Re: Warning in net/ipv4/af_inet.c:154

[Eric Dumazet <eric.dumazet@gmail.com>]

Le jeudi 27 mai 2010 à 13:56 +1000, Anton Blanchard a écrit :
> Hi Eric,
> 
> > You are 100% right David, maybe we should add a test when changing
> > sk_forward_alloc to test if socket is locked (lockdep only test), but
> > that's for 2.6.36 :)
> 
> Thanks for the patch, unfortunately I can still hit the WARN_ON. I'm somewhat
> confused by the two stage locking in the socket lock (ie sk_lock.slock and
> sk_lock.owned).
> 
> What state should the socket lock be in for serialising updates of
> sk_forward_alloc? In some cases we appear to update it with sk_lock.slock =
> unlocked, sk_lock.owned = 1:
> 
> NIP [c0000000005b4ad0] .sock_queue_rcv_skb
> LR [c0000000005b4acc] .sock_queue_rcv_skb
> Call Trace:
> [c0000000005f9fcc] .ip_queue_rcv_skb
> [c00000000061d604] .__udp_queue_rcv_skb
> [c0000000005b1a38] .release_sock
> [c0000000006205f0] .udp_sendmsg
> [c0000000006290d4] .inet_sendmsg
> [c0000000005abfb4] .sock_sendmsg
> [c0000000005ae9dc] .SyS_sendto
> [c0000000005ab6c0] .SyS_send
> 
> And other times we update it with sk_lock.slock = locked, sk_lock.owned = 0:
> 
> NIP [c0000000005b2b6c] .sock_rfree
> LR [c0000000005b2b68] .sock_rfree
> Call Trace:
> [c0000000005bca10] .skb_free_datagram_locked
> [c00000000061fe88] .udp_recvmsg
> [c0000000006285e8] .inet_recvmsg
> [c0000000005abe0c] .sock_recvmsg
> [c0000000005ae358] .SyS_recvfrom
> 
> I see we sometimes take sk_lock.slock then check the owned field, but we
> aren't doing that all the time.
> 

Old rule was :

A Process context was using 
lock + test_and_set_or_sleep + unlock (sk_lock.slock = unlocked,
sk_lock.owned = 1)

softirq handler was using :
(sk_lock.slock = locked, sk_lock.owned =0)

I added a shortcut, but it seems wrong as is

Process context :

lock only (sk_lock.slock = locked, sk_lock.owned = ???)

So I should add a test on owned. If set (by another thread), we should take the slow path.

Thanks !