[PATCH v2 8/8] drm/amdgpu: Prevent any job recoveries after device is unplugged.

Thu Nov 19 21:24:57 UTC 2020

On 11/19/20 10:29 AM, Daniel Vetter wrote:
> On Thu, Nov 19, 2020 at 10:02:28AM -0500, Andrey Grodzovsky wrote:
>> On 11/19/20 2:55 AM, Christian König wrote:
>>> Am 18.11.20 um 17:20 schrieb Andrey Grodzovsky:
>>>> On 11/18/20 7:01 AM, Christian König wrote:
>>>>> Am 18.11.20 um 08:39 schrieb Daniel Vetter:
>>>>>> On Tue, Nov 17, 2020 at 9:07 PM Andrey Grodzovsky
>>>>>> <Andrey.Grodzovsky at amd.com> wrote:
>>>>>>> On 11/17/20 2:49 PM, Daniel Vetter wrote:
>>>>>>>> On Tue, Nov 17, 2020 at 02:18:49PM -0500, Andrey Grodzovsky wrote:
>>>>>>>>> On 11/17/20 1:52 PM, Daniel Vetter wrote:
>>>>>>>>>> On Tue, Nov 17, 2020 at 01:38:14PM -0500, Andrey Grodzovsky wrote:
>>>>>>>>>>> On 6/22/20 5:53 AM, Daniel Vetter wrote:
>>>>>>>>>>>> On Sun, Jun 21, 2020 at 02:03:08AM -0400, Andrey Grodzovsky wrote:
>>>>>>>>>>>>> No point to try recovery if device is gone, just messes up things.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky at amd.com>
>>>>>>>>>>>>> ---
>>>>>>>>>>>>>       drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 16 ++++++++++++++++
>>>>>>>>>>>>>       drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 8 ++++++++
>>>>>>>>>>>>>       2 files changed, 24 insertions(+)
>>>>>>>>>>>>>
>>>>>>>>>>>>> diff --git
>>>>>>>>>>>>> a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>>>>>>>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>>>>>>>>>>>> index 6932d75..5d6d3d9 100644
>>>>>>>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>>>>>>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>>>>>>>>>>>> @@ -1129,12 +1129,28 @@ static
>>>>>>>>>>>>> int amdgpu_pci_probe(struct
>>>>>>>>>>>>> pci_dev *pdev,
>>>>>>>>>>>>>            return ret;
>>>>>>>>>>>>>       }
>>>>>>>>>>>>> +static void amdgpu_cancel_all_tdr(struct amdgpu_device *adev)
>>>>>>>>>>>>> +{
>>>>>>>>>>>>> +        int i;
>>>>>>>>>>>>> +
>>>>>>>>>>>>> +        for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
>>>>>>>>>>>>> +                struct amdgpu_ring *ring = adev->rings[i];
>>>>>>>>>>>>> +
>>>>>>>>>>>>> +                if (!ring || !ring->sched.thread)
>>>>>>>>>>>>> +                        continue;
>>>>>>>>>>>>> +
>>>>>>>>>>>>> + cancel_delayed_work_sync(&ring->sched.work_tdr);
>>>>>>>>>>>>> +        }
>>>>>>>>>>>>> +}
>>>>>>>>>>>> I think this is a function that's supposed to be in drm/scheduler, not
>>>>>>>>>>>> here. Might also just be your
>>>>>>>>>>>> cleanup code being ordered wrongly,
>>>>>>>>>>>> or your
>>>>>>>>>>>> split in one of the earlier patches not done quite right.
>>>>>>>>>>>> -Daniel
>>>>>>>>>>> This function iterates across all the
>>>>>>>>>>> schedulers per amdgpu device and
>>>>>>>>>>> accesses
>>>>>>>>>>> amdgpu specific structures ,
>>>>>>>>>>> drm/scheduler deals with single
>>>>>>>>>>> scheduler at most
>>>>>>>>>>> so looks to me like this is the right place for this function
>>>>>>>>>> I guess we could keep track of all schedulers somewhere in a list in
>>>>>>>>>> struct drm_device and wrap this up. That was kinda the idea.
>>>>>>>>>>
>>>>>>>>>> Minimally I think a tiny wrapper with docs for the
>>>>>>>>>> cancel_delayed_work_sync(&sched->work_tdr); which explains what you must
>>>>>>>>>> observe to make sure there's no race.
>>>>>>>>> Will do
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>> I'm not exactly sure there's no
>>>>>>>>>> guarantee here we won't get a new tdr work launched right afterwards at
>>>>>>>>>> least, so this looks a bit like a hack.
>>>>>>>>> Note that for any TDR work happening post amdgpu_cancel_all_tdr
>>>>>>>>> amdgpu_job_timedout->drm_dev_is_unplugged
>>>>>>>>> will return true and so it will return early. To make it water proof tight
>>>>>>>>> against race
>>>>>>>>> i can switch from drm_dev_is_unplugged to drm_dev_enter/exit
