]> git.openfabrics.org - ~shefty/rdma-dev.git/commitdiff
net: Fix the condition passed to sk_wait_event()
authorNagendra Tomar <tomer_iisc@yahoo.com>
Sat, 2 Oct 2010 23:45:06 +0000 (23:45 +0000)
committerDavid S. Miller <davem@davemloft.net>
Mon, 4 Oct 2010 03:41:32 +0000 (20:41 -0700)
This patch fixes the condition (3rd arg) passed to sk_wait_event() in
sk_stream_wait_memory(). The incorrect check in sk_stream_wait_memory()
causes the following soft lockup in tcp_sendmsg() when the global tcp
memory pool has exhausted.

>>> snip <<<

localhost kernel: BUG: soft lockup - CPU#3 stuck for 11s! [sshd:6429]
localhost kernel: CPU 3:
localhost kernel: RIP: 0010:[sk_stream_wait_memory+0xcd/0x200]  [sk_stream_wait_memory+0xcd/0x200] sk_stream_wait_memory+0xcd/0x200
localhost kernel:
localhost kernel: Call Trace:
localhost kernel:  [sk_stream_wait_memory+0x1b1/0x200] sk_stream_wait_memory+0x1b1/0x200
localhost kernel:  [<ffffffff802557c0>] autoremove_wake_function+0x0/0x40
localhost kernel:  [ipv6:tcp_sendmsg+0x6e6/0xe90] tcp_sendmsg+0x6e6/0xce0
localhost kernel:  [sock_aio_write+0x126/0x140] sock_aio_write+0x126/0x140
localhost kernel:  [xfs:do_sync_write+0xf1/0x130] do_sync_write+0xf1/0x130
localhost kernel:  [<ffffffff802557c0>] autoremove_wake_function+0x0/0x40
localhost kernel:  [hrtimer_start+0xe3/0x170] hrtimer_start+0xe3/0x170
localhost kernel:  [vfs_write+0x185/0x190] vfs_write+0x185/0x190
localhost kernel:  [sys_write+0x50/0x90] sys_write+0x50/0x90
localhost kernel:  [system_call+0x7e/0x83] system_call+0x7e/0x83

>>> snip <<<

What is happening is, that the sk_wait_event() condition passed from
sk_stream_wait_memory() evaluates to true for the case of tcp global memory
exhaustion. This is because both sk_stream_memory_free() and vm_wait are true
which causes sk_wait_event() to *not* call schedule_timeout().
Hence sk_stream_wait_memory() returns immediately to the caller w/o sleeping.
This causes the caller to again try allocation, which again fails and again
calls sk_stream_wait_memory(), and so on.

[ Bug introduced by commit c1cbe4b7ad0bc4b1d98ea708a3fecb7362aa4088
  ("[NET]: Avoid atomic xchg() for non-error case") -DaveM ]

Signed-off-by: Nagendra Singh Tomar <tomer_iisc@yahoo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/stream.c

index d959e0f41528ce71d69f4aafea0f6d028d72607a..f5df85dcd20bc7f790aec8f58967f55e02586444 100644 (file)
@@ -141,10 +141,10 @@ int sk_stream_wait_memory(struct sock *sk, long *timeo_p)
 
                set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
                sk->sk_write_pending++;
-               sk_wait_event(sk, &current_timeo, !sk->sk_err &&
-                                                 !(sk->sk_shutdown & SEND_SHUTDOWN) &&
-                                                 sk_stream_memory_free(sk) &&
-                                                 vm_wait);
+               sk_wait_event(sk, &current_timeo, sk->sk_err ||
+                                                 (sk->sk_shutdown & SEND_SHUTDOWN) ||
+                                                 (sk_stream_memory_free(sk) &&
+                                                 !vm_wait));
                sk->sk_write_pending--;
 
                if (vm_wait) {