freebsd-dev/sys/netinet/in_pcb.c

3019 lines
76 KiB
C
Raw Normal View History

/*-
* SPDX-License-Identifier: BSD-3-Clause
*
* Copyright (c) 1982, 1986, 1991, 1993, 1995
* The Regents of the University of California.
* Copyright (c) 2007-2009 Robert N. M. Watson
* Copyright (c) 2010-2011 Juniper Networks, Inc.
* All rights reserved.
1994-05-24 10:09:53 +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.
* 3. Neither the name of the University nor the names of its contributors
1994-05-24 10:09:53 +00:00
* 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.
*
* @(#)in_pcb.c 8.4 (Berkeley) 5/24/95
1994-05-24 10:09:53 +00:00
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_ddb.h"
#include "opt_ipsec.h"
#include "opt_inet.h"
#include "opt_inet6.h"
Implement kernel support for hardware rate limited sockets. - Add RATELIMIT kernel configuration keyword which must be set to enable the new functionality. - Add support for hardware driven, Receive Side Scaling, RSS aware, rate limited sendqueues and expose the functionality through the already established SO_MAX_PACING_RATE setsockopt(). The API support rates in the range from 1 to 4Gbytes/s which are suitable for regular TCP and UDP streams. The setsockopt(2) manual page has been updated. - Add rate limit function callback API to "struct ifnet" which supports the following operations: if_snd_tag_alloc(), if_snd_tag_modify(), if_snd_tag_query() and if_snd_tag_free(). - Add support to ifconfig to view, set and clear the IFCAP_TXRTLMT flag, which tells if a network driver supports rate limiting or not. - This patch also adds support for rate limiting through VLAN and LAGG intermediate network devices. - How rate limiting works: 1) The userspace application calls setsockopt() after accepting or making a new connection to set the rate which is then stored in the socket structure in the kernel. Later on when packets are transmitted a check is made in the transmit path for rate changes. A rate change implies a non-blocking ifp->if_snd_tag_alloc() call will be made to the destination network interface, which then sets up a custom sendqueue with the given rate limitation parameter. A "struct m_snd_tag" pointer is returned which serves as a "snd_tag" hint in the m_pkthdr for the subsequently transmitted mbufs. 2) When the network driver sees the "m->m_pkthdr.snd_tag" different from NULL, it will move the packets into a designated rate limited sendqueue given by the snd_tag pointer. It is up to the individual drivers how the rate limited traffic will be rate limited. 3) Route changes are detected by the NIC drivers in the ifp->if_transmit() routine when the ifnet pointer in the incoming snd_tag mismatches the one of the network interface. The network adapter frees the mbuf and returns EAGAIN which causes the ip_output() to release and clear the send tag. Upon next ip_output() a new "snd_tag" will be tried allocated. 4) When the PCB is detached the custom sendqueue will be released by a non-blocking ifp->if_snd_tag_free() call to the currently bound network interface. Reviewed by: wblock (manpages), adrian, gallatin, scottl (network) Differential Revision: https://reviews.freebsd.org/D3687 Sponsored by: Mellanox Technologies MFC after: 3 months
2017-01-18 13:31:17 +00:00
#include "opt_ratelimit.h"
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#include "opt_pcbgroup.h"
Several years after initial development, merge prototype support for linking NIC Receive Side Scaling (RSS) to the network stack's connection-group implementation. This prototype (and derived patches) are in use at Juniper and several other FreeBSD-using companies, so despite some reservations about its maturity, merge the patch to the base tree so that it can be iteratively refined in collaboration rather than maintained as a set of gradually diverging patch sets. (1) Merge a software implementation of the Toeplitz hash specified in RSS implemented by David Malone. This is used to allow suitable pcbgroup placement of connections before the first packet is received from the NIC. Software hashing is generally avoided, however, due to high cost of the hash on general-purpose CPUs. (2) In in_rss.c, maintain authoritative versions of RSS state intended to be pushed to each NIC, including keying material, hash algorithm/ configuration, and buckets. Provide software-facing interfaces to hash 2- and 4-tuples for IPv4 and IPv6 using both the RSS standardised Toeplitz and a 'naive' variation with a hash efficient in software but with poor distribution properties. Implement rss_m2cpuid()to be used by netisr and other load balancing code to look up the CPU on which an mbuf should be processed. (3) In the Ethernet link layer, allow netisr distribution using RSS as a source of policy as an alternative to source ordering; continue to default to direct dispatch (i.e., don't try and requeue packets for processing on the 'right' CPU if they arrive in a directly dispatchable context). (4) Allow RSS to control tuning of connection groups in order to align groups with RSS buckets. If a packet arrives on a protocol using connection groups, and contains a suitable hardware-generated hash, use that hash value to select the connection group for pcb lookup for both IPv4 and IPv6. If no hardware-generated Toeplitz hash is available, we fall back on regular PCB lookup risking contention rather than pay the cost of Toeplitz in software -- this is a less scalable but, at my last measurement, faster approach. As core counts go up, we may want to revise this strategy despite CPU overhead. Where device drivers suitably configure NICs, and connection groups / RSS are enabled, this should avoid both lock and line contention during connection lookup for TCP. This commit does not modify any device drivers to tune device RSS configuration to the global RSS configuration; patches are in circulation to do this for at least Chelsio T3 and Intel 1G/10G drivers. Currently, the KPI for device drivers is not particularly robust, nor aware of more advanced features such as runtime reconfiguration/rebalancing. This will hopefully prove a useful starting point for refinement. No MFC is scheduled as we will first want to nail down a more mature and maintainable KPI/KBI for device drivers. Sponsored by: Juniper Networks (original work) Sponsored by: EMC/Isilon (patch update and merge)
2014-03-15 00:57:50 +00:00
#include "opt_rss.h"
1994-05-24 10:09:53 +00:00
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/lock.h>
1994-05-24 10:09:53 +00:00
#include <sys/malloc.h>
#include <sys/mbuf.h>
#include <sys/callout.h>
#include <sys/eventhandler.h>
#include <sys/domain.h>
1994-05-24 10:09:53 +00:00
#include <sys/protosw.h>
#include <sys/rmlock.h>
1994-05-24 10:09:53 +00:00
#include <sys/socket.h>
#include <sys/socketvar.h>
Implement kernel support for hardware rate limited sockets. - Add RATELIMIT kernel configuration keyword which must be set to enable the new functionality. - Add support for hardware driven, Receive Side Scaling, RSS aware, rate limited sendqueues and expose the functionality through the already established SO_MAX_PACING_RATE setsockopt(). The API support rates in the range from 1 to 4Gbytes/s which are suitable for regular TCP and UDP streams. The setsockopt(2) manual page has been updated. - Add rate limit function callback API to "struct ifnet" which supports the following operations: if_snd_tag_alloc(), if_snd_tag_modify(), if_snd_tag_query() and if_snd_tag_free(). - Add support to ifconfig to view, set and clear the IFCAP_TXRTLMT flag, which tells if a network driver supports rate limiting or not. - This patch also adds support for rate limiting through VLAN and LAGG intermediate network devices. - How rate limiting works: 1) The userspace application calls setsockopt() after accepting or making a new connection to set the rate which is then stored in the socket structure in the kernel. Later on when packets are transmitted a check is made in the transmit path for rate changes. A rate change implies a non-blocking ifp->if_snd_tag_alloc() call will be made to the destination network interface, which then sets up a custom sendqueue with the given rate limitation parameter. A "struct m_snd_tag" pointer is returned which serves as a "snd_tag" hint in the m_pkthdr for the subsequently transmitted mbufs. 2) When the network driver sees the "m->m_pkthdr.snd_tag" different from NULL, it will move the packets into a designated rate limited sendqueue given by the snd_tag pointer. It is up to the individual drivers how the rate limited traffic will be rate limited. 3) Route changes are detected by the NIC drivers in the ifp->if_transmit() routine when the ifnet pointer in the incoming snd_tag mismatches the one of the network interface. The network adapter frees the mbuf and returns EAGAIN which causes the ip_output() to release and clear the send tag. Upon next ip_output() a new "snd_tag" will be tried allocated. 4) When the PCB is detached the custom sendqueue will be released by a non-blocking ifp->if_snd_tag_free() call to the currently bound network interface. Reviewed by: wblock (manpages), adrian, gallatin, scottl (network) Differential Revision: https://reviews.freebsd.org/D3687 Sponsored by: Mellanox Technologies MFC after: 3 months
2017-01-18 13:31:17 +00:00
#include <sys/sockio.h>
#include <sys/priv.h>
1994-05-24 10:09:53 +00:00
#include <sys/proc.h>
#include <sys/refcount.h>
This Implements the mumbled about "Jail" feature. This is a seriously beefed up chroot kind of thing. The process is jailed along the same lines as a chroot does it, but with additional tough restrictions imposed on what the superuser can do. For all I know, it is safe to hand over the root bit inside a prison to the customer living in that prison, this is what it was developed for in fact: "real virtual servers". Each prison has an ip number associated with it, which all IP communications will be coerced to use and each prison has its own hostname. Needless to say, you need more RAM this way, but the advantage is that each customer can run their own particular version of apache and not stomp on the toes of their neighbors. It generally does what one would expect, but setting up a jail still takes a little knowledge. A few notes: I have no scripts for setting up a jail, don't ask me for them. The IP number should be an alias on one of the interfaces. mount a /proc in each jail, it will make ps more useable. /proc/<pid>/status tells the hostname of the prison for jailed processes. Quotas are only sensible if you have a mountpoint per prison. There are no privisions for stopping resource-hogging. Some "#ifdef INET" and similar may be missing (send patches!) If somebody wants to take it from here and develop it into more of a "virtual machine" they should be most welcome! Tools, comments, patches & documentation most welcome. Have fun... Sponsored by: http://www.rndassociates.com/ Run for almost a year by: http://www.servetheweb.com/
1999-04-28 11:38:52 +00:00
#include <sys/jail.h>
#include <sys/kernel.h>
#include <sys/sysctl.h>
#ifdef DDB
#include <ddb/ddb.h>
#endif
#include <vm/uma.h>
1994-05-24 10:09:53 +00:00
#include <net/if.h>
#include <net/if_var.h>
#include <net/if_types.h>
#include <net/if_llatbl.h>
1994-05-24 10:09:53 +00:00
#include <net/route.h>
#include <net/rss_config.h>
#include <net/vnet.h>
1994-05-24 10:09:53 +00:00
#if defined(INET) || defined(INET6)
1994-05-24 10:09:53 +00:00
#include <netinet/in.h>
#include <netinet/in_pcb.h>
#include <netinet/ip_var.h>
#include <netinet/tcp_var.h>
#include <netinet/udp.h>
#include <netinet/udp_var.h>
#endif
#ifdef INET
#include <netinet/in_var.h>
#endif
#ifdef INET6
#include <netinet/ip6.h>
#include <netinet6/in6_pcb.h>
#include <netinet6/in6_var.h>
#include <netinet6/ip6_var.h>
#endif /* INET6 */
Merge projects/ipsec into head/. Small summary ------------- o Almost all IPsec releated code was moved into sys/netipsec. o New kernel modules added: ipsec.ko and tcpmd5.ko. New kernel option IPSEC_SUPPORT added. It enables support for loading and unloading of ipsec.ko and tcpmd5.ko kernel modules. o IPSEC_NAT_T option was removed. Now NAT-T support is enabled by default. The UDP_ENCAP_ESPINUDP_NON_IKE encapsulation type support was removed. Added TCP/UDP checksum handling for inbound packets that were decapsulated by transport mode SAs. setkey(8) modified to show run-time NAT-T configuration of SA. o New network pseudo interface if_ipsec(4) added. For now it is build as part of ipsec.ko module (or with IPSEC kernel). It implements IPsec virtual tunnels to create route-based VPNs. o The network stack now invokes IPsec functions using special methods. The only one header file <netipsec/ipsec_support.h> should be included to declare all the needed things to work with IPsec. o All IPsec protocols handlers (ESP/AH/IPCOMP protosw) were removed. Now these protocols are handled directly via IPsec methods. o TCP_SIGNATURE support was reworked to be more close to RFC. o PF_KEY SADB was reworked: - now all security associations stored in the single SPI namespace, and all SAs MUST have unique SPI. - several hash tables added to speed up lookups in SADB. - SADB now uses rmlock to protect access, and concurrent threads can do SA lookups in the same time. - many PF_KEY message handlers were reworked to reflect changes in SADB. - SADB_UPDATE message was extended to support new PF_KEY headers: SADB_X_EXT_NEW_ADDRESS_SRC and SADB_X_EXT_NEW_ADDRESS_DST. They can be used by IKE daemon to change SA addresses. o ipsecrequest and secpolicy structures were cardinally changed to avoid locking protection for ipsecrequest. Now we support only limited number (4) of bundled SAs, but they are supported for both INET and INET6. o INPCB security policy cache was introduced. Each PCB now caches used security policies to avoid SP lookup for each packet. o For inbound security policies added the mode, when the kernel does check for full history of applied IPsec transforms. o References counting rules for security policies and security associations were changed. The proper SA locking added into xform code. o xform code was also changed. Now it is possible to unregister xforms. tdb_xxx structures were changed and renamed to reflect changes in SADB/SPDB, and changed rules for locking and refcounting. Reviewed by: gnn, wblock Obtained from: Yandex LLC Relnotes: yes Sponsored by: Yandex LLC Differential Revision: https://reviews.freebsd.org/D9352
2017-02-06 08:49:57 +00:00
#include <netipsec/ipsec_support.h>
#include <security/mac/mac_framework.h>
static struct callout ipport_tick_callout;
/*
* These configure the range of local port addresses assigned to
* "unspecified" outgoing connections/packets/whatever.
*/
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, ipport_lowfirstauto) = IPPORT_RESERVED - 1; /* 1023 */
VNET_DEFINE(int, ipport_lowlastauto) = IPPORT_RESERVEDSTART; /* 600 */
VNET_DEFINE(int, ipport_firstauto) = IPPORT_EPHEMERALFIRST; /* 10000 */
VNET_DEFINE(int, ipport_lastauto) = IPPORT_EPHEMERALLAST; /* 65535 */
VNET_DEFINE(int, ipport_hifirstauto) = IPPORT_HIFIRSTAUTO; /* 49152 */
VNET_DEFINE(int, ipport_hilastauto) = IPPORT_HILASTAUTO; /* 65535 */
/*
* Reserved ports accessible only to root. There are significant
* security considerations that must be accounted for when changing these,
* but the security benefits can be great. Please be careful.
*/
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, ipport_reservedhigh) = IPPORT_RESERVED - 1; /* 1023 */
VNET_DEFINE(int, ipport_reservedlow);
/* Variables dealing with random ephemeral port allocation. */
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, ipport_randomized) = 1; /* user controlled via sysctl */
VNET_DEFINE(int, ipport_randomcps) = 10; /* user controlled via sysctl */
VNET_DEFINE(int, ipport_randomtime) = 45; /* user controlled via sysctl */
VNET_DEFINE(int, ipport_stoprandom); /* toggled by ipport_tick */
VNET_DEFINE(int, ipport_tcpallocs);
static VNET_DEFINE(int, ipport_tcplastcount);
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
#define V_ipport_tcplastcount VNET(ipport_tcplastcount)
static void in_pcbremlists(struct inpcb *inp);
#ifdef INET
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
static struct inpcb *in_pcblookup_hash_locked(struct inpcbinfo *pcbinfo,
struct in_addr faddr, u_int fport_arg,
struct in_addr laddr, u_int lport_arg,
int lookupflags, struct ifnet *ifp);
#define RANGECHK(var, min, max) \
if ((var) < (min)) { (var) = (min); } \
else if ((var) > (max)) { (var) = (max); }
static int
sysctl_net_ipport_check(SYSCTL_HANDLER_ARGS)
{
int error;
Permit buiding kernels with options VIMAGE, restricted to only a single active network stack instance. Turning on options VIMAGE at compile time yields the following changes relative to default kernel build: 1) V_ accessor macros for virtualized variables resolve to structure fields via base pointers, instead of being resolved as fields in global structs or plain global variables. As an example, V_ifnet becomes: options VIMAGE: ((struct vnet_net *) vnet_net)->_ifnet default build: vnet_net_0._ifnet options VIMAGE_GLOBALS: ifnet 2) INIT_VNET_* macros will declare and set up base pointers to be used by V_ accessor macros, instead of resolving to whitespace: INIT_VNET_NET(ifp->if_vnet); becomes struct vnet_net *vnet_net = (ifp->if_vnet)->mod_data[VNET_MOD_NET]; 3) Memory for vnet modules registered via vnet_mod_register() is now allocated at run time in sys/kern/kern_vimage.c, instead of per vnet module structs being declared as globals. If required, vnet modules can now request the framework to provide them with allocated bzeroed memory by filling in the vmi_size field in their vmi_modinfo structures. 4) structs socket, ifnet, inpcbinfo, tcpcb and syncache_head are extended to hold a pointer to the parent vnet. options VIMAGE builds will fill in those fields as required. 5) curvnet is introduced as a new global variable in options VIMAGE builds, always pointing to the default and only struct vnet. 6) struct sysctl_oid has been extended with additional two fields to store major and minor virtualization module identifiers, oid_v_subs and oid_v_mod. SYSCTL_V_* family of macros will fill in those fields accordingly, and store the offset in the appropriate vnet container struct in oid_arg1. In sysctl handlers dealing with virtualized sysctls, the SYSCTL_RESOLVE_V_ARG1() macro will compute the address of the target variable and make it available in arg1 variable for further processing. Unused fields in structs vnet_inet, vnet_inet6 and vnet_ipfw have been deleted. Reviewed by: bz, rwatson Approved by: julian (mentor)
2009-04-30 13:36:26 +00:00
error = sysctl_handle_int(oidp, arg1, arg2, req);
if (error == 0) {
RANGECHK(V_ipport_lowfirstauto, 1, IPPORT_RESERVED - 1);
RANGECHK(V_ipport_lowlastauto, 1, IPPORT_RESERVED - 1);
RANGECHK(V_ipport_firstauto, IPPORT_RESERVED, IPPORT_MAX);
RANGECHK(V_ipport_lastauto, IPPORT_RESERVED, IPPORT_MAX);
RANGECHK(V_ipport_hifirstauto, IPPORT_RESERVED, IPPORT_MAX);
RANGECHK(V_ipport_hilastauto, IPPORT_RESERVED, IPPORT_MAX);
}
return (error);
}
#undef RANGECHK
static SYSCTL_NODE(_net_inet_ip, IPPROTO_IP, portrange, CTLFLAG_RW, 0,
"IP Ports");
SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowfirst,
CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
&VNET_NAME(ipport_lowfirstauto), 0, &sysctl_net_ipport_check, "I", "");
SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowlast,
CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
&VNET_NAME(ipport_lowlastauto), 0, &sysctl_net_ipport_check, "I", "");
SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, first,
CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
&VNET_NAME(ipport_firstauto), 0, &sysctl_net_ipport_check, "I", "");
SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, last,
CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
&VNET_NAME(ipport_lastauto), 0, &sysctl_net_ipport_check, "I", "");
SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hifirst,
CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
&VNET_NAME(ipport_hifirstauto), 0, &sysctl_net_ipport_check, "I", "");
SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hilast,
CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW,
&VNET_NAME(ipport_hilastauto), 0, &sysctl_net_ipport_check, "I", "");
SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, reservedhigh,
CTLFLAG_VNET | CTLFLAG_RW | CTLFLAG_SECURE,
&VNET_NAME(ipport_reservedhigh), 0, "");
SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, reservedlow,
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
CTLFLAG_RW|CTLFLAG_SECURE, &VNET_NAME(ipport_reservedlow), 0, "");
SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomized,
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(ipport_randomized), 0, "Enable random port allocation");
SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomcps,
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(ipport_randomcps), 0, "Maximum number of random port "
"allocations before switching to a sequental one");
SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomtime,
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(ipport_randomtime), 0,
"Minimum time to keep sequental port "
"allocation before switching to a random one");
#endif /* INET */
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
/*
* in_pcb.c: manage the Protocol Control Blocks.
