drm/radeon: Fix regression with suspend/resume
diff mbox

Message ID 1423773026-5941-1-git-send-email-ross.zwisler@linux.intel.com
State New, archived
Headers show

Commit Message

Ross Zwisler Feb. 12, 2015, 8:30 p.m. UTC
This patch reverts the changes made in this commit:

  deadcb36f49b ("drm/radeon: Use two-ended allocation by size, v2")

That patch caused a regression on my system where the bottom of the
screen flickers after my laptop goes thorough a suspend and resume.
This is reproducible 100% of the time.

This patch applies cleanly to v3.19, and fixes the screen flicker issue
on my system.  Here is the hardware that I'm using (from lspci):

01:00.0 VGA compatible controller: Advanced Micro Devices, Inc.
[AMD/ATI] Seymour [Radeon HD 6400M/7400M Series]

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Lauri Kasanen <cand@gmx.com>
Cc: "Christian König" <christian.koenig@amd.com>
Cc: Dave Airlie <airlied@redhat.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: dri-devel@lists.freedesktop.org
---
 drivers/gpu/drm/radeon/radeon_object.c | 11 -----------
 1 file changed, 11 deletions(-)

Comments

Michel Dänzer Feb. 13, 2015, 2:41 a.m. UTC | #1
On 13.02.2015 05:30, Ross Zwisler wrote:
> This patch reverts the changes made in this commit:
> 
>   deadcb36f49b ("drm/radeon: Use two-ended allocation by size, v2")
> 
> That patch caused a regression on my system where the bottom of the
> screen flickers after my laptop goes thorough a suspend and resume.

What kind of flicker is it? E.g. does it only affect X or also console,
does it flicker all the time or only when there is activity, what does
it look like, ...
Ross Zwisler Feb. 14, 2015, 3:55 a.m. UTC | #2
On Fri, 2015-02-13 at 11:41 +0900, Michel Dänzer wrote:
> On 13.02.2015 05:30, Ross Zwisler wrote:
> > This patch reverts the changes made in this commit:
> > 
> >   deadcb36f49b ("drm/radeon: Use two-ended allocation by size, v2")
> > 
> > That patch caused a regression on my system where the bottom of the
> > screen flickers after my laptop goes thorough a suspend and resume.
> 
> What kind of flicker is it? E.g. does it only affect X or also console,
> does it flicker all the time or only when there is activity, what does
> it look like, ...

It's kind of hard to describe it precisely, so I made a video.  :)

http://youtu.be/ESm9SMnr0do

It only affects X, not the console, and it seems to go away if you log
out back to the login manager (I'm using GDM on Fedora 20) and back into
your window manager.

I've tested with OpenBox, Gnome and KDE, and it happens in all three, so
it doesn't appear to be related to the window manager.

It does appear to flicker more if you move the mouse, but it still does
flicker occasionally if you aren't doing anything.
Deucher, Alexander Feb. 14, 2015, 6:25 a.m. UTC | #3
> -----Original Message-----

> From: Ross Zwisler [mailto:ross.zwisler@linux.intel.com]

> Sent: Friday, February 13, 2015 10:55 PM

> To: Michel Dänzer

> Cc: linux-kernel@vger.kernel.org; dri-devel@lists.freedesktop.org; Deucher,

> Alexander; Dave Airlie; Lauri Kasanen; Koenig, Christian

> Subject: Re: [PATCH] drm/radeon: Fix regression with suspend/resume

> 

> On Fri, 2015-02-13 at 11:41 +0900, Michel Dänzer wrote:

> > On 13.02.2015 05:30, Ross Zwisler wrote:

> > > This patch reverts the changes made in this commit:

> > >

> > >   deadcb36f49b ("drm/radeon: Use two-ended allocation by size, v2")

> > >

> > > That patch caused a regression on my system where the bottom of the

> > > screen flickers after my laptop goes thorough a suspend and resume.

> >

> > What kind of flicker is it? E.g. does it only affect X or also console,

> > does it flicker all the time or only when there is activity, what does

> > it look like, ...

> 

> It's kind of hard to describe it precisely, so I made a video.  :)

> 

> http://youtu.be/ESm9SMnr0do

> 

> It only affects X, not the console, and it seems to go away if you log

> out back to the login manager (I'm using GDM on Fedora 20) and back into

> your window manager.


Does a VT switch or forcing a dpms cycle (sleep 5; xset dpms force off) also fix it?  It doesn't look related to the patch in question at all.  Is the flickering 100% reproducible or does it only happen periodically?

