Re: 回复: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Christian König]

Hi Monk,

yeah, that's what I can certainly agree on.

My primary concern is that I'm not convinced that we don't get problems 
at other places if we just add another band aid.


We already had this back and forth multiple times now and while we are 
currently under time pressure we will be under even more time pressure 
when a customer is running into other issues and we are still circling 
around the same fundamental problem.


Regards,
Christian.

Am 30.03.21 um 05:10 schrieb Liu, Monk:

[AMD Official Use Only - Internal Distribution Only]

Hi Christian,

We don't need to debate on the design's topic, each of us have our own opinion, 
it is hard to persuade others sometimes, again with more and more features and 
requirements it is pretty normal that an old design need to
Refine and or even rework to satisfy all those needs, so I'm not trying to 
argue with you that we don't need a better rework, that's also pleasure me .

In the moment, the more important thing I care is the solution because SRIOV 
project still try best to put all changes into upstreaming tree, we don't want 
to fork another tree unless no choice ...

Let's have a sync in another thread

Thanks for you help on this

--
Monk Liu | Cloud-GPU Core team
--

-Original Message-
From: Koenig, Christian 
Sent: Friday, March 26, 2021 10:51 PM
To: Liu, Monk ; Zhang, Jack (Jian) ; Grodzovsky, Andrey 
; Christian König ; dri-devel@lists.freedesktop.org; 
amd-...@lists.freedesktop.org; Deng, Emily ; Rob Herring ; Tomeu Vizoso 
; Steven Price 
Cc: Zhang, Andy ; Jiang, Jerry (SW) 
Subject: Re: 回复: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

Hi Monk,

I can't disagree more.

The fundamental problem here is that we have pushed a design without validating 
if it really fits into the concepts the Linux kernel mandates here.

My mistake was that I haven't pushed back hard enough on the initial design 
resulting in numerous cycles of trying to save the design while band aiding the 
flaws which became obvious after a while.

I haven't counted them but I think we are now already had over 10 patches which 
try to work around lifetime issues of the job object because I wasn't able to 
properly explain why this isn't going to work like this.

Because of this I will hard reject any attempt to band aid this issue even more 
which isn't starting over again with a design which looks like it is going to 
work.

Regards,
Christian.

Am 26.03.21 um 12:21 schrieb Liu, Monk:

[AMD Official Use Only - Internal Distribution Only]

Hi Christian

This is not correct or correct perspective, any design comes with its
pros and cons, otherwise it wouldn't comes to kernel tree in the very
beginning , it is just with time passed we have more and more
requirement and feature need to implement And those new requirement
drags many new solution or idea, and some idea you prefer need to
based on a new infrastructure, that's all

I don't why the job "should be" or not "should be" in the scheduler,
honestly speaking I can argue with you that the "scheduler" and the TDR feature which 
invented by AMD developer "should" never escalate to drm layer at all and by that 
assumption Those vendor's compatibilities headache right now won't happen at all.

Let's just focus on the issue so far.

The solution Andrey and Jack doing right now looks good to me, and it
can solve our problems without introducing regression from a surface
look, but it is fine if you need a neat solution,  since we have our
project pressure (which we always have) Either we implement the first
version with Jack's patch and do the revise in another series of
patches (that also my initial suggestion) or we rework anything you
mentioned, but since looks to me you are from time to time asking
people to rework Something in the stage that people already have a
solution, which frustrated people a lot,

I would like you do prepare a solution for us, which solves our
headaches ...  I really don't want to see you asked Jack to rework again and 
again If you are out of bandwidth or no interest in doing this ,please at least 
make your solution/proposal very detail and clear, jack told me he couldn't 
understand your point here.

Thanks very much, and please understand our painful here

/Monk


-邮件原件-
发件人: Koenig, Christian 
发送时间: 2021年3月26日 17:06
收件人: Zhang, Jack (Jian) ; Grodzovsky, Andrey
; Christian König
; dri-devel@lists.freedesktop.org;
amd-...@lists.freedesktop.org; Liu, Monk ; Deng,
Emily ; Rob Herring ; Tomeu
Vizoso ; Steven Price

主题: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid
memleak

Hi guys,

Am 26.03.21 um 03:23 schrieb Zhang, Jack (Jian):

[AMD Official Use Only - Internal Distribution Only]

Hi, Andrey,


how u handle non guilty singnaled jobs in drm_sched_stop, currently
looks like you don't call put for them and just explicitly free
them as

RE: 回复: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Liu, Monk]

[AMD Official Use Only - Internal Distribution Only]

Hi Christian,

We don't need to debate on the design's topic, each of us have our own opinion, 
it is hard to persuade others sometimes, again with more and more features and 
requirements it is pretty normal that an old design need to
Refine and or even rework to satisfy all those needs, so I'm not trying to 
argue with you that we don't need a better rework, that's also pleasure me .

In the moment, the more important thing I care is the solution because SRIOV 
project still try best to put all changes into upstreaming tree, we don't want 
to fork another tree unless no choice ... 

Let's have a sync in another thread 

Thanks for you help on this

--
Monk Liu | Cloud-GPU Core team
--

-Original Message-
From: Koenig, Christian  
Sent: Friday, March 26, 2021 10:51 PM
To: Liu, Monk ; Zhang, Jack (Jian) ; 
Grodzovsky, Andrey ; Christian König 
; dri-devel@lists.freedesktop.org; 
amd-...@lists.freedesktop.org; Deng, Emily ; Rob Herring 
; Tomeu Vizoso ; Steven Price 

Cc: Zhang, Andy ; Jiang, Jerry (SW) 
Subject: Re: 回复: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

Hi Monk,

I can't disagree more.

The fundamental problem here is that we have pushed a design without validating 
if it really fits into the concepts the Linux kernel mandates here.

My mistake was that I haven't pushed back hard enough on the initial design 
resulting in numerous cycles of trying to save the design while band aiding the 
flaws which became obvious after a while.

I haven't counted them but I think we are now already had over 10 patches which 
try to work around lifetime issues of the job object because I wasn't able to 
properly explain why this isn't going to work like this.

Because of this I will hard reject any attempt to band aid this issue even more 
which isn't starting over again with a design which looks like it is going to 
work.

Regards,
Christian.

Am 26.03.21 um 12:21 schrieb Liu, Monk:
> [AMD Official Use Only - Internal Distribution Only]
>
> Hi Christian
>
> This is not correct or correct perspective, any design comes with its 
> pros and cons, otherwise it wouldn't comes to kernel tree in the very 
> beginning , it is just with time passed we have more and more 
> requirement and feature need to implement And those new requirement 
> drags many new solution or idea, and some idea you prefer need to 
> based on a new infrastructure, that's all
>
> I don't why the job "should be" or not "should be" in the scheduler, 
> honestly speaking I can argue with you that the "scheduler" and the TDR 
> feature which invented by AMD developer "should" never escalate to drm layer 
> at all and by that assumption Those vendor's compatibilities headache right 
> now won't happen at all.
>
> Let's just focus on the issue so far.
>
> The solution Andrey and Jack doing right now looks good to me, and it 
> can solve our problems without introducing regression from a surface 
> look, but it is fine if you need a neat solution,  since we have our 
> project pressure (which we always have) Either we implement the first 
> version with Jack's patch and do the revise in another series of 
> patches (that also my initial suggestion) or we rework anything you 
> mentioned, but since looks to me you are from time to time asking 
> people to rework Something in the stage that people already have a 
> solution, which frustrated people a lot,
>
> I would like you do prepare a solution for us, which solves our 
> headaches ...  I really don't want to see you asked Jack to rework again and 
> again If you are out of bandwidth or no interest in doing this ,please at 
> least make your solution/proposal very detail and clear, jack told me he 
> couldn't understand your point here.
>
> Thanks very much, and please understand our painful here
>
> /Monk
>
>
> -邮件原件-
> 发件人: Koenig, Christian 
> 发送时间: 2021年3月26日 17:06
> 收件人: Zhang, Jack (Jian) ; Grodzovsky, Andrey 
> ; Christian König 
> ; dri-devel@lists.freedesktop.org; 
> amd-...@lists.freedesktop.org; Liu, Monk ; Deng, 
> Emily ; Rob Herring ; Tomeu 
> Vizoso ; Steven Price 
> 
> 主题: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid 
> memleak
>
> Hi guys,
>
> Am 26.03.21 um 03:23 schrieb Zhang, Jack (Jian):
>> [AMD Official Use Only - Internal Distribution Only]
>>
>> Hi, Andrey,
>>
>>>> how u handle non guilty singnaled jobs in drm_sched_stop, currently 
>>>> looks like you don't call put for them and just explicitly free 
>>>> them as before
>> Good point, I missed that place. Will cover that in my next patch.
>>
>>>> Also 

Re: 回复: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Christian König]

Hi Monk,

I can't disagree more.

The fundamental problem here is that we have pushed a design without 
validating if it really fits into the concepts the Linux kernel mandates 
here.


My mistake was that I haven't pushed back hard enough on the initial 
design resulting in numerous cycles of trying to save the design while 
band aiding the flaws which became obvious after a while.


I haven't counted them but I think we are now already had over 10 
patches which try to work around lifetime issues of the job object 
because I wasn't able to properly explain why this isn't going to work 
like this.


Because of this I will hard reject any attempt to band aid this issue 
even more which isn't starting over again with a design which looks like 
it is going to work.


Regards,
Christian.

Am 26.03.21 um 12:21 schrieb Liu, Monk:

[AMD Official Use Only - Internal Distribution Only]

Hi Christian

This is not correct or correct perspective, any design comes with its pros and 
cons, otherwise it wouldn't comes to kernel tree in the very beginning , it is 
just with time passed we have more and more requirement and feature need to 
implement
And those new requirement drags many new solution or idea, and some idea you 
prefer need to based on a new infrastructure, that's all

I don't why the job "should be" or not "should be" in the scheduler, honestly speaking I can argue 
with you that the "scheduler" and the TDR feature which invented by AMD developer "should" never 
escalate to drm layer at all and by that assumption
Those vendor's compatibilities headache right now won't happen at all.

Let's just focus on the issue so far.

The solution Andrey and Jack doing right now looks good to me, and it can solve 
our problems without introducing regression from a surface look, but it is fine 
if you need a neat solution,  since we have our project pressure (which we 
always have)
Either we implement the first version with Jack's patch and do the revise in 
another series of patches (that also my initial suggestion) or we rework 
anything you mentioned, but since looks to me you are from time to time asking 
people to rework
Something in the stage that people already have a solution, which frustrated 
people a lot,

I would like you do prepare a solution for us, which solves our headaches ...  
I really don't want to see you asked Jack to rework again and again
If you are out of bandwidth or no interest in doing this ,please at least make 
your solution/proposal very detail and clear, jack told me he couldn't 
understand your point here.

