Re: [Bugme-new] [Bug 10592] New: [old regression] performance drop for glx after vblank patch

[Andrew M. <ak...@li...>]

(switched to email.  Please respond via emailed reply-to-all, not via the
bugzilla web interface).

On Fri,  2 May 2008 10:36:33 -0700 (PDT)
bug...@bu... wrote:

> http://bugzilla.kernel.org/show_bug.cgi?id=10592
> 
>            Summary: [old regression] performance drop for glx after vblank
>                     patch
>            Product: Drivers
>            Version: 2.5
>      KernelVersion: 2.6.20-2.6.25
>           Platform: All
>         OS/Version: Linux
>               Tree: Mainline
>             Status: NEW
>           Severity: normal
>           Priority: P1
>          Component: Video(DRI)
>         AssignedTo: dri...@ke...
>         ReportedBy: bug...@fi...
> 
> 
> Latest working kernel version:2.6.19
> Earliest failing kernel version:2.6.20
> Distribution:ubuntu 8.04
> Hardware Environment:ASUS P5LD2-VM ( video i945g , Pentium D ) 
> Software Environment:
> Problem Description:
> 
> 
> This is old regression. Tested with glxgears i get this results:
> v2.6.19     _1413.112_FPS_
> v2.6.20      1280.060 FPS  
> v2.6.22      1300.116 FPS
> v2.6.24      1205.895 FPS
> v2.6.25-git* 1226.967 FPS
> 
> 
> Thirst regression was pasebly introduced by vblank patch in i915:
> commit ee2fae03d68e702866a8661fbee7ff2f2f3754d7
> Merge: e4ddc9c... f9841a8...
> Author: Linus Torvalds <tor...@wo...>
> Date:   Wed Dec 20 23:59:07 2006 -0800
> 
>     Merge branch 'drm-patches' of
> master.kernel.org:/pub/scm/linux/kernel/git/ai
> 
>     * 'drm-patches' of
> master.kernel.org:/pub/scm/linux/kernel/git/airlied/drm-2
>       drm: Stop defining pci_pretty_name
>       drm: r128: comment aligment with drm git
>       drm: make kernel context switch same as for drm git tree.
>       drm: fixup comment header style
>       drm: savage: compat fix from drm git.
>       drm: Unify radeon offset checking.
>       i915_vblank_tasklet: Try harder to avoid tearing.
>       DRM: handle pci_enable_device failure
>       drm: fix return value check
> 
> 
> I didn't get correct bisecting. Seems git-bisect can't correctly working with
> 2.9.19 tree or i did some thin really wrong, so i get this resoult:
> 
> commit e1036502e5263851259d147771226161e5ccc85a
> Author: Nicolas Pitre <ni...@ca...>
> Date:   Tue Dec 12 13:32:29 2006 -0500
> 
>     [PATCH] remove config ordering/dependency between ucb1400-ts and sound
> subsy
> 
>     Commit 2d4ba4a3b9aef95d328d74a17ae84f8d658059e2 introduced a dependency
>     that was never meant to exist when the ac97_bus.c module was created.
>     Move ac97_bus.c up the directory hierarchy to make sure it is built when
>     selected even if sound is configured out so things work as originally
>     intended.
> 
>     Signed-off-by: Nicolas Pitre <ni...@ca...>
>     Acked-by: Randy Dunlap <ran...@or...>
>     Signed-off-by: Linus Torvalds <tor...@os...>
> 

I assume that we're referring to

commit 3188a24c256bae0ed93d81d82db1f1bb6060d727
Author: =?utf-8?q?Michel_D=C3=A4nzer?= <mi...@tu...>
Date:   Mon Dec 11 18:32:27 2006 +1100

    i915_vblank_tasklet: Try harder to avoid tearing.
    
    Previously, if there were several buffer swaps scheduled for the same vertical
    blank, all but the first blit emitted stood a chance of exhibiting tearing. In
    order to avoid this, split the blits along slices of each output top to bottom.
    
    Signed-off-by: Dave Airlie <ai...@li...>

(and yup, that's how the Author: is recorded in git)

Unfrortunately that patch doesn't trivially revert, so I can't send you a
patch to test.