Alex

> 

> I've tested with OpenBox, Gnome and KDE, and it happens in all three, so

> it doesn't appear to be related to the window manager.

> 

> It does appear to flicker more if you move the mouse, but it still does

> flicker occasionally if you aren't doing anything.
Ross Zwisler Feb. 17, 2015, 5:49 p.m. UTC | #4
On Sat, 2015-02-14 at 06:25 +0000, Deucher, Alexander wrote:
> > -----Original Message-----
> > From: Ross Zwisler [mailto:ross.zwisler@linux.intel.com]
> > Sent: Friday, February 13, 2015 10:55 PM
> > To: Michel Dänzer
> > Cc: linux-kernel@vger.kernel.org; dri-devel@lists.freedesktop.org; Deucher,
> > Alexander; Dave Airlie; Lauri Kasanen; Koenig, Christian
> > Subject: Re: [PATCH] drm/radeon: Fix regression with suspend/resume
> > 
> > On Fri, 2015-02-13 at 11:41 +0900, Michel Dänzer wrote:
> > > On 13.02.2015 05:30, Ross Zwisler wrote:
> > > > This patch reverts the changes made in this commit:
> > > >
> > > >   deadcb36f49b ("drm/radeon: Use two-ended allocation by size, v2")
> > > >
> > > > That patch caused a regression on my system where the bottom of the
> > > > screen flickers after my laptop goes thorough a suspend and resume.
> > >
> > > What kind of flicker is it? E.g. does it only affect X or also console,
> > > does it flicker all the time or only when there is activity, what does
> > > it look like, ...
> > 
> > It's kind of hard to describe it precisely, so I made a video.  :)
> > 
> > http://youtu.be/ESm9SMnr0do
> > 
> > It only affects X, not the console, and it seems to go away if you log
> > out back to the login manager (I'm using GDM on Fedora 20) and back into
> > your window manager.
> 
> Does a VT switch or forcing a dpms cycle (sleep 5; xset dpms force off)
> also fix it?  It doesn't look related to the patch in question at all. 
> Is the flickering 100% reproducible or does it only happen
> periodically?

From kernels 3.14 or so (when the deadcb36f49b patch was introduced)
till 3.18 it happened 100% of the time.  With 3.19 it only seems to
happen maybe 50% of the time, but is still very easily reproducible.

It's entirely possible that the patch isn't the root cause, but it just
brought out a bug somewhere else.  All I know is that I did a bisect,
and with the commit before this the issue never happens, and after this
commit it happens 100% of the time. :)  Also, reverting that commit with
3.19 makes the issue go away.

Nope, "xset dpms force off" doesn't fix it. After the screen goes black
and comes back, the flicker is still there.

- Ross
Christian König Feb. 18, 2015, 12:02 p.m. UTC | #5
On 17.02.2015 18:49, Ross Zwisler wrote:
> On Sat, 2015-02-14 at 06:25 +0000, Deucher, Alexander wrote:
>>> -----Original Message-----
>>> From: Ross Zwisler [mailto:ross.zwisler@linux.intel.com]
>>> Sent: Friday, February 13, 2015 10:55 PM
>>> To: Michel Dänzer
>>> Cc: linux-kernel@vger.kernel.org; dri-devel@lists.freedesktop.org; Deucher,
>>> Alexander; Dave Airlie; Lauri Kasanen; Koenig, Christian
>>> Subject: Re: [PATCH] drm/radeon: Fix regression with suspend/resume
>>>
>>> On Fri, 2015-02-13 at 11:41 +0900, Michel Dänzer wrote:
>>>> On 13.02.2015 05:30, Ross Zwisler wrote:
>>>>> This patch reverts the changes made in this commit:
>>>>>
>>>>>    deadcb36f49b ("drm/radeon: Use two-ended allocation by size, v2")
>>>>>
>>>>> That patch caused a regression on my system where the bottom of the
>>>>> screen flickers after my laptop goes thorough a suspend and resume.
>>>> What kind of flicker is it? E.g. does it only affect X or also console,
>>>> does it flicker all the time or only when there is activity, what does
>>>> it look like, ...
>>> It's kind of hard to describe it precisely, so I made a video.  :)
>>>
>>> http://youtu.be/ESm9SMnr0do
>>>
>>> It only affects X, not the console, and it seems to go away if you log
>>> out back to the login manager (I'm using GDM on Fedora 20) and back into
>>> your window manager.
>> Does a VT switch or forcing a dpms cycle (sleep 5; xset dpms force off)
>> also fix it?  It doesn't look related to the patch in question at all.
>> Is the flickering 100% reproducible or does it only happen
>> periodically?
>  From kernels 3.14 or so (when the deadcb36f49b patch was introduced)
> till 3.18 it happened 100% of the time.  With 3.19 it only seems to
> happen maybe 50% of the time, but is still very easily reproducible.