Thanks very much, and please understand our painful here

/Monk


-邮件原件-
发件人: Koenig, Christian 
发送时间: 2021年3月26日 17:06
收件人: Zhang, Jack (Jian) ; Grodzovsky, Andrey ; Christian König 
; dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Liu, Monk 
; Deng, Emily ; Rob Herring ; Tomeu Vizoso 
; Steven Price 
主题: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

Hi guys,

Am 26.03.21 um 03:23 schrieb Zhang, Jack (Jian):

[AMD Official Use Only - Internal Distribution Only]

Hi, Andrey,


how u handle non guilty singnaled jobs in drm_sched_stop, currently
looks like you don't call put for them and just explicitly free them
as before

Good point, I missed that place. Will cover that in my next patch.


Also sched->free_guilty seems useless with the new approach.

Yes, I agree.


Do we even need the cleanup mechanism at drm_sched_get_cleanup_job with this 
approach...

I am not quite sure about that for now, let me think about this topic today.

Hi, Christian,
should I add a fence and get/put to that fence rather than using an explicit 
refcount?
And another concerns?

well let me re-iterate:

For the scheduler the job is just a temporary data structure used for 
scheduling the IBs to the hardware.

While pushing the job to the hardware we get a fence structure in return which 
represents the IBs executing on the hardware.

Unfortunately we have applied a design where the job structure is rather used 
for re-submitting the jobs to the hardware after a GPU reset and karma handling 
etc etc...

All that shouldn't have been pushed into the scheduler into the first place and 
we should now work on getting this cleaned up rather than making it an even 
bigger mess by applying halve backed solutions.

So in my opinion adding a reference count to the job is going into the 
completely wrong directly. What we should rather do is to fix the incorrect 
design decision to use jobs as vehicle in the scheduler for reset handling.

To fix this I suggest the following approach:
1. We add a pointer from the drm_sched_fence back to the drm_sched_job.
2. Instead of keeping the job around in the scheduler we keep the fence around. 
For this I suggest to replace the pending_list with a ring buffer.
3. The timedout_job callback is replaced with a timeout_fence callback.
4. The free_job callback is completed dropped.

回复: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Liu, Monk]

[AMD Official Use Only - Internal Distribution Only]

Hi Christian

This is not correct or correct perspective, any design comes with its pros and 
cons, otherwise it wouldn't comes to kernel tree in the very beginning , it is 
just with time passed we have more and more requirement and feature need to 
implement
And those new requirement drags many new solution or idea, and some idea you 
prefer need to based on a new infrastructure, that's all

I don't why the job "should be" or not "should be" in the scheduler, honestly 
speaking I can argue with you that the "scheduler" and the TDR feature which 
invented by AMD developer "should" never escalate to drm layer at all and by 
that assumption
Those vendor's compatibilities headache right now won't happen at all.

Let's just focus on the issue so far.

The solution Andrey and Jack doing right now looks good to me, and it can solve 
our problems without introducing regression from a surface look, but it is fine 
if you need a neat solution,  since we have our project pressure (which we 
always have)
Either we implement the first version with Jack's patch and do the revise in 
another series of patches (that also my initial suggestion) or we rework 
anything you mentioned, but since looks to me you are from time to time asking 
people to rework
Something in the stage that people already have a solution, which frustrated 
people a lot,

I would like you do prepare a solution for us, which solves our headaches ...  
I really don't want to see you asked Jack to rework again and again
If you are out of bandwidth or no interest in doing this ,please at least make 
your solution/proposal very detail and clear, jack told me he couldn't 
understand your point here.

Thanks very much, and please understand our painful here

/Monk


-邮件原件-
发件人: Koenig, Christian 
发送时间: 2021年3月26日 17:06
收件人: Zhang, Jack (Jian) ; Grodzovsky, Andrey 
; Christian König 
; dri-devel@lists.freedesktop.org; 
amd-...@lists.freedesktop.org; Liu, Monk ; Deng, Emily 
; Rob Herring ; Tomeu Vizoso 
; Steven Price 
主题: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

Hi guys,

Am 26.03.21 um 03:23 schrieb Zhang, Jack (Jian):
> [AMD Official Use Only - Internal Distribution Only]
>
> Hi, Andrey,
>
>>> how u handle non guilty singnaled jobs in drm_sched_stop, currently
>>> looks like you don't call put for them and just explicitly free them
>>> as before
> Good point, I missed that place. Will cover that in my next patch.
>
>>> Also sched->free_guilty seems useless with the new approach.
> Yes, I agree.
>
>>> Do we even need the cleanup mechanism at drm_sched_get_cleanup_job with 
>>> this approach...
> I am not quite sure about that for now, let me think about this topic today.
>
> Hi, Christian,
> should I add a fence and get/put to that fence rather than using an explicit 
> refcount?
> And another concerns?

well let me re-iterate:

For the scheduler the job is just a temporary data structure used for 
scheduling the IBs to the hardware.

While pushing the job to the hardware we get a fence structure in return which 
represents the IBs executing on the hardware.

Unfortunately we have applied a design where the job structure is rather used 
for re-submitting the jobs to the hardware after a GPU reset and karma handling 
etc etc...

All that shouldn't have been pushed into the scheduler into the first place and 
we should now work on getting this cleaned up rather than making it an even 
bigger mess by applying halve backed solutions.

So in my opinion adding a reference count to the job is going into the 
completely wrong directly. What we should rather do is to fix the incorrect 
design decision to use jobs as vehicle in the scheduler for reset handling.

To fix this I suggest the following approach:
1. We add a pointer from the drm_sched_fence back to the drm_sched_job.
2. Instead of keeping the job around in the scheduler we keep the fence around. 
For this I suggest to replace the pending_list with a ring buffer.
3. The timedout_job callback is replaced with a timeout_fence callback.
4. The free_job callback is completed dropped. Job lifetime is now handled in 
the driver, not the scheduler.

Regards,
Christian.

>
> Thanks,
> Jack
>
> -Original Message-
> From: Grodzovsky, Andrey 
> Sent: Friday, March 26, 2021 12:32 AM
> To: Zhang, Jack (Jian) ; Christian König
> ; dri-devel@lists.freedesktop.org;
> amd-...@lists.freedesktop.org; Koenig, Christian
> ; Liu, Monk ; Deng, Emily
> ; Rob Herring ; Tomeu Vizoso
> ; Steven Price 
> Subject: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid
> memleak
>
> There are a few issues here like - how u handle non guilty singnaled jobs in 
> drm_sched_stop, currently looks like you don't call put for them and

Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Steven Price]

On 26/03/2021 02:04, Zhang, Jack (Jian) wrote:

[AMD Official Use Only - Internal Distribution Only]

Hi, Steve,

Thank you for your detailed comments.

But currently the patch is not finalized.
We found some potential race condition even with this patch. The solution is 
under discussion and hopefully we could find an ideal one.
After that, I will start to consider other drm-driver if it will influence 
other drivers(except for amdgpu).


No problem. Please keep me CC'd, the suggestion of using reference 
counts may be beneficial for Panfrost as we already build a reference 
count on top of struct drm_sched_job. So there may be scope for cleaning 
up Panfrost afterwards even if your work doesn't directly affect it.


Thanks,

Steve


Best,
Jack

-Original Message-
From: Steven Price 
Sent: Monday, March 22, 2021 11:29 PM
To: Zhang, Jack (Jian) ; dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; 
Koenig, Christian ; Grodzovsky, Andrey ; Liu, Monk 
; Deng, Emily ; Rob Herring ; Tomeu Vizoso 

Subject: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

On 15/03/2021 05:23, Zhang, Jack (Jian) wrote:

[AMD Public Use]

Hi, Rob/Tomeu/Steven,

Would you please help to review this patch for panfrost driver?

Thanks,
Jack Zhang

-Original Message-
From: Jack Zhang 
Sent: Monday, March 15, 2021 1:21 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org;
Koenig, Christian ; Grodzovsky, Andrey
; Liu, Monk ; Deng, Emily

Cc: Zhang, Jack (Jian) 
Subject: [PATCH v3] drm/scheduler re-insert Bailing job to avoid
memleak

re-insert Bailing jobs to avoid memory leak.

V2: move re-insert step to drm/scheduler logic
V3: add panfrost's return value for bailing jobs in case it hits the
memleak issue.


This commit message could do with some work - it's really hard to decipher what 
the actual problem you're solving is.



Signed-off-by: Jack Zhang 
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
   drivers/gpu/drm/amd/amdgpu/amdgpu_job.c| 8 ++--
   drivers/gpu/drm/panfrost/panfrost_job.c| 4 ++--
   drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
   include/drm/gpu_scheduler.h| 1 +
   5 files changed, 19 insertions(+), 6 deletions(-)


[...]

diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c
b/drivers/gpu/drm/panfrost/panfrost_job.c
index 6003cfeb1322..e2cb4f32dae1 100644
--- a/drivers/gpu/drm/panfrost/panfrost_job.c
+++ b/drivers/gpu/drm/panfrost/panfrost_job.c
@@ -444,7 +444,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
* spurious. Bail out.
*/
   if (dma_fence_is_signaled(job->done_fence))
-return DRM_GPU_SCHED_STAT_NOMINAL;
+return DRM_GPU_SCHED_STAT_BAILING;

   dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, 
head=0x%x, tail=0x%x, sched_job=%p",
   js,
@@ -456,7 +456,7 @@ static enum drm_gpu_sched_stat
panfrost_job_timedout(struct drm_sched_job

   /* Scheduler is already stopped, nothing to do. */
   if (!panfrost_scheduler_stop(>js->queue[js], sched_job))
-return DRM_GPU_SCHED_STAT_NOMINAL;
+return DRM_GPU_SCHED_STAT_BAILING;

   /* Schedule a reset if there's no reset in progress. */
   if (!atomic_xchg(>reset.pending, 1))


This looks correct to me - in these two cases drm_sched_stop() is not called on 
the sched_job, so it looks like currently the job will be leaked.


diff --git a/drivers/gpu/drm/scheduler/sched_main.c
b/drivers/gpu/drm/scheduler/sched_main.c
index 92d8de24d0a1..a44f621fb5c4 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -314,6 +314,7 @@ static void drm_sched_job_timedout(struct work_struct *work)
   {
   struct drm_gpu_scheduler *sched;
   struct drm_sched_job *job;
+int ret;

   sched = container_of(work, struct drm_gpu_scheduler,
work_tdr.work);

@@ -331,8 +332,13 @@ static void drm_sched_job_timedout(struct work_struct 
*work)
   list_del_init(>list);
   spin_unlock(>job_list_lock);

-job->sched->ops->timedout_job(job);
+ret = job->sched->ops->timedout_job(job);

+if (ret == DRM_GPU_SCHED_STAT_BAILING) {
+spin_lock(>job_list_lock);
+list_add(>node, >ring_mirror_list);
+spin_unlock(>job_list_lock);
+}


I think we could really do with a comment somewhere explaining what "bailing" 
means in this context. For the Panfrost case we have two cases:

   * The GPU job actually finished while the timeout code was running 
(done_fence is signalled).

   * The GPU is already in the process of being reset (Panfrost has multiple 
queues, so mostly like a bad job in another queue).

