GitLab

commit 7d267278

 upstream.

Rainer Weikusat <rweikusat@mobileactivedefense.com> writes:
An AF_UNIX datagram socket being the client in an n:1 association with
some server socket is only allowed to send messages to the server if the
receive queue of this socket contains at most sk_max_ack_backlog
datagrams. This implies that prospective writers might be forced to go
to sleep despite none of the message presently enqueued on the server
receive queue were sent by them. In order to ensure that these will be
woken up once space becomes again available, the present unix_dgram_poll
routine does a second sock_poll_wait call with the peer_wait wait queue
of the server socket as queue argument (unix_dgram_recvmsg does a wake
up on this queue after a datagram was received). This is inherently
problematic because the server socket is only guaranteed to remain alive
for as long as the client still holds a reference to it. In case the
connection is dissolved via connect or by the dead peer detection logic
in unix_dgram_sendmsg, the server socket may be freed despite "the
polling mechanism" (in particular, epoll) still has a pointer to the
corresponding peer_wait queue. There's no way to forcibly deregister a
wait queue with epoll.

Based on an idea by Jason Baron, the patch below changes the code such
that a wait_queue_t belonging to the client socket is enqueued on the
peer_wait queue of the server whenever the peer receive queue full
condition is detected by either a sendmsg or a poll. A wake up on the
peer queue is then relayed to the ordinary wait queue of the client
socket via wake function. The connection to the peer wait queue is again
dissolved if either a wake up is about to be relayed or the client
socket reconnects or a dead peer is detected or the client socket is
itself closed. This enables removing the second sock_poll_wait from
unix_dgram_poll, thus avoiding the use-after-free, while still ensuring
that no blocked writer sleeps forever.
Signed-off-by: Rainer Weikusat <rweikusat@mobileactivedefense.com>
Fixes: ec0d215f

 ("af_unix: fix 'poll for write'/connected DGRAM sockets")
Reviewed-by: Jason Baron <jbaron@akamai.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Zefan Li <lizefan@huawei.com>

[ Upstream commit de144391 ]

Some applications didn't expect recvmsg() on a non blocking socket
could return -EINTR. This possibility was added as a side effect
of commit b3ca9b02 ("net: fix multithreaded signal handling in
unix recv routines").

To hit this bug, you need to be a bit unlucky, as the u->readlock
mutex is usually held for very small periods.

Fixes: b3ca9b02

 ("net: fix multithreaded signal handling in unix recv routines")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Rainer Weikusat <rweikusat@mobileactivedefense.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

[ Upstream commit 37ab4fa7

 ]

This is similar to the set_peek_off patch where calling bind while the
socket is stuck in unix_dgram_recvmsg() will block and cause a hung task
spew after a while.

This is also the last place that did a straightforward mutex_lock(), so
there shouldn't be any more of these patches.
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

[ Upstream commit 12663bfc

 ]

unix_dgram_recvmsg() will hold the readlock of the socket until recv
is complete.

In the same time, we may try to setsockopt(SO_PEEK_OFF) which will hang until
unix_dgram_recvmsg() will complete (which can take a while) without allowing
us to break out of it, triggering a hung task spew.

Instead, allow set_peek_off to fail, this way userspace will not hang.
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

[ Upstream commit f3d33426

 ]

This patch now always passes msg->msg_namelen as 0. recvmsg handlers must
set msg_namelen to the proper size <= sizeof(struct sockaddr_storage)
to return msg_name to the user.

This prevents numerous uninitialized memory leaks we had in the
recvmsg handlers and makes it harder for new code to accidentally leak
uninitialized memory.

Optimize for the case recvfrom is called with NULL as address. We don't
need to copy the address at all, so set it to NULL before invoking the
recvmsg handler. We can do so, because all the recvmsg handlers must
cope with the case a plain read() is called on them. read() also sets
msg_name to NULL.

Also document these changes in include/linux/net.h as suggested by David
Miller.

Changes since RFC:

Set msg->msg_name = NULL if user specified a NULL in msg_name but had a
non-null msg_namelen in verify_iovec/verify_compat_iovec. This doesn't
affect sendto as it would bail out earlier while trying to copy-in the
address. It also more naturally reflects the logic by the callers of
verify_iovec.

