completes the move to a fully parallel UDP transmit path by using
global read, rather than write, locking of inpcbinfo in further
semi-connected cases:
- Add macros to allow try-locking of inpcb and inpcbinfo.
- Always acquire an incpcb read lock in udp_output(), which stablizes the
local inpcb address and port bindings in order to determine what further
locking is required:
- If the inpcb is currently not bound (at all) and are implicitly
connecting, we require inpcbinfo and inpcb write locks, so drop the
read lock and re-acquire.
- If the inpcb is bound for at least one of the port or address, but an
explicit source or destination is requested, trylock the inpcbinfo
lock, and if that fails, drop the inpcb lock, lock the global lock,
and relock the inpcb lock.
- Otherwise, no further locking is required (common case).
- Update comments.
In practice, this means that the vast majority of consumers of UDP sockets
will not acquire any exclusive locks at the socket or UDP levels of the
network stack. This leads to a marked performance improvement in several
important workloads, including BIND, nsd, and memcached over UDP, as well
as significant improvements in pps microbenchmarks.
The plan is to MFC all of the rwlock changes to RELENG_7 once they have
settled for a weeks in the tree.
Tested by: ps, kris (older revision), bde
MFC after: 3 weeks
udp_output() so that argument validation occurs before jail processing.
Add additional comments explaining what's going on when we process
addresses and binding during udp_output().
MFC after: 3 weeks
2) Adds some __UserSpace__ on some of the common defines that
the user space code needs
3) Fixes a bug when we send up data to a user that failed. We
need to a) trim off the data chunk headers, if present, and
b) make sure the frag bit is communicated properly for the
msgs coming off the stream queues... i.e. we see if some
of the msg has been taken.
Obtained from: jeli contributed the VIMAGE changes on this pass Thanks Julain!
inpcb. When directly invoking udp_notify() from udp_ctlinput(), acquire
only a read lock; we may still see write locks in udp_notify() as the
in_pcbnotifyall() routine is shared with TCP and always uses a write lock
on the inpcb being notified.
MFC after: 1 month
some code paths, global or inpcb write locks are required, but for other
code paths, read locks or no locking at all are sufficient for the data
structures.
MFC after: 1 month
source or a specific destination address is requested as part of a send
on a UDP socket, read lock the inpcb rather than write lock it. This
will allow fully parallel transmit down to the IP layer when sending
simultaneously from multiple threads on a connected UDP socket.
Parallel transmit for more complex cases, such as when sendto(2) is
invoked with an address and there's already a local binding, will
follow.
MFC after: 1 month
possible to exhaust and garble stack with a packet that contains a couple
of hundreds nested encapsulation levels.
Submitted by: Ming Fu <fming@borderware.com>
Reviewed by: rwatson
PR: kern/85320
dispatched without Giant, and add NETISR_FORCEQUEUE, which allows specific
netisr handlers to always be dispatched via a queue (deferred). Mark the
usb and if_ppp netisr handlers as NETISR_FORCEQUEUE, and explicitly
acquire Giant in those handlers.
Previously, any netisr handler not marked NETISR_MPSAFE would necessarily
run deferred and with Giant acquired. This change removes Giant
scaffolding from the netisr infrastructure, but NETISR_FORCEQUEUE allows
non-MPSAFE handlers to continue to force deferred dispatch so as to avoid
lock order reversals between their acqusition of Giant and any calling
context.
It is likely we will be able to remove NETISR_FORCEQUEUE once
IFF_NEEDSGIANT is removed, as non-MPSAFE usb and if_ppp drivers will no
longer be supported.
Reviewed by: bz
MFC after: 1 month
X-MFC note: We can't remove NETISR_MPSAFE from stable/7 for KPI reasons,
but the rest can go back.
generating an RTM_MISS for every IP packet forwarded making user space
routing daemons unhappy.
PR: kern/123621, kern/124540, kern/122338
Reported by: Paul <paul gtcomm.net>, Mike Tancsa <mike sentex.net> on net@
Tested by: Paul and Mike
Reviewed by: andre
MFC after: 3 days
datagram-only protocols, such as UDP. This version removes use of
sblock(), which is not required due to an inability to interlace data
improperly with datagrams, as well as avoiding some of the larger loops
and state management that don't apply on datagram sockets.
This is experimental code, so hook it up only for UDPv4 for testing; if
there are problems we may need to revise it or turn it off by default,
but it offers *significant* performance improvements for threaded UDP
applications such as BIND9, nsd, and memcached using UDP.
Tested by: kris, ps
rather than write locking: while we need to maintain a valid reference
to the inpcb and fix its state, no protocol layer state is modified
during an IPv4 UDP receive -- there are only changes at the socket
layer, which is separately protected by socket locking.
