lkml.org 
[lkml]   [2019]   [Jan]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: possible deadlock in __wake_up_common_lock
On Wed, Jan 02, 2019 at 01:51:01PM +0100, Vlastimil Babka wrote:
> On 1/2/19 9:51 AM, syzbot wrote:
> > Hello,
> >
> > syzbot found the following crash on:
> >
> > HEAD commit: f346b0becb1b Merge branch 'akpm' (patches from Andrew)
> > git tree: upstream
> > console output: https://syzkaller.appspot.com/x/log.txt?x=1510cefd400000
> > kernel config: https://syzkaller.appspot.com/x/.config?x=c255c77ba370fe7c
> > dashboard link: https://syzkaller.appspot.com/bug?extid=93d94a001cfbce9e60e1
> > compiler: gcc (GCC) 8.0.1 20180413 (experimental)
> > userspace arch: i386
> >
> > Unfortunately, I don't have any reproducer for this crash yet.
> >
> > IMPORTANT: if you fix the bug, please add the following tag to the commit:
> > Reported-by: syzbot+93d94a001cfbce9e60e1@syzkaller.appspotmail.com
> >
> >
> > ======================================================
> > WARNING: possible circular locking dependency detected
> > 4.20.0+ #297 Not tainted
> > ------------------------------------------------------
> > syz-executor0/8529 is trying to acquire lock:
> > 000000005e7fb829 (&pgdat->kswapd_wait){....}, at:
> > __wake_up_common_lock+0x19e/0x330 kernel/sched/wait.c:120
>
> From the backtrace at the end of report I see it's coming from
>
> > wakeup_kswapd+0x5f0/0x930 mm/vmscan.c:3982
> > steal_suitable_fallback+0x538/0x830 mm/page_alloc.c:2217
>
> This wakeup_kswapd is new due to Mel's 1c30844d2dfe ("mm: reclaim small
> amounts of memory when an external fragmentation event occurs") so CC Mel.
>

New year new bugs :(

> > but task is already holding lock:
> > 000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: spin_lock
> > include/linux/spinlock.h:329 [inline]
> > 000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue_bulk
> > mm/page_alloc.c:2548 [inline]
> > 000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: __rmqueue_pcplist
> > mm/page_alloc.c:3021 [inline]
> > 000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue_pcplist
> > mm/page_alloc.c:3050 [inline]
> > 000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue
> > mm/page_alloc.c:3072 [inline]
> > 000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at:
> > get_page_from_freelist+0x1bae/0x52a0 mm/page_alloc.c:3491
> >
> > which lock already depends on the new lock.
>
> However, I don't understand why lockdep thinks it's a problem. IIRC it
> doesn't like that we are locking pgdat->kswapd_wait.lock while holding
> zone->lock. That means it has learned that the opposite order also
> exists, e.g. somebody would take zone->lock while manipulating the wait
> queue? I don't see where but I admit I'm not good at reading lockdep
> splats, so CCing Peterz and Ingo as well. Keeping rest of mail for
> reference.
>

I'm not sure I'm reading the output correctly because I'm having trouble
seeing the exact pattern that allows lockdep to conclude the lock ordering
is problematic.

I think it's hungup on the fact that mod_timer can allocate debug
objects for KASAN and somehow concludes that the waking of kswapd is
problematic because potentially a lock ordering exists that would trip.
I don't see how it's actually possible though due to either a lack of
imagination or maybe lockdep is being cautious as something could change
in the future that allows the lockup.

There are a few options I guess in order of preference.

1. Drop zone->lock for the call. It's not necessarily to keep track of
the IRQ flags as callers into that path already do things like treat
IRQ disabling and the spin lock separately.

2. Use another alloc_flag in steal_suitable_fallback that is set when a
wakeup is required but do the actual wakeup in rmqueue() after the
zone locks are dropped and the allocation request is completed

3. Always wakeup kswapd if watermarks are boosted. I like this the least
because it means doing wakeups that are unrelated to fragmentation
that occurred in the current context.

Any particular preference?

While I recognise there is no test case available, how often does this
trigger in syzbot as it would be nice to have some confirmation any
patch is really fixing the problem.

--
Mel Gorman
SUSE Labs

\
 
 \ /
  Last update: 2019-01-02 19:07    [W:0.114 / U:2.008 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site