*
* NOTE: It is assumed that most of these functions will be called with
* the pcbinfo lock held, and often, the inpcb lock held, as these utility
* functions often modify hash chains or addresses in pcbs.
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
*/
/*
* Different protocols initialize their inpcbs differently - giving
* different name to the lock. But they all are disposed the same.
*/
static void
inpcb_fini(void *mem, int size)
{
struct inpcb *inp = mem;
INP_LOCK_DESTROY(inp);
}
/*
* Initialize an inpcbinfo -- we should be able to reduce the number of
* arguments in time.
*/
void
in_pcbinfo_init(struct inpcbinfo *pcbinfo, const char *name,
struct inpcbhead *listhead, int hash_nelements, int porthash_nelements,
char *inpcbzone_name, uma_init inpcbzone_init, u_int hashfields)
{
INP_INFO_LOCK_INIT(pcbinfo, name);
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_HASH_LOCK_INIT(pcbinfo, "pcbinfohash"); /* XXXRW: argument? */
INP_LIST_LOCK_INIT(pcbinfo, "pcbinfolist");
#ifdef VIMAGE
pcbinfo->ipi_vnet = curvnet;
#endif
pcbinfo->ipi_listhead = listhead;
LIST_INIT(pcbinfo->ipi_listhead);
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
pcbinfo->ipi_count = 0;
pcbinfo->ipi_hashbase = hashinit(hash_nelements, M_PCB,
&pcbinfo->ipi_hashmask);
pcbinfo->ipi_porthashbase = hashinit(porthash_nelements, M_PCB,
&pcbinfo->ipi_porthashmask);
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#ifdef PCBGROUP
in_pcbgroup_init(pcbinfo, hashfields, hash_nelements);
#endif
pcbinfo->ipi_zone = uma_zcreate(inpcbzone_name, sizeof(struct inpcb),
NULL, NULL, inpcbzone_init, inpcb_fini, UMA_ALIGN_PTR, 0);
uma_zone_set_max(pcbinfo->ipi_zone, maxsockets);
uma_zone_set_warning(pcbinfo->ipi_zone,
"kern.ipc.maxsockets limit reached");
}
/*
* Destroy an inpcbinfo.
*/
void
in_pcbinfo_destroy(struct inpcbinfo *pcbinfo)
{
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
KASSERT(pcbinfo->ipi_count == 0,
("%s: ipi_count = %u", __func__, pcbinfo->ipi_count));
hashdestroy(pcbinfo->ipi_hashbase, M_PCB, pcbinfo->ipi_hashmask);
hashdestroy(pcbinfo->ipi_porthashbase, M_PCB,
pcbinfo->ipi_porthashmask);
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#ifdef PCBGROUP
in_pcbgroup_destroy(pcbinfo);
#endif
uma_zdestroy(pcbinfo->ipi_zone);
INP_LIST_LOCK_DESTROY(pcbinfo);
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_HASH_LOCK_DESTROY(pcbinfo);
INP_INFO_LOCK_DESTROY(pcbinfo);
}
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
/*
* Allocate a PCB and associate it with the socket.
* On success return with the PCB locked.
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
*/
1994-05-24 10:09:53 +00:00
int
in_pcballoc(struct socket *so, struct inpcbinfo *pcbinfo)
1994-05-24 10:09:53 +00:00
{
struct inpcb *inp;
int error;
#ifdef INVARIANTS
if (pcbinfo == &V_tcbinfo) {
INP_INFO_RLOCK_ASSERT(pcbinfo);
} else {
INP_INFO_WLOCK_ASSERT(pcbinfo);
}
#endif
error = 0;
inp = uma_zalloc(pcbinfo->ipi_zone, M_NOWAIT);
1994-05-24 10:09:53 +00:00
if (inp == NULL)
return (ENOBUFS);
bzero(&inp->inp_start_zero, inp_zero_size);
inp->inp_pcbinfo = pcbinfo;
1994-05-24 10:09:53 +00:00
inp->inp_socket = so;
inp->inp_cred = crhold(so->so_cred);
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
inp->inp_inc.inc_fibnum = so->so_fibnum;
#ifdef MAC
error = mac_inpcb_init(inp, M_NOWAIT);
if (error != 0)
goto out;
mac_inpcb_create(so, inp);
#endif
Merge projects/ipsec into head/. Small summary ------------- o Almost all IPsec releated code was moved into sys/netipsec. o New kernel modules added: ipsec.ko and tcpmd5.ko. New kernel option IPSEC_SUPPORT added. It enables support for loading and unloading of ipsec.ko and tcpmd5.ko kernel modules. o IPSEC_NAT_T option was removed. Now NAT-T support is enabled by default. The UDP_ENCAP_ESPINUDP_NON_IKE encapsulation type support was removed. Added TCP/UDP checksum handling for inbound packets that were decapsulated by transport mode SAs. setkey(8) modified to show run-time NAT-T configuration of SA. o New network pseudo interface if_ipsec(4) added. For now it is build as part of ipsec.ko module (or with IPSEC kernel). It implements IPsec virtual tunnels to create route-based VPNs. o The network stack now invokes IPsec functions using special methods. The only one header file <netipsec/ipsec_support.h> should be included to declare all the needed things to work with IPsec. o All IPsec protocols handlers (ESP/AH/IPCOMP protosw) were removed. Now these protocols are handled directly via IPsec methods. o TCP_SIGNATURE support was reworked to be more close to RFC. o PF_KEY SADB was reworked: - now all security associations stored in the single SPI namespace, and all SAs MUST have unique SPI. - several hash tables added to speed up lookups in SADB. - SADB now uses rmlock to protect access, and concurrent threads can do SA lookups in the same time. - many PF_KEY message handlers were reworked to reflect changes in SADB. - SADB_UPDATE message was extended to support new PF_KEY headers: SADB_X_EXT_NEW_ADDRESS_SRC and SADB_X_EXT_NEW_ADDRESS_DST. They can be used by IKE daemon to change SA addresses. o ipsecrequest and secpolicy structures were cardinally changed to avoid locking protection for ipsecrequest. Now we support only limited number (4) of bundled SAs, but they are supported for both INET and INET6. o INPCB security policy cache was introduced. Each PCB now caches used security policies to avoid SP lookup for each packet. o For inbound security policies added the mode, when the kernel does check for full history of applied IPsec transforms. o References counting rules for security policies and security associations were changed. The proper SA locking added into xform code. o xform code was also changed. Now it is possible to unregister xforms. tdb_xxx structures were changed and renamed to reflect changes in SADB/SPDB, and changed rules for locking and refcounting. Reviewed by: gnn, wblock Obtained from: Yandex LLC Relnotes: yes Sponsored by: Yandex LLC Differential Revision: https://reviews.freebsd.org/D9352
2017-02-06 08:49:57 +00:00
#if defined(IPSEC) || defined(IPSEC_SUPPORT)
error = ipsec_init_pcbpolicy(inp);
if (error != 0) {
#ifdef MAC
mac_inpcb_destroy(inp);
#endif
goto out;
}
#endif /*IPSEC*/
#ifdef INET6
if (INP_SOCKAF(so) == AF_INET6) {
inp->inp_vflag |= INP_IPV6PROTO;
if (V_ip6_v6only)
inp->inp_flags |= IN6P_IPV6_V6ONLY;
}
#endif
INP_WLOCK(inp);
INP_LIST_WLOCK(pcbinfo);
LIST_INSERT_HEAD(pcbinfo->ipi_listhead, inp, inp_list);
pcbinfo->ipi_count++;
1994-05-24 10:09:53 +00:00
so->so_pcb = (caddr_t)inp;
#ifdef INET6
if (V_ip6_auto_flowlabel)
inp->inp_flags |= IN6P_AUTOFLOWLABEL;
#endif
inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
refcount_init(&inp->inp_refcount, 1); /* Reference from inpcbinfo */
/*
* Routes in inpcb's can cache L2 as well; they are guaranteed
* to be cleaned up.
*/
inp->inp_route.ro_flags = RT_LLE_CACHE;
INP_LIST_WUNLOCK(pcbinfo);
#if defined(IPSEC) || defined(IPSEC_SUPPORT) || defined(MAC)
out:
if (error != 0) {
crfree(inp->inp_cred);
uma_zfree(pcbinfo->ipi_zone, inp);
}
#endif
return (error);
1994-05-24 10:09:53 +00:00
}
#ifdef INET
1994-05-24 10:09:53 +00:00
int
in_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
1994-05-24 10:09:53 +00:00
{
int anonport, error;
INP_WLOCK_ASSERT(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
INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
if (inp->inp_lport != 0 || inp->inp_laddr.s_addr != INADDR_ANY)
return (EINVAL);
anonport = nam == NULL || ((struct sockaddr_in *)nam)->sin_port == 0;
error = in_pcbbind_setup(inp, nam, &inp->inp_laddr.s_addr,
&inp->inp_lport, cred);
if (error)
return (error);
if (in_pcbinshash(inp) != 0) {
inp->inp_laddr.s_addr = INADDR_ANY;
inp->inp_lport = 0;
return (EAGAIN);
}
if (anonport)
inp->inp_flags |= INP_ANONPORT;
return (0);
}
#endif
/*
* Select a local port (number) to use.
*/
#if defined(INET) || defined(INET6)
int
2017-02-10 05:58:16 +00:00
in_pcb_lport(struct inpcb *inp, struct in_addr *laddrp, u_short *lportp,
struct ucred *cred, int lookupflags)
{
struct inpcbinfo *pcbinfo;
struct inpcb *tmpinp;
unsigned short *lastport;
int count, dorandom, error;
u_short aux, first, last, lport;
#ifdef INET
struct in_addr laddr;
#endif
pcbinfo = inp->inp_pcbinfo;
/*
* Because no actual state changes occur here, a global write lock on
* the pcbinfo isn't required.
*/
INP_LOCK_ASSERT(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
INP_HASH_LOCK_ASSERT(pcbinfo);
if (inp->inp_flags & INP_HIGHPORT) {
first = V_ipport_hifirstauto; /* sysctl */
last = V_ipport_hilastauto;
lastport = &pcbinfo->ipi_lasthi;
} else if (inp->inp_flags & INP_LOWPORT) {
error = priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0);
if (error)
return (error);
first = V_ipport_lowfirstauto; /* 1023 */
last = V_ipport_lowlastauto; /* 600 */
lastport = &pcbinfo->ipi_lastlow;
} else {
first = V_ipport_firstauto; /* sysctl */
last = V_ipport_lastauto;
lastport = &pcbinfo->ipi_lastport;
}
/*
* For UDP(-Lite), use random port allocation as long as the user
* allows it. For TCP (and as of yet unknown) connections,
* use random port allocation only if the user allows it AND
* ipport_tick() allows it.
*/
if (V_ipport_randomized &&
(!V_ipport_stoprandom || pcbinfo == &V_udbinfo ||
pcbinfo == &V_ulitecbinfo))
dorandom = 1;
else
dorandom = 0;
/*
* It makes no sense to do random port allocation if
* we have the only port available.
*/
if (first == last)
dorandom = 0;
/* Make sure to not include UDP(-Lite) packets in the count. */
if (pcbinfo != &V_udbinfo || pcbinfo != &V_ulitecbinfo)
V_ipport_tcpallocs++;
/*
* Instead of having two loops further down counting up or down
* make sure that first is always <= last and go with only one
* code path implementing all logic.
*/
if (first > last) {
aux = first;
first = last;
last = aux;
}
#ifdef INET
/* Make the compiler happy. */
laddr.s_addr = 0;
if ((inp->inp_vflag & (INP_IPV4|INP_IPV6)) == INP_IPV4) {
KASSERT(laddrp != NULL, ("%s: laddrp NULL for v4 inp %p",
__func__, inp));
laddr = *laddrp;
}
#endif
tmpinp = NULL; /* Make compiler happy. */
lport = *lportp;
if (dorandom)
*lastport = first + (arc4random() % (last - first));
count = last - first;
do {
if (count-- < 0) /* completely used? */
return (EADDRNOTAVAIL);
++*lastport;
if (*lastport < first || *lastport > last)
*lastport = first;
lport = htons(*lastport);
#ifdef INET6
2017-02-10 05:58:16 +00:00
if ((inp->inp_vflag & INP_IPV6) != 0)
tmpinp = in6_pcblookup_local(pcbinfo,
&inp->in6p_laddr, lport, lookupflags, cred);
#endif
#if defined(INET) && defined(INET6)
else
#endif
#ifdef INET
2017-02-10 05:58:16 +00:00
tmpinp = in_pcblookup_local(pcbinfo, laddr,
lport, lookupflags, cred);
#endif
} while (tmpinp != NULL);
#ifdef INET
if ((inp->inp_vflag & (INP_IPV4|INP_IPV6)) == INP_IPV4)
laddrp->s_addr = laddr.s_addr;
#endif
*lportp = lport;
return (0);
}
/*
* Return cached socket options.
*/
short
inp_so_options(const struct inpcb *inp)
{
short so_options;
so_options = 0;
if ((inp->inp_flags2 & INP_REUSEPORT) != 0)
so_options |= SO_REUSEPORT;
if ((inp->inp_flags2 & INP_REUSEADDR) != 0)
so_options |= SO_REUSEADDR;
return (so_options);
}
#endif /* INET || INET6 */
2014-07-10 03:10:56 +00:00
/*
* Check if a new BINDMULTI socket is allowed to be created.
*
* ni points to the new inp.
* oi points to the exisitng inp.
*
* This checks whether the existing inp also has BINDMULTI and
* whether the credentials match.
*/
int
2014-07-10 03:10:56 +00:00
in_pcbbind_check_bindmulti(const struct inpcb *ni, const struct inpcb *oi)
{
/* Check permissions match */
if ((ni->inp_flags2 & INP_BINDMULTI) &&
(ni->inp_cred->cr_uid !=
oi->inp_cred->cr_uid))
return (0);
/* Check the existing inp has BINDMULTI set */
if ((ni->inp_flags2 & INP_BINDMULTI) &&
((oi->inp_flags2 & INP_BINDMULTI) == 0))
return (0);
/*
* We're okay - either INP_BINDMULTI isn't set on ni, or
* it is and it matches the checks.
*/
return (1);
}
#ifdef INET
/*
* Set up a bind operation on a PCB, performing port allocation
* as required, but do not actually modify the PCB. Callers can
* either complete the bind by setting inp_laddr/inp_lport and
* calling in_pcbinshash(), or they can just use the resulting
* port and address to authorise the sending of a once-off packet.
*
* On error, the values of *laddrp and *lportp are not changed.