With this change in place I could remove "
if (!uaddr || msg_sys->msg_namelen == 0)
	msg->msg_name = NULL
".

This change does not alter the user visible error logic as we ignore
msg_namelen as long as msg_name is NULL.

Also remove two unnecessary curly brackets in ___sys_recvmsg and change
comments to netdev style.

Cc: David Miller <davem@davemloft.net>
Suggested-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

[ Upstream commit 90c6bd34 ]

In the case of credentials passing in unix stream sockets (dgram
sockets seem not affected), we get a rather sparse race after
commit 16e57262 ("af_unix: dont send SCM_CREDENTIALS by default").

We have a stream server on receiver side that requests credential
passing from senders (e.g. nc -U). Since we need to set SO_PASSCRED
on each spawned/accepted socket on server side to 1 first (as it's
not inherited), it can happen that in the time between accept() and
setsockopt() we get interrupted, the sender is being scheduled and
continues with passing data to our receiver. At that time SO_PASSCRED
is neither set on sender nor receiver side, hence in cmsg's
SCM_CREDENTIALS we get eventually pid:0, uid:65534, gid:65534
(== overflow{u,g}id) instead of what we actually would like to see.

On the sender side, here nc -U, the tests in maybe_add_creds()
invoked through unix_stream_sendmsg() would fail, as at that exact
time, as mentioned, the sender has neither SO_PASSCRED on his side
nor sees it on the server side, and we have a valid 'other' socket
in place. Thus, sender believes it would just look like a normal
connection, not needing/requesting SO_PASSCRED at that time.

As reverting 16e57262

 would not be an option due to the significant
performance regression reported when having creds always passed,
one way/trade-off to prevent that would be to set SO_PASSCRED on
the listener socket and allow inheriting these flags to the spawned
socket on server side in accept(). It seems also logical to do so
if we'd tell the listener socket to pass those flags onwards, and
would fix the race.

Before, strace:

recvmsg(4, {msg_name(0)=NULL, msg_iov(1)=[{"blub\n", 4096}],
        msg_controllen=32, {cmsg_len=28, cmsg_level=SOL_SOCKET,
        cmsg_type=SCM_CREDENTIALS{pid=0, uid=65534, gid=65534}},
        msg_flags=0}, 0) = 5

After, strace:

recvmsg(4, {msg_name(0)=NULL, msg_iov(1)=[{"blub\n", 4096}],
        msg_controllen=32, {cmsg_len=28, cmsg_level=SOL_SOCKET,
        cmsg_type=SCM_CREDENTIALS{pid=11580, uid=1000, gid=1000}},
        msg_flags=0}, 0) = 5
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

[ Upstream commit 6865d1e8

 ]

When filling the netlink message we miss to wipe the pad field,
therefore leak one byte of heap memory to userland. Fix this by
setting pad to 0.
Signed-off-by: Mathias Krause <minipli@googlemail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

[ Upstream commit 0e82e7f6 ]

It was reported that the following LSB test case failed
https://lsbbugs.linuxfoundation.org/attachment.cgi?id=2144

 because we
were not coallescing unix stream messages when the application was
expecting us to.

The problem was that the first send was before the socket was accepted
and thus sock->sk_socket was NULL in maybe_add_creds, and the second
send after the socket was accepted had a non-NULL value for sk->socket
and thus we could tell the credentials were not needed so we did not
bother.

The unnecessary credentials on the first message cause
unix_stream_recvmsg to start verifying that all messages had the same
credentials before coallescing and then the coallescing failed because
the second message had no credentials.

Ignoring credentials when we don't care in unix_stream_recvmsg fixes a
long standing pessimization which would fail to coallesce messages when
reading from a unix stream socket if the senders were different even if
we did not care about their credentials.

I have tested this and verified that the in the LSB test case mentioned
above that the messages do coallesce now, while the were failing to
coallesce without this change.
Reported-by: Karel Srot <ksrot@redhat.com>
Reported-by: Ding Tianhong <dingtianhong@huawei.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

[ Upstream commit ded34e0f

 ]

As reported by Jan, and others over the past few years, there is a
race condition caused by unix_release setting the sock->sk pointer
to NULL before properly marking the socket as dead/orphaned.  This
can cause a problem with the LSM hook security_unix_may_send() if
there is another socket attempting to write to this partially
released socket in between when sock->sk is set to NULL and it is
marked as dead/orphaned.  This patch fixes this by only setting
sock->sk to NULL after the socket has been marked as dead; I also
take the opportunity to make unix_release_sock() a void function
as it only ever returned 0/success.