Well, what the patch does is just changing where buffers are placed in 
memory. E.g. now we place the buffer at the end of memory as well.

So I can imagine at least three possible causes for the issues you see:
1. We haven't implemented all buffer placement restrictions correctly 
and without the patch everything just works fine by coincident.
2. Something is overwriting the buffer at it's new location. 
@Alex&Michel: Didn't we had a similar problem internally recently? Or 
was that just for APUs?
3. One of the memory chips on your hardware is faulty and without the 
patch the we just don't use the affected region (rather unlikely).

For testing could you try to limit the amount of VRAM used? E.g. give 
radeon.vramlimit=256 as kernel commandline to limit the VRAM to the 
first 256MB.

Regards,
Christian.

>
> It's entirely possible that the patch isn't the root cause, but it just
> brought out a bug somewhere else.  All I know is that I did a bisect,
> and with the commit before this the issue never happens, and after this
> commit it happens 100% of the time. :)  Also, reverting that commit with
> 3.19 makes the issue go away.
>
> Nope, "xset dpms force off" doesn't fix it. After the screen goes black
> and comes back, the flicker is still there.
>
> - Ross
>
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
Deucher, Alexander Feb. 18, 2015, 2:13 p.m. UTC | #6
> -----Original Message-----

> From: Christian König [mailto:deathsimple@vodafone.de]

> Sent: Wednesday, February 18, 2015 7:02 AM

> To: Ross Zwisler; Deucher, Alexander

> Cc: Michel Dänzer; linux-kernel@vger.kernel.org; dri-

> devel@lists.freedesktop.org; Dave Airlie; Lauri Kasanen; Koenig, Christian

> Subject: Re: [PATCH] drm/radeon: Fix regression with suspend/resume

> 

> On 17.02.2015 18:49, Ross Zwisler wrote:

> > On Sat, 2015-02-14 at 06:25 +0000, Deucher, Alexander wrote:

> >>> -----Original Message-----

> >>> From: Ross Zwisler [mailto:ross.zwisler@linux.intel.com]

> >>> Sent: Friday, February 13, 2015 10:55 PM

> >>> To: Michel Dänzer

> >>> Cc: linux-kernel@vger.kernel.org; dri-devel@lists.freedesktop.org;

> Deucher,

> >>> Alexander; Dave Airlie; Lauri Kasanen; Koenig, Christian

> >>> Subject: Re: [PATCH] drm/radeon: Fix regression with suspend/resume

> >>>

> >>> On Fri, 2015-02-13 at 11:41 +0900, Michel Dänzer wrote:

> >>>> On 13.02.2015 05:30, Ross Zwisler wrote:

> >>>>> This patch reverts the changes made in this commit:

> >>>>>

> >>>>>    deadcb36f49b ("drm/radeon: Use two-ended allocation by size, v2")

> >>>>>

> >>>>> That patch caused a regression on my system where the bottom of

> the

> >>>>> screen flickers after my laptop goes thorough a suspend and resume.

> >>>> What kind of flicker is it? E.g. does it only affect X or also console,

> >>>> does it flicker all the time or only when there is activity, what does

> >>>> it look like, ...

> >>> It's kind of hard to describe it precisely, so I made a video.  :)

> >>>

> >>> http://youtu.be/ESm9SMnr0do

> >>>

> >>> It only affects X, not the console, and it seems to go away if you log

> >>> out back to the login manager (I'm using GDM on Fedora 20) and back

> into

> >>> your window manager.

> >> Does a VT switch or forcing a dpms cycle (sleep 5; xset dpms force off)

> >> also fix it?  It doesn't look related to the patch in question at all.

> >> Is the flickering 100% reproducible or does it only happen

> >> periodically?

> >  From kernels 3.14 or so (when the deadcb36f49b patch was introduced)

> > till 3.18 it happened 100% of the time.  With 3.19 it only seems to

> > happen maybe 50% of the time, but is still very easily reproducible.

> 

> Well, what the patch does is just changing where buffers are placed in

> memory. E.g. now we place the buffer at the end of memory as well.

> 

> So I can imagine at least three possible causes for the issues you see:

> 1. We haven't implemented all buffer placement restrictions correctly

> and without the patch everything just works fine by coincident.

> 2. Something is overwriting the buffer at it's new location.

> @Alex&Michel: Didn't we had a similar problem internally recently? Or

> was that just for APUs?