*/
int
in_pcbbind_setup(struct inpcb *inp, struct sockaddr *nam, in_addr_t *laddrp,
u_short *lportp, struct ucred *cred)
{
struct socket *so = inp->inp_socket;
struct sockaddr_in *sin;
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
struct in_addr laddr;
1994-05-24 10:09:53 +00:00
u_short lport = 0;
int lookupflags = 0, reuseport = (so->so_options & SO_REUSEPORT);
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
int error;
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
* No state changes, so read locks are sufficient here.
*/
INP_LOCK_ASSERT(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
INP_HASH_LOCK_ASSERT(pcbinfo);
if (TAILQ_EMPTY(&V_in_ifaddrhead)) /* XXX broken! */
1994-05-24 10:09:53 +00:00
return (EADDRNOTAVAIL);
laddr.s_addr = *laddrp;
if (nam != NULL && laddr.s_addr != INADDR_ANY)
1994-05-24 10:09:53 +00:00
return (EINVAL);
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) == 0)
lookupflags = INPLOOKUP_WILDCARD;
2017-02-10 05:58:16 +00:00
if (nam == NULL) {
if ((error = prison_local_ip4(cred, &laddr)) != 0)
return (error);
} else {
sin = (struct sockaddr_in *)nam;
if (nam->sa_len != sizeof (*sin))
1994-05-24 10:09:53 +00:00
return (EINVAL);
#ifdef notdef
/*
* We should check the family, but old programs
* incorrectly fail to initialize it.
*/
if (sin->sin_family != AF_INET)
return (EAFNOSUPPORT);
#endif
error = prison_local_ip4(cred, &sin->sin_addr);
if (error)
return (error);
if (sin->sin_port != *lportp) {
/* Don't allow the port to change. */
if (*lportp != 0)
return (EINVAL);
lport = sin->sin_port;
}
/* NB: lport is left as 0 if the port isn't being changed. */
1994-05-24 10:09:53 +00:00
if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
/*
* Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
* allow complete duplication of binding if
* SO_REUSEPORT is set, or if SO_REUSEADDR is set
* and a multicast address is bound on both
* new and duplicated sockets.
*/
if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0)
1994-05-24 10:09:53 +00:00
reuseport = SO_REUSEADDR|SO_REUSEPORT;
} else if (sin->sin_addr.s_addr != INADDR_ANY) {
sin->sin_port = 0; /* yech... */
bzero(&sin->sin_zero, sizeof(sin->sin_zero));
/*
* Is the address a local IP address?
* If INP_BINDANY is set, then the socket may be bound
* to any endpoint address, local or not.
*/
if ((inp->inp_flags & INP_BINDANY) == 0 &&
ifa_ifwithaddr_check((struct sockaddr *)sin) == 0)
1994-05-24 10:09:53 +00:00
return (EADDRNOTAVAIL);
}
laddr = sin->sin_addr;
1994-05-24 10:09:53 +00:00
if (lport) {
struct inpcb *t;
struct tcptw *tw;
1994-05-24 10:09:53 +00:00
/* GROSS */
if (ntohs(lport) <= V_ipport_reservedhigh &&
ntohs(lport) >= V_ipport_reservedlow &&
priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT,
0))
return (EACCES);
if (!IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) &&
priv_check_cred(inp->inp_cred,
PRIV_NETINET_REUSEPORT, 0) != 0) {
t = in_pcblookup_local(pcbinfo, sin->sin_addr,
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
lport, INPLOOKUP_WILDCARD, cred);
/*
* XXX
* This entire block sorely needs a rewrite.
*/
if (t &&
2014-07-10 03:10:56 +00:00
((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
((t->inp_flags & INP_TIMEWAIT) == 0) &&
(so->so_type != SOCK_STREAM ||
ntohl(t->inp_faddr.s_addr) == INADDR_ANY) &&
(ntohl(sin->sin_addr.s_addr) != INADDR_ANY ||
ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
(t->inp_flags2 & INP_REUSEPORT) == 0) &&
(inp->inp_cred->cr_uid !=
t->inp_cred->cr_uid))
return (EADDRINUSE);
2014-07-10 03:10:56 +00:00
/*
* If the socket is a BINDMULTI socket, then
* the credentials need to match and the
* original socket also has to have been bound
* with BINDMULTI.
*/
if (t && (! in_pcbbind_check_bindmulti(inp, t)))
return (EADDRINUSE);
}
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
t = in_pcblookup_local(pcbinfo, sin->sin_addr,
lport, lookupflags, cred);
if (t && (t->inp_flags & INP_TIMEWAIT)) {
/*
* XXXRW: If an incpb has had its timewait
* state recycled, we treat the address as
* being in use (for now). This is better
* than a panic, but not desirable.
*/
tw = intotw(t);
if (tw == NULL ||
(reuseport & tw->tw_so_options) == 0)
return (EADDRINUSE);
2014-07-10 03:10:56 +00:00
} else if (t &&
((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
(reuseport & inp_so_options(t)) == 0) {
#ifdef INET6
if (ntohl(sin->sin_addr.s_addr) !=
INADDR_ANY ||
ntohl(t->inp_laddr.s_addr) !=
INADDR_ANY ||
(inp->inp_vflag & INP_IPV6PROTO) == 0 ||
(t->inp_vflag & INP_IPV6PROTO) == 0)
#endif
return (EADDRINUSE);
2014-07-10 03:10:56 +00:00
if (t && (! in_pcbbind_check_bindmulti(inp, t)))
return (EADDRINUSE);
}
1994-05-24 10:09:53 +00:00
}
}
if (*lportp != 0)
lport = *lportp;
if (lport == 0) {
2017-02-10 05:58:16 +00:00
error = in_pcb_lport(inp, &laddr, &lport, cred, lookupflags);
if (error != 0)
return (error);
}
*laddrp = laddr.s_addr;
*lportp = lport;
1994-05-24 10:09:53 +00:00
return (0);
}
/*
* Connect from a socket to a specified address.
* Both address and port must be specified in argument sin.
* If don't have a local address for this socket yet,
* then pick one.
*/
int
in_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
struct ucred *cred, struct mbuf *m)
{
u_short lport, fport;
in_addr_t laddr, faddr;
int anonport, error;
INP_WLOCK_ASSERT(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
INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
lport = inp->inp_lport;
laddr = inp->inp_laddr.s_addr;
anonport = (lport == 0);
error = in_pcbconnect_setup(inp, nam, &laddr, &lport, &faddr, &fport,
NULL, cred);
if (error)
return (error);
/* Do the initial binding of the local address if required. */
if (inp->inp_laddr.s_addr == INADDR_ANY && inp->inp_lport == 0) {
inp->inp_lport = lport;
inp->inp_laddr.s_addr = laddr;
if (in_pcbinshash(inp) != 0) {
inp->inp_laddr.s_addr = INADDR_ANY;
inp->inp_lport = 0;
return (EAGAIN);
}
}
/* Commit the remaining changes. */
inp->inp_lport = lport;
inp->inp_laddr.s_addr = laddr;
inp->inp_faddr.s_addr = faddr;
inp->inp_fport = fport;
in_pcbrehash_mbuf(inp, m);
if (anonport)
inp->inp_flags |= INP_ANONPORT;
return (0);
}
int
in_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
{
return (in_pcbconnect_mbuf(inp, nam, cred, NULL));
}
/*
* Do proper source address selection on an unbound socket in case
* of connect. Take jails into account as well.
*/
int
in_pcbladdr(struct inpcb *inp, struct in_addr *faddr, struct in_addr *laddr,
struct ucred *cred)
{
struct ifaddr *ifa;
struct sockaddr *sa;
struct sockaddr_in *sin;
struct route sro;
int error;
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
KASSERT(laddr != NULL, ("%s: laddr NULL", __func__));
/*
* Bypass source address selection and use the primary jail IP
* if requested.
*/
if (cred != NULL && !prison_saddrsel_ip4(cred, laddr))
return (0);
error = 0;
bzero(&sro, sizeof(sro));
sin = (struct sockaddr_in *)&sro.ro_dst;
sin->sin_family = AF_INET;
sin->sin_len = sizeof(struct sockaddr_in);
sin->sin_addr.s_addr = faddr->s_addr;
/*
* If route is known our src addr is taken from the i/f,
* else punt.
*
* Find out route to destination.
*/
if ((inp->inp_socket->so_options & SO_DONTROUTE) == 0)
in_rtalloc_ign(&sro, 0, inp->inp_inc.inc_fibnum);
/*
* If we found a route, use the address corresponding to
* the outgoing interface.
*
* Otherwise assume faddr is reachable on a directly connected
* network and try to find a corresponding interface to take
* the source address from.
*/
if (sro.ro_rt == NULL || sro.ro_rt->rt_ifp == NULL) {
struct in_ifaddr *ia;
struct ifnet *ifp;
ia = ifatoia(ifa_ifwithdstaddr((struct sockaddr *)sin,
inp->inp_socket->so_fibnum));
if (ia == NULL)
ia = ifatoia(ifa_ifwithnet((struct sockaddr *)sin, 0,
inp->inp_socket->so_fibnum));
if (ia == NULL) {
error = ENETUNREACH;
goto done;
}
if (cred == NULL || !prison_flag(cred, PR_IP4)) {
laddr->s_addr = ia->ia_addr.sin_addr.s_addr;
ifa_free(&ia->ia_ifa);
goto done;
}
ifp = ia->ia_ifp;
ifa_free(&ia->ia_ifa);
ia = NULL;
IF_ADDR_RLOCK(ifp);
TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
sa = ifa->ifa_addr;
if (sa->sa_family != AF_INET)
continue;
sin = (struct sockaddr_in *)sa;
if (prison_check_ip4(cred, &sin->sin_addr) == 0) {
ia = (struct in_ifaddr *)ifa;
break;
}
}
if (ia != NULL) {
laddr->s_addr = ia->ia_addr.sin_addr.s_addr;
IF_ADDR_RUNLOCK(ifp);
goto done;
}
IF_ADDR_RUNLOCK(ifp);
/* 3. As a last resort return the 'default' jail address. */
error = prison_get_ip4(cred, laddr);
goto done;
}
/*
* If the outgoing interface on the route found is not
* a loopback interface, use the address from that interface.
* In case of jails do those three steps:
* 1. check if the interface address belongs to the jail. If so use it.
* 2. check if we have any address on the outgoing interface
* belonging to this jail. If so use it.
* 3. as a last resort return the 'default' jail address.
*/
if ((sro.ro_rt->rt_ifp->if_flags & IFF_LOOPBACK) == 0) {
struct in_ifaddr *ia;
struct ifnet *ifp;
/* If not jailed, use the default returned. */
if (cred == NULL || !prison_flag(cred, PR_IP4)) {
ia = (struct in_ifaddr *)sro.ro_rt->rt_ifa;
laddr->s_addr = ia->ia_addr.sin_addr.s_addr;
goto done;
}
/* Jailed. */
/* 1. Check if the iface address belongs to the jail. */
sin = (struct sockaddr_in *)sro.ro_rt->rt_ifa->ifa_addr;
if (prison_check_ip4(cred, &sin->sin_addr) == 0) {
ia = (struct in_ifaddr *)sro.ro_rt->rt_ifa;
laddr->s_addr = ia->ia_addr.sin_addr.s_addr;
goto done;
}
/*
* 2. Check if we have any address on the outgoing interface
* belonging to this jail.
*/
ia = NULL;
ifp = sro.ro_rt->rt_ifp;
IF_ADDR_RLOCK(ifp);
TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
sa = ifa->ifa_addr;
if (sa->sa_family != AF_INET)
continue;
sin = (struct sockaddr_in *)sa;
if (prison_check_ip4(cred, &sin->sin_addr) == 0) {
ia = (struct in_ifaddr *)ifa;
break;
}
}
if (ia != NULL) {
laddr->s_addr = ia->ia_addr.sin_addr.s_addr;
IF_ADDR_RUNLOCK(ifp);
goto done;
}
IF_ADDR_RUNLOCK(ifp);
/* 3. As a last resort return the 'default' jail address. */
error = prison_get_ip4(cred, laddr);
goto done;
}
/*
* The outgoing interface is marked with 'loopback net', so a route
* to ourselves is here.
* Try to find the interface of the destination address and then
* take the address from there. That interface is not necessarily
* a loopback interface.
* In case of jails, check that it is an address of the jail
* and if we cannot find, fall back to the 'default' jail address.
*/
if ((sro.ro_rt->rt_ifp->if_flags & IFF_LOOPBACK) != 0) {
struct sockaddr_in sain;
struct in_ifaddr *ia;
bzero(&sain, sizeof(struct sockaddr_in));
sain.sin_family = AF_INET;
sain.sin_len = sizeof(struct sockaddr_in);
sain.sin_addr.s_addr = faddr->s_addr;
ia = ifatoia(ifa_ifwithdstaddr(sintosa(&sain),
inp->inp_socket->so_fibnum));
if (ia == NULL)
ia = ifatoia(ifa_ifwithnet(sintosa(&sain), 0,
inp->inp_socket->so_fibnum));
if (ia == NULL)
ia = ifatoia(ifa_ifwithaddr(sintosa(&sain)));
if (cred == NULL || !prison_flag(cred, PR_IP4)) {
if (ia == NULL) {
error = ENETUNREACH;
goto done;
}
laddr->s_addr = ia->ia_addr.sin_addr.s_addr;
ifa_free(&ia->ia_ifa);
goto done;
}
/* Jailed. */
if (ia != NULL) {
struct ifnet *ifp;
ifp = ia->ia_ifp;
ifa_free(&ia->ia_ifa);
ia = NULL;
IF_ADDR_RLOCK(ifp);
TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
sa = ifa->ifa_addr;
if (sa->sa_family != AF_INET)
continue;
sin = (struct sockaddr_in *)sa;
if (prison_check_ip4(cred,
&sin->sin_addr) == 0) {
ia = (struct in_ifaddr *)ifa;
break;
}
}
if (ia != NULL) {
laddr->s_addr = ia->ia_addr.sin_addr.s_addr;
IF_ADDR_RUNLOCK(ifp);
goto done;
}
IF_ADDR_RUNLOCK(ifp);
}
/* 3. As a last resort return the 'default' jail address. */
error = prison_get_ip4(cred, laddr);
goto done;
}
done:
if (sro.ro_rt != NULL)
RTFREE(sro.ro_rt);
return (error);
}
/*
* Set up for a connect from a socket to the specified address.
* On entry, *laddrp and *lportp should contain the current local
* address and port for the PCB; these are updated to the values
* that should be placed in inp_laddr and inp_lport to complete
* the connect.
*
* On success, *faddrp and *fportp will be set to the remote address
* and port. These are not updated in the error case.
*
* If the operation fails because the connection already exists,
* *oinpp will be set to the PCB of that connection so that the
* caller can decide to override it. In all other cases, *oinpp
* is set to NULL.
*/
int
in_pcbconnect_setup(struct inpcb *inp, struct sockaddr *nam,
in_addr_t *laddrp, u_short *lportp, in_addr_t *faddrp, u_short *fportp,
struct inpcb **oinpp, struct ucred *cred)
{
struct rm_priotracker in_ifa_tracker;
struct sockaddr_in *sin = (struct sockaddr_in *)nam;
1994-05-24 10:09:53 +00:00
struct in_ifaddr *ia;
struct inpcb *oinp;
struct in_addr laddr, faddr;
u_short lport, fport;
int error;
1994-05-24 10:09:53 +00:00
/*
* Because a global state change doesn't actually occur here, a read
* lock is sufficient.
*/
INP_LOCK_ASSERT(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
INP_HASH_LOCK_ASSERT(inp->inp_pcbinfo);
if (oinpp != NULL)
*oinpp = NULL;
if (nam->sa_len != sizeof (*sin))
1994-05-24 10:09:53 +00:00
return (EINVAL);
if (sin->sin_family != AF_INET)
return (EAFNOSUPPORT);
if (sin->sin_port == 0)
return (EADDRNOTAVAIL);
laddr.s_addr = *laddrp;
lport = *lportp;
faddr = sin->sin_addr;
fport = sin->sin_port;
if (!TAILQ_EMPTY(&V_in_ifaddrhead)) {
1994-05-24 10:09:53 +00:00
/*
* If the destination address is INADDR_ANY,
* use the primary local address.
* If the supplied address is INADDR_BROADCAST,
* and the primary interface supports broadcast,
* choose the broadcast address for that interface.
*/
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
if (faddr.s_addr == INADDR_ANY) {
IN_IFADDR_RLOCK(&in_ifa_tracker);
faddr =
IA_SIN(TAILQ_FIRST(&V_in_ifaddrhead))->sin_addr;
IN_IFADDR_RUNLOCK(&in_ifa_tracker);
if (cred != NULL &&
(error = prison_get_ip4(cred, &faddr)) != 0)
return (error);
} else if (faddr.s_addr == (u_long)INADDR_BROADCAST) {
IN_IFADDR_RLOCK(&in_ifa_tracker);
if (TAILQ_FIRST(&V_in_ifaddrhead)->ia_ifp->if_flags &
IFF_BROADCAST)
faddr = satosin(&TAILQ_FIRST(
&V_in_ifaddrhead)->ia_broadaddr)->sin_addr;
IN_IFADDR_RUNLOCK(&in_ifa_tracker);
}
1994-05-24 10:09:53 +00:00
}
if (laddr.s_addr == INADDR_ANY) {
error = in_pcbladdr(inp, &faddr, &laddr, cred);
1994-05-24 10:09:53 +00:00
/*
* If the destination address is multicast and an outgoing
* interface has been set as a multicast option, prefer the
1994-05-24 10:09:53 +00:00
* address of that interface as our source address.
