Re: 回复: [PATCH] drm/amdgpu: Ignore first evction failure during suspend

[Christian König <christian.koenig@xxxxxxx>]

    Am 14.09.23 um 15:59 schrieb Christian König:

Am 14.09.23 um 15:37 schrieb Felix Kuehling:

Userptr and SVM restore work is scheduled to the system WQ with schedule_delayed_work. See amdgpu_amdkfd_evict_userptr and svm_range_evict. This would need to use queue_delayed_work with the system_freezable_wq.

BO restoration is scheduled with queue_delayed_work on our own kfd_restore_wq that was allocated with alloc_ordered_workqueue. This would need to add the WQ_FREEZABLE flag when we create the wq in kfd_process_create_wq.

There is also evict_process_worker scheduled with schedule_delayed_work, which handles stopping of user mode queues, signaling of eviction fences and scheduling of restore work when BOs are evicted. I think that should not be freezable because it's needed to signal the eviction fences to allow suspend to evict BOs.

To make sure I'm not misunderstanding, I assume that freezing a freezable workqueue flushes work items in progress and prevents execution of more work until it is unfrozen. I assume work items are not frozen in the middle of execution, because that would not solve the problem.

I was wondering the exact same thing and to be honest I don't know that detail either and of hand can't find any documentation about it.

My suspicion is that a work item can freeze when it calls schedule(), e.g. when taking a look or similar.

    I've found some time to double check this. At least of hand it looks
    like freezing workqueues means that no more work items are scheduled
    and we wait for existing to finish.
    
    So using the freezable workqueues should solve the problem.
    
    Christian.

That would then indeed not work at all and we would need to make sure that the work is completed manually somehow.

Regards,
Christian.

Regards,
  Felix

On 2023-09-14 2:23, Christian König wrote:

[putting Harry on BCC, sorry for the noise]

Yeah, that is clearly a bug in the KFD.

During the second eviction the hw should already be disabled, so we don't have any SDMA or similar to evict BOs any more and can only copy them with the CPU.

@Felix what workqueue do you guys use for the restore work? I've just double checked and on the system workqueues you explicitly need to specify that stuff is freezable. E.g. use system_freezable_wq instead of system_wq.

Alternatively as Xinhui mentioned it might be necessary to flush all restore work before the first eviction phase or we have the chance that BOs are moved back into VRAM again.

Regards,
Christian.

Am 14.09.23 um 03:54 schrieb Pan, Xinhui:

[AMD Official Use Only - General]

I just make one debug patch to show busy BO’s alloc-trace when the eviction fails in suspend.

And dmesg log attached.

Looks like they are just kfd user Bos and locked by evict/restore work.

So in kfd suspend callback, it really need to flush the evict/restore work before HW fini as it do now.

That is why the first very early eviction fails and the second eviction succeed.

 

Thanks

xinhui

From: Pan, Xinhui
Sent: Thursday, September 14, 2023 8:02 AM
To: Koenig, Christian <Christian.Koenig@xxxxxxx>; Kuehling, Felix <Felix.Kuehling@xxxxxxx>; Christian König <ckoenig.leichtzumerken@xxxxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx; Wentland, Harry <Harry.Wentland@xxxxxxx>
Cc: Deucher, Alexander <Alexander.Deucher@xxxxxxx>; Fan, Shikang <Shikang.Fan@xxxxxxx>
Subject: RE: 回复: [PATCH] drm/amdgpu: Ignore first evction failure during suspend

 

Chris,

I can dump these busy BOs with their alloc/free stack later today.

 

BTW, the two evictions and the kfd suspend are all called before hw_fini. IOW, between phase 1 and phase 2. SDMA is turned only in phase2. So current code works fine maybe.

 

From: Koenig, Christian <Christian.Koenig@xxxxxxx>
Sent: Wednesday, September 13, 2023 10:29 PM
To: Kuehling, Felix <Felix.Kuehling@xxxxxxx>; Christian König <ckoenig.leichtzumerken@xxxxxxxxx>; Pan, Xinhui <Xinhui.Pan@xxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx; Wentland, Harry <Harry.Wentland@xxxxxxx>
Cc: Deucher, Alexander <Alexander.Deucher@xxxxxxx>; Fan, Shikang <Shikang.Fan@xxxxxxx>
Subject: Re: 回复: [PATCH] drm/amdgpu: Ignore first evction failure during suspend

 

[+Harry]

Am 13.09.23 um 15:54 schrieb Felix Kuehling:

On 2023-09-13 4:07, Christian König wrote:

[+Fleix]

Well that looks like quite a serious bug.

If I'm not completely mistaken the KFD work item tries to restore the process by moving BOs into memory even after the suspend freeze. Normally work items are frozen together with the user space processes unless explicitly marked as not freezable.

That this causes problem during the first eviction phase is just the tip of the iceberg here. If a BO is moved into invisible memory during this we wouldn't be able to get it out of that in the second phase because SDMA and hw is already turned off.

@Felix any idea how that can happen? Have you guys marked a work item / work queue as not freezable?

We don't set anything to non-freezable in KFD.

 

Regards,
  Felix

 

Or maybe the display guys?

Do you guys in the display do any delayed update in a work item which is marked as not-freezable?

Otherwise I have absolutely no idea what's going on here.

Thanks,
Christian.

@Xinhui please investigate what work item that is and where that is coming from. Something like "if (adev->in_suspend) dump_stack();" in the right place should probably do it.

Thanks,
Christian.