I'm also not convinced that (for Panfrost) it makes sense to be adding the jobs 
back to the list. For the first case above clearly the job could just be freed 
(it's complete). The second case is more interesting and Panfrost currently 
doesn't handle this well. In theory the driver could try to rescue the job 
('soft sto

Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Christian König]

Hi guys,

Am 26.03.21 um 03:23 schrieb Zhang, Jack (Jian):

[AMD Official Use Only - Internal Distribution Only]

Hi, Andrey,


how u handle non guilty singnaled jobs in drm_sched_stop, currently looks like 
you don't call put for them and just explicitly free them as before

Good point, I missed that place. Will cover that in my next patch.


Also sched->free_guilty seems useless with the new approach.

Yes, I agree.


Do we even need the cleanup mechanism at drm_sched_get_cleanup_job with this 
approach...

I am not quite sure about that for now, let me think about this topic today.

Hi, Christian,
should I add a fence and get/put to that fence rather than using an explicit 
refcount?
And another concerns?


well let me re-iterate:

For the scheduler the job is just a temporary data structure used for 
scheduling the IBs to the hardware.


While pushing the job to the hardware we get a fence structure in return 
which represents the IBs executing on the hardware.


Unfortunately we have applied a design where the job structure is rather 
used for re-submitting the jobs to the hardware after a GPU reset and 
karma handling etc etc...


All that shouldn't have been pushed into the scheduler into the first 
place and we should now work on getting this cleaned up rather than 
making it an even bigger mess by applying halve backed solutions.


So in my opinion adding a reference count to the job is going into the 
completely wrong directly. What we should rather do is to fix the 
incorrect design decision to use jobs as vehicle in the scheduler for 
reset handling.


To fix this I suggest the following approach:
1. We add a pointer from the drm_sched_fence back to the drm_sched_job.
2. Instead of keeping the job around in the scheduler we keep the fence 
around. For this I suggest to replace the pending_list with a ring buffer.

3. The timedout_job callback is replaced with a timeout_fence callback.
4. The free_job callback is completed dropped. Job lifetime is now 
handled in the driver, not the scheduler.


Regards,
Christian.



Thanks,
Jack

-Original Message-
From: Grodzovsky, Andrey 
Sent: Friday, March 26, 2021 12:32 AM
To: Zhang, Jack (Jian) ; Christian König ; 
dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, Christian ; Liu, Monk 
; Deng, Emily ; Rob Herring ; Tomeu Vizoso 
; Steven Price 
Subject: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

There are a few issues here like - how u handle non guilty singnaled jobs in 
drm_sched_stop, currently looks like you don't call put for them and just 
explicitly free them as before. Also sched->free_guilty seems useless with the 
new approach. Do we even need the cleanup mechanism at drm_sched_get_cleanup_job 
with this approach...

But first - We need Christian to express his opinion on this since I think he 
opposed refcounting jobs and that we should concentrate on fences instead.

Christian - can you chime in here ?

Andrey

On 2021-03-25 5:51 a.m., Zhang, Jack (Jian) wrote:

[AMD Official Use Only - Internal Distribution Only]


Hi, Andrey

Thank you for your good opinions.

I literally agree with you that the refcount could solve the
get_clean_up_up cocurrent job gracefully, and no need to re-insert the

job back anymore.

I quickly made a draft for this idea as follows:

How do you like it? I will start implement to it after I got your
acknowledge.

Thanks,

Jack

+void drm_job_get(struct drm_sched_job *s_job)

+{

+   kref_get(_job->refcount);

+}

+

+void drm_job_do_release(struct kref *ref)

+{

+   struct drm_sched_job *s_job;

+   struct drm_gpu_scheduler *sched;

+

+   s_job = container_of(ref, struct drm_sched_job, refcount);

+   sched = s_job->sched;

+   sched->ops->free_job(s_job);

+}

+

+void drm_job_put(struct drm_sched_job *s_job)

+{

+   kref_put(_job->refcount, drm_job_do_release);

+}

+

static void drm_sched_job_begin(struct drm_sched_job *s_job)

{

  struct drm_gpu_scheduler *sched = s_job->sched;

+   kref_init(_job->refcount);

+   drm_job_get(s_job);

  spin_lock(>job_list_lock);

  list_add_tail(_job->node, >ring_mirror_list);

  drm_sched_start_timeout(sched);

@@ -294,17 +316,16 @@ static void drm_sched_job_timedout(struct
work_struct *work)

   * drm_sched_cleanup_jobs. It will be reinserted back
after sched->thread

   * is parked at which point it's safe.

   */

-   list_del_init(>node);

+   drm_job_get(job);

  spin_unlock(>job_list_lock);

  job->sched->ops->timedout_job(job);

-

+   drm_job_put(job);

  /*

   * Guilty job did complete and hence needs to be
manually removed

   * See drm_sched_stop doc.

   */

  if (sched->free_guilty)

RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Zhang, Jack (Jian)]

[AMD Official Use Only - Internal Distribution Only]

Hi, Andrey,

>>how u handle non guilty singnaled jobs in drm_sched_stop, currently looks 
>>like you don't call put for them and just explicitly free them as before
Good point, I missed that place. Will cover that in my next patch.

>>Also sched->free_guilty seems useless with the new approach.
Yes, I agree.

>>Do we even need the cleanup mechanism at drm_sched_get_cleanup_job with this 
>>approach...
I am not quite sure about that for now, let me think about this topic today.

Hi, Christian,
should I add a fence and get/put to that fence rather than using an explicit 
refcount?
And another concerns?

Thanks,
Jack

-Original Message-
From: Grodzovsky, Andrey 
Sent: Friday, March 26, 2021 12:32 AM
To: Zhang, Jack (Jian) ; Christian König 
; dri-devel@lists.freedesktop.org; 
amd-...@lists.freedesktop.org; Koenig, Christian ; 
Liu, Monk ; Deng, Emily ; Rob Herring 
; Tomeu Vizoso ; Steven Price 

Subject: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

There are a few issues here like - how u handle non guilty singnaled jobs in 
drm_sched_stop, currently looks like you don't call put for them and just 
explicitly free them as before. Also sched->free_guilty seems useless with the 
new approach. Do we even need the cleanup mechanism at 
drm_sched_get_cleanup_job with this approach...

But first - We need Christian to express his opinion on this since I think he 
opposed refcounting jobs and that we should concentrate on fences instead.

Christian - can you chime in here ?

Andrey

On 2021-03-25 5:51 a.m., Zhang, Jack (Jian) wrote:
> [AMD Official Use Only - Internal Distribution Only]
>
>
> Hi, Andrey
>
> Thank you for your good opinions.
>
> I literally agree with you that the refcount could solve the
> get_clean_up_up cocurrent job gracefully, and no need to re-insert the
>
> job back anymore.
>
> I quickly made a draft for this idea as follows:
>
> How do you like it? I will start implement to it after I got your
> acknowledge.
>
> Thanks,
>
> Jack
>
> +void drm_job_get(struct drm_sched_job *s_job)
>
> +{
>
> +   kref_get(_job->refcount);
>
> +}
>
> +
>
> +void drm_job_do_release(struct kref *ref)
>
> +{
>
> +   struct drm_sched_job *s_job;
>
> +   struct drm_gpu_scheduler *sched;
>
> +
>
> +   s_job = container_of(ref, struct drm_sched_job, refcount);
>
> +   sched = s_job->sched;
>
> +   sched->ops->free_job(s_job);
>
> +}
>
> +
>
> +void drm_job_put(struct drm_sched_job *s_job)
>
> +{
>
> +   kref_put(_job->refcount, drm_job_do_release);
>
> +}
>
> +
>
> static void drm_sched_job_begin(struct drm_sched_job *s_job)
>
> {
>
>  struct drm_gpu_scheduler *sched = s_job->sched;
>
> +   kref_init(_job->refcount);
>
> +   drm_job_get(s_job);
>
>  spin_lock(>job_list_lock);
>
>  list_add_tail(_job->node, >ring_mirror_list);
>
>  drm_sched_start_timeout(sched);
>
> @@ -294,17 +316,16 @@ static void drm_sched_job_timedout(struct
> work_struct *work)
>
>   * drm_sched_cleanup_jobs. It will be reinserted back
> after sched->thread
>
>   * is parked at which point it's safe.
>
>   */
>
> -   list_del_init(>node);
>
> +   drm_job_get(job);
>
>  spin_unlock(>job_list_lock);
>
>  job->sched->ops->timedout_job(job);
>
> -
>
> +   drm_job_put(job);
>
>  /*
>
>   * Guilty job did complete and hence needs to be
> manually removed
>
>   * See drm_sched_stop doc.
>
>   */
>
>  if (sched->free_guilty) {
>
> -   job->sched->ops->free_job(job);
>
>  sched->free_guilty = false;
>
>  }
>
>  } else {
>
> @@ -355,20 +376,6 @@ void drm_sched_stop(struct drm_gpu_scheduler
> *sched, struct drm_sched_job *bad)
>
> -   /*
>
> -* Reinsert back the bad job here - now it's safe as
>
> -* drm_sched_get_cleanup_job cannot race against us and
> release the
>
> -* bad job at this point - we parked (waited for) any in
> progress
>
> -* (earlier) cleanups and drm_sched_get_cleanup_job will not
> be called
>
> -* now until the scheduler thread is unparked.
>
> -*/
>
> -   if (bad && bad->sched == sched)
>
> -   /*
>
> -* Add

RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Zhang, Jack (Jian)]

[AMD Official Use Only - Internal Distribution Only]

Hi, Steve,

Thank you for your detailed comments.

But currently the patch is not finalized.
We found some potential race condition even with this patch. The solution is 
under discussion and hopefully we could find an ideal one.
After that, I will start to consider other drm-driver if it will influence 
other drivers(except for amdgpu).

Best,
Jack

-Original Message-
From: Steven Price 
Sent: Monday, March 22, 2021 11:29 PM
To: Zhang, Jack (Jian) ; dri-devel@lists.freedesktop.org; 
amd-...@lists.freedesktop.org; Koenig, Christian ; 
Grodzovsky, Andrey ; Liu, Monk ; 
Deng, Emily ; Rob Herring ; Tomeu Vizoso 

Subject: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

On 15/03/2021 05:23, Zhang, Jack (Jian) wrote:
> [AMD Public Use]
>
> Hi, Rob/Tomeu/Steven,
>
> Would you please help to review this patch for panfrost driver?
>
> Thanks,
> Jack Zhang
>
> -Original Message-
> From: Jack Zhang 
> Sent: Monday, March 15, 2021 1:21 PM
> To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org;
> Koenig, Christian ; Grodzovsky, Andrey
> ; Liu, Monk ; Deng, Emily
> 
> Cc: Zhang, Jack (Jian) 
> Subject: [PATCH v3] drm/scheduler re-insert Bailing job to avoid
> memleak
>
> re-insert Bailing jobs to avoid memory leak.
>
> V2: move re-insert step to drm/scheduler logic
> V3: add panfrost's return value for bailing jobs in case it hits the
> memleak issue.

This commit message could do with some work - it's really hard to decipher what 
the actual problem you're solving is.