*/
if (IN_MULTICAST(ntohl(faddr.s_addr)) &&
1994-05-24 10:09:53 +00:00
inp->inp_moptions != NULL) {
struct ip_moptions *imo;
struct ifnet *ifp;
imo = inp->inp_moptions;
if (imo->imo_multicast_ifp != NULL) {
ifp = imo->imo_multicast_ifp;
IN_IFADDR_RLOCK(&in_ifa_tracker);
TAILQ_FOREACH(ia, &V_in_ifaddrhead, ia_link) {
if ((ia->ia_ifp == ifp) &&
(cred == NULL ||
prison_check_ip4(cred,
&ia->ia_addr.sin_addr) == 0))
1994-05-24 10:09:53 +00:00
break;
}
if (ia == NULL)
error = EADDRNOTAVAIL;
else {
laddr = ia->ia_addr.sin_addr;
error = 0;
}
IN_IFADDR_RUNLOCK(&in_ifa_tracker);
1994-05-24 10:09:53 +00:00
}
}
if (error)
return (error);
}
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
oinp = in_pcblookup_hash_locked(inp->inp_pcbinfo, faddr, fport,
laddr, lport, 0, NULL);
if (oinp != NULL) {
if (oinpp != NULL)
*oinpp = oinp;
1994-05-24 10:09:53 +00:00
return (EADDRINUSE);
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
}
if (lport == 0) {
error = in_pcbbind_setup(inp, NULL, &laddr.s_addr, &lport,
cred);
if (error)
return (error);
1994-05-24 10:09:53 +00:00
}
*laddrp = laddr.s_addr;
*lportp = lport;
*faddrp = faddr.s_addr;
*fportp = fport;
1994-05-24 10:09:53 +00:00
return (0);
}
void
in_pcbdisconnect(struct inpcb *inp)
1994-05-24 10:09:53 +00:00
{
INP_WLOCK_ASSERT(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
INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
1994-05-24 10:09:53 +00:00
inp->inp_faddr.s_addr = INADDR_ANY;
inp->inp_fport = 0;
in_pcbrehash(inp);
1994-05-24 10:09:53 +00:00
}
#endif /* INET */
1994-05-24 10:09:53 +00:00
/*
* in_pcbdetach() is responsibe for disassociating a socket from an inpcb.
* For most protocols, this will be invoked immediately prior to calling
* in_pcbfree(). However, with TCP the inpcb may significantly outlive the
* socket, in which case in_pcbfree() is deferred.
*/
void
in_pcbdetach(struct inpcb *inp)
1994-05-24 10:09:53 +00:00
{
KASSERT(inp->inp_socket != NULL, ("%s: inp_socket == NULL", __func__));
Implement kernel support for hardware rate limited sockets. - Add RATELIMIT kernel configuration keyword which must be set to enable the new functionality. - Add support for hardware driven, Receive Side Scaling, RSS aware, rate limited sendqueues and expose the functionality through the already established SO_MAX_PACING_RATE setsockopt(). The API support rates in the range from 1 to 4Gbytes/s which are suitable for regular TCP and UDP streams. The setsockopt(2) manual page has been updated. - Add rate limit function callback API to "struct ifnet" which supports the following operations: if_snd_tag_alloc(), if_snd_tag_modify(), if_snd_tag_query() and if_snd_tag_free(). - Add support to ifconfig to view, set and clear the IFCAP_TXRTLMT flag, which tells if a network driver supports rate limiting or not. - This patch also adds support for rate limiting through VLAN and LAGG intermediate network devices. - How rate limiting works: 1) The userspace application calls setsockopt() after accepting or making a new connection to set the rate which is then stored in the socket structure in the kernel. Later on when packets are transmitted a check is made in the transmit path for rate changes. A rate change implies a non-blocking ifp->if_snd_tag_alloc() call will be made to the destination network interface, which then sets up a custom sendqueue with the given rate limitation parameter. A "struct m_snd_tag" pointer is returned which serves as a "snd_tag" hint in the m_pkthdr for the subsequently transmitted mbufs. 2) When the network driver sees the "m->m_pkthdr.snd_tag" different from NULL, it will move the packets into a designated rate limited sendqueue given by the snd_tag pointer. It is up to the individual drivers how the rate limited traffic will be rate limited. 3) Route changes are detected by the NIC drivers in the ifp->if_transmit() routine when the ifnet pointer in the incoming snd_tag mismatches the one of the network interface. The network adapter frees the mbuf and returns EAGAIN which causes the ip_output() to release and clear the send tag. Upon next ip_output() a new "snd_tag" will be tried allocated. 4) When the PCB is detached the custom sendqueue will be released by a non-blocking ifp->if_snd_tag_free() call to the currently bound network interface. Reviewed by: wblock (manpages), adrian, gallatin, scottl (network) Differential Revision: https://reviews.freebsd.org/D3687 Sponsored by: Mellanox Technologies MFC after: 3 months
2017-01-18 13:31:17 +00:00
#ifdef RATELIMIT
if (inp->inp_snd_tag != NULL)
in_pcbdetach_txrtlmt(inp);
#endif
inp->inp_socket->so_pcb = NULL;
inp->inp_socket = NULL;
}
/*
* in_pcbref() bumps the reference count on an inpcb in order to maintain
* stability of an inpcb pointer despite the inpcb lock being released. This
* is used in TCP when the inpcbinfo lock needs to be acquired or upgraded,
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
* but where the inpcb lock may already held, or when acquiring a reference
* via a pcbgroup.
*
* in_pcbref() should be used only to provide brief memory stability, and
* must always be followed by a call to INP_WLOCK() and in_pcbrele() to
* garbage collect the inpcb if it has been in_pcbfree()'d from another
* context. Until in_pcbrele() has returned that the inpcb is still valid,
* lock and rele are the *only* safe operations that may be performed on the
* inpcb.
*
* While the inpcb will not be freed, releasing the inpcb lock means that the
* connection's state may change, so the caller should be careful to
* revalidate any cached state on reacquiring the lock. Drop the reference
* using in_pcbrele().
*/
void
in_pcbref(struct inpcb *inp)
{
KASSERT(inp->inp_refcount > 0, ("%s: refcount 0", __func__));
refcount_acquire(&inp->inp_refcount);
}
/*
* Drop a refcount on an inpcb elevated using in_pcbref(); because a call to
* in_pcbfree() may have been made between in_pcbref() and in_pcbrele(), we
* return a flag indicating whether or not the inpcb remains valid. If it is
* valid, we return with the inpcb lock held.
*
* Notice that, unlike in_pcbref(), the inpcb lock must be held to drop a
* reference on an inpcb. Historically more work was done here (actually, in
* in_pcbfree_internal()) but has been moved to in_pcbfree() to avoid the
* need for the pcbinfo lock in in_pcbrele(). Deferring the free is entirely
* about memory stability (and continued use of the write lock).
*/
int
in_pcbrele_rlocked(struct inpcb *inp)
{
struct inpcbinfo *pcbinfo;
KASSERT(inp->inp_refcount > 0, ("%s: refcount 0", __func__));
INP_RLOCK_ASSERT(inp);
if (refcount_release(&inp->inp_refcount) == 0) {
/*
* If the inpcb has been freed, let the caller know, even if
* this isn't the last reference.
*/
if (inp->inp_flags2 & INP_FREED) {
INP_RUNLOCK(inp);
return (1);
}
return (0);
}
KASSERT(inp->inp_socket == NULL, ("%s: inp_socket != NULL", __func__));
INP_RUNLOCK(inp);
pcbinfo = inp->inp_pcbinfo;
uma_zfree(pcbinfo->ipi_zone, inp);
return (1);
}
int
in_pcbrele_wlocked(struct inpcb *inp)
{
struct inpcbinfo *pcbinfo;
KASSERT(inp->inp_refcount > 0, ("%s: refcount 0", __func__));
INP_WLOCK_ASSERT(inp);
if (refcount_release(&inp->inp_refcount) == 0) {
/*
* If the inpcb has been freed, let the caller know, even if
* this isn't the last reference.
*/
if (inp->inp_flags2 & INP_FREED) {
INP_WUNLOCK(inp);
return (1);
}
return (0);
}
KASSERT(inp->inp_socket == NULL, ("%s: inp_socket != NULL", __func__));
INP_WUNLOCK(inp);
pcbinfo = inp->inp_pcbinfo;
uma_zfree(pcbinfo->ipi_zone, inp);
return (1);
}
/*
* Temporary wrapper.
*/
int
in_pcbrele(struct inpcb *inp)
{
return (in_pcbrele_wlocked(inp));
}
/*
* Unconditionally schedule an inpcb to be freed by decrementing its
* reference count, which should occur only after the inpcb has been detached
* from its socket. If another thread holds a temporary reference (acquired
* using in_pcbref()) then the free is deferred until that reference is
* released using in_pcbrele(), but the inpcb is still unlocked. Almost all
* work, including removal from global lists, is done in this context, where
* the pcbinfo lock is held.
*/
void
in_pcbfree(struct inpcb *inp)
{
struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
KASSERT(inp->inp_socket == NULL, ("%s: inp_socket != NULL", __func__));
#ifdef INVARIANTS
if (pcbinfo == &V_tcbinfo) {
INP_INFO_LOCK_ASSERT(pcbinfo);
} else {
INP_INFO_WLOCK_ASSERT(pcbinfo);
}
#endif
INP_WLOCK_ASSERT(inp);
/* XXXRW: Do as much as possible here. */
Merge projects/ipsec into head/. Small summary ------------- o Almost all IPsec releated code was moved into sys/netipsec. o New kernel modules added: ipsec.ko and tcpmd5.ko. New kernel option IPSEC_SUPPORT added. It enables support for loading and unloading of ipsec.ko and tcpmd5.ko kernel modules. o IPSEC_NAT_T option was removed. Now NAT-T support is enabled by default. The UDP_ENCAP_ESPINUDP_NON_IKE encapsulation type support was removed. Added TCP/UDP checksum handling for inbound packets that were decapsulated by transport mode SAs. setkey(8) modified to show run-time NAT-T configuration of SA. o New network pseudo interface if_ipsec(4) added. For now it is build as part of ipsec.ko module (or with IPSEC kernel). It implements IPsec virtual tunnels to create route-based VPNs. o The network stack now invokes IPsec functions using special methods. The only one header file <netipsec/ipsec_support.h> should be included to declare all the needed things to work with IPsec. o All IPsec protocols handlers (ESP/AH/IPCOMP protosw) were removed. Now these protocols are handled directly via IPsec methods. o TCP_SIGNATURE support was reworked to be more close to RFC. o PF_KEY SADB was reworked: - now all security associations stored in the single SPI namespace, and all SAs MUST have unique SPI. - several hash tables added to speed up lookups in SADB. - SADB now uses rmlock to protect access, and concurrent threads can do SA lookups in the same time. - many PF_KEY message handlers were reworked to reflect changes in SADB. - SADB_UPDATE message was extended to support new PF_KEY headers: SADB_X_EXT_NEW_ADDRESS_SRC and SADB_X_EXT_NEW_ADDRESS_DST. They can be used by IKE daemon to change SA addresses. o ipsecrequest and secpolicy structures were cardinally changed to avoid locking protection for ipsecrequest. Now we support only limited number (4) of bundled SAs, but they are supported for both INET and INET6. o INPCB security policy cache was introduced. Each PCB now caches used security policies to avoid SP lookup for each packet. o For inbound security policies added the mode, when the kernel does check for full history of applied IPsec transforms. o References counting rules for security policies and security associations were changed. The proper SA locking added into xform code. o xform code was also changed. Now it is possible to unregister xforms. tdb_xxx structures were changed and renamed to reflect changes in SADB/SPDB, and changed rules for locking and refcounting. Reviewed by: gnn, wblock Obtained from: Yandex LLC Relnotes: yes Sponsored by: Yandex LLC Differential Revision: https://reviews.freebsd.org/D9352
2017-02-06 08:49:57 +00:00
#if defined(IPSEC) || defined(IPSEC_SUPPORT)
if (inp->inp_sp != NULL)
ipsec_delete_pcbpolicy(inp);
#endif
INP_LIST_WLOCK(pcbinfo);
inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
in_pcbremlists(inp);
INP_LIST_WUNLOCK(pcbinfo);
#ifdef INET6
if (inp->inp_vflag & INP_IPV6PROTO) {
ip6_freepcbopts(inp->in6p_outputopts);
if (inp->in6p_moptions != NULL)
ip6_freemoptions(inp->in6p_moptions);
}
#endif
if (inp->inp_options)
(void)m_free(inp->inp_options);
#ifdef INET
if (inp->inp_moptions != NULL)
inp_freemoptions(inp->inp_moptions);
#endif
RO_RTFREE(&inp->inp_route);
if (inp->inp_route.ro_lle)
LLE_FREE(inp->inp_route.ro_lle); /* zeros ro_lle */
inp->inp_vflag = 0;
inp->inp_flags2 |= INP_FREED;
crfree(inp->inp_cred);
#ifdef MAC
mac_inpcb_destroy(inp);
#endif
if (!in_pcbrele_wlocked(inp))
INP_WUNLOCK(inp);
}
/*
* in_pcbdrop() removes an inpcb from hashed lists, releasing its address and
* port reservation, and preventing it from being returned by inpcb lookups.
*
* It is used by TCP to mark an inpcb as unused and avoid future packet
* delivery or event notification when a socket remains open but TCP has
* closed. This might occur as a result of a shutdown()-initiated TCP close
* or a RST on the wire, and allows the port binding to be reused while still
* maintaining the invariant that so_pcb always points to a valid inpcb until
* in_pcbdetach().
*
* XXXRW: Possibly in_pcbdrop() should also prevent future notifications by
* in_pcbnotifyall() and in_pcbpurgeif0()?
*/
void
in_pcbdrop(struct inpcb *inp)
{
INP_WLOCK_ASSERT(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
/*
* XXXRW: Possibly we should protect the setting of INP_DROPPED with
* the hash lock...?
*/
inp->inp_flags |= INP_DROPPED;
if (inp->inp_flags & INP_INHASHLIST) {
struct inpcbport *phd = inp->inp_phd;
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_HASH_WLOCK(inp->inp_pcbinfo);
LIST_REMOVE(inp, inp_hash);
LIST_REMOVE(inp, inp_portlist);
if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
LIST_REMOVE(phd, phd_hash);
free(phd, M_PCB);
}
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_HASH_WUNLOCK(inp->inp_pcbinfo);
inp->inp_flags &= ~INP_INHASHLIST;
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#ifdef PCBGROUP
in_pcbgroup_remove(inp);
#endif
}
}
#ifdef INET
/*
* Common routines to return the socket addresses associated with inpcbs.
*/
struct sockaddr *
in_sockaddr(in_port_t port, struct in_addr *addr_p)
{
struct sockaddr_in *sin;
sin = malloc(sizeof *sin, M_SONAME,
M_WAITOK | M_ZERO);
sin->sin_family = AF_INET;
sin->sin_len = sizeof(*sin);
sin->sin_addr = *addr_p;
sin->sin_port = port;
return (struct sockaddr *)sin;
}
int
in_getsockaddr(struct socket *so, struct sockaddr **nam)
1994-05-24 10:09:53 +00:00
{
struct inpcb *inp;
struct in_addr addr;
in_port_t port;
inp = sotoinpcb(so);
KASSERT(inp != NULL, ("in_getsockaddr: inp == NULL"));
INP_RLOCK(inp);
port = inp->inp_lport;
addr = inp->inp_laddr;
INP_RUNLOCK(inp);
*nam = in_sockaddr(port, &addr);
return 0;
1994-05-24 10:09:53 +00:00
}
int
in_getpeeraddr(struct socket *so, struct sockaddr **nam)
1994-05-24 10:09:53 +00:00
{
struct inpcb *inp;
struct in_addr addr;
in_port_t port;
inp = sotoinpcb(so);
KASSERT(inp != NULL, ("in_getpeeraddr: inp == NULL"));
INP_RLOCK(inp);
port = inp->inp_fport;
addr = inp->inp_faddr;
INP_RUNLOCK(inp);
*nam = in_sockaddr(port, &addr);
return 0;
1994-05-24 10:09:53 +00:00
}
void
in_pcbnotifyall(struct inpcbinfo *pcbinfo, struct in_addr faddr, int errno,
struct inpcb *(*notify)(struct inpcb *, int))
{
struct inpcb *inp, *inp_temp;
INP_INFO_WLOCK(pcbinfo);
LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
INP_WLOCK(inp);
#ifdef INET6
if ((inp->inp_vflag & INP_IPV4) == 0) {
INP_WUNLOCK(inp);
continue;
}
#endif
if (inp->inp_faddr.s_addr != faddr.s_addr ||
inp->inp_socket == NULL) {
INP_WUNLOCK(inp);
continue;
}
if ((*notify)(inp, errno))
INP_WUNLOCK(inp);
}
INP_INFO_WUNLOCK(pcbinfo);
}
void
in_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
{
struct inpcb *inp;
struct ip_moptions *imo;
int i, gap;
INP_INFO_WLOCK(pcbinfo);
LIST_FOREACH(inp, pcbinfo->ipi_listhead, inp_list) {
INP_WLOCK(inp);
imo = inp->inp_moptions;
if ((inp->inp_vflag & INP_IPV4) &&
imo != NULL) {
/*
* Unselect the outgoing interface if it is being
* detached.
