaled by 8). The following magic
* is equivalent to the smoothing algorithm in rfc793 with
* an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed
* point). Adjust rtt to origin 0.
*/
delta = ((rtt - 1) << TCP_DELTA_SHIFT)
- (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
if ((tp->t_srtt += delta) <= 0)
tp->t_srtt = 1;
/*
* We accumulate a smoothed rtt variance (actually, a
* smoothed mean difference), then set the retransmit
* timer to smoothed rtt + 4 times the smoothed variance.
* rttvar is stored as fixed point with 4 bits after the
* binary point (scaled by 16). The following is
* equivalent to rfc793 smoothing with an alpha of .75
* (rttvar = rttvar*3/4 + |delta| / 4). This replaces
* rfc793's wired-in beta.
*/
if (delta < 0)
delta = -delta;
delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
if ((tp->t_rttvar += delta) <= 0)
tp->t_rttvar = 1;
if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
} else {
/*
* No rtt measurement yet - use the unsmoothed rtt.
* Set the variance to half the rtt (so our first
* retransmit happens at 3*rtt).
*/
tp->t_srtt = rtt << TCP_RTT_SHIFT;
tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
}
tp->t_rtttime = 0;
1994-05-24 10:09:53 +00:00
tp->t_rxtshift = 0;
/*
* the retransmit should happen at rtt + 4 * rttvar.
* Because of the way we do the smoothing, srtt and rttvar
* will each average +1/2 tick of bias. When we compute
* the retransmit timer, we want 1/2 tick of rounding and
* 1 extra tick because of +-1/2 tick uncertainty in the
* firing of the timer. The bias will give us exactly the
* 1.5 tick we need. But, because the bias is
* statistical, we have to test that we don't drop below
* the minimum feasible timer (which is 2 ticks).
*/
TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX);
1995-05-30 08:16:23 +00:00
1994-05-24 10:09:53 +00:00
/*
* We received an ack for a packet that wasn't retransmitted;
* it is probably safe to discard any error indications we've
* received recently. This isn't quite right, but close enough
* for now (a route might have failed after we sent a segment,
* and the return path might not be symmetrical).
*/
tp->t_softerror = 0;
}
/*
* Determine a reasonable value for maxseg size.
* If the route is known, check route for mtu.
* If none, use an mss that can be handled on the outgoing interface
* without forcing IP to fragment. If no route is found, route has no mtu,
1994-05-24 10:09:53 +00:00
* or the destination isn't local, use a default, hopefully conservative
* size (usually 512 or the default IP max size, but no more than the mtu
* of the interface), as we can't discover anything about intervening
* gateways or networks. We also initialize the congestion/slow start
* window to be a single segment if the destination isn't local.
* While looking at the routing entry, we also initialize other path-dependent
* parameters from pre-set or cached values in the routing entry.
*
* NOTE that resulting t_maxseg doesn't include space for TCP options or
* IP options, e.g. IPSEC data, since length of this data may vary, and
* thus it is calculated for every segment separately in tcp_output().
*
* NOTE that this routine is only called when we process an incoming
* segment, or an ICMP need fragmentation datagram. Outgoing SYN/ACK MSS
* settings are handled in tcp_mssopt().
1994-05-24 10:09:53 +00:00
*/
void
tcp_mss_update(struct tcpcb *tp, int offer, int mtuoffer,
struct hc_metrics_lite *metricptr, struct tcp_ifcap *cap)
1994-05-24 10:09:53 +00:00
{
int mss = 0;
u_long maxmtu = 0;
struct inpcb *inp = tp->t_inpcb;
struct hc_metrics_lite metrics;
#ifdef INET6
int isipv6 = ((inp->inp_vflag & INP_IPV6) != 0) ? 1 : 0;
size_t min_protoh = isipv6 ?
sizeof (struct ip6_hdr) + sizeof (struct tcphdr) :
sizeof (struct tcpiphdr);
#else
const size_t min_protoh = sizeof(struct tcpiphdr);
#endif
INP_WLOCK_ASSERT(tp->t_inpcb);
if (mtuoffer != -1) {
KASSERT(offer == -1, ("%s: conflict", __func__));
offer = mtuoffer - min_protoh;
}
/* Initialize. */
#ifdef INET6
if (isipv6) {
maxmtu = tcp_maxmtu6(&inp->inp_inc, cap);
tp->t_maxseg = V_tcp_v6mssdflt;
}
#endif
#if defined(INET) && defined(INET6)
else
#endif
#ifdef INET
{
maxmtu = tcp_maxmtu(&inp->inp_inc, cap);
tp->t_maxseg = V_tcp_mssdflt;
1994-05-24 10:09:53 +00:00
}
#endif
1994-05-24 10:09:53 +00:00
/*
* No route to sender, stay with default mss and return.
