On 22/03/2021 16:43, Daniel Vetter wrote:
On Mon, Mar 22, 2021 at 4:31 PM Tvrtko Ursulin
<tvrtko.ursulin@xxxxxxxxxxxxxxx> wrote:
On 22/03/2021 14:57, Daniel Vetter wrote:
On Mon, Mar 22, 2021 at 3:33 PM Tvrtko Ursulin
<tvrtko.ursulin@xxxxxxxxxxxxxxx> wrote:
On 22/03/2021 14:09, Daniel Vetter wrote:
On Mon, Mar 22, 2021 at 11:22:01AM +0000, Tvrtko Ursulin wrote:
On 19/03/2021 22:38, Jason Ekstrand wrote:
This API allows one context to grab bits out of another context upon
creation. It can be used as a short-cut for setparam(getparam()) for
things like I915_CONTEXT_PARAM_VM. However, it's never been used by any
real userspace. It's used by a few IGT tests and that's it. Since it
doesn't add any real value (most of the stuff you can CLONE you can copy
in other ways), drop it.
No complaints to remove if it ended up unused outside IGT. Latter is a _big_
problem though, since it is much more that a few IGT tests. So I really
think there really needs to be an evaluation and a plan for that (we don't
want to lose 50% of the coverage over night).
There is one thing that this API allows you to clone which you cannot
clone via getparam/setparam: timelines. However, timelines are an
implementation detail of i915 and not really something that needs to be
Not really true timelines are i915 implementation detail. They are in fact a
dma-fence context:seqno concept, nothing more that than. I think you are
probably confusing struct intel_timeline with the timeline wording in the
uapi. Former is i915 implementation detail, but context:seqno are truly
userspace timelines.
I think you're both saying the same thing and talking a bit past each
another.
Yes the timeline is just a string of dma_fence, that's correct. Now
usually if you submit batches with execbuf, we have 3 ways to synchronize
concurrent submission: implicit sync, sync_file and drm_syncob. They all
map to different needs in different protocols/render apis.
Now in one additional case the kernel makes sure that batchbuffers are
ordered, and that's when you submit them to the same hw ctx. Because
there's only 1 hw context and you really can't have batchbuffers run on
that single hw context out of order. That's what the timeline object we
talk about here is. But that largely is an internal implementation detail,
which happens to also use most/all the same infrastructure as the
dma_fence uapi pieces above.
Now the internal implementation detail leaking here is that we exposed
this to userspace, without there being any need for this. What Jason
implements with syncobj in the next patch is essentially what userspace
should have been using for cross-engine sync. media userspace doesn't care
about interop with winsys/client apis, so they equally could have used
implicit sync or sync_file here (which I think is the solution now for the
new uapi prepped internally), since they all are about equally powerful
for stringing batchbuffers together.
Are you saying we exposed a single timeline of execution per hw context
via the single timeline flag?!
Nope.
Timelines of execution were always exposed. Any "engine" (ring
previously) in I915_EXEC_RING_MASK was a single timeline of execution.
It is completely the same with engine map engines, which are also
different indices into I915_EXEC_RING_MASK space.
Userspace was aware of these timelines forever as well. Media was
creating multiple contexts to have multiple timelines (so parallelism).
Everyone knew that engine-hopping submissions needs to be either
implicitly or explicitly synchronised, etc.
Yup, I think we're saying the same thing here.
So I really don't see that we have leaked timelines as a concept *now*.
What the patch has exposed to userspace is a new way to sync between
timelines and nothing more.
We've leaked it as something you can now share across hw context.
Okay so we agree on most things but apparently have different
definitions of what it means to leak internal implementation details.
While at the same time proof that we haven't leaked the internal
implementation details is that Jason was able to implement the single
timeline flag with a drm syncobj at the execbuf top level. (Well mostly,
ignoring the probably inconsequential difference of one vs multiple
fence contexts.)
It's not a matching implementation. It's only good enough for what
media needs, and essentially what media should have done to begin
with.
There's substantially different behaviour between SINGLE_TIMELINE and
what Jason has done here when you race concurrent execbuf calls:
Former guarantees total ordering, the latter doesn't even try. They
are not the same thing, but luckily userspace doesn't care about that
difference.
Sounds like a very important difference to stress in the commit message.
Secondly, I am unclear whether we have agreement on whether the single
timeline flag is leaking implementation details of the execlists
scheduler to userspace or not?
Regards,
Tvrtko
Aside, just to make sure this wont get lost: I do agree that we should
only allow this up to maybe ADL, and reject it on anything new (maybe
including dg1 while we're at it, since the pci ids for that aren't
even close to upstream yet).
-Daniel
Which is possible because of how it's internally implemented (I think
load balancer relies on that), but not really a synchronization
Virtual engine is a single timeline by definition and it is still that
regardless of the implementation details (execlists or GuC, in both
cases it is a single hardware context and a single timeline).
primitive we want to export as such to userspace. We have other
interfaces and concepts for that.
Yes, that is the only point to argue IMO. We can say it wasn't needed
and should have been avoided, but I still maintain we can't really say
we leaked anything backend specific to userspace via it.
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