Dave, I think this one should go on the -stable pile.

Special thanks to Jan for coming up with a reproducer for this
problem.
Reported-by: Jan Stancek <jan.stancek@gmail.com>
Signed-off-by: Paul Moore <pmoore@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

[ Upstream commit e0e3cea4 ]

Pablo Neira Ayuso discovered that avahi and
potentially NetworkManager accept spoofed Netlink messages because of a
kernel bug.  The kernel passes all-zero SCM_CREDENTIALS ancillary data
to the receiver if the sender did not provide such data, instead of not
including any such data at all or including the correct data from the
peer (as it is the case with AF_UNIX).

This bug was introduced in commit 16e57262


(af_unix: dont send SCM_CREDENTIALS by default)

This patch forces passing credentials for netlink, as
before the regression.

Another fix would be to not add SCM_CREDENTIALS in
netlink messages if not provided by the sender, but it
might break some programs.

With help from Florian Weimer & Petr Matousek

This issue is designated as CVE-2012-3520
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Petr Matousek <pmatouse@redhat.com>
Cc: Florian Weimer <fweimer@redhat.com>
Cc: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

In some cases the poll() implementation in a driver has to do different
things depending on the events the caller wants to poll for.  An example
is when a driver needs to start a DMA engine if the caller polls for
POLLIN, but doesn't want to do that if POLLIN is not requested but instead
only POLLOUT or POLLPRI is requested.  This is something that can happen
in the video4linux subsystem among others.

Unfortunately, the current epoll/poll/select implementation doesn't
provide that information reliably.  The poll_table_struct does have it: it
has a key field with the event mask.  But once a poll() call matches one
or more bits of that mask any following poll() calls are passed a NULL
poll_table pointer.

Also, the eventpoll implementation always left the key field at ~0 instead
of using the requested events mask.

This was changed in eventpoll.c so the key field now contains the actual
events that should be polled for as set by the caller.

The solution to the NULL poll_table pointer is to set the qproc field to
NULL in poll_table once poll() matches the events, not the poll_table
pointer itself.  That way drivers can obtain the mask through a new
poll_requested_events inline.

The poll_table_struct can still be NULL since some kernel code calls it
internally (netfs_state_poll() in ./drivers/staging/pohmelfs/netfs.h).  In
that case poll_requested_events() returns ~0 (i.e.  all events).

Very rarely drivers might want to know whether poll_wait will actually
wait.  If another earlier file descriptor in the set already matched the
events the caller wanted to wait for, then the kernel will return from the
select() call without waiting.  This might be useful information in order
to avoid doing expensive work.

A new helper function poll_does_not_wait() is added that drivers can use
to detect this situation.  This is now used in sock_poll_wait() in
include/net/sock.h.  This was the only place in the kernel that needed
this information.

Drivers should no longer access any of the poll_table internals, but use
the poll_requested_events() and poll_does_not_wait() access functions
instead.  In order to enforce that the poll_table fields are now prepended
with an underscore and a comment was added warning against using them
directly.

This required a change in unix_dgram_poll() in unix/af_unix.c which used
the key field to get the requested events.  It's been replaced by a call
to poll_requested_events().

For qproc it was especially important to change its name since the
behavior of that field changes with this patch since this function pointer
can now be NULL when that wasn't possible in the past.

Any driver accessing the qproc or key fields directly will now fail to compile.

Some notes regarding the correctness of this patch: the driver's poll()
function is called with a 'struct poll_table_struct *wait' argument.  This
pointer may or may not be NULL, drivers can never rely on it being one or
the other as that depends on whether or not an earlier file descriptor in
the select()'s fdset matched the requested events.

There are only three things a driver can do with the wait argument:

1) obtain the key field:

	events = wait ? wait->key : ~0;

   This will still work although it should be replaced with the new
   poll_requested_events() function (which does exactly the same).
   This will now even work better, since wait is no longer set to NULL
   unnecessarily.

2) use the qproc callback. This could be deadly since qproc can now be
   NULL. Renaming qproc should prevent this from happening. There are no
   kernel drivers that actually access this callback directly, BTW.