*/
if (maxmtu == 0) {
/*
* In case we return early we need to initialize metrics
* to a defined state as tcp_hc_get() would do for us
* if there was no cache hit.
*/
if (metricptr != NULL)
bzero(metricptr, sizeof(struct hc_metrics_lite));
return;
}
/* What have we got? */
switch (offer) {
case 0:
/*
* Offer == 0 means that there was no MSS on the SYN
* segment, in this case we use tcp_mssdflt as
* already assigned to t_maxseg above.
*/
offer = tp->t_maxseg;
break;
case -1:
/*
* Offer == -1 means that we didn't receive SYN yet.
*/
/* FALLTHROUGH */
default:
Limiters and sanity checks for TCP MSS (maximum segement size) resource exhaustion attacks. For network link optimization TCP can adjust its MSS and thus packet size according to the observed path MTU. This is done dynamically based on feedback from the remote host and network components along the packet path. This information can be abused to pretend an extremely low path MTU. The resource exhaustion works in two ways: o during tcp connection setup the advertized local MSS is exchanged between the endpoints. The remote endpoint can set this arbitrarily low (except for a minimum MTU of 64 octets enforced in the BSD code). When the local host is sending data it is forced to send many small IP packets instead of a large one. For example instead of the normal TCP payload size of 1448 it forces TCP payload size of 12 (MTU 64) and thus we have a 120 times increase in workload and packets. On fast links this quickly saturates the local CPU and may also hit pps processing limites of network components along the path. This type of attack is particularly effective for servers where the attacker can download large files (WWW and FTP). We mitigate it by enforcing a minimum MTU settable by sysctl net.inet.tcp.minmss defaulting to 256 octets. o the local host is reveiving data on a TCP connection from the remote host. The local host has no control over the packet size the remote host is sending. The remote host may chose to do what is described in the first attack and send the data in packets with an TCP payload of at least one byte. For each packet the tcp_input() function will be entered, the packet is processed and a sowakeup() is signalled to the connected process. For example an attack with 2 Mbit/s gives 4716 packets per second and the same amount of sowakeup()s to the process (and context switches). This type of attack is particularly effective for servers where the attacker can upload large amounts of data. Normally this is the case with WWW server where large POSTs can be made. We mitigate this by calculating the average MSS payload per second. If it goes below 'net.inet.tcp.minmss' and the pps rate is above 'net.inet.tcp.minmssoverload' defaulting to 1000 this particular TCP connection is resetted and dropped. MITRE CVE: CAN-2004-0002 Reviewed by: sam (mentor) MFC after: 1 day
2004-01-08 17:40:07 +00:00
/*
* Prevent DoS attack with too small MSS. Round up
* to at least minmss.
*/
offer = max(offer, V_tcp_minmss);
}
1994-05-24 10:09:53 +00:00
/*
* rmx information is now retrieved from tcp_hostcache.
1994-05-24 10:09:53 +00:00
*/
tcp_hc_get(&inp->inp_inc, &metrics);
if (metricptr != NULL)
bcopy(&metrics, metricptr, sizeof(struct hc_metrics_lite));
1994-05-24 10:09:53 +00:00
/*
2014-07-02 22:04:14 +00:00
* If there's a discovered mtu in tcp hostcache, use it.
* Else, use the link mtu.
1994-05-24 10:09:53 +00:00
*/
if (metrics.rmx_mtu)
mss = min(metrics.rmx_mtu, maxmtu) - min_protoh;
else {
#ifdef INET6
if (isipv6) {
mss = maxmtu - min_protoh;
if (!V_path_mtu_discovery &&
!in6_localaddr(&inp->in6p_faddr))
mss = min(mss, V_tcp_v6mssdflt);
}
#endif
#if defined(INET) && defined(INET6)
else
#endif
#ifdef INET
{
mss = maxmtu - min_protoh;
if (!V_path_mtu_discovery &&
!in_localaddr(inp->inp_faddr))
mss = min(mss, V_tcp_mssdflt);
}
#endif
/*
* XXX - The above conditional (mss = maxmtu - min_protoh)
* probably violates the TCP spec.
* The problem is that, since we don't know the
* other end's MSS, we are supposed to use a conservative
* default. But, if we do that, then MTU discovery will
* never actually take place, because the conservative
* default is much less than the MTUs typically seen
* on the Internet today. For the moment, we'll sweep
* this under the carpet.
*
* The conservative default might not actually be a problem
* if the only case this occurs is when sending an initial
* SYN with options and data to a host we've never talked
* to before. Then, they will reply with an MSS value which
* will get recorded and the new parameters should get
* recomputed. For Further Study.