*/
if (imo->imo_multicast_ifp == ifp)
imo->imo_multicast_ifp = NULL;
/*
* Drop multicast group membership if we joined
* through the interface being detached.
*/
for (i = 0, gap = 0; i < imo->imo_num_memberships;
i++) {
if (imo->imo_membership[i]->inm_ifp == ifp) {
in_delmulti(imo->imo_membership[i]);
gap++;
} else if (gap != 0)
imo->imo_membership[i - gap] =
imo->imo_membership[i];
}
imo->imo_num_memberships -= gap;
}
INP_WUNLOCK(inp);
}
INP_INFO_WUNLOCK(pcbinfo);
}
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +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
* Lookup a PCB based on the local address and port. Caller must hold the
* hash lock. No inpcb locks or references are acquired.
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
*/
#define INP_LOOKUP_MAPPED_PCB_COST 3
1994-05-24 10:09:53 +00:00
struct inpcb *
in_pcblookup_local(struct inpcbinfo *pcbinfo, struct in_addr laddr,
u_short lport, int lookupflags, struct ucred *cred)
1994-05-24 10:09:53 +00:00
{
struct inpcb *inp;
#ifdef INET6
int matchwild = 3 + INP_LOOKUP_MAPPED_PCB_COST;
#else
int matchwild = 3;
#endif
int wildcard;
KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
("%s: invalid lookup flags %d", __func__, lookupflags));
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_HASH_LOCK_ASSERT(pcbinfo);
if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
struct inpcbhead *head;
/*
* Look for an unconnected (wildcard foreign addr) PCB that
* matches the local address and port we're looking for.
*/
head = &pcbinfo->ipi_hashbase[INP_PCBHASH(INADDR_ANY, lport,
0, pcbinfo->ipi_hashmask)];
LIST_FOREACH(inp, head, inp_hash) {
#ifdef INET6
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
/* XXX inp locking */
1999-12-21 11:14:12 +00:00
if ((inp->inp_vflag & INP_IPV4) == 0)
continue;
#endif
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
if (inp->inp_faddr.s_addr == INADDR_ANY &&
inp->inp_laddr.s_addr == laddr.s_addr &&
inp->inp_lport == lport) {
/*
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
* Found?
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
*/
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
if (cred == NULL ||
prison_equal_ip4(cred->cr_prison,
inp->inp_cred->cr_prison))
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
return (inp);
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
}
}
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
/*
* Not found.
*/
return (NULL);
} else {
struct inpcbporthead *porthash;
struct inpcbport *phd;
struct inpcb *match = NULL;
/*
* Best fit PCB lookup.
*
* First see if this local port is in use by looking on the
* port hash list.
*/
porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
pcbinfo->ipi_porthashmask)];
LIST_FOREACH(phd, porthash, phd_hash) {
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
if (phd->phd_port == lport)
1994-05-24 10:09:53 +00:00
break;
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
}
if (phd != NULL) {
/*
* Port is in use by one or more PCBs. Look for best
* fit.
*/
LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
wildcard = 0;
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
if (cred != NULL &&
!prison_equal_ip4(inp->inp_cred->cr_prison,
cred->cr_prison))
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
continue;
#ifdef INET6
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
/* XXX inp locking */
1999-12-21 11:14:12 +00:00
if ((inp->inp_vflag & INP_IPV4) == 0)
continue;
/*
* We never select the PCB that has
* INP_IPV6 flag and is bound to :: if
* we have another PCB which is bound
* to 0.0.0.0. If a PCB has the
* INP_IPV6 flag, then we set its cost
* higher than IPv4 only PCBs.
*
* Note that the case only happens
* when a socket is bound to ::, under
* the condition that the use of the
* mapped address is allowed.
*/
if ((inp->inp_vflag & INP_IPV6) != 0)
wildcard += INP_LOOKUP_MAPPED_PCB_COST;
#endif
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
if (inp->inp_faddr.s_addr != INADDR_ANY)
wildcard++;
if (inp->inp_laddr.s_addr != INADDR_ANY) {
if (laddr.s_addr == INADDR_ANY)
wildcard++;
else if (inp->inp_laddr.s_addr != laddr.s_addr)
continue;
} else {
if (laddr.s_addr != INADDR_ANY)
wildcard++;
}
if (wildcard < matchwild) {
match = inp;
matchwild = wildcard;
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
if (matchwild == 0)
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
break;
}
}
1994-05-24 10:09:53 +00:00
}
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
return (match);
1994-05-24 10:09:53 +00:00
}
}
#undef INP_LOOKUP_MAPPED_PCB_COST
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#ifdef PCBGROUP
/*
* Lookup PCB in hash list, using pcbgroup tables.
*/
static struct inpcb *
in_pcblookup_group(struct inpcbinfo *pcbinfo, struct inpcbgroup *pcbgroup,
struct in_addr faddr, u_int fport_arg, struct in_addr laddr,
u_int lport_arg, int lookupflags, struct ifnet *ifp)
{
struct inpcbhead *head;
struct inpcb *inp, *tmpinp;
u_short fport = fport_arg, lport = lport_arg;
/*
* First look for an exact match.
*/
tmpinp = NULL;
INP_GROUP_LOCK(pcbgroup);
head = &pcbgroup->ipg_hashbase[INP_PCBHASH(faddr.s_addr, lport, fport,
pcbgroup->ipg_hashmask)];
LIST_FOREACH(inp, head, inp_pcbgrouphash) {
#ifdef INET6
/* XXX inp locking */
if ((inp->inp_vflag & INP_IPV4) == 0)
continue;
#endif
if (inp->inp_faddr.s_addr == faddr.s_addr &&
inp->inp_laddr.s_addr == laddr.s_addr &&
inp->inp_fport == fport &&
inp->inp_lport == lport) {
/*
* XXX We should be able to directly return
* the inp here, without any checks.
* Well unless both bound with SO_REUSEPORT?
*/
if (prison_flag(inp->inp_cred, PR_IP4))
goto found;
if (tmpinp == NULL)
tmpinp = inp;
}
}
if (tmpinp != NULL) {
inp = tmpinp;
goto found;
}
2014-07-10 03:10:56 +00:00
#ifdef RSS
/*
* For incoming connections, we may wish to do a wildcard
* match for an RSS-local socket.
*/
if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
struct inpcb *local_wild = NULL, *local_exact = NULL;
#ifdef INET6
struct inpcb *local_wild_mapped = NULL;
#endif
struct inpcb *jail_wild = NULL;
struct inpcbhead *head;
int injail;
/*
* Order of socket selection - we always prefer jails.
* 1. jailed, non-wild.
* 2. jailed, wild.
* 3. non-jailed, non-wild.
* 4. non-jailed, wild.
*/
head = &pcbgroup->ipg_hashbase[INP_PCBHASH(INADDR_ANY,
lport, 0, pcbgroup->ipg_hashmask)];
LIST_FOREACH(inp, head, inp_pcbgrouphash) {
#ifdef INET6
/* XXX inp locking */
if ((inp->inp_vflag & INP_IPV4) == 0)
continue;
#endif
if (inp->inp_faddr.s_addr != INADDR_ANY ||
inp->inp_lport != lport)
continue;
injail = prison_flag(inp->inp_cred, PR_IP4);
if (injail) {
if (prison_check_ip4(inp->inp_cred,
&laddr) != 0)
continue;
} else {
if (local_exact != NULL)
continue;
}
if (inp->inp_laddr.s_addr == laddr.s_addr) {
if (injail)
goto found;
else
local_exact = inp;
} else if (inp->inp_laddr.s_addr == INADDR_ANY) {
#ifdef INET6
/* XXX inp locking, NULL check */
if (inp->inp_vflag & INP_IPV6PROTO)
local_wild_mapped = inp;
else
#endif
if (injail)
jail_wild = inp;
else
local_wild = inp;
}
} /* LIST_FOREACH */
inp = jail_wild;
if (inp == NULL)
inp = local_exact;
if (inp == NULL)
inp = local_wild;
#ifdef INET6
if (inp == NULL)
inp = local_wild_mapped;
#endif
if (inp != NULL)
goto found;
}
#endif
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
/*
* Then look for a wildcard match, if requested.
*/
if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
struct inpcb *local_wild = NULL, *local_exact = NULL;
#ifdef INET6
struct inpcb *local_wild_mapped = NULL;
#endif
struct inpcb *jail_wild = NULL;
struct inpcbhead *head;
int injail;
/*
* Order of socket selection - we always prefer jails.
* 1. jailed, non-wild.
* 2. jailed, wild.
* 3. non-jailed, non-wild.
* 4. non-jailed, wild.
*/
head = &pcbinfo->ipi_wildbase[INP_PCBHASH(INADDR_ANY, lport,
0, pcbinfo->ipi_wildmask)];
LIST_FOREACH(inp, head, inp_pcbgroup_wild) {
#ifdef INET6
/* XXX inp locking */
if ((inp->inp_vflag & INP_IPV4) == 0)
continue;
#endif
if (inp->inp_faddr.s_addr != INADDR_ANY ||
inp->inp_lport != lport)
continue;
injail = prison_flag(inp->inp_cred, PR_IP4);
if (injail) {
if (prison_check_ip4(inp->inp_cred,
&laddr) != 0)
continue;
} else {
if (local_exact != NULL)
continue;
}
if (inp->inp_laddr.s_addr == laddr.s_addr) {
if (injail)
goto found;
else
local_exact = inp;
} else if (inp->inp_laddr.s_addr == INADDR_ANY) {
#ifdef INET6
/* XXX inp locking, NULL check */
if (inp->inp_vflag & INP_IPV6PROTO)
local_wild_mapped = inp;
else
#endif
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
if (injail)
jail_wild = inp;
else
local_wild = inp;
}
} /* LIST_FOREACH */
inp = jail_wild;
if (inp == NULL)
inp = local_exact;
if (inp == NULL)
inp = local_wild;
#ifdef INET6
if (inp == NULL)
inp = local_wild_mapped;
#endif
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
if (inp != NULL)
goto found;
} /* if (lookupflags & INPLOOKUP_WILDCARD) */
INP_GROUP_UNLOCK(pcbgroup);
return (NULL);
found:
in_pcbref(inp);
INP_GROUP_UNLOCK(pcbgroup);
if (lookupflags & INPLOOKUP_WLOCKPCB) {
INP_WLOCK(inp);
if (in_pcbrele_wlocked(inp))
return (NULL);
} else if (lookupflags & INPLOOKUP_RLOCKPCB) {
INP_RLOCK(inp);
if (in_pcbrele_rlocked(inp))
return (NULL);
} else
panic("%s: locking bug", __func__);
return (inp);
}
#endif /* PCBGROUP */
/*
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
* Lookup PCB in hash list, using pcbinfo tables. This variation assumes
* that the caller has locked the hash list, and will not perform any further
* locking or reference operations on either the hash list or the connection.
*/
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
static struct inpcb *
in_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in_addr faddr,
u_int fport_arg, struct in_addr laddr, u_int lport_arg, int lookupflags,
struct ifnet *ifp)
{
struct inpcbhead *head;
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
struct inpcb *inp, *tmpinp;
u_short fport = fport_arg, lport = lport_arg;
KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
("%s: invalid lookup flags %d", __func__, lookupflags));
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_HASH_LOCK_ASSERT(pcbinfo);
/*
* First look for an exact match.
*/
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
tmpinp = NULL;
head = &pcbinfo->ipi_hashbase[INP_PCBHASH(faddr.s_addr, lport, fport,
pcbinfo->ipi_hashmask)];
LIST_FOREACH(inp, head, inp_hash) {
#ifdef INET6
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
/* XXX inp locking */
1999-12-21 11:14:12 +00:00
if ((inp->inp_vflag & INP_IPV4) == 0)
continue;
#endif
if (inp->inp_faddr.s_addr == faddr.s_addr &&
inp->inp_laddr.s_addr == laddr.s_addr &&
inp->inp_fport == fport &&
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
inp->inp_lport == lport) {
/*
* XXX We should be able to directly return
* the inp here, without any checks.
* Well unless both bound with SO_REUSEPORT?
*/
if (prison_flag(inp->inp_cred, PR_IP4))
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
return (inp);
if (tmpinp == NULL)
tmpinp = inp;
}
}
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
if (tmpinp != NULL)
return (tmpinp);
/*
* Then look for a wildcard match, if requested.
*/
if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
struct inpcb *local_wild = NULL, *local_exact = NULL;
#ifdef INET6
struct inpcb *local_wild_mapped = NULL;
#endif
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
struct inpcb *jail_wild = NULL;
int injail;
/*
* Order of socket selection - we always prefer jails.
* 1. jailed, non-wild.
* 2. jailed, wild.
* 3. non-jailed, non-wild.
* 4. non-jailed, wild.
*/
head = &pcbinfo->ipi_hashbase[INP_PCBHASH(INADDR_ANY, lport,
0, pcbinfo->ipi_hashmask)];
LIST_FOREACH(inp, head, inp_hash) {
#ifdef INET6
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
/* XXX inp locking */
1999-12-21 11:14:12 +00:00
if ((inp->inp_vflag & INP_IPV4) == 0)
continue;
#endif
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
if (inp->inp_faddr.s_addr != INADDR_ANY ||
inp->inp_lport != lport)
continue;
injail = prison_flag(inp->inp_cred, PR_IP4);
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
if (injail) {
if (prison_check_ip4(inp->inp_cred,
&laddr) != 0)
continue;
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
} else {
if (local_exact != NULL)
continue;
}
if (inp->inp_laddr.s_addr == laddr.s_addr) {
if (injail)
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
return (inp);
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
else
local_exact = inp;
} else if (inp->inp_laddr.s_addr == INADDR_ANY) {
#ifdef INET6
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
/* XXX inp locking, NULL check */
if (inp->inp_vflag & INP_IPV6PROTO)
local_wild_mapped = inp;
else
#endif
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
if (injail)
jail_wild = inp;
else
local_wild = inp;
}
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
} /* LIST_FOREACH */
if (jail_wild != NULL)
return (jail_wild);
if (local_exact != NULL)
return (local_exact);
if (local_wild != NULL)
return (local_wild);
#ifdef INET6
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
if (local_wild_mapped != NULL)
return (local_wild_mapped);
#endif
} /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
MFp4: Bring in updated jail support from bz_jail branch. This enhances the current jail implementation to permit multiple addresses per jail. In addtion to IPv4, IPv6 is supported as well. Due to updated checks it is even possible to have jails without an IP address at all, which basically gives one a chroot with restricted process view, no networking,.. SCTP support was updated and supports IPv6 in jails as well. Cpuset support permits jails to be bound to specific processor sets after creation. Jails can have an unrestricted (no duplicate protection, etc.) name in addition to the hostname. The jail name cannot be changed from within a jail and is considered to be used for management purposes or as audit-token in the future. DDB 'show jails' command was added to aid debugging. Proper compat support permits 32bit jail binaries to be used on 64bit systems to manage jails. Also backward compatibility was preserved where possible: for jail v1 syscalls, as well as with user space management utilities. Both jail as well as prison version were updated for the new features. A gap was intentionally left as the intermediate versions had been used by various patches floating around the last years. Bump __FreeBSD_version for the afore mentioned and in kernel changes. Special thanks to: - Pawel Jakub Dawidek (pjd) for his multi-IPv4 patches and Olivier Houchard (cognet) for initial single-IPv6 patches. - Jeff Roberson (jeff) and Randall Stewart (rrs) for their help, ideas and review on cpuset and SCTP support. - Robert Watson (rwatson) for lots and lots of help, discussions, suggestions and review of most of the patch at various stages. - John Baldwin (jhb) for his help. - Simon L. Nielsen (simon) as early adopter testing changes on cluster machines as well as all the testers and people who provided feedback the last months on freebsd-jail and other channels. - My employer, CK Software GmbH, for the support so I could work on this. Reviewed by: (see above) MFC after: 3 months (this is just so that I get the mail) X-MFC Before: 7.2-RELEASE if possible
2008-11-29 14:32:14 +00:00
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
return (NULL);
}
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
/*
* Lookup PCB in hash list, using pcbinfo tables. This variation locks the
* hash list lock, and will return the inpcb locked (i.e., requires
* INPLOOKUP_LOCKPCB).
*/
static struct inpcb *
in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
u_int fport, struct in_addr laddr, u_int lport, int lookupflags,
struct ifnet *ifp)
{
struct inpcb *inp;
INP_HASH_RLOCK(pcbinfo);
inp = in_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
(lookupflags & ~(INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)), ifp);
if (inp != NULL) {
in_pcbref(inp);
INP_HASH_RUNLOCK(pcbinfo);
if (lookupflags & INPLOOKUP_WLOCKPCB) {
INP_WLOCK(inp);
if (in_pcbrele_wlocked(inp))
return (NULL);
} else if (lookupflags & INPLOOKUP_RLOCKPCB) {
INP_RLOCK(inp);
if (in_pcbrele_rlocked(inp))
return (NULL);
} else
panic("%s: locking bug", __func__);
} else
INP_HASH_RUNLOCK(pcbinfo);
return (inp);
}
/*
* Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
* from which a pre-calculated hash value may be extracted.