Am 13.09.23 um 07:13 schrieb Pan, Xinhui:

[AMD Official Use Only - General]

 

I notice that only user space process are frozen on my side.  kthread and workqueue  keeps running. Maybe some kernel configs are not enabled.

I made one module which just prints something like i++ with mutex lock both in workqueue and kthread. I paste some logs below.

[438619.696196] XH: 14 from workqueue

[438619.700193] XH: 15 from kthread

[438620.394335] PM: suspend entry (deep)

[438620.399619] Filesystems sync: 0.001 seconds

[438620.403887] PM: Preparing system for sleep (deep)

[438620.409299] Freezing user space processes

[438620.414862] Freezing user space processes completed (elapsed 0.001 seconds)

[438620.421881] OOM killer disabled.

[438620.425197] Freezing remaining freezable tasks

[438620.430890] Freezing remaining freezable tasks completed (elapsed 0.001 seconds)

[438620.438348] PM: Suspending system (deep)

.....

[438623.746038] PM: suspend of devices complete after 3303.137 msecs

[438623.752125] PM: start suspend of devices complete after 3309.713 msecs

[438623.758722] PM: suspend debug: Waiting for 5 second(s).

[438623.792166] XH: 22 from kthread

[438623.824140] XH: 23 from workqueue

 

 

So BOs definitely can be in use during suspend.

Even if kthread or workqueue can be stopped with one special kernel config. I think suspend can only stop the workqueue with its callback finish.

otherwise something like below makes things crazy.

LOCK BO

do something

    -> schedule or wait, anycode might sleep.  Stopped by suspend now? no, i think.

UNLOCK BO

 

I do tests  with  cmds below.

echo devices  > /sys/power/pm_test

echo 0  > /sys/power/pm_async

echo 1  > /sys/power/pm_print_times

echo 1 > /sys/power/pm_debug_messages

echo 1 > /sys/module/amdgpu/parameters/debug_evictions

./kfd.sh --gtest_filter=KFDEvictTest.BasicTest

pm-suspend

 

thanks

xinhui

 

 

---

发件人: Christian König <ckoenig.leichtzumerken@xxxxxxxxx>
发送时间: 2023年9月12日 17:01
收件人: Pan, Xinhui <Xinhui.Pan@xxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx <amd-gfx@xxxxxxxxxxxxxxxxxxxxx>
抄送: Deucher, Alexander <Alexander.Deucher@xxxxxxx>; Koenig, Christian <Christian.Koenig@xxxxxxx>; Fan, Shikang <Shikang.Fan@xxxxxxx>
主题: Re: [PATCH] drm/amdgpu: Ignore first evction failure during suspend

 

When amdgpu_device_suspend() is called processes should be frozen
already. In other words KFD queues etc... should already be idle.

So when the eviction fails here we missed something previously and that
in turn can cause tons amount of problems.

So ignoring those errors is most likely not a good idea at all.

Regards,
Christian.

Am 12.09.23 um 02:21 schrieb Pan, Xinhui:
> [AMD Official Use Only - General]
>
> Oh yep, Pinned BO is moved to other LRU list, So eviction fails because of other reason.
> I will change the comments in the patch.
> The problem is eviction fails as many reasons, say, BO is locked.
> ASAIK, kfd will stop the queues and flush some evict/restore work in its suspend callback. SO the first eviction before kfd callback likely fails.
>
> -----Original Message-----
> From: Christian König <ckoenig.leichtzumerken@xxxxxxxxx>
> Sent: Friday, September 8, 2023 2:49 PM
> To: Pan, Xinhui <Xinhui.Pan@xxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx
> Cc: Deucher, Alexander <Alexander.Deucher@xxxxxxx>; Koenig, Christian <Christian.Koenig@xxxxxxx>; Fan, Shikang <Shikang.Fan@xxxxxxx>
> Subject: Re: [PATCH] drm/amdgpu: Ignore first evction failure during suspend
>
> Am 08.09.23 um 05:39 schrieb xinhui pan:
>> Some BOs might be pinned. So the first eviction's failure will abort
>> the suspend sequence. These pinned BOs will be unpined afterwards
>> during suspend.
> That doesn't make much sense since pinned BOs don't cause eviction failure here.
>
> What exactly is the error code you see?
>
> Christian.
>
>> Actaully it has evicted most BOs, so that should stil work fine in
>> sriov full access mode.
>>
>> Fixes: 47ea20762bb7 ("drm/amdgpu: Add an extra evict_resource call
>> during device_suspend.")
>> Signed-off-by: xinhui pan <xinhui.pan@xxxxxxx>
>> ---
>>    drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 9 +++++----
>>    1 file changed, 5 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> index 5c0e2b766026..39af526cdbbe 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> @@ -4148,10 +4148,11 @@ int amdgpu_device_suspend(struct drm_device
>> *dev, bool fbcon)
>>
>>        adev->in_suspend = true;
>>
>> -     /* Evict the majority of BOs before grabbing the full access */
>> -     r = amdgpu_device_evict_resources(adev);
>> -     if (r)
>> -             return r;
>> +     /* Try to evict the majority of BOs before grabbing the full access
>> +      * Ignore the ret val at first place as we will unpin some BOs if any
>> +      * afterwards.
>> +      */
>> +     (void)amdgpu_device_evict_resources(adev);
>>
>>        if (amdgpu_sriov_vf(adev)) {
>>                amdgpu_virt_fini_data_exchange(adev);