sched/fair: Fix post_init_entity_util_avg() serialization
authorPeter Zijlstra <peterz@infradead.org>
Thu, 9 Jun 2016 13:07:50 +0000 (15:07 +0200)
committerIngo Molnar <mingo@kernel.org>
Tue, 14 Jun 2016 08:58:34 +0000 (10:58 +0200)
commitb7fa30c9cc48c4f55663420472505d3b4f6e1705
tree38f0c717a7a3044c2c1d285f06bd3d136cf746c4
parentdb06d759d6cf903aeda8c107fd3abd366dd80200
sched/fair: Fix post_init_entity_util_avg() serialization

Chris Wilson reported a divide by 0 at:

 post_init_entity_util_avg():

 >    725 if (cfs_rq->avg.util_avg != 0) {
 >    726 sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
 > -> 727 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
 >    728
 >    729 if (sa->util_avg > cap)
 >    730 sa->util_avg = cap;
 >    731 } else {

Which given the lack of serialization, and the code generated from
update_cfs_rq_load_avg() is entirely possible:

if (atomic_long_read(&cfs_rq->removed_load_avg)) {
s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
sa->load_avg = max_t(long, sa->load_avg - r, 0);
sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
removed_load = 1;
}

turns into:

  ffffffff81087064:       49 8b 85 98 00 00 00    mov    0x98(%r13),%rax
  ffffffff8108706b:       48 85 c0                test   %rax,%rax
  ffffffff8108706e:       74 40                   je     ffffffff810870b0
  ffffffff81087070:       4c 89 f8                mov    %r15,%rax
  ffffffff81087073:       49 87 85 98 00 00 00    xchg   %rax,0x98(%r13)
  ffffffff8108707a:       49 29 45 70             sub    %rax,0x70(%r13)
  ffffffff8108707e:       4c 89 f9                mov    %r15,%rcx
  ffffffff81087081:       bb 01 00 00 00          mov    $0x1,%ebx
  ffffffff81087086:       49 83 7d 70 00          cmpq   $0x0,0x70(%r13)
  ffffffff8108708b:       49 0f 49 4d 70          cmovns 0x70(%r13),%rcx

Which you'll note ends up with 'sa->load_avg - r' in memory at
ffffffff8108707a.

By calling post_init_entity_util_avg() under rq->lock we're sure to be
fully serialized against PELT updates and cannot observe intermediate
state like this.

Reported-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Yuyang Du <yuyang.du@intel.com>
Cc: bsegall@google.com
Cc: morten.rasmussen@arm.com
Cc: pjt@google.com
Cc: steve.muckle@linaro.org
Fixes: 2b8c41daba32 ("sched/fair: Initiate a new task's util avg to a bounded value")
Link: http://lkml.kernel.org/r/20160609130750.GQ30909@twins.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/sched/core.c
kernel/sched/fair.c