While parallel concurrent receive on a single UDP socket is currently
relatively unusual, introducing read locking in the transmit path,
allowing concurrent receive and transmit, will significantly improve
performance for loads such as BIND, memcached, etc.
MFC after: 2 months
Tested by: gnn, kris, ps
in_ifaddrhashtbl in in_ifinit because error handler in in_control removes
entries only for AF_INET addresses. If in_ifinit is called for the cloned
inteface that has just been created its address family is not AF_INET and
therefor LIST_REMOVE is not called for respective LIST_INSERT_HEAD and
freed entries remain in in_ifaddrhashtbl and lead to memory corruption.
PR: kern/124384
This is needed for correct behavior when packets are lost or reordered.
PR: kern/123950
Reviewed by: andre@, silby@
Reported by: Yahoo!, Wang Jin
MFC after: 1 week
- only one functino to destroy an SCTP stack sctp_finish()
- Make it so this function also arranges for any threads
created by the image to do a kthread_exit()
- Vimage prep - these are major restructures to move
all global variables to be accessed via a macro or two.
The variables all go into a single structure.
- Asconf address addition tweaks (add_or_del Interfaces)
- Fix rwnd calcualtion to be more conservative.
- Support SACK_IMMEDIATE flag to skip delayed sack
by demand of peer.
- Comment updates in the sack mapping calculations
- Invarients panic added.
- Pre-support for UDP tunneling (we can do this on
MAC but will need added support from UDP to
get a "pipe" of UDP packets in.
- clear trace buffer sysctl added when local tracing on.
Note the majority of this huge patch is all the vimage prep stuff :-)
template, use an M_TEMP malloc(9) allocation rather than an mbuf
with mtod(9) and dtom(9). This eliminates the last use of
dtom(9) in TCP.
MFC after: 3 weeks
to reduce performance degradation under heavy outgoing scan/flood.
Scalability is now much more important then several kilobytes of RAM.
Remove unneded TCP-specific expiration handeling. Before this connected
TCP sessions could never expire. Now connected TCP sessions will expire
after 24hours of inactivity.
Simplify HouseKeeping() to avoid several mul/div-s per packet. Taking into
account increased LINK_TABLE_OUT_SIZE, precision is still much more then
required.
- to increase performance do not reallocate mbuf when possible,
- to support up to 16K packets (was 2K max) use mbuf cluster of proper size.
This change depends on recent ng_nat and ip_fw_nat changes.
This fixes packet fragmentation handeling.
Pass really available buffer size to libalias instead of MCLBYTES constant.
MCLBYTES constant were used with believe that m_megapullup() always moves
date into a fresh cluster that sometimes may become not so.
some cases, add explicit inpcb locking rather than relying on the global
lock, as we dereference inp_socket, but also allowing us to drop the
global lock more quickly.
MFC after: 1 week
monitoring UDP connections using sysctls. In some cases, add
previously missing locking of inpcbs, as inp_socket is followed,
which also allows us to drop global locks more quickly.
MFC after: 1 week
ip6_savecontrol in preparation for udp_append() to no longer
need an WLOCK as we will no longer be modifying socket options.
Requested by: rwatson
Reviewed by: gnn
MFC after: 10 days
routines for those modules, rather than in the raw socket code. This
each privilege check to occur in exactly once place and avoids
duplicate checks across layers.
MFC after: 3 weeks
Sponsored by: nCircle Network Security, Inc.
Change so that we save off a type field for display and
NULL inp just for good measure.
- sctp_output.c - Fix it so in sending to the loopback we use the
src address of the inbound INIT. We don't want
to do this for non local addresses since otherwise
we might be ingressed filtered so we need to use
the best src address and list the address sent to.
Obtained from: time bug - Neil Wilson
MFC after: 1 week
- Adds some prepwork (Not all yet) for vimage in particular
support the delete the sctppcbinfo.xx structs. There is
still a leak in here if it were to be called plus we stil
need the regrouping (From Me and Michael Tuexen)
- Adds support for UDP tunneling. For BSD there is no
socket yet setup so its disabled, but major argument
changes are in here to emcompass the passing of the port
number (zero when you don't have a udp tunnel, the default
for BSD). Will add some hooks in UDP here shortly (discussed
with Robert) that will allow easy tunneling. (Mainly from
Peter Lei and Michael Tuexen with some BSD work from me :-D)
- Some ease for windows, evidently leave is reserved by their
compile move label leave: -> out:
MFC after: 1 week
- Bug in CA that does not get us incrementing the PBA properly which
made us more conservative.
- comment updated in sctp_input.c
- memsets added before we log
- added arg to hmac id's
MFC after: 2 weeks
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
syncache that has an invalid SEQ instead of only doing it when we suceed
in mallocing space for the log message.
MFC after: 1 week
Reviewed by: sam, bz
