tcp: fix possible socket refcount problem
authorEric Dumazet <edumazet@google.com>
Mon, 20 Aug 2012 00:22:46 +0000 (00:22 +0000)
committerDavid S. Miller <davem@davemloft.net>
Tue, 21 Aug 2012 21:42:23 +0000 (14:42 -0700)
commit144d56e91044181ec0ef67aeca91e9a8b5718348
treee379655603668d66dbec27f446b68479cb18e6e2
parentfae6ef87faeb8853896920c68ee703d715799d28
tcp: fix possible socket refcount problem

Commit 6f458dfb40 (tcp: improve latencies of timer triggered events)
added bug leading to following trace :

[ 2866.131281] IPv4: Attempt to release TCP socket in state 1 ffff880019ec0000
[ 2866.131726]
[ 2866.132188] =========================
[ 2866.132281] [ BUG: held lock freed! ]
[ 2866.132281] 3.6.0-rc1+ #622 Not tainted
[ 2866.132281] -------------------------
[ 2866.132281] kworker/0:1/652 is freeing memory ffff880019ec0000-ffff880019ec0a1f, with a lock still held there!
[ 2866.132281]  (sk_lock-AF_INET-RPC){+.+...}, at: [<ffffffff81903619>] tcp_sendmsg+0x29/0xcc6
[ 2866.132281] 4 locks held by kworker/0:1/652:
[ 2866.132281]  #0:  (rpciod){.+.+.+}, at: [<ffffffff81083567>] process_one_work+0x1de/0x47f
[ 2866.132281]  #1:  ((&task->u.tk_work)){+.+.+.}, at: [<ffffffff81083567>] process_one_work+0x1de/0x47f
[ 2866.132281]  #2:  (sk_lock-AF_INET-RPC){+.+...}, at: [<ffffffff81903619>] tcp_sendmsg+0x29/0xcc6
[ 2866.132281]  #3:  (&icsk->icsk_retransmit_timer){+.-...}, at: [<ffffffff81078017>] run_timer_softirq+0x1ad/0x35f
[ 2866.132281]
[ 2866.132281] stack backtrace:
[ 2866.132281] Pid: 652, comm: kworker/0:1 Not tainted 3.6.0-rc1+ #622
[ 2866.132281] Call Trace:
[ 2866.132281]  <IRQ>  [<ffffffff810bc527>] debug_check_no_locks_freed+0x112/0x159
[ 2866.132281]  [<ffffffff818a0839>] ? __sk_free+0xfd/0x114
[ 2866.132281]  [<ffffffff811549fa>] kmem_cache_free+0x6b/0x13a
[ 2866.132281]  [<ffffffff818a0839>] __sk_free+0xfd/0x114
[ 2866.132281]  [<ffffffff818a08c0>] sk_free+0x1c/0x1e
[ 2866.132281]  [<ffffffff81911e1c>] tcp_write_timer+0x51/0x56
[ 2866.132281]  [<ffffffff81078082>] run_timer_softirq+0x218/0x35f
[ 2866.132281]  [<ffffffff81078017>] ? run_timer_softirq+0x1ad/0x35f
[ 2866.132281]  [<ffffffff810f5831>] ? rb_commit+0x58/0x85
[ 2866.132281]  [<ffffffff81911dcb>] ? tcp_write_timer_handler+0x148/0x148
[ 2866.132281]  [<ffffffff81070bd6>] __do_softirq+0xcb/0x1f9
[ 2866.132281]  [<ffffffff81a0a00c>] ? _raw_spin_unlock+0x29/0x2e
[ 2866.132281]  [<ffffffff81a1227c>] call_softirq+0x1c/0x30
[ 2866.132281]  [<ffffffff81039f38>] do_softirq+0x4a/0xa6
[ 2866.132281]  [<ffffffff81070f2b>] irq_exit+0x51/0xad
[ 2866.132281]  [<ffffffff81a129cd>] do_IRQ+0x9d/0xb4
[ 2866.132281]  [<ffffffff81a0a3ef>] common_interrupt+0x6f/0x6f
[ 2866.132281]  <EOI>  [<ffffffff8109d006>] ? sched_clock_cpu+0x58/0xd1
[ 2866.132281]  [<ffffffff81a0a172>] ? _raw_spin_unlock_irqrestore+0x4c/0x56
[ 2866.132281]  [<ffffffff81078692>] mod_timer+0x178/0x1a9
[ 2866.132281]  [<ffffffff818a00aa>] sk_reset_timer+0x19/0x26
[ 2866.132281]  [<ffffffff8190b2cc>] tcp_rearm_rto+0x99/0xa4
[ 2866.132281]  [<ffffffff8190dfba>] tcp_event_new_data_sent+0x6e/0x70
[ 2866.132281]  [<ffffffff8190f7ea>] tcp_write_xmit+0x7de/0x8e4
[ 2866.132281]  [<ffffffff818a565d>] ? __alloc_skb+0xa0/0x1a1
