[PATCH v2] Documentation/gpu: VM_BIND locking document

[Thomas Hellström]

Hi, Boris

On 9/6/23 13:09, Boris Brezillon wrote:
> On Wed, 6 Sep 2023 10:32:24 +0200
> Thomas Hellström <thomas.hellstrom at linux.intel.com> wrote:
>
>
>>>>>> +Introducing external (or shared) buffer objects
>>>>>> +===============================================
>>>>>> +
>>>>>> +Since shared buffer objects may be shared by multiple gpu_vm's they
>>>>>> +can't share their reservation object with a single gpu_vm, but
>>>>>> will rather
>>>>>> +have a reservation object of their own. The shared objects bound to a
>>>>>> +gpu_vm using one or many
>>>>>> +gpu_vmas are therefore typically put on a per-gpu_vm list which is
>>>>>> +protected by the gpu_vm lock. One could in theory protect it also
>>>>>> with
>>>>>> +the ``gpu_vm->resv``, but since the list of dma_resvs to take is
>>>>>> typically
>>>>>> +built before the ``gpu_vm->resv`` is locked due to a limitation in
>>>>>> +the current locking helpers, that is typically not done. Also see
>>>>>> +below for userptr gpu_vmas.
>>>>>> +
>>>>>> +At eviction time we now need to invalidate *all* gpu_vmas of a shared
>>>>>> +object, but we can no longer be certain that we hold the gpu_vm's
>>>>>> +dma_resv of all the object's gpu_vmas. We can only be certain that we
>>>>> I need to think a bit more about locking of extobj and evicted
>>>>> object tracking
>>>>> in the case of processing 'drm_gpuva_ops' directly through callbacks
>>>>> within the
>>>>> fence signalling critical path as mentioend in [1].
>>>>>
>>>>> In order to support that, we'd need to protect extobjs with a
>>>>> separate lock,
>>>>> and while iterating extobjs to acquire the dma-resv lock drop the
>>>>> lock within
>>>>> the loop before we actually acquire the dma-resv lock. Maple tree
>>>>> supports that
>>>>> already and this can be fully done within the GPUVA manager; no need
>>>>> for the
>>>>> driver to care about that.
>>>> So do I understand correctly that this because you want to update the
>>>> gpuvm state while operations are progressing asynchronously?
>>>>
>>>> If so, I wonder whether that could really be done? For example to
>>>> allocate enough memory for page-tables etc, you need to know the
>>>> details of the operations at IOCTL execution time, and to know the
>>>> details you need to know the state from the previous operation?
>>>
>>> Right, sync and async bind can't run fully concurrently, but you could
>>> "inject" a
>>> sync one between two async ones such that the sync ones executed from
>>> the IOCTL
>>> directly while async execution is stalled meanwhile. This would be
>>> possible because
>>> the actual drm_gpuva_ops would be calculated within the async
>>> execution path rather
>>> than in the IOCTL. But yes, page-table management must be desinged to
>>> support that.
> FWIW, the panthor driver is designed this way (note that I'm not
> supporting GEM eviction yet, so there might be subtleties I missed).

The problem is that once you've published your VM_BIND out-fence, any 
code path required to signal that fence may notallocate memory nor or 
grab any locks that allows allocating memory while held including 
dma_resv locks, and that means all required page-table memory needs to 
be allocated synchronously in the IOCTL, and all evicted bos need to be 
made resident in the IOCTL, and at least in the xe driver the amount of 
memory we need to allocate depends on the vm state, so we can't really 
update the vm state asynchronously either.

But as long as any async binding work required for signalling the 
VM_BIND out-fence is properly annotated with 
dma_fence_begin_signalling() and dma_fence_end_signalling() and there 
aren't any lockdep splats, things should be good. It would trigger on 
both memory allocation and attempts to grab a dma_resv lock.


>
>> OK, well one of the main motivations for Xe is to be able to pipeline
>> interleaving binds and execs if needed, like so:
>>
>> - Bind vmas for scene 1.
>> - Submit scene 1.
>> - Unbind vmas for scene 1.
>> - Bind vmas for scene 2.
>> - Submit scene 2.
>> - Unbind vmas for scene 2.
>>
>> And being able to *submit* all of the above while the async binding of
>> vmas for scene (step 1) has not yet completed.
>> I can't really see how this could be done, while obeying dma-fence
>> rules, unless state is updated synchronously while submitting?
> The idea in this case is to detect when a GPU job dependency is a
> VM_BIND out-fence, turn drm_sched_fence->parent into an
> xxx_vm_bind_job_fence object that's holding the GEM that's about to be
> mapped (AFAICT, we don't need to do anything for unmap operations), and
> then add our GPU job fence to this BO. This should not only guarantee
> that the GEMs we depend on are mapped before the GPU job is executed
> (the fence wait does that), but also that such yet-to-be-mapped GEMs
> won't be evicted just after they've been mapped and before the GPU had
> a chance to execute (unless I'm missing something, adding our GPU job
> fence to the BO being targeted by a pending VM_BIND(async,map) operation
> solves this problem).

Yes, we're essentially doing the same. The issue here is that when we, 
for example *submit* Bind vmas for scene 2,
we need to know how much page-table memory to allocate, and what BOs to 
make resident to be able to publish the out-fence. That means we need to 
know what the VM state would look like at the end of "Unbind vmas for 
scene 1". If the VM state is updated at submission time, that's all ok 
but if it's updated at execution time, we'd have to guess what resources 
to pre-allocate.

/Thomas