873799 - I think we need an additional barrier before __sync_lock_test_and_set in Atomics.h

Status: RESOLVED FIXED

(Core :: MFBT, defect)

Description

[Justin Lebar (not reading bugmail)]

__sync_lock_test_and_set is only an acquire barrier.  I think we need another barrier.  But you tell me if this is wrong.

Jeff, let me know if you're not ok with Nathan reviewing these patches.

Comment 1

[Justin Lebar (not reading bugmail)]

Attachment: Patch, v1

Comment 2

[Jeff Walden [:Waldo]]

This was brought up on bug 732043 comment 55, and responded to in bug 732043 comment 60.  I'm not sure I grok this quite fully enough to say anything one way or the other here, fully confidently, but it's totally fine with me playing it safe.  :-)

Comment 3

[Jeff Walden [:Waldo]]

...which is to say, the fix sort of makes sense if I squint, but I might not actually be squinting fully correctly here, so I take that with a bit of a grain of salt.

Comment 4

[Justin Lebar (not reading bugmail)]

I don't grok this well enough to say with confidence, either.  I'm happy to defer to froydnj.  But at the very least, a comment would be nice.

Comment 5

[Justin Lebar (not reading bugmail)]

(I also feel like sacrificing perf for strict correctness here is probably the right trade-off, but again, I'm not so comfortable with all this.)

Comment 6

[Nathan Froyd [:froydnj]]

Comment on attachment 751403 [details] [diff] [review]
Patch, v1

Review of attachment 751403 [details] [diff] [review]:
-----------------------------------------------------------------

It doesn't matter one way or the other to me; I can see arguments both directions.  If we start replacing PR_ATOMIC_SET and people start yelling about performance, I guess we can revisit this.

::: mfbt/Atomics.h
@@ +375,5 @@
> +
> +      // __sync_lock_test_and_set is an "acquire" barrier, so loads and stores
> +      // can't be moved above it (into the space between the beforeStore() and
> +      // the __sync_lock_test_and_set).
> +

It makes more sense to me to move this comment prior to the call to beforeStore, something like:

// Since __sync_lock_test_and_set is only an acquire barrier, we need this
// barrier to ensure that loads and stores don't get moved after the call to it.

Comment 7

[Nathan Froyd [:froydnj]]

Comment on attachment 751403 [details] [diff] [review]
Patch, v1

Review of attachment 751403 [details] [diff] [review]:
-----------------------------------------------------------------

Here, let's actually r+ that.

Comment 8

[Justin Lebar (not reading bugmail)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/9d74599c1c00

Comment 9

[Phil Ringnalda (:philor)]

https://hg.mozilla.org/mozilla-central/rev/9d74599c1c00