>
> Signed-off-by: Jack Zhang 
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
>   drivers/gpu/drm/amd/amdgpu/amdgpu_job.c| 8 ++--
>   drivers/gpu/drm/panfrost/panfrost_job.c| 4 ++--
>   drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
>   include/drm/gpu_scheduler.h| 1 +
>   5 files changed, 19 insertions(+), 6 deletions(-)
>
[...]
> diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c
> b/drivers/gpu/drm/panfrost/panfrost_job.c
> index 6003cfeb1322..e2cb4f32dae1 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_job.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_job.c
> @@ -444,7 +444,7 @@ static enum drm_gpu_sched_stat 
> panfrost_job_timedout(struct drm_sched_job
>* spurious. Bail out.
>*/
>   if (dma_fence_is_signaled(job->done_fence))
> -return DRM_GPU_SCHED_STAT_NOMINAL;
> +return DRM_GPU_SCHED_STAT_BAILING;
>
>   dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, 
> head=0x%x, tail=0x%x, sched_job=%p",
>   js,
> @@ -456,7 +456,7 @@ static enum drm_gpu_sched_stat
> panfrost_job_timedout(struct drm_sched_job
>
>   /* Scheduler is already stopped, nothing to do. */
>   if (!panfrost_scheduler_stop(>js->queue[js], sched_job))
> -return DRM_GPU_SCHED_STAT_NOMINAL;
> +return DRM_GPU_SCHED_STAT_BAILING;
>
>   /* Schedule a reset if there's no reset in progress. */
>   if (!atomic_xchg(>reset.pending, 1))

This looks correct to me - in these two cases drm_sched_stop() is not called on 
the sched_job, so it looks like currently the job will be leaked.

> diff --git a/drivers/gpu/drm/scheduler/sched_main.c
> b/drivers/gpu/drm/scheduler/sched_main.c
> index 92d8de24d0a1..a44f621fb5c4 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -314,6 +314,7 @@ static void drm_sched_job_timedout(struct work_struct 
> *work)
>   {
>   struct drm_gpu_scheduler *sched;
>   struct drm_sched_job *job;
> +int ret;
>
>   sched = container_of(work, struct drm_gpu_scheduler,
> work_tdr.work);
>
> @@ -331,8 +332,13 @@ static void drm_sched_job_timedout(struct work_struct 
> *work)
>   list_del_init(>list);
>   spin_unlock(>job_list_lock);
>
> -job->sched->ops->timedout_job(job);
> +ret = job->sched->ops->timedout_job(job);
>
> +if (ret == DRM_GPU_SCHED_STAT_BAILING) {
> +spin_lock(>job_list_lock);
> +list_add(>node, >ring_mirror_list);
> +spin_unlock(>job_list_lock);
> +}

I think we could really do with a comment somewhere explaining what "bailing" 
means in this context. For the Panfrost case we have two cases:

  * The GPU job actually finished while the timeout code was running 
(done_fence is signalled).

  * The GPU is already in the process of being reset (Panfrost has multiple 
queues, so mostly like a bad job in another queue).

I'm also not convinced that (for Panfrost) it makes sense to be adding the jobs 
back to the list. For the first case above clearly the job could just be freed 
(it's complete). The second case is more interesting and Panfrost currently 
doesn't handle this well. In theory the dri

Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Andrey Grodzovsky]

 Deng, Emily 
; Rob Herring ; Tomeu Vizoso 
; Steven Price 
*Subject:* Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid 
memleak


On 2021-03-18 6:41 a.m., Zhang, Jack (Jian) wrote:

[AMD Official Use Only - Internal Distribution Only]

Hi, Andrey

Let me summarize the background of this patch:

In TDR resubmit step “amdgpu_device_recheck_guilty_jobs,

It will submit first jobs of each ring and do guilty job re-check.

At that point, We had to make sure each job is in the mirror list(or
re-inserted back already).

But we found the current code never re-insert the job to mirror list
in the 2^nd , 3^rd job_timeout thread(Bailing TDR thread).

This not only will cause memleak of the bailing jobs. What’s more
important, the 1^st tdr thread can never iterate the bailing job and
set its guilty status to a correct status.

Therefore, we had to re-insert the job(or even not delete node) for
bailing job.

For the above V3 patch, the racing condition in my mind is:

we cannot make sure all bailing jobs are finished before we do
amdgpu_device_recheck_guilty_jobs.

Yes,that race i missed - so you say that for 2nd, baling thread who 
extracted the job, even if he reinsert it right away back after driver 
callback return DRM_GPU_SCHED_STAT_BAILING, there is small time slot 
where the job is not in mirror list and so the 1st TDR might miss it and 
not find that  2nd job is the actual guilty job, right ? But, still this 
job will get back into mirror list, and since it's really the bad job, 
it will never signal completion and so on the next timeout cycle it will 
be caught (of course there is a starvation scenario here if more TDRs 
kick in and it bails out again but this is really unlikely).


Based on this insight, I think we have two options to solve this issue:

 1. Skip delete node in tdr thread2, thread3, 4 … (using mutex or
atomic variable)
 2. Re-insert back bailing job, and meanwhile use semaphore in each
tdr thread to keep the sequence as expected and ensure each job
is in the mirror list when do resubmit step.

For Option1, logic is simpler and we need only one global atomic
variable:

What do you think about this plan?

Option1 should look like the following logic:

+static atomic_t in_reset; //a global atomic var for
synchronization

static void drm_sched_process_job(struct dma_fence *f, struct
dma_fence_cb *cb);

  /**

@@ -295,6 +296,12 @@ static void drm_sched_job_timedout(struct
work_struct *work)

  * drm_sched_cleanup_jobs. It will be reinserted
back after sched->thread

  * is parked at which point it's safe.

  */

+   if (atomic_cmpxchg(_reset, 0, 1) != 0) {  //skip
delete node if it’s thead1,2,3,….

+   spin_unlock(>job_list_lock);

+   drm_sched_start_timeout(sched);

+   return;

+   }

+

     list_del_init(>node);

     spin_unlock(>job_list_lock);

@@ -320,6 +327,7 @@ static void drm_sched_job_timedout(struct
work_struct *work)

     spin_lock(>job_list_lock);

     drm_sched_start_timeout(sched);

     spin_unlock(>job_list_lock);

+   atomic_set(_reset, 0); //reset in_reset when the first
thread finished tdr

}

Technically looks like it should work as you don't access the job 
pointer any longer and so no risk that if signaled it will be freed by 
drm_sched_get_cleanup_job but,you can't just use one global variable an 
by this bailing from TDR when different drivers run their TDR threads in 
parallel, and even for amdgpu, if devices in different XGMI hives or 2 
independent devices in non XGMI setup. There should be defined some kind 
of GPU reset group structure on drm_scheduler level for which this 
variable would be used.


P.S I wonder why we can't just ref-count the job so that even if 
drm_sched_get_cleanup_job would delete it before we had a chance to stop 
the scheduler thread, we wouldn't crash. This would avoid all the dance 
with deletion and reinsertion.


Andrey

Thanks,

Jack

*From:* amd-gfx 
<mailto:amd-gfx-boun...@lists.freedesktop.org> *On Behalf Of *Zhang,
Jack (Jian)
*Sent:* Wednesday, March 17, 2021 11:11 PM
*To:* Christian König 
<mailto:ckoenig.leichtzumer...@gmail.com>;
dri-devel@lists.freedesktop.org
<mailto:dri-devel@lists.freedesktop.org>;
amd-...@lists.freedesktop.org
<mailto:amd-...@lists.freedesktop.org>; Koenig, Christian
 <mailto:christian.koe...@amd.com>; Liu,
Monk  <mailto:monk@amd.com>; Deng, Emily
 <mailto:emily.d...@amd.com>; Rob Herring
 <mailto:r...@kernel.org>; Tomeu Vizoso
 <mailto:tomeu.viz...@col

RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Zhang, Jack (Jian)]

[AMD Official Use Only - Internal Distribution Only]

Hi, Andrey

Thank you for your good opinions.
I literally agree with you that the refcount could solve the get_clean_up_up 
cocurrent job gracefully, and no need to re-insert the
job back anymore.

I quickly made a draft for this idea as follows:
How do you like it? I will start implement to it after I got your acknowledge.

Thanks,
Jack

+void drm_job_get(struct drm_sched_job *s_job)
+{
+   kref_get(_job->refcount);
+}
+
+void drm_job_do_release(struct kref *ref)
+{
+   struct drm_sched_job *s_job;
+   struct drm_gpu_scheduler *sched;
+
+   s_job = container_of(ref, struct drm_sched_job, refcount);
+   sched = s_job->sched;
+   sched->ops->free_job(s_job);
+}
+
+void drm_job_put(struct drm_sched_job *s_job)
+{
+   kref_put(_job->refcount, drm_job_do_release);
+}
+
static void drm_sched_job_begin(struct drm_sched_job *s_job)
{
struct drm_gpu_scheduler *sched = s_job->sched;
+   kref_init(_job->refcount);
+   drm_job_get(s_job);
spin_lock(>job_list_lock);
list_add_tail(_job->node, >ring_mirror_list);
drm_sched_start_timeout(sched);
@@ -294,17 +316,16 @@ static void drm_sched_job_timedout(struct work_struct 
*work)
 * drm_sched_cleanup_jobs. It will be reinserted back after 
sched->thread
 * is parked at which point it's safe.
 */
-   list_del_init(>node);
+   drm_job_get(job);
spin_unlock(>job_list_lock);
job->sched->ops->timedout_job(job);
-
+   drm_job_put(job);
/*
 * Guilty job did complete and hence needs to be manually 
removed
 * See drm_sched_stop doc.
 */
if (sched->free_guilty) {
-   job->sched->ops->free_job(job);
sched->free_guilty = false;
}
} else {
@@ -355,20 +376,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, 
struct drm_sched_job *bad)
-   /*
-* Reinsert back the bad job here - now it's safe as
-* drm_sched_get_cleanup_job cannot race against us and release the
-* bad job at this point - we parked (waited for) any in progress
-* (earlier) cleanups and drm_sched_get_cleanup_job will not be called
-* now until the scheduler thread is unparked.
-*/
-   if (bad && bad->sched == sched)
-   /*
-* Add at the head of the queue to reflect it was the earliest
-* job extracted.
-*/
-   list_add(>node, >ring_mirror_list);
-
/*
 * Iterate the job list from later to  earlier one and either deactive
 * their HW callbacks or remove them from mirror list if they already
@@ -774,7 +781,7 @@ static int drm_sched_main(void *param)
 kthread_should_stop());
if (cleanup_job) {
-   sched->ops->free_job(cleanup_job);
+   drm_job_put(cleanup_job);
/* queue timeout for next job */
drm_sched_start_timeout(sched);
}
diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index 5a1f068af1c2..b80513eec90f 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -188,6 +188,7 @@ struct drm_sched_fence *to_drm_sched_fence(struct dma_fence 
*f);
  * to schedule the job.
  */
struct drm_sched_job {
+   struct kref refcount;
struct spsc_nodequeue_node;
struct drm_gpu_scheduler*sched;
struct drm_sched_fence  *s_fence;
@@ -198,6 +199,7 @@ struct drm_sched_job {
enum drm_sched_priority s_priority;
struct drm_sched_entity  *entity;
struct dma_fence_cb cb;
+
};

From: Grodzovsky, Andrey 
Sent: Friday, March 19, 2021 12:17 AM
To: Zhang, Jack (Jian) ; Christian König 
; dri-devel@lists.freedesktop.org; 
amd-...@lists.freedesktop.org; Koenig, Christian ; 
Liu, Monk ; Deng, Emily ; Rob Herring 
; Tomeu Vizoso ; Steven Price 

Subject: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak



On 2021-03-18 6:41 a.m., Zhang, Jack (Jian) wrote:

[AMD Official Use Only - Internal Distribution Only]

Hi, Andrey

Let me summarize the background of this patch:

In TDR resubmit step “amdgpu_device_recheck_guilty_jobs,
It will submit first jobs of each ring and do guilty job re-check.
At that point, We had to make sure each job is in the mirror list(or 
re-inserted back already).