[ 2866.132281]  [<ffffffff8190f952>] __tcp_push_pending_frames+0x2e/0x8a
[ 2866.132281]  [<ffffffff81904122>] tcp_sendmsg+0xb32/0xcc6
[ 2866.132281]  [<ffffffff819229c2>] inet_sendmsg+0xaa/0xd5
[ 2866.132281]  [<ffffffff81922918>] ? inet_autobind+0x5f/0x5f
[ 2866.132281]  [<ffffffff810ee7f1>] ? trace_clock_local+0x9/0xb
[ 2866.132281]  [<ffffffff8189adab>] sock_sendmsg+0xa3/0xc4
[ 2866.132281]  [<ffffffff810f5de6>] ? rb_reserve_next_event+0x26f/0x2d5
[ 2866.132281]  [<ffffffff8103e6a9>] ? native_sched_clock+0x29/0x6f
[ 2866.132281]  [<ffffffff8103e6f8>] ? sched_clock+0x9/0xd
[ 2866.132281]  [<ffffffff810ee7f1>] ? trace_clock_local+0x9/0xb
[ 2866.132281]  [<ffffffff8189ae03>] kernel_sendmsg+0x37/0x43
[ 2866.132281]  [<ffffffff8199ce49>] xs_send_kvec+0x77/0x80
[ 2866.132281]  [<ffffffff8199cec1>] xs_sendpages+0x6f/0x1a0
[ 2866.132281]  [<ffffffff8107826d>] ? try_to_del_timer_sync+0x55/0x61
[ 2866.132281]  [<ffffffff8199d0d2>] xs_tcp_send_request+0x55/0xf1
[ 2866.132281]  [<ffffffff8199bb90>] xprt_transmit+0x89/0x1db
[ 2866.132281]  [<ffffffff81999bcd>] ? call_connect+0x3c/0x3c
[ 2866.132281]  [<ffffffff81999d92>] call_transmit+0x1c5/0x20e
[ 2866.132281]  [<ffffffff819a0d55>] __rpc_execute+0x6f/0x225
[ 2866.132281]  [<ffffffff81999bcd>] ? call_connect+0x3c/0x3c
[ 2866.132281]  [<ffffffff819a0f33>] rpc_async_schedule+0x28/0x34
[ 2866.132281]  [<ffffffff810835d6>] process_one_work+0x24d/0x47f
[ 2866.132281]  [<ffffffff81083567>] ? process_one_work+0x1de/0x47f
[ 2866.132281]  [<ffffffff819a0f0b>] ? __rpc_execute+0x225/0x225
[ 2866.132281]  [<ffffffff81083a6d>] worker_thread+0x236/0x317
[ 2866.132281]  [<ffffffff81083837>] ? process_scheduled_works+0x2f/0x2f
[ 2866.132281]  [<ffffffff8108b7b8>] kthread+0x9a/0xa2
[ 2866.132281]  [<ffffffff81a12184>] kernel_thread_helper+0x4/0x10
[ 2866.132281]  [<ffffffff81a0a4b0>] ? retint_restore_args+0x13/0x13
[ 2866.132281]  [<ffffffff8108b71e>] ? __init_kthread_worker+0x5a/0x5a
[ 2866.132281]  [<ffffffff81a12180>] ? gs_change+0x13/0x13
[ 2866.308506] IPv4: Attempt to release TCP socket in state 1 ffff880019ec0000
[ 2866.309689] =============================================================================
[ 2866.310254] BUG TCP (Not tainted): Object already free
[ 2866.310254] -----------------------------------------------------------------------------
[ 2866.310254]

The bug comes from the fact that timer set in sk_reset_timer() can run
before we actually do the sock_hold(). socket refcount reaches zero and
we free the socket too soon.

timer handler is not allowed to reduce socket refcnt if socket is owned
by the user, or we need to change sk_reset_timer() implementation.

We should take a reference on the socket in case TCP_DELACK_TIMER_DEFERRED
or TCP_DELACK_TIMER_DEFERRED bit are set in tsq_flags

Also fix a typo in tcp_delack_timer(), where TCP_WRITE_TIMER_DEFERRED
was used instead of TCP_DELACK_TIMER_DEFERRED.

For consistency, use same socket refcount change for TCP_MTU_REDUCED_DEFERRED,
even if not fired from a timer.

Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Tested-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_ipv4.c
net/ipv4/tcp_output.c
net/ipv4/tcp_timer.c