*/
}
mss = min(mss, offer);
/*
* Sanity check: make sure that maxseg will be large
* enough to allow some data on segments even if the
* all the option space is used (40bytes). Otherwise
* funny things may happen in tcp_output.
*
* XXXGL: shouldn't we reserve space for IP/IPv6 options?
*/
mss = max(mss, 64);
tp->t_maxseg = mss;
}
void
tcp_mss(struct tcpcb *tp, int offer)
{
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
int mss;
u_long bufsize;
struct inpcb *inp;
struct socket *so;
struct hc_metrics_lite metrics;
struct tcp_ifcap cap;
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
KASSERT(tp != NULL, ("%s: tp == NULL", __func__));
bzero(&cap, sizeof(cap));
tcp_mss_update(tp, offer, -1, &metrics, &cap);
mss = tp->t_maxseg;
inp = tp->t_inpcb;
1994-05-24 10:09:53 +00:00
/*
* If there's a pipesize, change the socket buffer to that size,
* don't change if sb_hiwat is different than default (then it
* has been changed on purpose with setsockopt).
* Make the socket buffers an integral number of mss units;
* if the mss is larger than the socket buffer, decrease the mss.
1994-05-24 10:09:53 +00:00
*/
so = inp->inp_socket;
SOCKBUF_LOCK(&so->so_snd);
if ((so->so_snd.sb_hiwat == V_tcp_sendspace) && metrics.rmx_sendpipe)
bufsize = metrics.rmx_sendpipe;
else
bufsize = so->so_snd.sb_hiwat;
if (bufsize < mss)
mss = bufsize;
else {
bufsize = roundup(bufsize, mss);
if (bufsize > sb_max)
bufsize = sb_max;
if (bufsize > so->so_snd.sb_hiwat)
(void)sbreserve_locked(&so->so_snd, bufsize, so, NULL);
}
SOCKBUF_UNLOCK(&so->so_snd);
tp->t_maxseg = mss;
1994-05-24 10:09:53 +00:00
SOCKBUF_LOCK(&so->so_rcv);
if ((so->so_rcv.sb_hiwat == V_tcp_recvspace) && metrics.rmx_recvpipe)
bufsize = metrics.rmx_recvpipe;
else
bufsize = so->so_rcv.sb_hiwat;
if (bufsize > mss) {
bufsize = roundup(bufsize, mss);
if (bufsize > sb_max)
bufsize = sb_max;
if (bufsize > so->so_rcv.sb_hiwat)
(void)sbreserve_locked(&so->so_rcv, bufsize, so, NULL);
1994-05-24 10:09:53 +00:00
}
SOCKBUF_UNLOCK(&so->so_rcv);
/* Check the interface for TSO capabilities. */
if (cap.ifcap & CSUM_TSO) {
tp->t_flags |= TF_TSO;
tp->t_tsomax = cap.tsomax;
tp->t_tsomaxsegcount = cap.tsomaxsegcount;
tp->t_tsomaxsegsize = cap.tsomaxsegsize;
}
}
/*
* Determine the MSS option to send on an outgoing SYN.
*/
int
tcp_mssopt(struct in_conninfo *inc)
{
int mss = 0;
u_long maxmtu = 0;
u_long thcmtu = 0;
size_t min_protoh;
KASSERT(inc != NULL, ("tcp_mssopt with NULL in_conninfo pointer"));
#ifdef INET6
if (inc->inc_flags & INC_ISIPV6) {
mss = V_tcp_v6mssdflt;
maxmtu = tcp_maxmtu6(inc, NULL);
min_protoh = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
}
#endif
#if defined(INET) && defined(INET6)
else
#endif
#ifdef INET
{
mss = V_tcp_mssdflt;
maxmtu = tcp_maxmtu(inc, NULL);
min_protoh = sizeof(struct tcpiphdr);
}
#endif
#if defined(INET6) || defined(INET)
thcmtu = tcp_hc_getmtu(inc); /* IPv4 and IPv6 */
#endif
if (maxmtu && thcmtu)
mss = min(maxmtu, thcmtu) - min_protoh;
else if (maxmtu || thcmtu)
mss = max(maxmtu, thcmtu) - min_protoh;
return (mss);
1994-05-24 10:09:53 +00:00
}
/*
* On a partial ack arrives, force the retransmission of the
* next unacknowledged segment. Do not clear tp->t_dupacks.
* By setting snd_nxt to ti_ack, this forces retransmission timer to
* be started again.