But we found the current code never re-insert the job to mirror list in the 
2nd, 3rd job_timeout thread(Bailing TDR thread).
This not only will cause memleak of the bailing jobs. What’s more important, 
the 1st tdr t

Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Steven Price]

On 15/03/2021 05:23, Zhang, Jack (Jian) wrote:

[AMD Public Use]

Hi, Rob/Tomeu/Steven,

Would you please help to review this patch for panfrost driver?

Thanks,
Jack Zhang

-Original Message-
From: Jack Zhang 
Sent: Monday, March 15, 2021 1:21 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, Christian 
; Grodzovsky, Andrey ; Liu, Monk 
; Deng, Emily 
Cc: Zhang, Jack (Jian) 
Subject: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

re-insert Bailing jobs to avoid memory leak.

V2: move re-insert step to drm/scheduler logic
V3: add panfrost's return value for bailing jobs
in case it hits the memleak issue.


This commit message could do with some work - it's really hard to 
decipher what the actual problem you're solving is.




Signed-off-by: Jack Zhang 
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_job.c| 8 ++--
  drivers/gpu/drm/panfrost/panfrost_job.c| 4 ++--
  drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
  include/drm/gpu_scheduler.h| 1 +
  5 files changed, 19 insertions(+), 6 deletions(-)


[...]

diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c 
b/drivers/gpu/drm/panfrost/panfrost_job.c
index 6003cfeb1322..e2cb4f32dae1 100644
--- a/drivers/gpu/drm/panfrost/panfrost_job.c
+++ b/drivers/gpu/drm/panfrost/panfrost_job.c
@@ -444,7 +444,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 * spurious. Bail out.
 */
if (dma_fence_is_signaled(job->done_fence))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
  
  	dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",

js,
@@ -456,7 +456,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
  
  	/* Scheduler is already stopped, nothing to do. */

if (!panfrost_scheduler_stop(>js->queue[js], sched_job))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
  
  	/* Schedule a reset if there's no reset in progress. */

if (!atomic_xchg(>reset.pending, 1))


This looks correct to me - in these two cases drm_sched_stop() is not 
called on the sched_job, so it looks like currently the job will be leaked.



diff --git a/drivers/gpu/drm/scheduler/sched_main.c 
b/drivers/gpu/drm/scheduler/sched_main.c
index 92d8de24d0a1..a44f621fb5c4 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -314,6 +314,7 @@ static void drm_sched_job_timedout(struct work_struct *work)
  {
struct drm_gpu_scheduler *sched;
struct drm_sched_job *job;
+   int ret;
  
  	sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
  
@@ -331,8 +332,13 @@ static void drm_sched_job_timedout(struct work_struct *work)

list_del_init(>list);
spin_unlock(>job_list_lock);
  
-		job->sched->ops->timedout_job(job);

+   ret = job->sched->ops->timedout_job(job);
  
+		if (ret == DRM_GPU_SCHED_STAT_BAILING) {

+   spin_lock(>job_list_lock);
+   list_add(>node, >ring_mirror_list);
+   spin_unlock(>job_list_lock);
+   }


I think we could really do with a comment somewhere explaining what 
"bailing" means in this context. For the Panfrost case we have two cases:


 * The GPU job actually finished while the timeout code was running 
(done_fence is signalled).


 * The GPU is already in the process of being reset (Panfrost has 
multiple queues, so mostly like a bad job in another queue).


I'm also not convinced that (for Panfrost) it makes sense to be adding 
the jobs back to the list. For the first case above clearly the job 
could just be freed (it's complete). The second case is more interesting 
and Panfrost currently doesn't handle this well. In theory the driver 
could try to rescue the job ('soft stop' in Mali language) so that it 
could be resubmitted. Panfrost doesn't currently support that, so 
attempting to resubmit the job is almost certainly going to fail.


It's on my TODO list to look at improving Panfrost in this regard, but 
sadly still quite far down.


Steve


/*
 * Guilty job did complete and hence needs to be manually 
removed
 * See drm_sched_stop doc.
diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index 4ea8606d91fe..8093ac2427ef 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -210,6 +210,7 @@ enum drm_gpu_sched_stat {
DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
DRM_GPU_SCHED_STAT_NOMINAL,
DRM_GPU_SCHED_STAT_ENODEV,
+   DRM_GPU_SCHED_STAT_BAILING,
  };
  
  /**




___

Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Andrey Grodzovsky]

On 2021-03-18 6:41 a.m., Zhang, Jack (Jian) wrote:


[AMD Official Use Only - Internal Distribution Only]


Hi, Andrey

Let me summarize the background of this patch:

In TDR resubmit step “amdgpu_device_recheck_guilty_jobs,

It will submit first jobs of each ring and do guilty job re-check.

At that point, We had to make sure each job is in the mirror list(or 
re-inserted back already).


But we found the current code never re-insert the job to mirror list 
in the 2^nd , 3^rd job_timeout thread(Bailing TDR thread).


This not only will cause memleak of the bailing jobs. What’s more 
important, the 1^st tdr thread can never iterate the bailing job and 
set its guilty status to a correct status.


Therefore, we had to re-insert the job(or even not delete node) for 
bailing job.


For the above V3 patch, the racing condition in my mind is:

we cannot make sure all bailing jobs are finished before we do 
amdgpu_device_recheck_guilty_jobs.




Yes,that race i missed - so you say that for 2nd, baling thread who 
extracted the job, even if he reinsert it right away back after driver 
callback return DRM_GPU_SCHED_STAT_BAILING, there is small time slot 
where the job is not in mirror list and so the 1st TDR might miss it and 
not find that  2nd job is the actual guilty job, right ? But, still this 
job will get back into mirror list, and since it's really the bad job, 
it will never signal completion and so on the next timeout cycle it will 
be caught (of course there is a starvation scenario here if more TDRs 
kick in and it bails out again but this is really unlikely).




Based on this insight, I think we have two options to solve this issue:

 1. Skip delete node in tdr thread2, thread3, 4 … (using mutex or
atomic variable)
 2. Re-insert back bailing job, and meanwhile use semaphore in each
tdr thread to keep the sequence as expected and ensure each job is
in the mirror list when do resubmit step.

For Option1, logic is simpler and we need only one global atomic variable:

What do you think about this plan?

Option1 should look like the following logic:

+static atomic_t in_reset; //a global atomic var for synchronization

static void drm_sched_process_job(struct dma_fence *f, struct 
dma_fence_cb *cb);


 /**

@@ -295,6 +296,12 @@ static void drm_sched_job_timedout(struct 
work_struct *work)


 * drm_sched_cleanup_jobs. It will be reinserted back 
after sched->thread


 * is parked at which point it's safe.

 */

+   if (atomic_cmpxchg(_reset, 0, 1) != 0) {  //skip 
delete node if it’s thead1,2,3,….


+ spin_unlock(>job_list_lock);

+ drm_sched_start_timeout(sched);

+   return;

+   }

+

list_del_init(>node);

spin_unlock(>job_list_lock);

@@ -320,6 +327,7 @@ static void drm_sched_job_timedout(struct 
work_struct *work)


spin_lock(>job_list_lock);

    drm_sched_start_timeout(sched);

spin_unlock(>job_list_lock);

+   atomic_set(_reset, 0); //reset in_reset when the first 
thread finished tdr


}



Technically looks like it should work as you don't access the job 
pointer any longer and so no risk that if signaled it will be freed by 
drm_sched_get_cleanup_job but,you can't just use one global variable an 
by this bailing from TDR when different drivers run their TDR threads in 
parallel, and even for amdgpu, if devices in different XGMI hives or 2 
independent devices in non XGMI setup. There should be defined some kind 
of GPU reset group structure on drm_scheduler level for which this 
variable would be used.


P.S I wonder why we can't just ref-count the job so that even if 
drm_sched_get_cleanup_job would delete it before we had a chance to stop 
the scheduler thread, we wouldn't crash. This would avoid all the dance 
with deletion and reinsertion.


Andrey



Thanks,

Jack

*From:* amd-gfx  *On Behalf Of 
*Zhang, Jack (Jian)

*Sent:* Wednesday, March 17, 2021 11:11 PM
*To:* Christian König ; 
dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; 
Koenig, Christian ; Liu, Monk 
; Deng, Emily ; Rob Herring 
; Tomeu Vizoso ; Steven 
Price ; Grodzovsky, Andrey 

*Subject:* Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid 
memleak


[AMD Official Use Only - Internal Distribution Only]

[AMD Official Use Only - Internal Distribution Only]

Hi，Andrey，

Good catch，I will expore this corner case and give feedback soon～

Best，

Jack



*From:*Grodzovsky, Andrey <mailto:andrey.grodzov...@amd.com>>

*Sent:* Wednesday, March 17, 2021 10:50:59 PM
*To:* Christian König <mailto:ckoenig.leichtzumer...@gmail.com>>; Zhang, Jack (Jian) 
mailto:jack.zha...@amd.com>>; 
dri-devel@lists.freedesktop.org 
<mailto:dri-devel@lists.freedesktop.org> 
<mailto:dri-devel@lists.freedesktop.org>>; 
amd-...@lists.freedesktop.org <mailto:amd-...@lists.freedesktop.org&

RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Zhang, Jack (Jian)]

[AMD Official Use Only - Internal Distribution Only]

Hi, Andrey

Let me summarize the background of this patch:

In TDR resubmit step “amdgpu_device_recheck_guilty_jobs,
It will submit first jobs of each ring and do guilty job re-check.
At that point, We had to make sure each job is in the mirror list(or 
re-inserted back already).

But we found the current code never re-insert the job to mirror list in the 
2nd, 3rd job_timeout thread(Bailing TDR thread).
This not only will cause memleak of the bailing jobs. What’s more important, 
the 1st tdr thread can never iterate the bailing job and set its guilty status 
to a correct status.

Therefore, we had to re-insert the job(or even not delete node) for bailing job.

