1277450 - Crashes in mozilla::DisplayItemScrollClip::PickAncestor on youtube.com videos, on June 1 builds, on bad AMD CPUs

Status: RESOLVED WORKSFORME

(Core :: Web Painting, defect)

Description

[David Baron :dbaron: (⌚️UTC-5, no longer working on Mozilla)]

This bug was filed from the Socorro interface and is 
report bp-5c8a4f7b-75e7-40ca-8141-5a2b62160601.
=============================================================

This is the new #1 topcrash in Aurora, starting in today's builds (June 1).  Clearly multiple users.  I suspect related to something being backported, though it could be related to a site change.

Almost all of the crashes have a URL field that is a youtube.com video.

Comment 1

[David Baron :dbaron: (⌚️UTC-5, no longer working on Mozilla)]

Regression window:
https://hg.mozilla.org/releases/mozilla-aurora/pushloghtml?fromchange=3c7a33fe0b6472bf47264c2390c5bc19f1a6dce4&tochange=3cc4a98864ad31001292b38c97df30962b9e94d0

Comment 2

[Timothy Nikkel (:tnikkel)]

(In reply to David Baron :dbaron: ⌚️UTC-7 (review requests must explain patch) from comment #1)
> Regression window:
> https://hg.mozilla.org/releases/mozilla-aurora/
> pushloghtml?fromchange=3c7a33fe0b6472bf47264c2390c5bc19f1a6dce4&tochange=3cc4
> a98864ad31001292b38c97df30962b9e94d0

Nothing in there seems plausible as having caused this.

Comment 3

[David Baron :dbaron: (⌚️UTC-5, no longer working on Mozilla)]

The regression window doesn't make sense to me, so needinfoing all the not-quite-making-sense possibilities.

Comment 4

[David Baron :dbaron: (⌚️UTC-5, no longer working on Mozilla)]

Interesting characteristics of these crashes:
 * adapter vendor ID is always 0x1002 (AMD, I think)
 * adapter device ID is either 0x9802, 0x9803, 0x9804, or 0x9807
 * crash address is always 0x408b6051

Comment 5

[Chris Pearce [:cpearce (Not reading bugmail)]]

This regression range includes us blacklisting some ATI graphics driver versions for D3D11 DXVA. But we should have fallen back to D3D9 hardware acceleration for these users. Of the the handful of crash reports that I looked into, not all were actively playing video, and those that were playing video weren't using hardware accelerated video decoding; they were using WebM software decoding.

So it's hard to see how my changeset was to blame.

Comment 6

[David Baron :dbaron: (⌚️UTC-5, no longer working on Mozilla)]

Sorry, should have seen it sooner; this is the AMD CPU bug.

Comment 7

[David Baron :dbaron: (⌚️UTC-5, no longer working on Mozilla)]

(Which implies that it will go away in tomorrow's builds.)

The giveaway, from the report in comment 0, was that the instruction pointer was not instruction-aligned.  Then I remembered to aggregate on "cpu info", which I should have done before.