*/
void
tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th)
{
tcp_seq onxt = tp->snd_nxt;
u_long ocwnd = tp->snd_cwnd;
u_int maxseg = tcp_maxseg(tp);
INP_WLOCK_ASSERT(tp->t_inpcb);
tcp_timer_activate(tp, TT_REXMT, 0);
tp->t_rtttime = 0;
tp->snd_nxt = th->th_ack;
/*
* Set snd_cwnd to one segment beyond acknowledged offset.
* (tp->snd_una has not yet been updated when this function is called.)
*/
tp->snd_cwnd = maxseg + BYTES_THIS_ACK(tp, th);
tp->t_flags |= TF_ACKNOW;
(void) tp->t_fb->tfb_tcp_output(tp);
tp->snd_cwnd = ocwnd;
if (SEQ_GT(onxt, tp->snd_nxt))
tp->snd_nxt = onxt;
/*
* Partial window deflation. Relies on fact that tp->snd_una
* not updated yet.
*/
This commit marks the first formal contribution of the "Five New TCP Congestion Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details about the project are available at: http://caia.swin.edu.au/freebsd/5cc/ - Add a KPI and supporting infrastructure to allow modular congestion control algorithms to be used in the net stack. Algorithms can maintain per-connection state if required, and connections maintain their own algorithm pointer, which allows different connections to concurrently use different algorithms. The TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to programmatically query or change the congestion control algorithm respectively from within an application at runtime. - Integrate the framework with the TCP stack in as least intrusive a manner as possible. Care was also taken to develop the framework in a way that should allow integration with other congestion aware transport protocols (e.g. SCTP) in the future. The hope is that we will one day be able to share a single set of congestion control algorithm modules between all congestion aware transport protocols. - Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack and use it to decouple the meaning of recovery from a congestion event and recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based congestion control protocols don't generally need to recover from packet loss and need a different way to note a congestion recovery episode within the stack. - Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code and ensures the stack always uses the appropriate mechanisms for recovering from packet loss during a congestion recovery episode. - Extract the NewReno congestion control algorithm from the TCP stack and massage it into module form. NewReno is always built into the kernel and will remain the default algorithm for the forseeable future. Implementations of additional different algorithms will become available in the near future. - Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code that relies on the size of "struct tcpcb" is required. Many thanks go to the Cisco University Research Program Fund at Community Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work at the Centre for Advanced Internet Architectures, Swinburne University of Technology is greatly appreciated. In collaboration with: David Hayes <dahayes at swin edu au> and Grenville Armitage <garmitage at swin edu au> Sponsored by: Cisco URP, FreeBSD Foundation Reviewed by: rpaulo Tested by: David Hayes (and many others over the years) MFC after: 3 months
2010-11-12 06:41:55 +00:00
if (tp->snd_cwnd > BYTES_THIS_ACK(tp, th))
tp->snd_cwnd -= BYTES_THIS_ACK(tp, th);
else
tp->snd_cwnd = 0;
tp->snd_cwnd += maxseg;
}
Calculate the correct amount of bytes that are in-flight for a connection as suggested by RFC 6675. Currently differnt places in the stack tries to guess this in suboptimal ways. The main problem is that current calculations don't take sacked bytes into account. Sacked bytes are the bytes receiver acked via SACK option. This is suboptimal because it assumes that network has more outstanding (unacked) bytes than the actual value and thus sends less data by setting congestion window lower than what's possible which in turn may cause slower recovery from losses. As an example, one of the current calculations looks something like this: snd_nxt - snd_fack + sackhint.sack_bytes_rexmit New proposal from RFC 6675 is: snd_max - snd_una - sackhint.sacked_bytes + sackhint.sack_bytes_rexmit which takes sacked bytes into account which is a new addition to the sackhint struct. Only thing we are missing from RFC 6675 is isLost() i.e. segment being considered lost and thus adjusting pipe based on that which makes this calculation a bit on conservative side. The approach is very simple. We already process each ack with sack info in tcp_sack_doack() and extract sack blocks/holes out of it. We'd now also track this new variable sacked_bytes which keeps track of total sacked bytes reported. One downside to this approach is that we may get incorrect count of sacked_bytes if the other end decides to drop sack info in the ack because of memory pressure or some other reasons. But in this (not very likely) case also the pipe calculation would be conservative which is okay as opposed to being aggressive in sending packets into the network. Next step is to use this more accurate pipe estimation to drive congestion window adjustments. In collaboration with: rrs Reviewed by: jason_eggnet dot com, rrs MFC after: 2 weeks Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D3971
2015-10-28 22:57:51 +00:00
int
tcp_compute_pipe(struct tcpcb *tp)
{
return (tp->snd_max - tp->snd_una +
tp->sackhint.sack_bytes_rexmit -
tp->sackhint.sacked_bytes);
}