On Wed, 2006-11-29 at 18:49 +1100, Herbert Xu wrote:
> Peter Zijlstra <[EMAIL PROTECTED]> wrote:
> >
> > =========================================================
> > [ INFO: possible irq lock inversion dependency detected ]
> > 2.6.19-rc6 #4
> > ---------------------------------------------------------
> > nc/1854 just changed the state of lock:
> > (af_callback_keys + sk->sk_family#2){-.-?}, at: [<c0268a7f>]
> > sock_def_error_report+0x1f/0x90
> > but this lock was taken by another, soft-irq-safe lock in the past:
> > (slock-AF_INET){-+..}
> >
> > and interrupts could create inverse lock ordering between them.
>
> I think this is bogus. The slock is not a standard lock. When we
> hold it in process context we don't actually hold the spin lock part
> of it. However, it does prevent the softirq path from running in
> critical sections which also prevents any attempt to grab the
> callback lock from softirq context.
>
> If you still think there is a problem, please show an actual scenario
> where it dead locks.
process context does lock_sock(sk) which is basically a sleeping lock
and sets an owner field when acquired.
BH context does bh_lock_sock(sk); which spins on the spinlock protecting
the owner field; and checks for an owner under this lock. When an owner
is found it will stick the skb on a queue for later processing.
This scheme does indeed seem to avoid the reported deadlock scenario -
although I didn't audit all code paths.
However I'm not quite sure yet how to teach lockdep about this. The
proposed patch will shut it up though.
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
