Re: packet_sendmsg_spkt sleeping from invalid context

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Dec 14, 2009 at 08:24:23PM +0100, Eric Dumazet wrote:
> Le 14/12/2009 18:52, Frederic Weisbecker a écrit :
> > Hi,
> > 
> > I don't know if it has been reported already.
> > I get the following warning on boot, with latest upstream tree:
> > 
> > [   32.776502] sched: BUG: sleeping function called from invalid context at mm/slab.c:3032
> > [   32.802173] sched: in_atomic(): 1, irqs_disabled(): 0, pid: 3555, name: dhclient3
> > [   32.821141] 1 lock held by dhclient3/3555:
> > [   32.821147]  #0:  (rcu_read_lock){.+.+.+}, at: [<ffffffff815d177d>] packet_sendmsg_spkt+0x7d/0x2c0
> > [   32.821174] Pid: 3555, comm: dhclient3 Tainted: G        W  2.6.32-tip+ #134
> > [   32.821181] Call Trace:
> > [   32.821194]  [<ffffffff810811c3>] ? __debug_show_held_locks+0x13/0x30
> > [   32.821207]  [<ffffffff8103d6b8>] __might_sleep+0x118/0x140
> > [   32.821219]  [<ffffffff81110c23>] kmem_cache_alloc+0x173/0x190
> > [   32.821231]  [<ffffffff815483d9>] __alloc_skb+0x49/0x170
> > [   32.821241]  [<ffffffff81542238>] sock_wmalloc+0x38/0x80
> > [   32.821250]  [<ffffffff815d182b>] packet_sendmsg_spkt+0x12b/0x2c0
> > [   32.821260]  [<ffffffff815d177d>] ? packet_sendmsg_spkt+0x7d/0x2c0
> > [   32.821272]  [<ffffffff8153ded7>] sock_sendmsg+0x127/0x140
> > [   32.821285]  [<ffffffff8106f4a0>] ? autoremove_wake_function+0x0/0x40
> > [   32.821297]  [<ffffffff810f467b>] ? might_fault+0x7b/0xd0
> > [   32.821306]  [<ffffffff810f467b>] ? might_fault+0x7b/0xd0
> > [   32.821318]  [<ffffffff815401aa>] ? move_addr_to_kernel+0x6a/0x70
> > [   32.821328]  [<ffffffff8154029f>] sys_sendto+0xef/0x120
> > [   32.821340]  [<ffffffff81131479>] ? mntput_no_expire+0x29/0x110
> > [   32.821355]  [<ffffffff810027db>] system_call_fastpath+0x16/0x1b
> > 
> 
> Thanks for the report Frederic.
> 
> We could partly revert the original commit, but as we wanted to avoid touching 
> device refcount, and af_packet might be the only real abuser, we could
> try following patch instead.
> 
> Thanks



I also wonder. Are you using PREEMPT_RCU ?
That may explain why you haven't seen this issue because
might_sleep() doesn't see you are in a rcu read locked
section as preemption is not disabled, but it is illegal to
voluntarily sleep in such area (although it's fine with
preempt rcu) as doing so with non-prempt RCU config would barf.

I'm trying a patch to handle that.