On Thu, Mar 25, 2021 at 10:48 AM Tvrtko Ursulin tvrtko.ursulin@linux.intel.com wrote:
On 24/03/2021 17:18, Jason Ekstrand wrote:
On Wed, Mar 24, 2021 at 6:36 AM Tvrtko Ursulin tvrtko.ursulin@linux.intel.com wrote:
On 24/03/2021 09:52, Daniel Vetter wrote:
On Wed, Mar 24, 2021 at 09:28:58AM +0000, Tvrtko Ursulin wrote:
On 23/03/2021 17:51, Jason Ekstrand wrote:
This API is entirely unnecessary and I'd love to get rid of it. If userspace wants a single timeline across multiple contexts, they can either use implicit synchronization or a syncobj, both of which existed at the time this feature landed. The justification given at the time was that it would help GL drivers which are inherently single-timeline. However, neither of our GL drivers actually wanted the feature. i965 was already in maintenance mode at the time and iris uses syncobj for everything.
Unfortunately, as much as I'd love to get rid of it, it is used by the media driver so we can't do that. We can, however, do the next-best thing which is to embed a syncobj in the context and do exactly what we'd expect from userspace internally. This isn't an entirely identical implementation because it's no longer atomic if userspace races with itself by calling execbuffer2 twice simultaneously from different threads. It won't crash in that case; it just doesn't guarantee any ordering between those two submits.
Moving SINGLE_TIMELINE to a syncobj emulation has a couple of technical advantages beyond mere annoyance. One is that intel_timeline is no longer an api-visible object and can remain entirely an implementation detail. This may be advantageous as we make scheduler changes going forward. Second is that, together with deleting the CLONE_CONTEXT API, we should now have a 1:1 mapping between intel_context and intel_timeline which may help us reduce locking.
Much, much better commit message although I still fail to understand where do you see implementation details leaking out. So for me this is still something I'd like to get to the bottom of.
I didn't say "leaking". I said it's no longer API-visible. That's just true. It used to be a kernel object that userspace was unaware of, then we added SINGLE_TIMELINE and userspace now has some influence over the object. With this patch, it's hidden again. I don't get why that's confusing.
I am definitely glad we moved on from leaking to less dramatic wording, but it is still not API "visible" in so much struct file_operations * is not user visible in any negative way when you do open(2), being just implementation detail. But I give up.
I would also mention the difference regarding fence context change.
There are no fence context changes. The fence that we stuff in the syncobj is an i915 fence and the fence we pull back out is an i915 fence. A syncobj is just a fancy wrapper for a dma_buf pointer.
Change is in the dma_fence->context.
Current code single timeline is one fence context. With this patch that is no longer the case.
Not sure that it has any practical implications for the internal dma_fence code like is_later checks , haven't analysed it.
We merge fewer fences at the higher levels and rely more on the fence dependency tracking of the scheduler frontend to sort stuff out.
And sync fence info ioctl exposes this value to userspace which probably has no practical implications. Unless some indirect effect when merging fences.
Userspace can use that to do fence merging. Which is always only a strict optimization. I'm not even sure whether Android does that or not.
Main point is that I think these are the things which need to be discussed in the commit message.
Yeah makes sense to add these as impact, next to the "we don't deal with races anymore" part. -Daniel
And in general I would maintain this patch as part of a series which ends up demonstrating the "mays" and "shoulds".
I disagree. The past few years we've merged way too much patches and features without carefully answering the high level questions of
- do we really need to solve this problem
- and if so, are we really solving this problem in the right place
Now we're quite in a hole, and we're not going to get out of this hole if we keep applying the same standards that got us here. Anything that does not clearly and without reservation the above two questions with "yes" needs to be removed or walled off, just so we can eventually see which complexity we really need, and what is actually superflous.
I understand your general point but when I apply it to this specific patch, even if it is simple, it is neither removing the uapi or walling it off. So I see it as the usual review standard to ask to see what "mays" and "shoulds" actually get us in concrete terms.
Instead of focusing on the term "leak", let's focus on this line of the commit message instead:
Second is that, together with deleting the CLONE_CONTEXT API, we should now have a 1:1 mapping between intel_context and intel_timeline which may help us reduce locking.
Now, I've not written any patches yet which actually reduce the locking. I can try to look into that today. I CC'd Maarten on the first send of this because I was hoping he would have good intuition about this. It may be that this object will always have to require some amount of locking if the scheduler has to touch them in parallel with other stuff. What I can say concretely, however, is that this does reduce the sharing of an internal object even if it doesn't get rid of it completely. The one thing that is shared all over the place with this patch is a syncobj which is explicitly designed for exactly this sort of case and can be used and abused by as many threads as you'd like. That seems like it's going in the right direction.
I can further weasel-word the commit message to make the above more prominent if you'd like.
I am not interested in making you weasel-word anything but reaching a consensus and what is actually true and accurate.
Regards,
Tvrtko