On Tue, Jul 25, 2017 at 11:11:35AM +0200, Christian König wrote:

Am 24.07.2017 um 13:57 schrieb Daniel Vetter:

...

On Mon, Jul 24, 2017 at 11:51 AM, Christian König deathsimple@vodafone.de wrote:

...

Am 24.07.2017 um 10:33 schrieb Daniel Vetter:

...

On Fri, Jul 21, 2017 at 06:20:01PM +0200, Christian König wrote:

...

From: Christian König christian.koenig@amd.com

With hardware resets in mind it is possible that all shared fences are signaled, but the exlusive isn't. Fix waiting for everything in this situation.

How did you end up with both shared and exclusive fences on the same reservation object? At least I thought the point of exclusive was that it's exclusive (and has an implicit barrier on all previous shared fences). Same for shared fences, they need to wait for the exclusive one (and replace it).

Is this fallout from the amdgpu trickery where by default you do all shared fences? I thought we've aligned semantics a while back ...

No, that is perfectly normal even for other drivers. Take a look at the reservation code.

The exclusive fence replaces all shared fences, but adding a shared fence doesn't replace the exclusive fence. That actually makes sense, cause when you want to add move shared fences those need to wait for the last exclusive fence as well.

Hm right.

...

Now normally I would agree that when you have shared fences it is sufficient to wait for all of them cause those operations can't start before the exclusive one finishes. But with GPU reset and/or the ability to abort already submitted operations it is perfectly possible that you end up with an exclusive fence which isn't signaled and a shared fence which is signaled in the same reservation object.

How does that work? The batch(es) with the shared fence are all supposed to wait for the exclusive fence before they start, which means even if you gpu reset and restart/cancel certain things, they shouldn't be able to complete out of order.

Assume the following:

1. The exclusive fence is some move operation by the kernel which executes

on a DMA engine. 2. The shared fence is a 3D operation submitted by userspace which executes on the 3D engine.

Now we found the 3D engine to be hung and needs a reset, all currently submitted jobs are aborted, marked with an error code and their fences put into the signaled state.

Since we only reset the 3D engine, the move operation (fortunately) isn't affected by this.

I think this applies to all drivers and isn't something amdgpu specific.

Not i915 because: - At first we only had system wide gpu reset that killed everything, which means all requests will be completed, not just on a single engine.

- Now we have per-engine reset, but combined with replaying them (the offending one gets a no-op batch to avoid re-hanging), to make sure the depency tree doesn't fall apart.

Now I see that doing this isn't all that simple, and either way we still have the case where one driver resets but not the other (in multi-gpu), but I'm not exactly sure how to best handle this.

What exactly is the downside of not dropping this assumption, i.e. why do you want this patch? What blows up? -Daniel

Regards, Christian.

...

If you outright cancel a fence then you're supposed to first call dma_fence_set_error(-EIO) and then complete it. Note that atm that part might be slightly overengineered and I'm not sure about how we expose stuff to userspace, e.g. dma_fence_set_error(-EAGAIN) is (or soon, has been) used by i915 for it's internal book-keeping, which might not be the best to leak to other consumers. But completing fences (at least exported ones, where userspace or other drivers can get at them) shouldn't be possible. -Daniel