For the above V3 patch, the racing condition in my mind is:
we cannot make sure all bailing jobs are finished before we do 
amdgpu_device_recheck_guilty_jobs.

Based on this insight, I think we have two options to solve this issue:

  1.  Skip delete node in tdr thread2, thread3, 4 … (using mutex or atomic 
variable)
  2.  Re-insert back bailing job, and meanwhile use semaphore in each tdr 
thread to keep the sequence as expected and ensure each job is in the mirror 
list when do resubmit step.

For Option1, logic is simpler and we need only one global atomic variable:
What do you think about this plan?

Option1 should look like the following logic:


+static atomic_t in_reset; //a global atomic var for synchronization
static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
 /**
@@ -295,6 +296,12 @@ static void drm_sched_job_timedout(struct work_struct 
*work)
 * drm_sched_cleanup_jobs. It will be reinserted back after 
sched->thread
 * is parked at which point it's safe.
 */
+   if (atomic_cmpxchg(_reset, 0, 1) != 0) {  //skip delete node 
if it’s thead1,2,3,….
+   spin_unlock(>job_list_lock);
+   drm_sched_start_timeout(sched);
+   return;
+   }
+
list_del_init(>node);
spin_unlock(>job_list_lock);
@@ -320,6 +327,7 @@ static void drm_sched_job_timedout(struct work_struct *work)
spin_lock(>job_list_lock);
drm_sched_start_timeout(sched);
spin_unlock(>job_list_lock);
+   atomic_set(_reset, 0); //reset in_reset when the first thread 
finished tdr
}


Thanks,
Jack
From: amd-gfx  On Behalf Of Zhang, Jack 
(Jian)
Sent: Wednesday, March 17, 2021 11:11 PM
To: Christian König ; 
dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, 
Christian ; Liu, Monk ; Deng, Emily 
; Rob Herring ; Tomeu Vizoso 
; Steven Price ; Grodzovsky, 
Andrey 
Subject: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak


[AMD Official Use Only - Internal Distribution Only]


[AMD Official Use Only - Internal Distribution Only]

Hi，Andrey，

Good catch，I will expore this corner case and give feedback soon～
Best，
Jack


From: Grodzovsky, Andrey 
mailto:andrey.grodzov...@amd.com>>
Sent: Wednesday, March 17, 2021 10:50:59 PM
To: Christian König 
mailto:ckoenig.leichtzumer...@gmail.com>>; 
Zhang, Jack (Jian) mailto:jack.zha...@amd.com>>; 
dri-devel@lists.freedesktop.org<mailto:dri-devel@lists.freedesktop.org> 
mailto:dri-devel@lists.freedesktop.org>>; 
amd-...@lists.freedesktop.org<mailto:amd-...@lists.freedesktop.org> 
mailto:amd-...@lists.freedesktop.org>>; Koenig, 
Christian mailto:christian.koe...@amd.com>>; Liu, 
Monk mailto:monk@amd.com>>; Deng, Emily 
mailto:emily.d...@amd.com>>; Rob Herring 
mailto:r...@kernel.org>>; Tomeu Vizoso 
mailto:tomeu.viz...@collabora.com>>; Steven Price 
mailto:steven.pr...@arm.com>>
Subject: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

I actually have a race condition concern here - see bellow -

On 2021-03-17 3:43 a.m., Christian König wrote:
> I was hoping Andrey would take a look since I'm really busy with other
> work right now.
>
> Regards,
> Christian.
>
> Am 17.03.21 um 07:46 schrieb Zhang, Jack (Jian):
>> Hi, Andrey/Crhistian and Team,
>>
>> I didn't receive the reviewer's message from maintainers on panfrost
>> driver for several days.
>> Due to this patch is urgent for my current working project.
>> Would you please help to give some review ideas?
>>
>> Many Thanks,
>> Jack
>> -Original Message-
>> From: Zhang, Jack (Jian)
>> Sent: Tuesday, March 16, 2021 3:20 PM
>> To: dri-devel@lists.freedesktop.org<mailto:dri-devel@lists.freedesktop.org>; 
>> amd-...@lists.freedesktop.org<mailto:amd-...@lists.freedesktop.org>;
>> Koenig, Christian 
>> mailto:christian.koe...@amd.com>>; Grodzovsky, 
>> Andrey
>> mailto:andrey.gro

Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Zhang, Jack (Jian)]

[AMD Official Use Only - Internal Distribution Only]

Hi，Andrey，

Good catch，I will expore this corner case and give feedback soon～

Best，
Jack


From: Grodzovsky, Andrey 
Sent: Wednesday, March 17, 2021 10:50:59 PM
To: Christian König ; Zhang, Jack (Jian) 
; dri-devel@lists.freedesktop.org 
; amd-...@lists.freedesktop.org 
; Koenig, Christian ; 
Liu, Monk ; Deng, Emily ; Rob Herring 
; Tomeu Vizoso ; Steven Price 

Subject: Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

I actually have a race condition concern here - see bellow -

On 2021-03-17 3:43 a.m., Christian König wrote:
> I was hoping Andrey would take a look since I'm really busy with other
> work right now.
>
> Regards,
> Christian.
>
> Am 17.03.21 um 07:46 schrieb Zhang, Jack (Jian):
>> Hi, Andrey/Crhistian and Team,
>>
>> I didn't receive the reviewer's message from maintainers on panfrost
>> driver for several days.
>> Due to this patch is urgent for my current working project.
>> Would you please help to give some review ideas?
>>
>> Many Thanks,
>> Jack
>> -Original Message-
>> From: Zhang, Jack (Jian)
>> Sent: Tuesday, March 16, 2021 3:20 PM
>> To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org;
>> Koenig, Christian ; Grodzovsky, Andrey
>> ; Liu, Monk ; Deng,
>> Emily ; Rob Herring ; Tomeu
>> Vizoso ; Steven Price 
>> Subject: RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid
>> memleak
>>
>> [AMD Public Use]
>>
>> Ping
>>
>> -Original Message-
>> From: Zhang, Jack (Jian)
>> Sent: Monday, March 15, 2021 1:24 PM
>> To: Jack Zhang ;
>> dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org;
>> Koenig, Christian ; Grodzovsky, Andrey
>> ; Liu, Monk ; Deng,
>> Emily ; Rob Herring ; Tomeu
>> Vizoso ; Steven Price 
>> Subject: RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid
>> memleak
>>
>> [AMD Public Use]
>>
>> Hi, Rob/Tomeu/Steven,
>>
>> Would you please help to review this patch for panfrost driver?
>>
>> Thanks,
>> Jack Zhang
>>
>> -----Original Message-
>> From: Jack Zhang 
>> Sent: Monday, March 15, 2021 1:21 PM
>> To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org;
>> Koenig, Christian ; Grodzovsky, Andrey
>> ; Liu, Monk ; Deng,
>> Emily 
>> Cc: Zhang, Jack (Jian) 
>> Subject: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak
>>
>> re-insert Bailing jobs to avoid memory leak.
>>
>> V2: move re-insert step to drm/scheduler logic
>> V3: add panfrost's return value for bailing jobs in case it hits the
>> memleak issue.
>>
>> Signed-off-by: Jack Zhang 
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_job.c| 8 ++--
>>   drivers/gpu/drm/panfrost/panfrost_job.c| 4 ++--
>>   drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
>>   include/drm/gpu_scheduler.h| 1 +
>>   5 files changed, 19 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> index 79b9cc73763f..86463b0f936e 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> @@ -4815,8 +4815,10 @@ int amdgpu_device_gpu_recover(struct
>> amdgpu_device *adev,
>>   job ? job->base.id : -1);
>> /* even we skipped this reset, still need to set the job
>> to guilty */
>> -if (job)
>> +if (job) {
>>   drm_sched_increase_karma(>base);
>> +r = DRM_GPU_SCHED_STAT_BAILING;
>> +}
>>   goto skip_recovery;
>>   }
>>   diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>> index 759b34799221..41390bdacd9e 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>> @@ -34,6 +34,7 @@ static enum drm_gpu_sched_stat
>> amdgpu_job_timedout(struct drm_sched_job *s_job)
>>   struct amdgpu_job *job = to_amdgpu_job(s_job);
>>   struct amdgpu_task_info ti;
>>   struct amdgpu_device *adev = ring->adev;
>> +int ret;
>> memset(, 0, sizeof(struct amdgpu_task_info));
>>   @@ -52,8 +53,11 @@ static enum drm_gpu_sched_stat
>> amdgpu_job_timedout(struct drm_sched_job *s_job)
>> ti.process_name, ti.

Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Andrey Grodzovsky]

I actually have a race condition concern here - see bellow -

On 2021-03-17 3:43 a.m., Christian König wrote:
I was hoping Andrey would take a look since I'm really busy with other 
work right now.


Regards,
Christian.

Am 17.03.21 um 07:46 schrieb Zhang, Jack (Jian):

Hi, Andrey/Crhistian and Team,

I didn't receive the reviewer's message from maintainers on panfrost 
driver for several days.

Due to this patch is urgent for my current working project.
Would you please help to give some review ideas?

Many Thanks,
Jack
-Original Message-
From: Zhang, Jack (Jian)
Sent: Tuesday, March 16, 2021 3:20 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; 
Koenig, Christian ; Grodzovsky, Andrey 
; Liu, Monk ; Deng, 
Emily ; Rob Herring ; Tomeu 
Vizoso ; Steven Price 
Subject: RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid 
memleak


[AMD Public Use]

Ping

-Original Message-
From: Zhang, Jack (Jian)
Sent: Monday, March 15, 2021 1:24 PM
To: Jack Zhang ; 
dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; 
Koenig, Christian ; Grodzovsky, Andrey 
; Liu, Monk ; Deng, 
Emily ; Rob Herring ; Tomeu 
Vizoso ; Steven Price 
Subject: RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid 
memleak


[AMD Public Use]

Hi, Rob/Tomeu/Steven,

Would you please help to review this patch for panfrost driver?

Thanks,
Jack Zhang

-Original Message-
From: Jack Zhang 
Sent: Monday, March 15, 2021 1:21 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; 
Koenig, Christian ; Grodzovsky, Andrey 
; Liu, Monk ; Deng, 
Emily 

Cc: Zhang, Jack (Jian) 
Subject: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

re-insert Bailing jobs to avoid memory leak.

V2: move re-insert step to drm/scheduler logic
V3: add panfrost's return value for bailing jobs in case it hits the 
memleak issue.


