Bug 93147

Created attachment 120187 [details]
dmesg from kernel built at "bad" commit

Building Linux from drm-next or drm-fixes gives me a kernel which causes what appears to be stuttering/hesitations in test applications (particularly severe in 'prboom-plus') which were not present in Linux 4.3.


    HARDWARE
    GPU:  Radeon HD 7850 (Pitcairn)
    CPU:  64-bit AMD FX-8320E Eight-Core Processor
chipset:  AMD 990FX
 mboard:  ASUS M5A99FX PRO R2.0

    SOFTWARE
xf86-video-ati:  7.6.99+ (git master at commit 10b7c3de, Nov 17 2015)
   xorg-server:  1.18.0+ (git master at commit 51984ddd, Nov 18 2015)
          mesa:  11.1.0+ (git master at commit d8c26969, Nov 20 2015)
        libdrm:  2.4.65+ (git master at commit c3deddd9, Nov 10 2015)
        kernel:  drm-fixes (at commit 2d591ab1, Nov 20 2015)
  distribution:  Debian unstable


Reproducing steps:

I apply future Radeon DRM and core DRM code to latest stable Linux kernels, and
I use several applications to test the results.  I use a less demanding game,
'prboom-plus', as a kind of canary; I use more demanding games like 'alien-arena' and 'torcs' as well.  All 3 of the games mentioned show a kind of stuttering or hesitation if using the current git HEAD of drm-next or drm-fixes, but 'prboom-plus' is affected to the point of being extremely annoying, almost unplayable.

To reproduce, boot Linux 4.3 and play a few moments of 'prboom-plus'; then boot
a kernel built from drm-next or drm-fixes, then play 'prboom-plus' again.  The
effects are quite pronounced and obvious.  (The effects on 'alien-arena' and
'torcs' are much more subtle, and might easily go unnoticed, though someone
looking for a difference should easily be able to notice.)  None of the games crash, but they are simply affected (to varying degrees) by a kind of stuttering lag caused by the commit identified below by bisection.


Additional info:

I have reported this against DRI/Radeon but problem commit is touching code in
'drivers/gpu/drm/drm_irq.c', which does not seem limited to affecting only Radeon.  (I am not a developer, so any comment I make about code can probably be ignored.)  I would appreciate if the developers reading this would consider whether I have reported the bug against the wrong product and/or component, and make changes as appropriate!

In 'xorg.conf' I usually have "SwapbuffersWait" disabled with a stanza like this:

    Section "Device"
      Identifier	"Configured Video Device"
      Driver		"radeon"
      Option		"DRI"              "3"
      Option		"SwapbuffersWait"  "off"
    EndSection

Changing the "SwapbuffersWait" setting has no effect on the symptoms of this bug, however.


Bisection:

[I test future Radeon code in advance by cherry picking from drm-next and drm-fixes changesets which touch Radeon DRM and core DRM.  Because of the likelihood of error on my part, I never report bugs on one of these Frankenstein kernels of my own making.  If I bisect a bug, I confirm that the same buggy behavior is present upstream before reporting it.]

I first noticed this problem in early October, when I had made a local branch from Linux 4.2.3 and cherry picked from DRM 4.3 and DRM 4.4 (up to cf648305).  I noticed the bug described earlier immediately.  I didn't have much time to troubleshoot, and assumed the error was my own, so I made a second local branch from 4.2.3 and only cherry-picked from DRM 4.3 (up to 30c64664, my previously working list of commits).

After Linux 4.3 was released, I tried again.  I made a local branch based on 4.3 and applied DRM 4.4 up to commit 5481c8fb.  The process went very smooth, but the buggy behavior of my test programs (particularly 'prboom-plus') was back again.  I decided to bisect my 4.3 + DRM 4.4 tree to identify where things were going wrong.

The HEAD of my tree was clearly "bad" so I had a place to start.  I built Linux 4.3 without any cherry picks, and it was "good".  After 6 more builds, 'git bisect' pointed at the following commit:

    commit 4dfd64862ff852df7b1198d667dda778715ee88f
    Author: Ville Syrjälä <ville.syrjala@linux.intel.com>
    Date:   Mon Sep 14 22:43:51 2015 +0300

        drm: Use vblank timestamps to guesstimate how many vblanks were missed

Since the problem could still have been only in my Frankenstein local tree, I built kernels from drm-next at commits 4dfd6486 and 4dfd6486^ (= 1b2eb710).  I found that 4dfd6486 showed the buggy behavior, while 1b2eb710 did not.


I did notice that a regression against this same commit was noticed and apparently fixed in this commit:

    commit fa4270d8e0257b4b76f11baa2866f4313d29aaf5
    Author: Ville Syrjälä <ville.syrjala@linux.intel.com>
    Date:   Wed Sep 30 19:21:34 2015 +0300

        drm: Don't zero vblank timestamps from the irq handler

        [...]    
        This fixes a regression from
        4dfd64862ff852df drm: Use vblank timestamps to guesstimate how many
        vblanks were missed

This commit was present in my local Frankenstein tree, as well as both drm-next
and drm-fixes, and (obviously) does not fix the bug I am reporting.  The problem
commit no doubt had other effects that were fixed by fa4270d8, but the stuttering I am seeing was not fixed by it.