[PATCH 0/4] Refine GPU recovery sequence to enhance its stability

[Christian König]

Am 08.04.21 um 22:39 schrieb Andrey Grodzovsky:
>
> On 2021-04-08 2:58 p.m., Christian König wrote:
>> Am 08.04.21 um 18:08 schrieb Andrey Grodzovsky:
>>> On 2021-04-08 4:32 a.m., Christian König wrote:
>>>> Am 08.04.21 um 10:22 schrieb Christian König:
>>>>> [SNIP]
>>>>>>>
>>>>>>>
>>>>>>>> Beyond blocking all delayed works and scheduler threads we also 
>>>>>>>> need to guarantee no  IOCTL can access MMIO post device unplug 
>>>>>>>> OR in flight IOCTLs are done before we finish pci_remove 
>>>>>>>> (amdgpu_pci_remove for us).
>>>>>>>> For this I suggest we do something like what we worked on with 
>>>>>>>> Takashi Iwai the ALSA maintainer recently when he helped 
>>>>>>>> implementing PCI BARs move support for snd_hda_intel. Take a 
>>>>>>>> look at
>>>>>>>> https://cgit.freedesktop.org/~agrodzov/linux/commit/?h=yadro/pcie_hotplug/movable_bars_v9.1&id=cbaa324799718e2b828a8c7b5b001dd896748497 
>>>>>>>> and
>>>>>>>> https://cgit.freedesktop.org/~agrodzov/linux/commit/?h=yadro/pcie_hotplug/movable_bars_v9.1&id=e36365d9ab5bbc30bdc221ab4b3437de34492440 
>>>>>>>>
>>>>>>>> We also had same issue there, how to prevent MMIO accesses 
>>>>>>>> while the BARs are migrating. What was done there is a refcount 
>>>>>>>> was added to count all IOCTLs in flight, for any in flight  
>>>>>>>> IOCTL the BAR migration handler would
>>>>>>>> block for the refcount to drop to 0 before it would proceed, 
>>>>>>>> for any later IOCTL it stops and wait if device is in migration 
>>>>>>>> state. We even don't need the wait part, nothing to wait for, 
>>>>>>>> we just return with -ENODEV for this case.
>>>>>>>>
>>>>>>>
>>>>>>> This is essentially what the DRM SRCU is doing as well.
>>>>>>>
>>>>>>> For the hotplug case we could do this in the toplevel since we 
>>>>>>> can signal the fence and don't need to block memory management.
>>>>>>
>>>>>>
>>>>>> To make SRCU 'wait for' all IOCTLs in flight we would need to 
>>>>>> wrap every IOCTL ( practically - just drm_ioctl function) with 
>>>>>> drm_dev_enter/drm_dev_exit - can we do it ?
>>>>>>
>>>>
>>>> Sorry totally missed this question.
>>>>
>>>> Yes, exactly that. As discussed for the hotplug case we can do this.
>>>
>>>
>>> Thinking more about it - assuming we are  treating synchronize_srcu 
>>> as a 'wait for completion' of any in flight {drm_dev_enter, 
>>> drm_dev_exit} scope, some of those scopes might do dma_fence_wait 
>>> inside. Since we haven't force signaled the fences yet we will end 
>>> up a deadlock. We have to signal all the various fences before doing 
>>> the 'wait for'. But we can't signal the fences before setting 
>>> 'dev->unplugged = true' to reject further CS and other stuff which 
>>> might create more fences we were supposed-to force signal and now 
>>> missed them. Effectively setting 'dev->unplugged = true' and doing 
>>> synchronize_srcu in one call like drm_dev_unplug does without 
>>> signalling all the fences in the device in between these two steps 
>>> looks luck a possible deadlock to me - what do you think ?
>>>
>>
>> Indeed, that is a really good argument to handle it the same way as 
>> the reset lock.
>>
>> E.g. not taking it at the high level IOCTL, but rather when the 
>> frontend of the driver has acquired all the necessary locks (BO resv, 
>> VM lock etc...) before calling into the backend to actually do things 
>> with the hardware.
>>
>> Christian.
>
> From what you said I understand that you want to solve this problem by 
> using drm_dev_enter/exit brackets low enough in the code such that it 
> will not include and fence wait.
>
> But inserting dmr_dev_enter/exit on the highest level in drm_ioctl is 
> much less effort and less room for error then going through each IOCTL 
> and trying to identify at what point (possibly multiple points) they 
> are about to access HW, some of this is hidden deep in HAL layers such 
> as DC layer in display driver or the multi layers of powerplay/SMU 
> libraries. Also, we can't only limit our-self to back-end if by this 
> you mean ASIC specific functions which access registers. We also need 
> to take care of any MMIO kernel BO (VRAM BOs) where we may access 
> directly MMIO space by pointer from the front end of the driver (HW 
> agnostic) and TTM/DRM layers.

Exactly, yes. The key point is we need to identify such places anyway 
for GPU reset to work properly. So we could just piggy back hotplug on 
top of that work and are done.

>
> Our problem here is how to signal all the existing  fences on one hand 
> and on the other prevent any new dma_fence waits after we finished 
> signaling existing fences. Once we solved this then there is no 
> problem using drm_dev_unplug in conjunction with drm_dev_enter/exit at 
> the highest level of drm_ioctl to flush any IOCTLs in flight and block 
> any new ones.
>
> IMHO when we speak about signalling all fences we don't mean ALL the 
> currently existing dma_fence structs (they are spread all over the 
> place) but rather signal all the HW fences because HW is what's gone 
> and we can't expect for those fences to be ever signaled. All the rest 
> such as: scheduler fences,  user fences, drm_gem reservation objects 
> e.t.c. are either dependent on those HW fences and hence signaling the 
> HW fences will in turn signal them or, are not impacted by the HW 
> being gone and hence can still be waited on and will complete. If this 
> assumption is correct then I think that we should use some flag to 
> prevent any new submission to HW which creates HW fences (somewhere 
> around amdgpu_fence_emit), then traverse all existing HW fences 
> (currently they are spread in a few places so maybe we need to track 
> them in a list) and signal them. After that it's safe to cal 
> drm_dev_unplug and be sure synchronize_srcu won't stall because of of 
> dma_fence_wait. After that we can proceed to canceling work items, 
> stopping schedulers e.t.c.

That is problematic as well since you need to make sure that the 
scheduler is not creating a new hardware fence in the moment you try to 
signal all of them. It would require another SRCU or lock for this.

Christian.

>
> Andrey
>
>
>>
>>> Andrey
>>>
>>>
>>