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
*
* Possibly more of this logic should be in in_pcbgroup.c.
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
*/
struct inpcb *
in_pcblookup(struct inpcbinfo *pcbinfo, struct in_addr faddr, u_int fport,
struct in_addr laddr, u_int lport, int lookupflags, struct ifnet *ifp)
{
Several years after initial development, merge prototype support for linking NIC Receive Side Scaling (RSS) to the network stack's connection-group implementation. This prototype (and derived patches) are in use at Juniper and several other FreeBSD-using companies, so despite some reservations about its maturity, merge the patch to the base tree so that it can be iteratively refined in collaboration rather than maintained as a set of gradually diverging patch sets. (1) Merge a software implementation of the Toeplitz hash specified in RSS implemented by David Malone. This is used to allow suitable pcbgroup placement of connections before the first packet is received from the NIC. Software hashing is generally avoided, however, due to high cost of the hash on general-purpose CPUs. (2) In in_rss.c, maintain authoritative versions of RSS state intended to be pushed to each NIC, including keying material, hash algorithm/ configuration, and buckets. Provide software-facing interfaces to hash 2- and 4-tuples for IPv4 and IPv6 using both the RSS standardised Toeplitz and a 'naive' variation with a hash efficient in software but with poor distribution properties. Implement rss_m2cpuid()to be used by netisr and other load balancing code to look up the CPU on which an mbuf should be processed. (3) In the Ethernet link layer, allow netisr distribution using RSS as a source of policy as an alternative to source ordering; continue to default to direct dispatch (i.e., don't try and requeue packets for processing on the 'right' CPU if they arrive in a directly dispatchable context). (4) Allow RSS to control tuning of connection groups in order to align groups with RSS buckets. If a packet arrives on a protocol using connection groups, and contains a suitable hardware-generated hash, use that hash value to select the connection group for pcb lookup for both IPv4 and IPv6. If no hardware-generated Toeplitz hash is available, we fall back on regular PCB lookup risking contention rather than pay the cost of Toeplitz in software -- this is a less scalable but, at my last measurement, faster approach. As core counts go up, we may want to revise this strategy despite CPU overhead. Where device drivers suitably configure NICs, and connection groups / RSS are enabled, this should avoid both lock and line contention during connection lookup for TCP. This commit does not modify any device drivers to tune device RSS configuration to the global RSS configuration; patches are in circulation to do this for at least Chelsio T3 and Intel 1G/10G drivers. Currently, the KPI for device drivers is not particularly robust, nor aware of more advanced features such as runtime reconfiguration/rebalancing. This will hopefully prove a useful starting point for refinement. No MFC is scheduled as we will first want to nail down a more mature and maintainable KPI/KBI for device drivers. Sponsored by: Juniper Networks (original work) Sponsored by: EMC/Isilon (patch update and merge)
2014-03-15 00:57:50 +00:00
#if defined(PCBGROUP) && !defined(RSS)
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
struct inpcbgroup *pcbgroup;
#endif
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
KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
("%s: invalid lookup flags %d", __func__, lookupflags));
KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
("%s: LOCKPCB not set", __func__));
Several years after initial development, merge prototype support for linking NIC Receive Side Scaling (RSS) to the network stack's connection-group implementation. This prototype (and derived patches) are in use at Juniper and several other FreeBSD-using companies, so despite some reservations about its maturity, merge the patch to the base tree so that it can be iteratively refined in collaboration rather than maintained as a set of gradually diverging patch sets. (1) Merge a software implementation of the Toeplitz hash specified in RSS implemented by David Malone. This is used to allow suitable pcbgroup placement of connections before the first packet is received from the NIC. Software hashing is generally avoided, however, due to high cost of the hash on general-purpose CPUs. (2) In in_rss.c, maintain authoritative versions of RSS state intended to be pushed to each NIC, including keying material, hash algorithm/ configuration, and buckets. Provide software-facing interfaces to hash 2- and 4-tuples for IPv4 and IPv6 using both the RSS standardised Toeplitz and a 'naive' variation with a hash efficient in software but with poor distribution properties. Implement rss_m2cpuid()to be used by netisr and other load balancing code to look up the CPU on which an mbuf should be processed. (3) In the Ethernet link layer, allow netisr distribution using RSS as a source of policy as an alternative to source ordering; continue to default to direct dispatch (i.e., don't try and requeue packets for processing on the 'right' CPU if they arrive in a directly dispatchable context). (4) Allow RSS to control tuning of connection groups in order to align groups with RSS buckets. If a packet arrives on a protocol using connection groups, and contains a suitable hardware-generated hash, use that hash value to select the connection group for pcb lookup for both IPv4 and IPv6. If no hardware-generated Toeplitz hash is available, we fall back on regular PCB lookup risking contention rather than pay the cost of Toeplitz in software -- this is a less scalable but, at my last measurement, faster approach. As core counts go up, we may want to revise this strategy despite CPU overhead. Where device drivers suitably configure NICs, and connection groups / RSS are enabled, this should avoid both lock and line contention during connection lookup for TCP. This commit does not modify any device drivers to tune device RSS configuration to the global RSS configuration; patches are in circulation to do this for at least Chelsio T3 and Intel 1G/10G drivers. Currently, the KPI for device drivers is not particularly robust, nor aware of more advanced features such as runtime reconfiguration/rebalancing. This will hopefully prove a useful starting point for refinement. No MFC is scheduled as we will first want to nail down a more mature and maintainable KPI/KBI for device drivers. Sponsored by: Juniper Networks (original work) Sponsored by: EMC/Isilon (patch update and merge)
2014-03-15 00:57:50 +00:00
/*
* When not using RSS, use connection groups in preference to the
* reservation table when looking up 4-tuples. When using RSS, just
* use the reservation table, due to the cost of the Toeplitz hash
* in software.
*
* XXXRW: This policy belongs in the pcbgroup code, as in principle
* we could be doing RSS with a non-Toeplitz hash that is affordable
* in software.
*/
#if defined(PCBGROUP) && !defined(RSS)
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
if (in_pcbgroup_enabled(pcbinfo)) {
pcbgroup = in_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
fport);
return (in_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
laddr, lport, lookupflags, ifp));
}
#endif
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
return (in_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
lookupflags, ifp));
}
struct inpcb *
in_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in_addr faddr,
u_int fport, struct in_addr laddr, u_int lport, int lookupflags,
struct ifnet *ifp, struct mbuf *m)
{
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#ifdef PCBGROUP
struct inpcbgroup *pcbgroup;
#endif
KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
("%s: invalid lookup flags %d", __func__, lookupflags));
KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
("%s: LOCKPCB not set", __func__));
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#ifdef PCBGROUP
Several years after initial development, merge prototype support for linking NIC Receive Side Scaling (RSS) to the network stack's connection-group implementation. This prototype (and derived patches) are in use at Juniper and several other FreeBSD-using companies, so despite some reservations about its maturity, merge the patch to the base tree so that it can be iteratively refined in collaboration rather than maintained as a set of gradually diverging patch sets. (1) Merge a software implementation of the Toeplitz hash specified in RSS implemented by David Malone. This is used to allow suitable pcbgroup placement of connections before the first packet is received from the NIC. Software hashing is generally avoided, however, due to high cost of the hash on general-purpose CPUs. (2) In in_rss.c, maintain authoritative versions of RSS state intended to be pushed to each NIC, including keying material, hash algorithm/ configuration, and buckets. Provide software-facing interfaces to hash 2- and 4-tuples for IPv4 and IPv6 using both the RSS standardised Toeplitz and a 'naive' variation with a hash efficient in software but with poor distribution properties. Implement rss_m2cpuid()to be used by netisr and other load balancing code to look up the CPU on which an mbuf should be processed. (3) In the Ethernet link layer, allow netisr distribution using RSS as a source of policy as an alternative to source ordering; continue to default to direct dispatch (i.e., don't try and requeue packets for processing on the 'right' CPU if they arrive in a directly dispatchable context). (4) Allow RSS to control tuning of connection groups in order to align groups with RSS buckets. If a packet arrives on a protocol using connection groups, and contains a suitable hardware-generated hash, use that hash value to select the connection group for pcb lookup for both IPv4 and IPv6. If no hardware-generated Toeplitz hash is available, we fall back on regular PCB lookup risking contention rather than pay the cost of Toeplitz in software -- this is a less scalable but, at my last measurement, faster approach. As core counts go up, we may want to revise this strategy despite CPU overhead. Where device drivers suitably configure NICs, and connection groups / RSS are enabled, this should avoid both lock and line contention during connection lookup for TCP. This commit does not modify any device drivers to tune device RSS configuration to the global RSS configuration; patches are in circulation to do this for at least Chelsio T3 and Intel 1G/10G drivers. Currently, the KPI for device drivers is not particularly robust, nor aware of more advanced features such as runtime reconfiguration/rebalancing. This will hopefully prove a useful starting point for refinement. No MFC is scheduled as we will first want to nail down a more mature and maintainable KPI/KBI for device drivers. Sponsored by: Juniper Networks (original work) Sponsored by: EMC/Isilon (patch update and merge)
2014-03-15 00:57:50 +00:00
/*
* If we can use a hardware-generated hash to look up the connection
* group, use that connection group to find the inpcb. Otherwise
* fall back on a software hash -- or the reservation table if we're
* using RSS.
*
* XXXRW: As above, that policy belongs in the pcbgroup code.
*/
if (in_pcbgroup_enabled(pcbinfo) &&
!(M_HASHTYPE_TEST(m, M_HASHTYPE_NONE))) {
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
pcbgroup = in_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
m->m_pkthdr.flowid);
if (pcbgroup != NULL)
return (in_pcblookup_group(pcbinfo, pcbgroup, faddr,
fport, laddr, lport, lookupflags, ifp));
Several years after initial development, merge prototype support for linking NIC Receive Side Scaling (RSS) to the network stack's connection-group implementation. This prototype (and derived patches) are in use at Juniper and several other FreeBSD-using companies, so despite some reservations about its maturity, merge the patch to the base tree so that it can be iteratively refined in collaboration rather than maintained as a set of gradually diverging patch sets. (1) Merge a software implementation of the Toeplitz hash specified in RSS implemented by David Malone. This is used to allow suitable pcbgroup placement of connections before the first packet is received from the NIC. Software hashing is generally avoided, however, due to high cost of the hash on general-purpose CPUs. (2) In in_rss.c, maintain authoritative versions of RSS state intended to be pushed to each NIC, including keying material, hash algorithm/ configuration, and buckets. Provide software-facing interfaces to hash 2- and 4-tuples for IPv4 and IPv6 using both the RSS standardised Toeplitz and a 'naive' variation with a hash efficient in software but with poor distribution properties. Implement rss_m2cpuid()to be used by netisr and other load balancing code to look up the CPU on which an mbuf should be processed. (3) In the Ethernet link layer, allow netisr distribution using RSS as a source of policy as an alternative to source ordering; continue to default to direct dispatch (i.e., don't try and requeue packets for processing on the 'right' CPU if they arrive in a directly dispatchable context). (4) Allow RSS to control tuning of connection groups in order to align groups with RSS buckets. If a packet arrives on a protocol using connection groups, and contains a suitable hardware-generated hash, use that hash value to select the connection group for pcb lookup for both IPv4 and IPv6. If no hardware-generated Toeplitz hash is available, we fall back on regular PCB lookup risking contention rather than pay the cost of Toeplitz in software -- this is a less scalable but, at my last measurement, faster approach. As core counts go up, we may want to revise this strategy despite CPU overhead. Where device drivers suitably configure NICs, and connection groups / RSS are enabled, this should avoid both lock and line contention during connection lookup for TCP. This commit does not modify any device drivers to tune device RSS configuration to the global RSS configuration; patches are in circulation to do this for at least Chelsio T3 and Intel 1G/10G drivers. Currently, the KPI for device drivers is not particularly robust, nor aware of more advanced features such as runtime reconfiguration/rebalancing. This will hopefully prove a useful starting point for refinement. No MFC is scheduled as we will first want to nail down a more mature and maintainable KPI/KBI for device drivers. Sponsored by: Juniper Networks (original work) Sponsored by: EMC/Isilon (patch update and merge)
2014-03-15 00:57:50 +00:00
#ifndef RSS
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
pcbgroup = in_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
fport);
return (in_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
laddr, lport, lookupflags, ifp));
Several years after initial development, merge prototype support for linking NIC Receive Side Scaling (RSS) to the network stack's connection-group implementation. This prototype (and derived patches) are in use at Juniper and several other FreeBSD-using companies, so despite some reservations about its maturity, merge the patch to the base tree so that it can be iteratively refined in collaboration rather than maintained as a set of gradually diverging patch sets. (1) Merge a software implementation of the Toeplitz hash specified in RSS implemented by David Malone. This is used to allow suitable pcbgroup placement of connections before the first packet is received from the NIC. Software hashing is generally avoided, however, due to high cost of the hash on general-purpose CPUs. (2) In in_rss.c, maintain authoritative versions of RSS state intended to be pushed to each NIC, including keying material, hash algorithm/ configuration, and buckets. Provide software-facing interfaces to hash 2- and 4-tuples for IPv4 and IPv6 using both the RSS standardised Toeplitz and a 'naive' variation with a hash efficient in software but with poor distribution properties. Implement rss_m2cpuid()to be used by netisr and other load balancing code to look up the CPU on which an mbuf should be processed. (3) In the Ethernet link layer, allow netisr distribution using RSS as a source of policy as an alternative to source ordering; continue to default to direct dispatch (i.e., don't try and requeue packets for processing on the 'right' CPU if they arrive in a directly dispatchable context). (4) Allow RSS to control tuning of connection groups in order to align groups with RSS buckets. If a packet arrives on a protocol using connection groups, and contains a suitable hardware-generated hash, use that hash value to select the connection group for pcb lookup for both IPv4 and IPv6. If no hardware-generated Toeplitz hash is available, we fall back on regular PCB lookup risking contention rather than pay the cost of Toeplitz in software -- this is a less scalable but, at my last measurement, faster approach. As core counts go up, we may want to revise this strategy despite CPU overhead. Where device drivers suitably configure NICs, and connection groups / RSS are enabled, this should avoid both lock and line contention during connection lookup for TCP. This commit does not modify any device drivers to tune device RSS configuration to the global RSS configuration; patches are in circulation to do this for at least Chelsio T3 and Intel 1G/10G drivers. Currently, the KPI for device drivers is not particularly robust, nor aware of more advanced features such as runtime reconfiguration/rebalancing. This will hopefully prove a useful starting point for refinement. No MFC is scheduled as we will first want to nail down a more mature and maintainable KPI/KBI for device drivers. Sponsored by: Juniper Networks (original work) Sponsored by: EMC/Isilon (patch update and merge)
2014-03-15 00:57:50 +00:00
#endif
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
}
#endif
return (in_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
lookupflags, ifp));
}
#endif /* INET */
/*
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
* Insert PCB onto various hash lists.
*/
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
static int
in_pcbinshash_internal(struct inpcb *inp, int do_pcbgroup_update)
{
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
struct inpcbhead *pcbhash;
struct inpcbporthead *pcbporthash;
struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
struct inpcbport *phd;
u_int32_t hashkey_faddr;
INP_WLOCK_ASSERT(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
INP_HASH_WLOCK_ASSERT(pcbinfo);
KASSERT((inp->inp_flags & INP_INHASHLIST) == 0,
("in_pcbinshash: INP_INHASHLIST"));
#ifdef INET6
if (inp->inp_vflag & INP_IPV6)
hashkey_faddr = INP6_PCBHASHKEY(&inp->in6p_faddr);
else
#endif
hashkey_faddr = inp->inp_faddr.s_addr;
pcbhash = &pcbinfo->ipi_hashbase[INP_PCBHASH(hashkey_faddr,
inp->inp_lport, inp->inp_fport, pcbinfo->ipi_hashmask)];
pcbporthash = &pcbinfo->ipi_porthashbase[
INP_PCBPORTHASH(inp->inp_lport, pcbinfo->ipi_porthashmask)];
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
/*
* Go through port list and look for a head for this lport.
*/
LIST_FOREACH(phd, pcbporthash, phd_hash) {
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
if (phd->phd_port == inp->inp_lport)
break;
}
/*
* If none exists, malloc one and tack it on.
*/
if (phd == NULL) {
phd = malloc(sizeof(struct inpcbport), M_PCB, M_NOWAIT);
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
if (phd == NULL) {
return (ENOBUFS); /* XXX */
}
phd->phd_port = inp->inp_lport;
LIST_INIT(&phd->phd_pcblist);
LIST_INSERT_HEAD(pcbporthash, phd, phd_hash);
}
inp->inp_phd = phd;
LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist);
LIST_INSERT_HEAD(pcbhash, inp, inp_hash);
inp->inp_flags |= INP_INHASHLIST;
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#ifdef PCBGROUP
if (do_pcbgroup_update)
in_pcbgroup_update(inp);
#endif
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
return (0);
}
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
/*
* For now, there are two public interfaces to insert an inpcb into the hash
* lists -- one that does update pcbgroups, and one that doesn't. The latter
* is used only in the TCP syncache, where in_pcbinshash is called before the
* full 4-tuple is set for the inpcb, and we don't want to install in the
* pcbgroup until later.