3) test whether wait == NULL to determine whether poll would return without
   waiting. This is no longer sufficient as the correct test is now
   wait == NULL || wait->_qproc == NULL.

   However, the worst that can happen here is a slight performance hit in
   the case where wait != NULL and wait->_qproc == NULL. In that case the
   driver will assume that poll_wait() will actually add the fd to the set
   of waiting file descriptors. Of course, poll_wait() will not do that
   since it tests for wait->_qproc. This will not break anything, though.

   There is only one place in the whole kernel where this happens
   (sock_poll_wait() in include/net/sock.h) and that code will be replaced
   by a call to poll_does_not_wait() in the next patch.

   Note that even if wait->_qproc != NULL drivers cannot rely on poll_wait()
   actually waiting. The next file descriptor from the set might match the
   event mask and thus any possible waits will never happen.
Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Reviewed-by: Jonathan Corbet <corbet@lwn.net>
Reviewed-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Cc: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

Davem considers that the argument list of this interface is getting
out of control. This patch tries to address this issue following
his proposal:

struct netlink_dump_control c = { .dump = dump, .done = done, ... };

netlink_dump_start(..., &c);

Suggested by David S. Miller.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

Piergiorgio Beruto expressed the need to fetch size of first datagram in
queue for AF_UNIX sockets and suggested a patch against SIOCINQ ioctl.

I suggested instead to implement MSG_TRUNC support as a recv() input
flag, as already done for RAW, UDP & NETLINK sockets.

len = recv(fd, &byte, 1, MSG_PEEK | MSG_TRUNC);

MSG_TRUNC asks recv() to return the real length of the packet, even when
is was longer than the passed buffer.

There is risk that a userland application used MSG_TRUNC by accident
(since it had no effect on af_unix sockets) and this might break after
this patch.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Tested-by: Piergiorgio Beruto <piergiorgio.beruto@gmail.com>
CC: Michael Kerrisk <mtk.manpages@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The same here -- we can protect the sk_peek_off manipulations with
the unix_sk->readlock mutex.

The peeking of data from a stream socket is done in the datagram style,
i.e. even if there's enough room for more data in the user buffer, only
the head skb's data is copied in there. This feature is preserved when
peeking data from a given offset -- the data is read till the nearest
skb's boundary.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The sk_peek_off manipulations are protected with the unix_sk->readlock mutex.
This mutex is enough since all we need is to syncronize setting the offset
vs reading the queue head. The latter is fully covered with the mentioned lock.

The recently added __skb_recv_datagram's offset is used to pick the skb to
read the data from.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Commit 0884d7aa

 (AF_UNIX: Fix poll blocking problem when reading from
a stream socket) added a regression for epoll() in Edge Triggered mode
(EPOLLET)

Appropriate fix is to use skb_peek()/skb_unlink() instead of
skb_dequeue(), and only call skb_unlink() when skb is fully consumed.

This remove the need to requeue a partial skb into sk_receive_queue head
and the extra sk->sk_data_ready() calls that added the regression.

This is safe because once skb is given to sk_receive_queue, it is not
modified by a writer, and readers are serialized by u->readlock mutex.

This also reduce number of spinlock acquisition for small reads or
MSG_PEEK users so should improve overall performance.
Reported-by: Nick Mathewson <nickm@freehaven.net>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Alexey Moiseytsev <himeraster@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Signed-off-by: David S. Miller <davem@davemloft.net>

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

While it's not too late fix the recently added RQLEN diag extension
to report rqlen and wqlen in the same way as TCP does.

I.e. for listening sockets the ack backlog length (which is the input
queue length for socket) in rqlen and the max ack backlog length in
wqlen, and what the CINQ/OUTQ ioctls do for established.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Currently tcp diag reports rqlen and wqlen values similar to how
the CINQ/COUTQ iotcls do. To make unix diag report these values
in the same way move the respective code into helpers.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[ Fix indentation of sock_diag*() calls. -DaveM ]
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Otherwise we leave uninitialized kernel memory in there.
Reported-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The NLA_PUT macro should accept the actual attribute length, not
the amount of elements in array :(
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Otherwise getting

 | net/unix/diag.c:312:16: error: expected declaration specifiers or ‘...’ before string constant
 | net/unix/diag.c:313:1: error: expected declaration specifiers or ‘...’ before string constant
Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

When establishing a unix connection on stream sockets the
server end receives an skb with socket in its receive queue.

