]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/commitdiff
sched: Fix data-race in wakeup
authorPeter Zijlstra <peterz@infradead.org>
Tue, 17 Nov 2020 08:08:41 +0000 (09:08 +0100)
committerPeter Zijlstra <peterz@infradead.org>
Tue, 17 Nov 2020 12:15:27 +0000 (13:15 +0100)
Mel reported that on some ARM64 platforms loadavg goes bananas and
Will tracked it down to the following race:

  CPU0 CPU1

  schedule()
    prev->sched_contributes_to_load = X;
    deactivate_task(prev);

try_to_wake_up()
  if (p->on_rq &&) // false
  if (smp_load_acquire(&p->on_cpu) && // true
      ttwu_queue_wakelist())
        p->sched_remote_wakeup = Y;

    smp_store_release(prev->on_cpu, 0);

where both p->sched_contributes_to_load and p->sched_remote_wakeup are
in the same word, and thus the stores X and Y race (and can clobber
one another's data).

Whereas prior to commit c6e7bd7afaeb ("sched/core: Optimize ttwu()
spinning on p->on_cpu") the p->on_cpu handoff serialized access to
p->sched_remote_wakeup (just as it still does with
p->sched_contributes_to_load) that commit broke that by calling
ttwu_queue_wakelist() with p->on_cpu != 0.

However, due to

  p->XXX = X ttwu()
  schedule()   if (p->on_rq && ...) // false
    smp_mb__after_spinlock()   if (smp_load_acquire(&p->on_cpu) &&
    deactivate_task()       ttwu_queue_wakelist())
      p->on_rq = 0;         p->sched_remote_wakeup = Y;

We can be sure any 'current' store is complete and 'current' is
guaranteed asleep. Therefore we can move p->sched_remote_wakeup into
the current flags word.

Note: while the observed failure was loadavg accounting gone wrong due
to ttwu() cobbering p->sched_contributes_to_load, the reverse problem
is also possible where schedule() clobbers p->sched_remote_wakeup,
this could result in enqueue_entity() wrecking ->vruntime and causing
scheduling artifacts.

Fixes: c6e7bd7afaeb ("sched/core: Optimize ttwu() spinning on p->on_cpu")
Reported-by: Mel Gorman <mgorman@techsingularity.net>
Debugged-by: Will Deacon <will@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201117083016.GK3121392@hirez.programming.kicks-ass.net
include/linux/sched.h

index d383cf09e78f5b12947b3377510a64c72462029b..0e91b451d2a241c9d0dfd7ccf55b399797849fbb 100644 (file)
@@ -769,7 +769,6 @@ struct task_struct {
        unsigned                        sched_reset_on_fork:1;
        unsigned                        sched_contributes_to_load:1;
        unsigned                        sched_migrated:1;
-       unsigned                        sched_remote_wakeup:1;
 #ifdef CONFIG_PSI
        unsigned                        sched_psi_wake_requeue:1;
 #endif
@@ -779,6 +778,21 @@ struct task_struct {
 
        /* Unserialized, strictly 'current' */
 
+       /*
+        * This field must not be in the scheduler word above due to wakelist
+        * queueing no longer being serialized by p->on_cpu. However:
+        *
+        * p->XXX = X;                  ttwu()
+        * schedule()                     if (p->on_rq && ..) // false
+        *   smp_mb__after_spinlock();    if (smp_load_acquire(&p->on_cpu) && //true
+        *   deactivate_task()                ttwu_queue_wakelist())
+        *     p->on_rq = 0;                    p->sched_remote_wakeup = Y;
+        *
+        * guarantees all stores of 'current' are visible before
+        * ->sched_remote_wakeup gets used, so it can be in this word.
+        */
+       unsigned                        sched_remote_wakeup:1;
+
        /* Bit to tell LSMs we're in execve(): */
        unsigned                        in_execve:1;
        unsigned                        in_iowait:1;