*
* XXXRW: This seems like a misfeature. in_pcbinshash should always update
* connection groups, and partially initialised inpcbs should not be exposed
* to either reservation hash tables or pcbgroups.
*/
int
in_pcbinshash(struct inpcb *inp)
{
return (in_pcbinshash_internal(inp, 1));
}
int
in_pcbinshash_nopcbgroup(struct inpcb *inp)
{
return (in_pcbinshash_internal(inp, 0));
}
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
/*
* Move PCB to the proper hash bucket when { faddr, fport } have been
* changed. NOTE: This does not handle the case of the lport changing (the
* hashed port list would have to be updated as well), so the lport must
* not change after in_pcbinshash() has been called.
*/
void
in_pcbrehash_mbuf(struct inpcb *inp, struct mbuf *m)
{
struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
struct inpcbhead *head;
u_int32_t hashkey_faddr;
INP_WLOCK_ASSERT(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
INP_HASH_WLOCK_ASSERT(pcbinfo);
KASSERT(inp->inp_flags & INP_INHASHLIST,
("in_pcbrehash: !INP_INHASHLIST"));
#ifdef INET6
if (inp->inp_vflag & INP_IPV6)
hashkey_faddr = INP6_PCBHASHKEY(&inp->in6p_faddr);
else
#endif
hashkey_faddr = inp->inp_faddr.s_addr;
head = &pcbinfo->ipi_hashbase[INP_PCBHASH(hashkey_faddr,
inp->inp_lport, inp->inp_fport, pcbinfo->ipi_hashmask)];
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
LIST_REMOVE(inp, inp_hash);
LIST_INSERT_HEAD(head, inp, inp_hash);
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#ifdef PCBGROUP
if (m != NULL)
in_pcbgroup_update_mbuf(inp, m);
else
in_pcbgroup_update(inp);
#endif
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
}
void
in_pcbrehash(struct inpcb *inp)
{
in_pcbrehash_mbuf(inp, NULL);
}
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
/*
* Remove PCB from various lists.
*/
static void
in_pcbremlists(struct inpcb *inp)
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
{
struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
#ifdef INVARIANTS
if (pcbinfo == &V_tcbinfo) {
INP_INFO_RLOCK_ASSERT(pcbinfo);
} else {
INP_INFO_WLOCK_ASSERT(pcbinfo);
}
#endif
INP_WLOCK_ASSERT(inp);
INP_LIST_WLOCK_ASSERT(pcbinfo);
inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
if (inp->inp_flags & INP_INHASHLIST) {
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
struct inpcbport *phd = inp->inp_phd;
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_HASH_WLOCK(pcbinfo);
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
LIST_REMOVE(inp, inp_hash);
LIST_REMOVE(inp, inp_portlist);
if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
LIST_REMOVE(phd, phd_hash);
free(phd, M_PCB);
}
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_HASH_WUNLOCK(pcbinfo);
inp->inp_flags &= ~INP_INHASHLIST;
Improved connection establishment performance by doing local port lookups via a hashed port list. In the new scheme, in_pcblookup() goes away and is replaced by a new routine, in_pcblookup_local() for doing the local port check. Note that this implementation is space inefficient in that the PCB struct is now too large to fit into 128 bytes. I might deal with this in the future by using the new zone allocator, but I wanted these changes to be extensively tested in their current form first. Also: 1) Fixed off-by-one errors in the port lookup loops in in_pcbbind(). 2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash() to do the initialial hash insertion. 3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability. 4) Added a new routine, in_pcbremlists() to remove the PCB from the various hash lists. 5) Added/deleted comments where appropriate. 6) Removed unnecessary splnet() locking. In general, the PCB functions should be called at splnet()...there are unfortunately a few exceptions, however. 7) Reorganized a few structs for better cache line behavior. 8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in the future, however. These changes have been tested on wcarchive for more than a month. In tests done here, connection establishment overhead is reduced by more than 50 times, thus getting rid of one of the major networking scalability problems. Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult. WARNING: Anything that knows about inpcb and tcpcb structs will have to be recompiled; at the very least, this includes netstat(1).
1998-01-27 09:15:13 +00:00
}
LIST_REMOVE(inp, inp_list);
pcbinfo->ipi_count--;
Implement a CPU-affine TCP and UDP connection lookup data structure, struct inpcbgroup. pcbgroups, or "connection groups", supplement the existing inpcbinfo connection hash table, which when pcbgroups are enabled, might now be thought of more usefully as a per-protocol 4-tuple reservation table. Connections are assigned to connection groups base on a hash of their 4-tuple; wildcard sockets require special handling, and are members of all connection groups. During a connection lookup, a per-connection group lock is employed rather than the global pcbinfo lock. By aligning connection groups with input path processing, connection groups take on an effective CPU affinity, especially when aligned with RSS work placement (see a forthcoming commit for details). This eliminates cache line migration associated with global, protocol-layer data structures in steady state TCP and UDP processing (with the exception of protocol-layer statistics; further commit to follow). Elements of this approach were inspired by Willman, Rixner, and Cox's 2006 USENIX paper, "An Evaluation of Network Stack Parallelization Strategies in Modern Operating Systems". However, there are also significant differences: we maintain the inpcb lock, rather than using the connection group lock for per-connection state. Likewise, the focus of this implementation is alignment with NIC packet distribution strategies such as RSS, rather than pure software strategies. Despite that focus, software distribution is supported through the parallel netisr implementation, and works well in configurations where the number of hardware threads is greater than the number of NIC input queues, such as in the RMI XLR threaded MIPS architecture. Another important difference is the continued maintenance of existing hash tables as "reservation tables" -- these are useful both to distinguish the resource allocation aspect of protocol name management and the more common-case lookup aspect. In configurations where connection tables are aligned with hardware hashes, it is desirable to use the traditional lookup tables for loopback or encapsulated traffic rather than take the expense of hardware hashes that are hard to implement efficiently in software (such as RSS Toeplitz). Connection group support is enabled by compiling "options PCBGROUP" into your kernel configuration; for the time being, this is an experimental feature, and hence is not enabled by default. Subject to the limited MFCability of change dependencies in inpcb, and its change to the inpcbinfo init function signature, this change in principle could be merged to FreeBSD 8.x. Reviewed by: bz Sponsored by: Juniper Networks, Inc.
2011-06-06 12:55:02 +00:00
#ifdef PCBGROUP
in_pcbgroup_remove(inp);
#endif
}
This Implements the mumbled about "Jail" feature. This is a seriously beefed up chroot kind of thing. The process is jailed along the same lines as a chroot does it, but with additional tough restrictions imposed on what the superuser can do. For all I know, it is safe to hand over the root bit inside a prison to the customer living in that prison, this is what it was developed for in fact: "real virtual servers". Each prison has an ip number associated with it, which all IP communications will be coerced to use and each prison has its own hostname. Needless to say, you need more RAM this way, but the advantage is that each customer can run their own particular version of apache and not stomp on the toes of their neighbors. It generally does what one would expect, but setting up a jail still takes a little knowledge. A few notes: I have no scripts for setting up a jail, don't ask me for them. The IP number should be an alias on one of the interfaces. mount a /proc in each jail, it will make ps more useable. /proc/<pid>/status tells the hostname of the prison for jailed processes. Quotas are only sensible if you have a mountpoint per prison. There are no privisions for stopping resource-hogging. Some "#ifdef INET" and similar may be missing (send patches!) If somebody wants to take it from here and develop it into more of a "virtual machine" they should be most welcome! Tools, comments, patches & documentation most welcome. Have fun... Sponsored by: http://www.rndassociates.com/ Run for almost a year by: http://www.servetheweb.com/
1999-04-28 11:38:52 +00:00
/*
* Check for alternatives when higher level complains
* about service problems. For now, invalidate cached
* routing information. If the route was created dynamically
* (by a redirect), time to try a default gateway again.
*/
void
in_losing(struct inpcb *inp)
{
RO_RTFREE(&inp->inp_route);
if (inp->inp_route.ro_lle)
LLE_FREE(inp->inp_route.ro_lle); /* zeros ro_lle */
return;
}
/*
* A set label operation has occurred at the socket layer, propagate the
* label change into the in_pcb for the socket.
*/
void
in_pcbsosetlabel(struct socket *so)
{
#ifdef MAC
struct inpcb *inp;
inp = sotoinpcb(so);
KASSERT(inp != NULL, ("in_pcbsosetlabel: so->so_pcb == NULL"));
INP_WLOCK(inp);
SOCK_LOCK(so);
mac_inpcb_sosetlabel(so, inp);
SOCK_UNLOCK(so);
INP_WUNLOCK(inp);
#endif
}
/*
* ipport_tick runs once per second, determining if random port allocation
* should be continued. If more than ipport_randomcps ports have been
* allocated in the last second, then we return to sequential port
* allocation. We return to random allocation only once we drop below
* ipport_randomcps for at least ipport_randomtime seconds.
*/
static void
ipport_tick(void *xtp)
{
VNET_ITERATOR_DECL(vnet_iter);
VNET_LIST_RLOCK_NOSLEEP();
VNET_FOREACH(vnet_iter) {
CURVNET_SET(vnet_iter); /* XXX appease INVARIANTS here */
if (V_ipport_tcpallocs <=
V_ipport_tcplastcount + V_ipport_randomcps) {
if (V_ipport_stoprandom > 0)
V_ipport_stoprandom--;
} else
V_ipport_stoprandom = V_ipport_randomtime;
V_ipport_tcplastcount = V_ipport_tcpallocs;
CURVNET_RESTORE();
}
VNET_LIST_RUNLOCK_NOSLEEP();
callout_reset(&ipport_tick_callout, hz, ipport_tick, NULL);
}
static void
ip_fini(void *xtp)
{
callout_stop(&ipport_tick_callout);
}
/*
* The ipport_callout should start running at about the time we attach the
* inet or inet6 domains.
*/
static void
ipport_tick_init(const void *unused __unused)
{
/* Start ipport_tick. */
callout_init(&ipport_tick_callout, 1);
callout_reset(&ipport_tick_callout, 1, ipport_tick, NULL);
EVENTHANDLER_REGISTER(shutdown_pre_sync, ip_fini, NULL,
SHUTDOWN_PRI_DEFAULT);
}
SYSINIT(ipport_tick_init, SI_SUB_PROTO_DOMAIN, SI_ORDER_MIDDLE,
ipport_tick_init, NULL);
void
inp_wlock(struct inpcb *inp)
{
INP_WLOCK(inp);
}
void
inp_wunlock(struct inpcb *inp)
{
INP_WUNLOCK(inp);
}
void
inp_rlock(struct inpcb *inp)
{
INP_RLOCK(inp);
}
void
inp_runlock(struct inpcb *inp)
{
INP_RUNLOCK(inp);
}
#ifdef INVARIANT_SUPPORT
void
inp_lock_assert(struct inpcb *inp)
{
INP_WLOCK_ASSERT(inp);
}
void
inp_unlock_assert(struct inpcb *inp)
{
INP_UNLOCK_ASSERT(inp);
}
#endif
void
inp_apply_all(void (*func)(struct inpcb *, void *), void *arg)
{
struct inpcb *inp;
INP_INFO_WLOCK(&V_tcbinfo);
LIST_FOREACH(inp, V_tcbinfo.ipi_listhead, inp_list) {
INP_WLOCK(inp);
func(inp, arg);
INP_WUNLOCK(inp);
}
INP_INFO_WUNLOCK(&V_tcbinfo);
}
struct socket *
inp_inpcbtosocket(struct inpcb *inp)
{
INP_WLOCK_ASSERT(inp);
return (inp->inp_socket);
}
struct tcpcb *
inp_inpcbtotcpcb(struct inpcb *inp)
{
INP_WLOCK_ASSERT(inp);
return ((struct tcpcb *)inp->inp_ppcb);
}
int
inp_ip_tos_get(const struct inpcb *inp)
{
return (inp->inp_ip_tos);
}
void
inp_ip_tos_set(struct inpcb *inp, int val)
{
inp->inp_ip_tos = val;
}
void
inp_4tuple_get(struct inpcb *inp, uint32_t *laddr, uint16_t *lp,
uint32_t *faddr, uint16_t *fp)
{
INP_LOCK_ASSERT(inp);
*laddr = inp->inp_laddr.s_addr;
*faddr = inp->inp_faddr.s_addr;
*lp = inp->inp_lport;
*fp = inp->inp_fport;
}
struct inpcb *
so_sotoinpcb(struct socket *so)
{
return (sotoinpcb(so));
}
struct tcpcb *
so_sototcpcb(struct socket *so)
{
return (sototcpcb(so));
}
/*
* Create an external-format (``xinpcb'') structure using the information in
* the kernel-format in_pcb structure pointed to by inp. This is done to
* reduce the spew of irrelevant information over this interface, to isolate
* user code from changes in the kernel structure, and potentially to provide
* information-hiding if we decide that some of this information should be
* hidden from users.
