sched/fair: Fix cfs_rq avg tracking underflow
authorPeter Zijlstra <peterz@infradead.org>
Thu, 16 Jun 2016 08:50:40 +0000 (10:50 +0200)
committerIngo Molnar <mingo@kernel.org>
Mon, 20 Jun 2016 09:29:09 +0000 (11:29 +0200)
As per commit:

  b7fa30c9cc48 ("sched/fair: Fix post_init_entity_util_avg() serialization")

> the code generated from update_cfs_rq_load_avg():
>
>  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 <update_blocked_averages+0xc0>
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.

So I _should_ have looked at other unserialized users of ->load_avg,
but alas. Luckily nikbor reported a similar /0 from task_h_load() which
instantly triggered recollection of this here problem.

Aside from the intermediate value hitting memory and causing problems,
there's another problem: the underflow detection relies on the signed
bit. This reduces the effective width of the variables, IOW its
effectively the same as having these variables be of signed type.

This patch changes to a different means of unsigned underflow
detection to not rely on the signed bit. This allows the variables to
use the 'full' unsigned range. And it does so with explicit LOAD -
STORE to ensure any intermediate value will never be visible in
memory, allowing these unserialized loads.

Note: GCC generates crap code for this, might warrant a look later.

Note2: I say 'full' above, if we end up at U*_MAX we'll still explode;
       maybe we should do clamping on add too.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
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: kernel@kyup.com
Cc: morten.rasmussen@arm.com
Cc: pjt@google.com
Cc: steve.muckle@linaro.org
Fixes: 9d89c257dfb9 ("sched/fair: Rewrite runnable load and utilization average tracking")
Link: http://lkml.kernel.org/r/20160617091948.GJ30927@twins.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/sched/fair.c

index a2348deab7a30a0500b0e7facd7584a199d0aec5..2ae68f0e3bf5a4919c069c9ce1873d6a1817f89a 100644 (file)
@@ -2904,6 +2904,23 @@ static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
        }
 }
 
+/*
+ * Unsigned subtract and clamp on underflow.
+ *
+ * Explicitly do a load-store to ensure the intermediate value never hits
+ * memory. This allows lockless observations without ever seeing the negative
+ * values.
+ */
+#define sub_positive(_ptr, _val) do {                          \
+       typeof(_ptr) ptr = (_ptr);                              \
+       typeof(*ptr) val = (_val);                              \
+       typeof(*ptr) res, var = READ_ONCE(*ptr);                \
+       res = var - val;                                        \
+       if (res > var)                                          \
+               res = 0;                                        \
+       WRITE_ONCE(*ptr, res);                                  \
+} while (0)
+
 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
 static inline int
 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
@@ -2913,15 +2930,15 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
 
        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);
+               sub_positive(&sa->load_avg, r);
+               sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
                removed_load = 1;
        }
 
        if (atomic_long_read(&cfs_rq->removed_util_avg)) {
                long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
-               sa->util_avg = max_t(long, sa->util_avg - r, 0);
-               sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
+               sub_positive(&sa->util_avg, r);
+               sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
                removed_util = 1;
        }
 
@@ -2994,10 +3011,10 @@ static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *s
                          &se->avg, se->on_rq * scale_load_down(se->load.weight),
                          cfs_rq->curr == se, NULL);
 
-       cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
-       cfs_rq->avg.load_sum = max_t(s64,  cfs_rq->avg.load_sum - se->avg.load_sum, 0);
-       cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
-       cfs_rq->avg.util_sum = max_t(s32,  cfs_rq->avg.util_sum - se->avg.util_sum, 0);
+       sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
+       sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
+       sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
+       sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
 
        cfs_rq_util_change(cfs_rq);
 }