YouTube causes 755s to crash hard, and 745s to look funny

[geofft]

A user reported that cluster computers would crash after a minute or two of watching a YouTube? video. We reproduced this on his account on opus, which is an OptiPlex? 755 running 10.04.2 debathena-workstation i686: the screen goes blank and the machine no longer pings.

[geofft]

Jon and I can reproduce, although it's not immediate and may require some clicking around to trigger. Thanks to the magic of  netconsole, I have a backtrace from opus:

[  837.584013] [drm:radeon_fence_wait] *ERROR* fence(f6861be0:0x00000EE1) 508ms timeout going to reset GPU
[  837.584021] radeon 0000:01:00.0: GPU softreset 
[  837.584025] radeon 0000:01:00.0:   R_008010_GRBM_STATUS=0xE53024E0
[  837.584029] radeon 0000:01:00.0:   R_008014_GRBM_STATUS2=0x00110003
[  837.584033] radeon 0000:01:00.0:   R_000E50_SRBM_STATUS=0x200000C0
[  837.584042] radeon 0000:01:00.0:   R_008020_GRBM_SOFT_RESET=0x00007FEE
[  837.584096] radeon 0000:01:00.0: R_008020_GRBM_SOFT_RESET=0x00000001
[  837.584154] radeon 0000:01:00.0:   R_000E60_SRBM_SOFT_RESET=0x00000402

[jdreed]

This bug manifests itself differently on 745s (RV516 cards), in that YouTube? works, but the video is purple tinted and there's some ghosting.

I cannot reproduce on a 760. Other people should test 760s and 780s please.

This seems to be very specifically YouTube?'s fault, in that TechTV works just fine, as does Hulu.

[jdreed]

Interestingly enough, videos which require an ad first on youtube work just fine right up until the video itself is loaded. (The ad displays fine).

[jdreed]

This is entirely Flash's fault. Downgrading to 10.0.45.2 results in normal behavior.

[jdreed]

Bob points us to  http://ubuntuforums.org/showthread.php?p=10519535 and  http://www.webupd8.org/2011/03/fix-pinkred-youtube-videos-bug-using.html however the /etc/adobe/mms.cfg method documented there did not appear to work for me (the "hardware acceleration" checkbox was still checked, even after clearing various caches).

[rbasch]

 LP:729307 seems relevant, and the Adobe bug  https://bugs.adobe.com/jira/browse/FP-6327.

I think the suggestion of disabling OverrideGPUValidation in /etc/adobe/mms.cfg won't work generally, as that option (as the name suggests) only bypasses Flash's check of the card/driver before enabling acceleration. I don't see any way of disabling acceleration via mms.cfg.

[kaduk]

We had a user trigger this in m66-080 (hardware appears to be a 780 from remote reports), reportedly losing three hours of work on a paper.
Do we want to pull back to an older flash?

[kaduk]

Upstream has pushed flash 10.2.153, and with flashplugin-nonfree-10.2.153.1ubuntu0.10.04.1 installed on opus.mit.edu, I could not produce the crash.

(We can probably pull the gmotd?)

[jdreed]

with flashplugin-nonfree-10.2.153.1ubuntu0.10.04.1 installed on opus.mit.edu, I could not produce the crash.

Confirmed. I'll ask Ops to pull the motd.