Report who is waiting for these ends to be accepted for
listening sockets via NLA.

There's a lokcing issue with this -- the unix sk state lock is
required to access the peer, and it is taken under the listening
sk's queue lock. Strictly speaking the queue lock should be taken
inside the state lock, but since in this case these two sockets
are different it shouldn't lead to deadlock.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Report the peer socket inode ID as NLA. With this it's finally
possible to find out the other end of an interesting unix connection.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Actually, the socket path if it's not anonymous doesn't give
a clue to which file the socket is bound to. Even if the path
is absolute, it can be unlinked and then new socket can be
bound to it.

With this NLA it's possible to check which file a particular
socket is really bound to.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Report the sun_path when requested as NLA. With leading '\0' if
present but without the leading AF_UNIX bits.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The socket inode is used as a key for lookup. This is effectively
the only really unique ID of a unix socket, but using this for
search currently has one problem -- it is O(number of sockets) :(

Does it worth fixing this lookup or inventing some other ID for
unix sockets?
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Walk the unix sockets table and fill the core response structure,
which includes type, state and inode.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Includes basic module_init/_exit functionality, dump/get_exact stubs
and declares the basic API structures for request and response.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

poll() call may be blocked by concurrent reading from the same stream
socket.
Signed-off-by: Alexey Moiseytsev <himeraster@gmail.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Since commit 7361c36c (af_unix: Allow credentials to work across
user and pid namespaces) af_unix performance dropped a lot.

This is because we now take a reference on pid and cred in each write(),
and release them in read(), usually done from another process,
eventually from another cpu. This triggers false sharing.

# Events: 154K cycles
#
# Overhead  Command       Shared Object        Symbol
# ........  .......  ..................  .........................
#
    10.40%  hackbench  [kernel.kallsyms]   [k] put_pid
     8.60%  hackbench  [kernel.kallsyms]   [k] unix_stream_recvmsg
     7.87%  hackbench  [kernel.kallsyms]   [k] unix_stream_sendmsg
     6.11%  hackbench  [kernel.kallsyms]   [k] do_raw_spin_lock
     4.95%  hackbench  [kernel.kallsyms]   [k] unix_scm_to_skb
     4.87%  hackbench  [kernel.kallsyms]   [k] pid_nr_ns
     4.34%  hackbench  [kernel.kallsyms]   [k] cred_to_ucred
     2.39%  hackbench  [kernel.kallsyms]   [k] unix_destruct_scm
   ...

This reverts commit 0856a304

.

As requested by Eric Dumazet, it has various ref-counting
problems and has introduced regressions.  Eric will add
a more suitable version of this performance fix.
Signed-off-by: David S. Miller <davem@davemloft.net>

Patch series 109f6e39..7361c36c

 back in 2.6.36 added functionality to
allow credentials to work across pid namespaces for packets sent via
UNIX sockets.  However, the atomic reference counts on pid and
credentials caused plenty of cache bouncing when there are numerous
threads of the same pid sharing a UNIX socket.  This patch mitigates the
problem by eliminating extraneous reference counts on pid and
credentials on both send and receive path of UNIX sockets. I found a 2x
improvement in hackbench's threaded case.

On the receive path in unix_dgram_recvmsg, currently there is an
increment of reference count on pid and credentials in scm_set_cred.
Then there are two decrement of the reference counts.  Once in scm_recv
and once when skb_free_datagram call skb->destructor function
unix_destruct_scm.  One pair of increment and decrement of ref count on
pid and credentials can be eliminated from the receive path.  Until we
destroy the skb, we already set a reference when we created the skb on
the send side.

On the send path, there are two increments of ref count on pid and
credentials, once in scm_send and once in unix_scm_to_skb.  Then there
is a decrement of the reference counts in scm_destroy's call to
scm_destroy_cred at the end of unix_dgram_sendmsg functions.   One pair
of increment and decrement of the reference counts can be removed so we
only need to increment the ref counts once.

By incorporating these changes, for hackbench running on a 4 socket
NHM-EX machine with 40 cores, the execution of hackbench on
50 groups of 20 threads sped up by factor of 2.

Hackbench command used for testing:
./hackbench 50 thread 2000
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>