net: fix sk_sockets_allocated_read_positive
Eric Dumazet [Sat, 28 Apr 2012 23:21:56 +0000 (23:21 +0000)]
Denys Fedoryshchenko reported frequent crashes on a proxy server and kindly
provided a lockdep report that explains it all :

  [  762.903868]
  [  762.903880] =================================
  [  762.903890] [ INFO: inconsistent lock state ]
  [  762.903903] 3.3.4-build-0061 #8 Not tainted
  [  762.904133] ---------------------------------
  [  762.904344] inconsistent {IN-SOFTIRQ-W} -> {SOFTIRQ-ON-W} usage.
  [  762.904542] squid/1603 [HC0[0]:SC0[0]:HE1:SE1] takes:
  [  762.904542]  (key#3){+.?...}, at: [<c0232cc4>]
__percpu_counter_sum+0xd/0x58
  [  762.904542] {IN-SOFTIRQ-W} state was registered at:
  [  762.904542]   [<c0158b84>] __lock_acquire+0x284/0xc26
  [  762.904542]   [<c01598e8>] lock_acquire+0x71/0x85
  [  762.904542]   [<c0349765>] _raw_spin_lock+0x33/0x40
  [  762.904542]   [<c0232c93>] __percpu_counter_add+0x58/0x7c
  [  762.904542]   [<c02cfde1>] sk_clone_lock+0x1e5/0x200
  [  762.904542]   [<c0303ee4>] inet_csk_clone_lock+0xe/0x78
  [  762.904542]   [<c0315778>] tcp_create_openreq_child+0x1b/0x404
  [  762.904542]   [<c031339c>] tcp_v4_syn_recv_sock+0x32/0x1c1
  [  762.904542]   [<c031615a>] tcp_check_req+0x1fd/0x2d7
  [  762.904542]   [<c0313f77>] tcp_v4_do_rcv+0xab/0x194
  [  762.904542]   [<c03153bb>] tcp_v4_rcv+0x3b3/0x5cc
  [  762.904542]   [<c02fc0c4>] ip_local_deliver_finish+0x13a/0x1e9
  [  762.904542]   [<c02fc539>] NF_HOOK.clone.11+0x46/0x4d
  [  762.904542]   [<c02fc652>] ip_local_deliver+0x41/0x45
  [  762.904542]   [<c02fc4d1>] ip_rcv_finish+0x31a/0x33c
  [  762.904542]   [<c02fc539>] NF_HOOK.clone.11+0x46/0x4d
  [  762.904542]   [<c02fc857>] ip_rcv+0x201/0x23e
  [  762.904542]   [<c02daa3a>] __netif_receive_skb+0x319/0x368
  [  762.904542]   [<c02dac07>] netif_receive_skb+0x4e/0x7d
  [  762.904542]   [<c02dacf6>] napi_skb_finish+0x1e/0x34
  [  762.904542]   [<c02db122>] napi_gro_receive+0x20/0x24
  [  762.904542]   [<f85d1743>] e1000_receive_skb+0x3f/0x45 [e1000e]
  [  762.904542]   [<f85d3464>] e1000_clean_rx_irq+0x1f9/0x284 [e1000e]
  [  762.904542]   [<f85d3926>] e1000_clean+0x62/0x1f4 [e1000e]
  [  762.904542]   [<c02db228>] net_rx_action+0x90/0x160
  [  762.904542]   [<c012a445>] __do_softirq+0x7b/0x118
  [  762.904542] irq event stamp: 156915469
  [  762.904542] hardirqs last  enabled at (156915469): [<c019b4f4>]
__slab_alloc.clone.58.clone.63+0xc4/0x2de
  [  762.904542] hardirqs last disabled at (156915468): [<c019b452>]
__slab_alloc.clone.58.clone.63+0x22/0x2de
  [  762.904542] softirqs last  enabled at (156915466): [<c02ce677>]
lock_sock_nested+0x64/0x6c
  [  762.904542] softirqs last disabled at (156915464): [<c0349914>]
_raw_spin_lock_bh+0xe/0x45
  [  762.904542]
  [  762.904542] other info that might help us debug this:
  [  762.904542]  Possible unsafe locking scenario:
  [  762.904542]
  [  762.904542]        CPU0
  [  762.904542]        ----
  [  762.904542]   lock(key#3);
  [  762.904542]   <Interrupt>
  [  762.904542]     lock(key#3);
  [  762.904542]
  [  762.904542]  *** DEADLOCK ***
  [  762.904542]
  [  762.904542] 1 lock held by squid/1603:
  [  762.904542]  #0:  (sk_lock-AF_INET){+.+.+.}, at: [<c03055c0>]
lock_sock+0xa/0xc
  [  762.904542]
  [  762.904542] stack backtrace:
  [  762.904542] Pid: 1603, comm: squid Not tainted 3.3.4-build-0061 #8
  [  762.904542] Call Trace:
  [  762.904542]  [<c0347b73>] ? printk+0x18/0x1d
  [  762.904542]  [<c015873a>] valid_state+0x1f6/0x201
  [  762.904542]  [<c0158816>] mark_lock+0xd1/0x1bb
  [  762.904542]  [<c015876b>] ? mark_lock+0x26/0x1bb
  [  762.904542]  [<c015805d>] ? check_usage_forwards+0x77/0x77
  [  762.904542]  [<c0158bf8>] __lock_acquire+0x2f8/0xc26
  [  762.904542]  [<c0159b8e>] ? mark_held_locks+0x5d/0x7b
  [  762.904542]  [<c0159cf6>] ? trace_hardirqs_on+0xb/0xd
  [  762.904542]  [<c0158dd4>] ? __lock_acquire+0x4d4/0xc26
  [  762.904542]  [<c01598e8>] lock_acquire+0x71/0x85
  [  762.904542]  [<c0232cc4>] ? __percpu_counter_sum+0xd/0x58
  [  762.904542]  [<c0349765>] _raw_spin_lock+0x33/0x40
  [  762.904542]  [<c0232cc4>] ? __percpu_counter_sum+0xd/0x58
  [  762.904542]  [<c0232cc4>] __percpu_counter_sum+0xd/0x58
  [  762.904542]  [<c02cebc4>] __sk_mem_schedule+0xdd/0x1c7
  [  762.904542]  [<c02d178d>] ? __alloc_skb+0x76/0x100
  [  762.904542]  [<c0305e8e>] sk_wmem_schedule+0x21/0x2d
  [  762.904542]  [<c0306370>] sk_stream_alloc_skb+0x42/0xaa
  [  762.904542]  [<c0306567>] tcp_sendmsg+0x18f/0x68b
  [  762.904542]  [<c031f3dc>] ? ip_fast_csum+0x30/0x30
  [  762.904542]  [<c0320193>] inet_sendmsg+0x53/0x5a
  [  762.904542]  [<c02cb633>] sock_aio_write+0xd2/0xda
  [  762.904542]  [<c015876b>] ? mark_lock+0x26/0x1bb
  [  762.904542]  [<c01a1017>] do_sync_write+0x9f/0xd9
  [  762.904542]  [<c01a2111>] ? file_free_rcu+0x2f/0x2f
  [  762.904542]  [<c01a17a1>] vfs_write+0x8f/0xab
  [  762.904542]  [<c01a284d>] ? fget_light+0x75/0x7c
  [  762.904542]  [<c01a1900>] sys_write+0x3d/0x5e
  [  762.904542]  [<c0349ec9>] syscall_call+0x7/0xb
  [  762.904542]  [<c0340000>] ? rp_sidt+0x41/0x83