Signed-off-by: Jack Zhang 
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_job.c    | 8 ++--
  drivers/gpu/drm/panfrost/panfrost_job.c    | 4 ++--
  drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
  include/drm/gpu_scheduler.h    | 1 +
  5 files changed, 19 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c

index 79b9cc73763f..86463b0f936e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4815,8 +4815,10 @@ int amdgpu_device_gpu_recover(struct 
amdgpu_device *adev,

  job ? job->base.id : -1);
    /* even we skipped this reset, still need to set the job 
to guilty */

-    if (job)
+    if (job) {
  drm_sched_increase_karma(>base);
+    r = DRM_GPU_SCHED_STAT_BAILING;
+    }
  goto skip_recovery;
  }
  diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c

index 759b34799221..41390bdacd9e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -34,6 +34,7 @@ static enum drm_gpu_sched_stat 
amdgpu_job_timedout(struct drm_sched_job *s_job)

  struct amdgpu_job *job = to_amdgpu_job(s_job);
  struct amdgpu_task_info ti;
  struct amdgpu_device *adev = ring->adev;
+    int ret;
    memset(, 0, sizeof(struct amdgpu_task_info));
  @@ -52,8 +53,11 @@ static enum drm_gpu_sched_stat 
amdgpu_job_timedout(struct drm_sched_job *s_job)

    ti.process_name, ti.tgid, ti.task_name, ti.pid);
    if (amdgpu_device_should_recover_gpu(ring->adev)) {
-    amdgpu_device_gpu_recover(ring->adev, job);
-    return DRM_GPU_SCHED_STAT_NOMINAL;
+    ret = amdgpu_device_gpu_recover(ring->adev, job);
+    if (ret == DRM_GPU_SCHED_STAT_BAILING)
+    return DRM_GPU_SCHED_STAT_BAILING;
+    else
+    return DRM_GPU_SCHED_STAT_NOMINAL;
  } else {
  drm_sched_suspend_timeout(>sched);
  if (amdgpu_sriov_vf(adev))
diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c 
b/drivers/gpu/drm/panfrost/panfrost_job.c

index 6003cfeb1322..e2cb4f32dae1 100644
--- a/drivers/gpu/drm/panfrost/panfrost_job.c
+++ b/drivers/gpu/drm/panfrost/panfrost_job.c
@@ -444,7 +444,7 @@ static enum drm_gpu_sched_stat 
panfrost_job_timedout(struct drm_sched_job

   * spurious. Bail out.
   */
  if (dma_fence_is_signaled(job->done_fence))
-    return DRM_GPU_SCHED_STAT_NOMINAL;
+    return DRM_GPU_SCHED_STAT_BAILING;
    dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, 
status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",

  js,
@@ -456,7 +456,7 @@ static enum drm_gpu_sched_stat 
panfrost_job_timedout(struct drm_sched_job

    /* Scheduler is already stopped, nothing to do. */
  if (!panfrost_scheduler_stop(>js->queue[js], sched_job))
-    return DRM_GPU_SCHED_STAT_NOMINAL;

Re: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Christian König]

I was hoping Andrey would take a look since I'm really busy with other 
work right now.


Regards,
Christian.

Am 17.03.21 um 07:46 schrieb Zhang, Jack (Jian):

Hi, Andrey/Crhistian and Team,

I didn't receive the reviewer's message from maintainers on panfrost driver for 
several days.
Due to this patch is urgent for my current working project.
Would you please help to give some review ideas?

Many Thanks,
Jack
-Original Message-
From: Zhang, Jack (Jian)
Sent: Tuesday, March 16, 2021 3:20 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, Christian ; 
Grodzovsky, Andrey ; Liu, Monk ; Deng, Emily 
; Rob Herring ; Tomeu Vizoso ; Steven 
Price 
Subject: RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[AMD Public Use]

Ping

-Original Message-
From: Zhang, Jack (Jian)
Sent: Monday, March 15, 2021 1:24 PM
To: Jack Zhang ; dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, Christian 
; Grodzovsky, Andrey ; Liu, Monk ; 
Deng, Emily ; Rob Herring ; Tomeu Vizoso ; 
Steven Price 
Subject: RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[AMD Public Use]

Hi, Rob/Tomeu/Steven,

Would you please help to review this patch for panfrost driver?

Thanks,
Jack Zhang

-Original Message-
From: Jack Zhang 
Sent: Monday, March 15, 2021 1:21 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, Christian 
; Grodzovsky, Andrey ; Liu, Monk 
; Deng, Emily 
Cc: Zhang, Jack (Jian) 
Subject: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

re-insert Bailing jobs to avoid memory leak.

V2: move re-insert step to drm/scheduler logic
V3: add panfrost's return value for bailing jobs in case it hits the memleak 
issue.

Signed-off-by: Jack Zhang 
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_job.c| 8 ++--
  drivers/gpu/drm/panfrost/panfrost_job.c| 4 ++--
  drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
  include/drm/gpu_scheduler.h| 1 +
  5 files changed, 19 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 79b9cc73763f..86463b0f936e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4815,8 +4815,10 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
job ? job->base.id : -1);
  
  		/* even we skipped this reset, still need to set the job to guilty */

-   if (job)
+   if (job) {
drm_sched_increase_karma(>base);
+   r = DRM_GPU_SCHED_STAT_BAILING;
+   }
goto skip_recovery;
}
  
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c

index 759b34799221..41390bdacd9e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -34,6 +34,7 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct 
drm_sched_job *s_job)
struct amdgpu_job *job = to_amdgpu_job(s_job);
struct amdgpu_task_info ti;
struct amdgpu_device *adev = ring->adev;
+   int ret;
  
  	memset(, 0, sizeof(struct amdgpu_task_info));
  
@@ -52,8 +53,11 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)

  ti.process_name, ti.tgid, ti.task_name, ti.pid);
  
  	if (amdgpu_device_should_recover_gpu(ring->adev)) {

-   amdgpu_device_gpu_recover(ring->adev, job);
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   ret = amdgpu_device_gpu_recover(ring->adev, job);
+   if (ret == DRM_GPU_SCHED_STAT_BAILING)
+   return DRM_GPU_SCHED_STAT_BAILING;
+   else
+   return DRM_GPU_SCHED_STAT_NOMINAL;
} else {
drm_sched_suspend_timeout(>sched);
if (amdgpu_sriov_vf(adev))
diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c 
b/drivers/gpu/drm/panfrost/panfrost_job.c
index 6003cfeb1322..e2cb4f32dae1 100644
--- a/drivers/gpu/drm/panfrost/panfrost_job.c
+++ b/drivers/gpu/drm/panfrost/panfrost_job.c
@@ -444,7 +444,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 * spurious. Bail out.
 */
if (dma_fence_is_signaled(job->done_fence))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
  
  	dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",

js,
@@ -456,7 +456,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
  
  	/* Scheduler is already stopped, nothing to do. */

if (!panfrost_scheduler_stop(>js->queue[js], sched_job))
- 

RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Zhang, Jack (Jian)]

Hi, Andrey/Crhistian and Team,

I didn't receive the reviewer's message from maintainers on panfrost driver for 
several days.
Due to this patch is urgent for my current working project.
Would you please help to give some review ideas?

Many Thanks,
Jack
-Original Message-
From: Zhang, Jack (Jian) 
Sent: Tuesday, March 16, 2021 3:20 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, 
Christian ; Grodzovsky, Andrey 
; Liu, Monk ; Deng, Emily 
; Rob Herring ; Tomeu Vizoso 
; Steven Price 
Subject: RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[AMD Public Use]

Ping

-Original Message-
From: Zhang, Jack (Jian) 
Sent: Monday, March 15, 2021 1:24 PM
To: Jack Zhang ; dri-devel@lists.freedesktop.org; 
amd-...@lists.freedesktop.org; Koenig, Christian ; 
Grodzovsky, Andrey ; Liu, Monk ; 
Deng, Emily ; Rob Herring ; Tomeu Vizoso 
; Steven Price 
Subject: RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[AMD Public Use]

Hi, Rob/Tomeu/Steven,

Would you please help to review this patch for panfrost driver?

Thanks,
Jack Zhang

-Original Message-
From: Jack Zhang 
Sent: Monday, March 15, 2021 1:21 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, 
Christian ; Grodzovsky, Andrey 
; Liu, Monk ; Deng, Emily 

Cc: Zhang, Jack (Jian) 
Subject: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

re-insert Bailing jobs to avoid memory leak.

V2: move re-insert step to drm/scheduler logic
V3: add panfrost's return value for bailing jobs in case it hits the memleak 
issue.

Signed-off-by: Jack Zhang 
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c| 8 ++--
 drivers/gpu/drm/panfrost/panfrost_job.c| 4 ++--
 drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
 include/drm/gpu_scheduler.h| 1 +
 5 files changed, 19 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 79b9cc73763f..86463b0f936e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4815,8 +4815,10 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
job ? job->base.id : -1);
 
/* even we skipped this reset, still need to set the job to 
guilty */
-   if (job)
+   if (job) {
drm_sched_increase_karma(>base);
+   r = DRM_GPU_SCHED_STAT_BAILING;
+   }
goto skip_recovery;
}
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index 759b34799221..41390bdacd9e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -34,6 +34,7 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct 
drm_sched_job *s_job)
struct amdgpu_job *job = to_amdgpu_job(s_job);
struct amdgpu_task_info ti;
struct amdgpu_device *adev = ring->adev;
+   int ret;
 
memset(, 0, sizeof(struct amdgpu_task_info));
 
@@ -52,8 +53,11 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct 
drm_sched_job *s_job)
  ti.process_name, ti.tgid, ti.task_name, ti.pid);
 
if (amdgpu_device_should_recover_gpu(ring->adev)) {
-   amdgpu_device_gpu_recover(ring->adev, job);
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   ret = amdgpu_device_gpu_recover(ring->adev, job);
+   if (ret == DRM_GPU_SCHED_STAT_BAILING)
+   return DRM_GPU_SCHED_STAT_BAILING;
+   else
+   return DRM_GPU_SCHED_STAT_NOMINAL;
} else {
drm_sched_suspend_timeout(>sched);
if (amdgpu_sriov_vf(adev))
diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c 
b/drivers/gpu/drm/panfrost/panfrost_job.c
index 6003cfeb1322..e2cb4f32dae1 100644
--- a/drivers/gpu/drm/panfrost/panfrost_job.c
+++ b/drivers/gpu/drm/panfrost/panfrost_job.c
@@ -444,7 +444,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 * spurious. Bail out.
 */
if (dma_fence_is_signaled(job->done_fence))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
 
dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, 
status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",
js,
@@ -456,7 +456,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 
/* Scheduler is already stopped, nothing to do. */
if (!panfrost_scheduler_stop(>js->queue[js], sched_job))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
 
/* Sched

RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Zhang, Jack (Jian)]

[AMD Public Use]

Ping

-Original Message-
From: Zhang, Jack (Jian) 
Sent: Monday, March 15, 2021 1:24 PM
To: Jack Zhang ; dri-devel@lists.freedesktop.org; 
amd-...@lists.freedesktop.org; Koenig, Christian ; 
Grodzovsky, Andrey ; Liu, Monk ; 
Deng, Emily ; Rob Herring ; Tomeu Vizoso 
; Steven Price 
Subject: RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[AMD Public Use]

Hi, Rob/Tomeu/Steven,

Would you please help to review this patch for panfrost driver?

Thanks,
Jack Zhang

-Original Message-
From: Jack Zhang 
Sent: Monday, March 15, 2021 1:21 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, 
Christian ; Grodzovsky, Andrey 
; Liu, Monk ; Deng, Emily 

Cc: Zhang, Jack (Jian) 
Subject: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

re-insert Bailing jobs to avoid memory leak.