I think that was just two engines trying to use the same memory location for writeback since we had not properly allocated a writeback slot.

What's odd to me is that the region is actively flickering, it looks almost like a display timing issue.  I would expect static corruption unless something else is actively writing to the region.

Alex

> 3. One of the memory chips on your hardware is faulty and without the

> patch the we just don't use the affected region (rather unlikely).

> 

> For testing could you try to limit the amount of VRAM used? E.g. give

> radeon.vramlimit=256 as kernel commandline to limit the VRAM to the

> first 256MB.

> 

> Regards,

> Christian.

> 

> >

> > It's entirely possible that the patch isn't the root cause, but it just

> > brought out a bug somewhere else.  All I know is that I did a bisect,

> > and with the commit before this the issue never happens, and after this

> > commit it happens 100% of the time. :)  Also, reverting that commit with

> > 3.19 makes the issue go away.

> >

> > Nope, "xset dpms force off" doesn't fix it. After the screen goes black

> > and comes back, the flicker is still there.

> >

> > - Ross

> >

> > _______________________________________________

> > dri-devel mailing list

> > dri-devel@lists.freedesktop.org

> > http://lists.freedesktop.org/mailman/listinfo/dri-devel
Ross Zwisler Feb. 21, 2015, 11:18 p.m. UTC | #7
On Wed, 2015-02-18 at 13:02 +0100, Christian König wrote:
> Well, what the patch does is just changing where buffers are placed in 
> memory. E.g. now we place the buffer at the end of memory as well.
> 
> So I can imagine at least three possible causes for the issues you see:
> 1. We haven't implemented all buffer placement restrictions correctly 
> and without the patch everything just works fine by coincident.
> 2. Something is overwriting the buffer at it's new location. 
> @Alex&Michel: Didn't we had a similar problem internally recently? Or 
> was that just for APUs?
> 3. One of the memory chips on your hardware is faulty and without the 
> patch the we just don't use the affected region (rather unlikely).
> 
> For testing could you try to limit the amount of VRAM used? E.g. give 
> radeon.vramlimit=256 as kernel commandline to limit the VRAM to the 
> first 256MB.

Tried with the kernel parameter radeon.vramlimit=256, and it seemed to have
the exact same behavior.  The flicker was still there, same size, same
frequency.

Thanks,
- Ross
Christian König Feb. 22, 2015, 10:44 a.m. UTC | #8
On 22.02.2015 00:18, Ross Zwisler wrote:
> On Wed, 2015-02-18 at 13:02 +0100, Christian König wrote:
>> Well, what the patch does is just changing where buffers are placed in
>> memory. E.g. now we place the buffer at the end of memory as well.
>>
>> So I can imagine at least three possible causes for the issues you see:
>> 1. We haven't implemented all buffer placement restrictions correctly
>> and without the patch everything just works fine by coincident.
>> 2. Something is overwriting the buffer at it's new location.
>> @Alex&Michel: Didn't we had a similar problem internally recently? Or
>> was that just for APUs?
>> 3. One of the memory chips on your hardware is faulty and without the
>> patch the we just don't use the affected region (rather unlikely).
>>
>> For testing could you try to limit the amount of VRAM used? E.g. give
>> radeon.vramlimit=256 as kernel commandline to limit the VRAM to the
>> first 256MB.
> Tried with the kernel parameter radeon.vramlimit=256, and it seemed to have
> the exact same behavior.  The flicker was still there, same size, same
> frequency.

Well how much memory does the card have in the first place? 256MB was 
just an example.

Please provide a complete dmesg output with that paramater.

Regards,
Christian.

>
> Thanks,
> - Ross
>
>

Patch
diff mbox

diff --git a/drivers/gpu/drm/radeon/radeon_object.c b/drivers/gpu/drm/radeon/radeon_object.c
index 86fc564..dea1baf 100644
--- a/drivers/gpu/drm/radeon/radeon_object.c
+++ b/drivers/gpu/drm/radeon/radeon_object.c
@@ -173,17 +173,6 @@  void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
 		else
 			rbo->placements[i].lpfn = 0;
 	}
-
-	/*
-	 * Use two-ended allocation depending on the buffer size to
-	 * improve fragmentation quality.
-	 * 512kb was measured as the most optimal number.
-	 */
-	if (rbo->tbo.mem.size > 512 * 1024) {
-		for (i = 0; i < c; i++) {
-			rbo->placements[i].flags |= TTM_PL_FLAG_TOPDOWN;
-		}
-	}
 }
 
 int radeon_bo_create(struct radeon_device *rdev,