Bug is that sk_sockets_allocated_read_positive() calls
percpu_counter_sum_positive() without BH being disabled.

This bug was added in commit 180d8cd942ce33
(foundations of per-cgroup memory pressure controlling.), since previous
code was using percpu_counter_read_positive() which is IRQ safe.

In __sk_mem_schedule() we dont need the precise count of allocated
sockets and can revert to previous behavior.

Reported-by: Denys Fedoryshchenko <denys@visp.net.lb>
Sined-off-by: Eric Dumazet <edumazet@google.com>
Cc: Glauber Costa <glommer@parallels.com>
Acked-by: Neal Cardwell <ncardwell@google.com>

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

include/net/sock.h

index 188532e..5a0a58a 100644 (file)
@@ -1129,9 +1129,9 @@ sk_sockets_allocated_read_positive(struct sock *sk)
        struct proto *prot = sk->sk_prot;
 
        if (mem_cgroup_sockets_enabled && sk->sk_cgrp)
-               return percpu_counter_sum_positive(sk->sk_cgrp->sockets_allocated);
+               return percpu_counter_read_positive(sk->sk_cgrp->sockets_allocated);
 
-       return percpu_counter_sum_positive(prot->sockets_allocated);
+       return percpu_counter_read_positive(prot->sockets_allocated);
 }
 
 static inline int