>>>>>>>> Hm that's confusing. You do a work_cancel_sync, so that at least looks
>>>>>>>> like "tdr work must not run after this point"
>>>>>>>>
>>>>>>>> If you only rely on drm_dev_enter/exit check with the tdr work, then
>>>>>>>> there's no need to cancel anything.
>>>>>>> Agree, synchronize_srcu from drm_dev_unplug should play the role
>>>>>>> of 'flushing' any earlier (in progress) tdr work which is
>>>>>>> using drm_dev_enter/exit pair. Any later arising tdr
>>>>>>> will terminate early when
>>>>>>> drm_dev_enter
>>>>>>> returns false.
>>>>>> Nope, anything you put into the work itself cannot close this race.
>>>>>> It's the schedule_work that matters here. Or I'm missing something ...
>>>>>> I thought that the tdr work you're cancelling here is launched by
>>>>>> drm/scheduler code, not by the amd callback?
>>>>
>>>> My bad, you are right, I am supposed to put drm_dev_enter.exit pair
>>>> into drm_sched_job_timedout
>>>>
>>>>
>>>>> Yes that is correct. Canceling the work item is not the right
>>>>> approach at all, nor is adding dev_enter/exit pair in the
>>>>> recovery handler.
>>>>
>>>> Without adding the dev_enter/exit guarding pair in the recovery
>>>> handler you are ending up with GPU reset starting while
>>>> the device is already unplugged, this leads to multiple errors and general mess.
>>>>
>>>>
>>>>> What we need to do here is to stop the scheduler thread and then
>>>>> wait for any timeout handling to have finished.
>>>>>
>>>>> Otherwise it can scheduler a new timeout just after we have canceled this one.
>>>>>
>>>>> Regards,
>>>>> Christian.
>>>>
>>>> Schedulers are stopped from amdgpu_driver_unload_kms which indeed
>>>> happens after drm_dev_unplug
>>>> so yes, there is still a chance for new work being scheduler and
>>>> timeout armed after but, once i fix the code
>>>> to place drm_dev_enter/exit pair into drm_sched_job_timeout I don't
>>>> see why that not a good solution ?
>>> Yeah that should work as well, but then you also don't need to cancel
>>> the work item from the driver.
>>
>> Indeed, as Daniel pointed out no need and I dropped it. One correction - I
>> previously said that w/o
>> dev_enter/exit guarding pair in scheduler's TO handler you will get GPU
>> reset starting while device already gone -
>> of course this is not fully preventing this as the device can be extracted
>> at any moment just after we
>> already entered GPU recovery. But it does saves us processing a futile  GPU
>> recovery which always
>> starts once you unplug the device if there are active gobs in progress at
>> the moment and so I think it's
>> still justifiable to keep the dev_enter/exit guarding pair there.
> Yeah sprinkling drm_dev_enter/exit over the usual suspect code paths like
> tdr to make the entire unloading much faster makes sense. Waiting for
> enormous amounts of mmio ops to time out isn't fun. A comment might be
> good for that though, to explain why we're doing that.
> -Daniel

Will do, I also tried to insert drm_dev_enter/exit in all MMIO accessors in amdgpu
to try and avoid at that level but didn't get good results for unclear reason, 
will probably get
to this as a follow up work to again avoid expanding the scope of current work 
too much.

Andrey

>
>> Andrey
>>
>>
>>>
>>>> Any tdr work started after drm_dev_unplug finished will simply abort
>>>> on entry to drm_sched_job_timedout
>>>> because drm_dev_enter will be false and the function will return
>>>> without rearming the timeout timer and
>>>> so will have no impact.
>>>>
>>>> The only issue i see here now is of possible use after free if some
>>>> late tdr work will try to execute after
>>>> drm device already gone, for this we probably should add
>>>> cancel_delayed_work_sync(sched.work_tdr)
>>>> to drm_sched_fini after sched->thread is stopped there.
>>> Good point, that is indeed missing as far as I can see.
>>>
>>> Christian.
>>>
>>>> Andrey