https://bugs.freedesktop.org/show_bug.cgi?id=33867
Summary: [bisected] Graphics corruption related to pageflip ioctl support in 2.6.38-rc* Product: DRI Version: DRI CVS Platform: x86-64 (AMD64) OS/Version: Linux (All) Status: NEW Severity: normal Priority: medium Component: DRM/Radeon AssignedTo: dri-devel@lists.freedesktop.org ReportedBy: dawitbro@sbcglobal.net
Created an attachment (id=42887) --> (https://bugs.freedesktop.org/attachment.cgi?id=42887) Commits not present in 2.6.37 applied to local branch
I am troubleshooting some graphics corruption I noticed when testing the post-2.6.37 commits from drm-airlied git (drm-fixes in this case). I was trying to produce a kernel as close as possible to the latest stable release (since 2.6.38 is very early in the rc stage) with all of the newest radeon-related features.
This is a preliminary report, and may turn out to be invalid, because the kernel I am using is actually a local branch from v2.6.37 with only the commits from drm-airlied relevant to my hardware individually cherry-picked. I have bisected the problem down to a specific commit, but if I made any errors during the cherry-pick process then this report is useless. I plan to confirm that the problem is real tomorrow, by building directly from the drm-airlied/drm-fixes tree (which was the most up-to-date tree I could find today).
As I will report below, there is some similarity with
https://bugs.freedesktop.org/show_bug.cgi?id=33515
and that is the only reason I decided to report my findings before I am really sure there is a problem. Hopefully this will help Michel and other devs save some time and trouble if my guess (that my bug is related) turns out to be correct. ----------------------
OK, here is what I have done so far:
1. I made a local git branch based on v2.6.37. 2. I identified commits I wanted from drm-airlied/drm-fixes (Feb. 2, 2011) 3. Because of GPU lockups recently cured by Alex Deucher, I picked one particular commit first (1e644d6d, "drm/radeon/kms: re-emit full context state for evergreen blits"); without this, testing would be pointless since my GPU would just lock up when trying and 3D app. 4. I then cherry-picked the rest of the commits I had chosen, in order from oldest to newest (according to 'git log'). See attached file, 'applied-cherry-picks.txt', if interested in specifics.
Everything actually seemed to be working fine; I only happened to notice a small glitch. I use a locally-built game called 'prboom-plus' with my original Ultimate DOOM WAD file to test Radeon support (in kernels, Mesa, xf86-video-ati, etc.). When a game begins, there is a melt-down animation which transitions into the new game; there is a similar (I would have guessed identical before) melting effect when the player is killed but hits the space bar to restart the game. In this second melting effect, after being killed, the "melted" part of the screen was all black, a clear regression.
Before today's testing, I had been using a straight 2.6.37 kernel with the patch from Alex I mentioned in item 3 above. With that kernel, both melting effects work fine. That seems to rule out a problem with xorg-server, mesa, libdrm, or xf86-video-ati.
I decided to try bisecting the issue. I built kernels from the first and last commits listed in the attached file: 1e644d6d and dca0d612. The former was "good" and the latter "bad". The bisect jumped to 204663c4 and 18007401 next -- both were "bad". The next jump, 65705962, caused 'prboom-plus' to hang during the second kind of melting; I was able to SSH into the machine and 'kill -9' the game.
Since there was a series of TTM related changes in that series of commits, I used 'git bisect skip' until I was presented with a commit before or after the TTM series; I tested the kernels (all hung on the second melt), but told git to skip them; the ones I pretended to skip were eba67093, 95762c2b, and 702adba2. I was offered 147666fb, and it was "bad" but did not hang during the second kind of melt.
Continuing, the bisect took me back to 2357cbe5, which hung X so that it could not be killed with 'kill -9'; I had to reboot via SSH.
I skipped ecf7ace9 and 68c4fa31 without building kernels. The bisect then went to d6ea8886 and b6724405, which hung (so I pretended to skip them). I skipped 96726fe50 without building. The bisect moved to 3e4ea742, which hung, and 6f34be50, which also hung.
Interestingly, f5a80209 was perfectly OK. This meant that the last commit I tested was the first bad commit. I had been skipping the commits causing hangs because I expected the hang to be a temporary problem resolved somewhere in the middle of the list; I believed this because of the fact the first and last commits in my list did _not_ produce hanging kernels. Of course, I was actually building most of the kernels that I "skipped", so I knew they were "bad" even if git did not.
To sum up. the first three cherry-picks (applied to v2.6.37) were fine:
1e644d6dce366a7bae22484f60133b61ba322911 drm/radeon/kms: re-emit full context state for evergreen blits
27641c3f003e7f3b6585c01d8a788883603eb262 drm/vblank: Add support for precise vblank timestamping.
f5a8020903932624cf020dc72455a10a3e005087 drm/kms/radeon: Add support for precise vblank timestamping.
The commit which introduced the hangs -- and these happened entirely predictably and consistently: when the second kind of "melt" in 'prboom-plus' occurs -- was
6f34be50bd1bdd2ff3c955940e033a80d05f248a drm/radeon/kms: add pageflip ioctl support (v3)
After a series of several TTM-related commits, the hanging was resolved and the second kind of melt would display in all black instead, beginning at this commit:
147666fb3b93b8c484f562da33a37f886ddff768 drm/radeon: Use the ttm execbuf utilities
As I mentioned above, these results are preliminary; I will try builds directly from drm-airlied tomorrow since we are snowed-in here in Michigan, and I don't have to work tomorrow.
My problems with hangs sound very similar to f.d.o. bug #33515. This may also be related to #33418, since that user (Erdem) is using 2.6.38-rc1.
Hardware: Radeon HD 5750 (Evergreen JUNIPER)
Software: kernel 2.6.37 + commits as described above libdrm 2.4.23 xorg-server 1.9.3.902 xf86-video-ati 6.13.99 (git commit 3dc28c86 of Jan. 27) mesa 7.11.0 (git commit 11b15c4d of Jan 30), r600g