Re: net/atm: warning in alloc_tx/__might_sleep

[Chas Williams <3chas3@gmail.com>]

On Wed, 2017-01-11 at 20:36 -0800, Cong Wang wrote:
> On Wed, Jan 11, 2017 at 11:46 AM, Michal Hocko <mhocko@kernel.org> wrote:
> > On Wed 11-01-17 20:45:25, Michal Hocko wrote:
> >> On Wed 11-01-17 09:37:06, Chas Williams wrote:
> >> > On Mon, 2017-01-09 at 18:20 +0100, Andrey Konovalov wrote:
> >> > > Hi!
> >> > >
> >> > > I've got the following error report while running the syzkaller fuzzer.
> >> > >
> >> > > On commit a121103c922847ba5010819a3f250f1f7fc84ab8 (4.10-rc3).
> >> > >
> >> > > A reproducer is attached.
> >> > >
> >> > > ------------[ cut here ]------------
> >> > > WARNING: CPU: 0 PID: 4114 at kernel/sched/core.c:7737 __might_sleep+0x149/0x1a0
> >> > > do not call blocking ops when !TASK_RUNNING; state=1 set at
> >> > > [<ffffffff813fcb22>] prepare_to_wait+0x182/0x530
> >> > > Modules linked in:
> >> > > CPU: 0 PID: 4114 Comm: a.out Not tainted 4.10.0-rc3+ #59
> >> > > Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
> >> > > Call Trace:
> >> > >  __dump_stack lib/dump_stack.c:15
> >> > >  dump_stack+0x292/0x398 lib/dump_stack.c:51
> >> > >  __warn+0x19f/0x1e0 kernel/panic.c:547
> >> > >  warn_slowpath_fmt+0xc5/0x110 kernel/panic.c:562
> >> > >  __might_sleep+0x149/0x1a0 kernel/sched/core.c:7732
> >> > >  slab_pre_alloc_hook mm/slab.h:408
> >> > >  slab_alloc_node mm/slub.c:2634
> >> > >  kmem_cache_alloc_node+0x14a/0x280 mm/slub.c:2744
> >> > >  __alloc_skb+0x10f/0x800 net/core/skbuff.c:219
> >> > >  alloc_skb ./include/linux/skbuff.h:926
> >> > >  alloc_tx net/atm/common.c:75
> >> >
> >> > This is likely alloc_skb(..., GFP_KERNEL) in alloc_tx().  The simplest
> >> > fix for this would be simply to switch this GFP_ATOMIC.  See if this is
> >> > any better.
> >> >
> >> > diff --git a/net/atm/common.c b/net/atm/common.c
> >> > index a3ca922..d84220c 100644
> >> > --- a/net/atm/common.c
> >> > +++ b/net/atm/common.c
> >> > @@ -72,7 +72,7 @@ static struct sk_buff *alloc_tx(struct atm_vcc *vcc, unsigned int size)
> >> >                      sk_wmem_alloc_get(sk), size, sk->sk_sndbuf);
> >> >             return NULL;
> >> >     }
> >> > -   while (!(skb = alloc_skb(size, GFP_KERNEL)))
> >> > +   while (!(skb = alloc_skb(size, GFP_ATOMIC)))
> >> >             schedule();
> >> >     pr_debug("%d += %d\n", sk_wmem_alloc_get(sk), skb->truesize);
> >> >     atomic_add(skb->truesize, &sk->sk_wmem_alloc);
> >>
> >> Blee, this code is just horrendous. But the "fix" is obviously broken!
> >> schedule() is just a noop if you do not change the task state and what
> >> you are just asking for is a never failing non sleeping allocation - aka
> >> a busy loop in the kernel!
> >
> > And btw. this while loop should be really turned into GFP_KERNEL |
> > __GFP_NOFAIL with and explanation why this allocation cannot possibly
> > fail.
> 
> I think a nested loop is quite unnecessary, probably due to the code itself
> is pretty old. The alloc_tx() is in the outer loop, the alloc_skb() is
> in the inner
> loop, both seem to wait for a successful GFP allocation. The inner one
> is even more unnecessary.
> 
> Of course, I am not surprised MM may already have a mechanism to do
> the similar logic.
> 
> There maybe some reason ATM needs such a logic, although other proto
> could handle skb allocation failure quite well in ->sendmsg().


I can't think of any particular reason that it needs this loop here.  I suspect
that the loop for alloc_tx() predates the wait logic in ->sendmsg() and that the
original looping was in alloc_tx() initially and was simply never removed.  Changes
here would date back to before the git conversion.