V2: move re-insert step to drm/scheduler logic
V3: add panfrost's return value for bailing jobs in case it hits the memleak 
issue.

Signed-off-by: Jack Zhang 
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c| 8 ++--
 drivers/gpu/drm/panfrost/panfrost_job.c| 4 ++--
 drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
 include/drm/gpu_scheduler.h| 1 +
 5 files changed, 19 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 79b9cc73763f..86463b0f936e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4815,8 +4815,10 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
job ? job->base.id : -1);
 
/* even we skipped this reset, still need to set the job to 
guilty */
-   if (job)
+   if (job) {
drm_sched_increase_karma(>base);
+   r = DRM_GPU_SCHED_STAT_BAILING;
+   }
goto skip_recovery;
}
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index 759b34799221..41390bdacd9e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -34,6 +34,7 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct 
drm_sched_job *s_job)
struct amdgpu_job *job = to_amdgpu_job(s_job);
struct amdgpu_task_info ti;
struct amdgpu_device *adev = ring->adev;
+   int ret;
 
memset(, 0, sizeof(struct amdgpu_task_info));
 
@@ -52,8 +53,11 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct 
drm_sched_job *s_job)
  ti.process_name, ti.tgid, ti.task_name, ti.pid);
 
if (amdgpu_device_should_recover_gpu(ring->adev)) {
-   amdgpu_device_gpu_recover(ring->adev, job);
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   ret = amdgpu_device_gpu_recover(ring->adev, job);
+   if (ret == DRM_GPU_SCHED_STAT_BAILING)
+   return DRM_GPU_SCHED_STAT_BAILING;
+   else
+   return DRM_GPU_SCHED_STAT_NOMINAL;
} else {
drm_sched_suspend_timeout(>sched);
if (amdgpu_sriov_vf(adev))
diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c 
b/drivers/gpu/drm/panfrost/panfrost_job.c
index 6003cfeb1322..e2cb4f32dae1 100644
--- a/drivers/gpu/drm/panfrost/panfrost_job.c
+++ b/drivers/gpu/drm/panfrost/panfrost_job.c
@@ -444,7 +444,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 * spurious. Bail out.
 */
if (dma_fence_is_signaled(job->done_fence))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
 
dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, 
status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",
js,
@@ -456,7 +456,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 
/* Scheduler is already stopped, nothing to do. */
if (!panfrost_scheduler_stop(>js->queue[js], sched_job))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
 
/* Schedule a reset if there's no reset in progress. */
if (!atomic_xchg(>reset.pending, 1)) diff --git 
a/drivers/gpu/drm/scheduler/sched_main.c 
b/drivers/gpu/drm/scheduler/sched_main.c
index 92d8de24d0a1..a44f621fb5c4 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -314,6 +314,7 @@ static void drm_sched_job_timedout(struct work_struct 
*work)  {
struct drm_gpu_scheduler *sched;
struct drm_sched_job *job;
+   int ret;
 
sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 
@@ -331,8 +332,13 @@ static void dr

RE: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Zhang, Jack (Jian)]

[AMD Public Use]

Hi, Rob/Tomeu/Steven,

Would you please help to review this patch for panfrost driver?

Thanks,
Jack Zhang

-Original Message-
From: Jack Zhang  
Sent: Monday, March 15, 2021 1:21 PM
To: dri-devel@lists.freedesktop.org; amd-...@lists.freedesktop.org; Koenig, 
Christian ; Grodzovsky, Andrey 
; Liu, Monk ; Deng, Emily 

Cc: Zhang, Jack (Jian) 
Subject: [PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

re-insert Bailing jobs to avoid memory leak.

V2: move re-insert step to drm/scheduler logic
V3: add panfrost's return value for bailing jobs
in case it hits the memleak issue.

Signed-off-by: Jack Zhang 
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c| 8 ++--
 drivers/gpu/drm/panfrost/panfrost_job.c| 4 ++--
 drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
 include/drm/gpu_scheduler.h| 1 +
 5 files changed, 19 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 79b9cc73763f..86463b0f936e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4815,8 +4815,10 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
job ? job->base.id : -1);
 
/* even we skipped this reset, still need to set the job to 
guilty */
-   if (job)
+   if (job) {
drm_sched_increase_karma(>base);
+   r = DRM_GPU_SCHED_STAT_BAILING;
+   }
goto skip_recovery;
}
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index 759b34799221..41390bdacd9e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -34,6 +34,7 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct 
drm_sched_job *s_job)
struct amdgpu_job *job = to_amdgpu_job(s_job);
struct amdgpu_task_info ti;
struct amdgpu_device *adev = ring->adev;
+   int ret;
 
memset(, 0, sizeof(struct amdgpu_task_info));
 
@@ -52,8 +53,11 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct 
drm_sched_job *s_job)
  ti.process_name, ti.tgid, ti.task_name, ti.pid);
 
if (amdgpu_device_should_recover_gpu(ring->adev)) {
-   amdgpu_device_gpu_recover(ring->adev, job);
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   ret = amdgpu_device_gpu_recover(ring->adev, job);
+   if (ret == DRM_GPU_SCHED_STAT_BAILING)
+   return DRM_GPU_SCHED_STAT_BAILING;
+   else
+   return DRM_GPU_SCHED_STAT_NOMINAL;
} else {
drm_sched_suspend_timeout(>sched);
if (amdgpu_sriov_vf(adev))
diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c 
b/drivers/gpu/drm/panfrost/panfrost_job.c
index 6003cfeb1322..e2cb4f32dae1 100644
--- a/drivers/gpu/drm/panfrost/panfrost_job.c
+++ b/drivers/gpu/drm/panfrost/panfrost_job.c
@@ -444,7 +444,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 * spurious. Bail out.
 */
if (dma_fence_is_signaled(job->done_fence))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
 
dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, 
status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",
js,
@@ -456,7 +456,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 
/* Scheduler is already stopped, nothing to do. */
if (!panfrost_scheduler_stop(>js->queue[js], sched_job))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
 
/* Schedule a reset if there's no reset in progress. */
if (!atomic_xchg(>reset.pending, 1))
diff --git a/drivers/gpu/drm/scheduler/sched_main.c 
b/drivers/gpu/drm/scheduler/sched_main.c
index 92d8de24d0a1..a44f621fb5c4 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -314,6 +314,7 @@ static void drm_sched_job_timedout(struct work_struct *work)
 {
struct drm_gpu_scheduler *sched;
struct drm_sched_job *job;
+   int ret;
 
sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 
@@ -331,8 +332,13 @@ static void drm_sched_job_timedout(struct work_struct 
*work)
list_del_init(>list);
spin_unlock(>job_list_lock);
 
-   job->sched->ops->timedout_job(job);
+   ret = job->sched->ops->timedout_job(job);
 
+   if (ret == DRM_GPU_SCHED_STAT_BAILING) {
+   spin_lock(>job_list_lock);
+ 

[PATCH v3] drm/scheduler re-insert Bailing job to avoid memleak

[Jack Zhang]

re-insert Bailing jobs to avoid memory leak.

V2: move re-insert step to drm/scheduler logic
V3: add panfrost's return value for bailing jobs
in case it hits the memleak issue.

Signed-off-by: Jack Zhang 
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c| 8 ++--
 drivers/gpu/drm/panfrost/panfrost_job.c| 4 ++--
 drivers/gpu/drm/scheduler/sched_main.c | 8 +++-
 include/drm/gpu_scheduler.h| 1 +
 5 files changed, 19 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 79b9cc73763f..86463b0f936e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4815,8 +4815,10 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
job ? job->base.id : -1);
 
/* even we skipped this reset, still need to set the job to 
guilty */
-   if (job)
+   if (job) {
drm_sched_increase_karma(>base);
+   r = DRM_GPU_SCHED_STAT_BAILING;
+   }
goto skip_recovery;
}
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index 759b34799221..41390bdacd9e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -34,6 +34,7 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct 
drm_sched_job *s_job)
struct amdgpu_job *job = to_amdgpu_job(s_job);
struct amdgpu_task_info ti;
struct amdgpu_device *adev = ring->adev;
+   int ret;
 
memset(, 0, sizeof(struct amdgpu_task_info));
 
@@ -52,8 +53,11 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct 
drm_sched_job *s_job)
  ti.process_name, ti.tgid, ti.task_name, ti.pid);
 
if (amdgpu_device_should_recover_gpu(ring->adev)) {
-   amdgpu_device_gpu_recover(ring->adev, job);
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   ret = amdgpu_device_gpu_recover(ring->adev, job);
+   if (ret == DRM_GPU_SCHED_STAT_BAILING)
+   return DRM_GPU_SCHED_STAT_BAILING;
+   else
+   return DRM_GPU_SCHED_STAT_NOMINAL;
} else {
drm_sched_suspend_timeout(>sched);
if (amdgpu_sriov_vf(adev))
diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c 
b/drivers/gpu/drm/panfrost/panfrost_job.c
index 6003cfeb1322..e2cb4f32dae1 100644
--- a/drivers/gpu/drm/panfrost/panfrost_job.c
+++ b/drivers/gpu/drm/panfrost/panfrost_job.c
@@ -444,7 +444,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 * spurious. Bail out.
 */
if (dma_fence_is_signaled(job->done_fence))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
 
dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, 
status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",
js,
@@ -456,7 +456,7 @@ static enum drm_gpu_sched_stat panfrost_job_timedout(struct 
drm_sched_job
 
/* Scheduler is already stopped, nothing to do. */
if (!panfrost_scheduler_stop(>js->queue[js], sched_job))
-   return DRM_GPU_SCHED_STAT_NOMINAL;
+   return DRM_GPU_SCHED_STAT_BAILING;
 
/* Schedule a reset if there's no reset in progress. */
if (!atomic_xchg(>reset.pending, 1))
diff --git a/drivers/gpu/drm/scheduler/sched_main.c 
b/drivers/gpu/drm/scheduler/sched_main.c
index 92d8de24d0a1..a44f621fb5c4 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -314,6 +314,7 @@ static void drm_sched_job_timedout(struct work_struct *work)
 {
struct drm_gpu_scheduler *sched;
struct drm_sched_job *job;
+   int ret;
 
sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
 
@@ -331,8 +332,13 @@ static void drm_sched_job_timedout(struct work_struct 
*work)
list_del_init(>list);
spin_unlock(>job_list_lock);
 
-   job->sched->ops->timedout_job(job);
+   ret = job->sched->ops->timedout_job(job);
 
+   if (ret == DRM_GPU_SCHED_STAT_BAILING) {
+   spin_lock(>job_list_lock);
+   list_add(>node, >ring_mirror_list);
+   spin_unlock(>job_list_lock);
+   }
/*
 * Guilty job did complete and hence needs to be manually 
removed
 * See drm_sched_stop doc.
diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index 4ea8606d91fe..8093ac2427ef 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -210,6 +210,7 @@ enum drm_gpu_sched_stat {
DRM_GPU_SCHED_STAT_NONE, /*