*/
void
in_pcbtoxinpcb(const struct inpcb *inp, struct xinpcb *xi)
{
xi->xi_len = sizeof(struct xinpcb);
if (inp->inp_socket)
sotoxsocket(inp->inp_socket, &xi->xi_socket);
else
bzero(&xi->xi_socket, sizeof(struct xsocket));
bcopy(&inp->inp_inc, &xi->inp_inc, sizeof(struct in_conninfo));
xi->inp_gencnt = inp->inp_gencnt;
xi->inp_ppcb = inp->inp_ppcb;
xi->inp_flow = inp->inp_flow;
xi->inp_flowid = inp->inp_flowid;
xi->inp_flowtype = inp->inp_flowtype;
xi->inp_flags = inp->inp_flags;
xi->inp_flags2 = inp->inp_flags2;
xi->inp_rss_listen_bucket = inp->inp_rss_listen_bucket;
xi->in6p_cksum = inp->in6p_cksum;
xi->in6p_hops = inp->in6p_hops;
xi->inp_ip_tos = inp->inp_ip_tos;
xi->inp_vflag = inp->inp_vflag;
xi->inp_ip_ttl = inp->inp_ip_ttl;
xi->inp_ip_p = inp->inp_ip_p;
xi->inp_ip_minttl = inp->inp_ip_minttl;
}
#ifdef DDB
static void
db_print_indent(int indent)
{
int i;
for (i = 0; i < indent; i++)
db_printf(" ");
}
static void
db_print_inconninfo(struct in_conninfo *inc, const char *name, int indent)
{
char faddr_str[48], laddr_str[48];
db_print_indent(indent);
db_printf("%s at %p\n", name, inc);
indent += 2;
#ifdef INET6
if (inc->inc_flags & INC_ISIPV6) {
/* IPv6. */
ip6_sprintf(laddr_str, &inc->inc6_laddr);
ip6_sprintf(faddr_str, &inc->inc6_faddr);
} else
#endif
{
/* IPv4. */
inet_ntoa_r(inc->inc_laddr, laddr_str);
inet_ntoa_r(inc->inc_faddr, faddr_str);
}
db_print_indent(indent);
db_printf("inc_laddr %s inc_lport %u\n", laddr_str,
ntohs(inc->inc_lport));
db_print_indent(indent);
db_printf("inc_faddr %s inc_fport %u\n", faddr_str,
ntohs(inc->inc_fport));
}
static void
db_print_inpflags(int inp_flags)
{
int comma;
comma = 0;
if (inp_flags & INP_RECVOPTS) {
db_printf("%sINP_RECVOPTS", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_RECVRETOPTS) {
db_printf("%sINP_RECVRETOPTS", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_RECVDSTADDR) {
db_printf("%sINP_RECVDSTADDR", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_ORIGDSTADDR) {
db_printf("%sINP_ORIGDSTADDR", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_HDRINCL) {
db_printf("%sINP_HDRINCL", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_HIGHPORT) {
db_printf("%sINP_HIGHPORT", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_LOWPORT) {
db_printf("%sINP_LOWPORT", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_ANONPORT) {
db_printf("%sINP_ANONPORT", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_RECVIF) {
db_printf("%sINP_RECVIF", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_MTUDISC) {
db_printf("%sINP_MTUDISC", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_RECVTTL) {
db_printf("%sINP_RECVTTL", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_DONTFRAG) {
db_printf("%sINP_DONTFRAG", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_RECVTOS) {
db_printf("%sINP_RECVTOS", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_IPV6_V6ONLY) {
db_printf("%sIN6P_IPV6_V6ONLY", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_PKTINFO) {
db_printf("%sIN6P_PKTINFO", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_HOPLIMIT) {
db_printf("%sIN6P_HOPLIMIT", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_HOPOPTS) {
db_printf("%sIN6P_HOPOPTS", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_DSTOPTS) {
db_printf("%sIN6P_DSTOPTS", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_RTHDR) {
db_printf("%sIN6P_RTHDR", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_RTHDRDSTOPTS) {
db_printf("%sIN6P_RTHDRDSTOPTS", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_TCLASS) {
db_printf("%sIN6P_TCLASS", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_AUTOFLOWLABEL) {
db_printf("%sIN6P_AUTOFLOWLABEL", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_TIMEWAIT) {
db_printf("%sINP_TIMEWAIT", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_ONESBCAST) {
db_printf("%sINP_ONESBCAST", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_DROPPED) {
db_printf("%sINP_DROPPED", comma ? ", " : "");
comma = 1;
}
if (inp_flags & INP_SOCKREF) {
db_printf("%sINP_SOCKREF", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_RFC2292) {
db_printf("%sIN6P_RFC2292", comma ? ", " : "");
comma = 1;
}
if (inp_flags & IN6P_MTU) {
db_printf("IN6P_MTU%s", comma ? ", " : "");
comma = 1;
}
}
static void
db_print_inpvflag(u_char inp_vflag)
{
int comma;
comma = 0;
if (inp_vflag & INP_IPV4) {
db_printf("%sINP_IPV4", comma ? ", " : "");
comma = 1;
}
if (inp_vflag & INP_IPV6) {
db_printf("%sINP_IPV6", comma ? ", " : "");
comma = 1;
}
if (inp_vflag & INP_IPV6PROTO) {
db_printf("%sINP_IPV6PROTO", comma ? ", " : "");
comma = 1;
}
}
static void
db_print_inpcb(struct inpcb *inp, const char *name, int indent)
{
db_print_indent(indent);
db_printf("%s at %p\n", name, inp);
indent += 2;
db_print_indent(indent);
db_printf("inp_flow: 0x%x\n", inp->inp_flow);
db_print_inconninfo(&inp->inp_inc, "inp_conninfo", indent);
db_print_indent(indent);
db_printf("inp_ppcb: %p inp_pcbinfo: %p inp_socket: %p\n",
inp->inp_ppcb, inp->inp_pcbinfo, inp->inp_socket);
db_print_indent(indent);
db_printf("inp_label: %p inp_flags: 0x%x (",
inp->inp_label, inp->inp_flags);
db_print_inpflags(inp->inp_flags);
db_printf(")\n");
db_print_indent(indent);
db_printf("inp_sp: %p inp_vflag: 0x%x (", inp->inp_sp,
inp->inp_vflag);
db_print_inpvflag(inp->inp_vflag);
db_printf(")\n");
db_print_indent(indent);
db_printf("inp_ip_ttl: %d inp_ip_p: %d inp_ip_minttl: %d\n",
inp->inp_ip_ttl, inp->inp_ip_p, inp->inp_ip_minttl);
db_print_indent(indent);
#ifdef INET6
if (inp->inp_vflag & INP_IPV6) {
db_printf("in6p_options: %p in6p_outputopts: %p "
"in6p_moptions: %p\n", inp->in6p_options,
inp->in6p_outputopts, inp->in6p_moptions);
db_printf("in6p_icmp6filt: %p in6p_cksum %d "
"in6p_hops %u\n", inp->in6p_icmp6filt, inp->in6p_cksum,
inp->in6p_hops);
} else
#endif
{
db_printf("inp_ip_tos: %d inp_ip_options: %p "
"inp_ip_moptions: %p\n", inp->inp_ip_tos,
inp->inp_options, inp->inp_moptions);
}
db_print_indent(indent);
db_printf("inp_phd: %p inp_gencnt: %ju\n", inp->inp_phd,
(uintmax_t)inp->inp_gencnt);
}
DB_SHOW_COMMAND(inpcb, db_show_inpcb)
{
struct inpcb *inp;
if (!have_addr) {
db_printf("usage: show inpcb <addr>\n");
return;
}
inp = (struct inpcb *)addr;
db_print_inpcb(inp, "inpcb", 0);
}
#endif /* DDB */
Implement kernel support for hardware rate limited sockets. - Add RATELIMIT kernel configuration keyword which must be set to enable the new functionality. - Add support for hardware driven, Receive Side Scaling, RSS aware, rate limited sendqueues and expose the functionality through the already established SO_MAX_PACING_RATE setsockopt(). The API support rates in the range from 1 to 4Gbytes/s which are suitable for regular TCP and UDP streams. The setsockopt(2) manual page has been updated. - Add rate limit function callback API to "struct ifnet" which supports the following operations: if_snd_tag_alloc(), if_snd_tag_modify(), if_snd_tag_query() and if_snd_tag_free(). - Add support to ifconfig to view, set and clear the IFCAP_TXRTLMT flag, which tells if a network driver supports rate limiting or not. - This patch also adds support for rate limiting through VLAN and LAGG intermediate network devices. - How rate limiting works: 1) The userspace application calls setsockopt() after accepting or making a new connection to set the rate which is then stored in the socket structure in the kernel. Later on when packets are transmitted a check is made in the transmit path for rate changes. A rate change implies a non-blocking ifp->if_snd_tag_alloc() call will be made to the destination network interface, which then sets up a custom sendqueue with the given rate limitation parameter. A "struct m_snd_tag" pointer is returned which serves as a "snd_tag" hint in the m_pkthdr for the subsequently transmitted mbufs. 2) When the network driver sees the "m->m_pkthdr.snd_tag" different from NULL, it will move the packets into a designated rate limited sendqueue given by the snd_tag pointer. It is up to the individual drivers how the rate limited traffic will be rate limited. 3) Route changes are detected by the NIC drivers in the ifp->if_transmit() routine when the ifnet pointer in the incoming snd_tag mismatches the one of the network interface. The network adapter frees the mbuf and returns EAGAIN which causes the ip_output() to release and clear the send tag. Upon next ip_output() a new "snd_tag" will be tried allocated. 4) When the PCB is detached the custom sendqueue will be released by a non-blocking ifp->if_snd_tag_free() call to the currently bound network interface. Reviewed by: wblock (manpages), adrian, gallatin, scottl (network) Differential Revision: https://reviews.freebsd.org/D3687 Sponsored by: Mellanox Technologies MFC after: 3 months
2017-01-18 13:31:17 +00:00
#ifdef RATELIMIT
/*
* Modify TX rate limit based on the existing "inp->inp_snd_tag",
* if any.
*/
int
in_pcbmodify_txrtlmt(struct inpcb *inp, uint32_t max_pacing_rate)
{
union if_snd_tag_modify_params params = {
.rate_limit.max_rate = max_pacing_rate,
};
struct m_snd_tag *mst;
struct ifnet *ifp;
int error;
mst = inp->inp_snd_tag;
if (mst == NULL)
return (EINVAL);
ifp = mst->ifp;
if (ifp == NULL)
return (EINVAL);
if (ifp->if_snd_tag_modify == NULL) {
error = EOPNOTSUPP;
} else {
error = ifp->if_snd_tag_modify(mst, &params);
}
return (error);
}
/*
* Query existing TX rate limit based on the existing
* "inp->inp_snd_tag", if any.
*/
int
in_pcbquery_txrtlmt(struct inpcb *inp, uint32_t *p_max_pacing_rate)
{
union if_snd_tag_query_params params = { };
struct m_snd_tag *mst;
struct ifnet *ifp;
int error;
mst = inp->inp_snd_tag;
if (mst == NULL)
return (EINVAL);
ifp = mst->ifp;
if (ifp == NULL)
return (EINVAL);
if (ifp->if_snd_tag_query == NULL) {
error = EOPNOTSUPP;
} else {
error = ifp->if_snd_tag_query(mst, &params);
if (error == 0 && p_max_pacing_rate != NULL)
*p_max_pacing_rate = params.rate_limit.max_rate;
}
return (error);
}
/*
* Query existing TX queue level based on the existing
* "inp->inp_snd_tag", if any.
*/
int
in_pcbquery_txrlevel(struct inpcb *inp, uint32_t *p_txqueue_level)
{
union if_snd_tag_query_params params = { };
struct m_snd_tag *mst;
struct ifnet *ifp;
int error;
mst = inp->inp_snd_tag;
if (mst == NULL)
return (EINVAL);
ifp = mst->ifp;
if (ifp == NULL)
return (EINVAL);
if (ifp->if_snd_tag_query == NULL)
return (EOPNOTSUPP);
error = ifp->if_snd_tag_query(mst, &params);
if (error == 0 && p_txqueue_level != NULL)
*p_txqueue_level = params.rate_limit.queue_level;
return (error);
}
Implement kernel support for hardware rate limited sockets. - Add RATELIMIT kernel configuration keyword which must be set to enable the new functionality. - Add support for hardware driven, Receive Side Scaling, RSS aware, rate limited sendqueues and expose the functionality through the already established SO_MAX_PACING_RATE setsockopt(). The API support rates in the range from 1 to 4Gbytes/s which are suitable for regular TCP and UDP streams. The setsockopt(2) manual page has been updated. - Add rate limit function callback API to "struct ifnet" which supports the following operations: if_snd_tag_alloc(), if_snd_tag_modify(), if_snd_tag_query() and if_snd_tag_free(). - Add support to ifconfig to view, set and clear the IFCAP_TXRTLMT flag, which tells if a network driver supports rate limiting or not. - This patch also adds support for rate limiting through VLAN and LAGG intermediate network devices. - How rate limiting works: 1) The userspace application calls setsockopt() after accepting or making a new connection to set the rate which is then stored in the socket structure in the kernel. Later on when packets are transmitted a check is made in the transmit path for rate changes. A rate change implies a non-blocking ifp->if_snd_tag_alloc() call will be made to the destination network interface, which then sets up a custom sendqueue with the given rate limitation parameter. A "struct m_snd_tag" pointer is returned which serves as a "snd_tag" hint in the m_pkthdr for the subsequently transmitted mbufs. 2) When the network driver sees the "m->m_pkthdr.snd_tag" different from NULL, it will move the packets into a designated rate limited sendqueue given by the snd_tag pointer. It is up to the individual drivers how the rate limited traffic will be rate limited. 3) Route changes are detected by the NIC drivers in the ifp->if_transmit() routine when the ifnet pointer in the incoming snd_tag mismatches the one of the network interface. The network adapter frees the mbuf and returns EAGAIN which causes the ip_output() to release and clear the send tag. Upon next ip_output() a new "snd_tag" will be tried allocated. 4) When the PCB is detached the custom sendqueue will be released by a non-blocking ifp->if_snd_tag_free() call to the currently bound network interface. Reviewed by: wblock (manpages), adrian, gallatin, scottl (network) Differential Revision: https://reviews.freebsd.org/D3687 Sponsored by: Mellanox Technologies MFC after: 3 months
2017-01-18 13:31:17 +00:00
/*
* Allocate a new TX rate limit send tag from the network interface
* given by the "ifp" argument and save it in "inp->inp_snd_tag":
*/
int
in_pcbattach_txrtlmt(struct inpcb *inp, struct ifnet *ifp,
uint32_t flowtype, uint32_t flowid, uint32_t max_pacing_rate)
{
union if_snd_tag_alloc_params params = {
.rate_limit.hdr.type = (max_pacing_rate == -1U) ?
IF_SND_TAG_TYPE_UNLIMITED : IF_SND_TAG_TYPE_RATE_LIMIT,
Implement kernel support for hardware rate limited sockets. - Add RATELIMIT kernel configuration keyword which must be set to enable the new functionality. - Add support for hardware driven, Receive Side Scaling, RSS aware, rate limited sendqueues and expose the functionality through the already established SO_MAX_PACING_RATE setsockopt(). The API support rates in the range from 1 to 4Gbytes/s which are suitable for regular TCP and UDP streams. The setsockopt(2) manual page has been updated. - Add rate limit function callback API to "struct ifnet" which supports the following operations: if_snd_tag_alloc(), if_snd_tag_modify(), if_snd_tag_query() and if_snd_tag_free(). - Add support to ifconfig to view, set and clear the IFCAP_TXRTLMT flag, which tells if a network driver supports rate limiting or not. - This patch also adds support for rate limiting through VLAN and LAGG intermediate network devices. - How rate limiting works: 1) The userspace application calls setsockopt() after accepting or making a new connection to set the rate which is then stored in the socket structure in the kernel. Later on when packets are transmitted a check is made in the transmit path for rate changes. A rate change implies a non-blocking ifp->if_snd_tag_alloc() call will be made to the destination network interface, which then sets up a custom sendqueue with the given rate limitation parameter. A "struct m_snd_tag" pointer is returned which serves as a "snd_tag" hint in the m_pkthdr for the subsequently transmitted mbufs. 2) When the network driver sees the "m->m_pkthdr.snd_tag" different from NULL, it will move the packets into a designated rate limited sendqueue given by the snd_tag pointer. It is up to the individual drivers how the rate limited traffic will be rate limited. 3) Route changes are detected by the NIC drivers in the ifp->if_transmit() routine when the ifnet pointer in the incoming snd_tag mismatches the one of the network interface. The network adapter frees the mbuf and returns EAGAIN which causes the ip_output() to release and clear the send tag. Upon next ip_output() a new "snd_tag" will be tried allocated. 4) When the PCB is detached the custom sendqueue will be released by a non-blocking ifp->if_snd_tag_free() call to the currently bound network interface. Reviewed by: wblock (manpages), adrian, gallatin, scottl (network) Differential Revision: https://reviews.freebsd.org/D3687 Sponsored by: Mellanox Technologies MFC after: 3 months
2017-01-18 13:31:17 +00:00
.rate_limit.hdr.flowid = flowid,
.rate_limit.hdr.flowtype = flowtype,
.rate_limit.max_rate = max_pacing_rate,
};
int error;
INP_WLOCK_ASSERT(inp);
if (inp->inp_snd_tag != NULL)
return (EINVAL);
if (ifp->if_snd_tag_alloc == NULL) {
error = EOPNOTSUPP;
} else {
error = ifp->if_snd_tag_alloc(ifp, &params, &inp->inp_snd_tag);
/*
* At success increment the refcount on
* the send tag's network interface:
*/
if (error == 0)
if_ref(inp->inp_snd_tag->ifp);
}
return (error);
}
/*
* Free an existing TX rate limit tag based on the "inp->inp_snd_tag",
* if any:
*/
void
in_pcbdetach_txrtlmt(struct inpcb *inp)
{
struct m_snd_tag *mst;
struct ifnet *ifp;
INP_WLOCK_ASSERT(inp);
mst = inp->inp_snd_tag;
inp->inp_snd_tag = NULL;
if (mst == NULL)
return;
ifp = mst->ifp;
if (ifp == NULL)
return;
/*
* If the device was detached while we still had reference(s)
* on the ifp, we assume if_snd_tag_free() was replaced with
* stubs.
*/
ifp->if_snd_tag_free(mst);
/* release reference count on network interface */
if_rele(ifp);
}
/*
* This function should be called when the INP_RATE_LIMIT_CHANGED flag
* is set in the fast path and will attach/detach/modify the TX rate
* limit send tag based on the socket's so_max_pacing_rate value.
*/
void
in_pcboutput_txrtlmt(struct inpcb *inp, struct ifnet *ifp, struct mbuf *mb)
{
struct socket *socket;
uint32_t max_pacing_rate;
bool did_upgrade;
int error;
if (inp == NULL)
return;
socket = inp->inp_socket;
if (socket == NULL)
return;
if (!INP_WLOCKED(inp)) {
/*
* NOTE: If the write locking fails, we need to bail
* out and use the non-ratelimited ring for the
* transmit until there is a new chance to get the
* write lock.
*/
if (!INP_TRY_UPGRADE(inp))
return;
did_upgrade = 1;
} else {
did_upgrade = 0;
}
/*
* NOTE: The so_max_pacing_rate value is read unlocked,
* because atomic updates are not required since the variable
* is checked at every mbuf we send. It is assumed that the
* variable read itself will be atomic.
*/
max_pacing_rate = socket->so_max_pacing_rate;
/*
* NOTE: When attaching to a network interface a reference is
* made to ensure the network interface doesn't go away until
* all ratelimit connections are gone. The network interface
* pointers compared below represent valid network interfaces,
* except when comparing towards NULL.
*/
if (max_pacing_rate == 0 && inp->inp_snd_tag == NULL) {
error = 0;
} else if (!(ifp->if_capenable & IFCAP_TXRTLMT)) {
if (inp->inp_snd_tag != NULL)
in_pcbdetach_txrtlmt(inp);
error = 0;
} else if (inp->inp_snd_tag == NULL) {
/*
* In order to utilize packet pacing with RSS, we need
* to wait until there is a valid RSS hash before we
* can proceed:
*/
if (M_HASHTYPE_GET(mb) == M_HASHTYPE_NONE) {
error = EAGAIN;
} else {
error = in_pcbattach_txrtlmt(inp, ifp, M_HASHTYPE_GET(mb),
mb->m_pkthdr.flowid, max_pacing_rate);
}
} else {
error = in_pcbmodify_txrtlmt(inp, max_pacing_rate);
}
if (error == 0 || error == EOPNOTSUPP)
inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
if (did_upgrade)
INP_DOWNGRADE(inp);
}
/*
* Track route changes for TX rate limiting.
*/
void
in_pcboutput_eagain(struct inpcb *inp)
{
struct socket *socket;
bool did_upgrade;
if (inp == NULL)
return;
socket = inp->inp_socket;
if (socket == NULL)
return;
if (inp->inp_snd_tag == NULL)
return;
if (!INP_WLOCKED(inp)) {
/*
* NOTE: If the write locking fails, we need to bail
* out and use the non-ratelimited ring for the
* transmit until there is a new chance to get the
* write lock.
*/
if (!INP_TRY_UPGRADE(inp))
return;
did_upgrade = 1;
} else {
did_upgrade = 0;
}
/* detach rate limiting */
in_pcbdetach_txrtlmt(inp);
/* make sure new mbuf send tag allocation is made */
inp->inp_flags2 |= INP_RATE_LIMIT_CHANGED;
if (did_upgrade)
INP_DOWNGRADE(inp);
}
#endif /* RATELIMIT */