On Thu, 6 Apr 2023 at 18:58, Matthew Brost <matthew.brost@xxxxxxxxx> wrote: > > On Thu, Apr 06, 2023 at 08:32:59AM +0200, Daniel Vetter wrote: > > On Wed, Apr 05, 2023 at 11:58:44PM +0000, Matthew Brost wrote: > > > On Wed, Apr 05, 2023 at 03:09:08PM +0200, Daniel Vetter wrote: > > > > On Tue, Apr 04, 2023 at 07:48:27PM +0000, Matthew Brost wrote: > > > > > On Tue, Apr 04, 2023 at 09:25:52PM +0200, Daniel Vetter wrote: > > > > > > On Tue, Apr 04, 2023 at 07:02:23PM +0000, Matthew Brost wrote: > > > > > > > On Tue, Apr 04, 2023 at 08:14:01PM +0200, Thomas Hellström (Intel) wrote: > > > > > > > > > > > > > > > > On 4/4/23 15:10, Christian König wrote: > > > > > > > > > Am 04.04.23 um 14:54 schrieb Thomas Hellström: > > > > > > > > > > Hi, Christian, > > > > > > > > > > > > > > > > > > > > On 4/4/23 11:09, Christian König wrote: > > > > > > > > > > > Am 04.04.23 um 02:22 schrieb Matthew Brost: > > > > > > > > > > > > From: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx> > > > > > > > > > > > > > > > > > > > > > > > > For long-running workloads, drivers either need to open-code > > > > > > > > > > > > completion > > > > > > > > > > > > waits, invent their own synchronization primitives or internally use > > > > > > > > > > > > dma-fences that do not obey the cross-driver dma-fence protocol, but > > > > > > > > > > > > without any lockdep annotation all these approaches are error prone. > > > > > > > > > > > > > > > > > > > > > > > > So since for example the drm scheduler uses dma-fences it is > > > > > > > > > > > > desirable for > > > > > > > > > > > > a driver to be able to use it for throttling and error > > > > > > > > > > > > handling also with > > > > > > > > > > > > internal dma-fences tha do not obey the cros-driver > > > > > > > > > > > > dma-fence protocol. > > > > > > > > > > > > > > > > > > > > > > > > Introduce long-running completion fences in form of > > > > > > > > > > > > dma-fences, and add > > > > > > > > > > > > lockdep annotation for them. In particular: > > > > > > > > > > > > > > > > > > > > > > > > * Do not allow waiting under any memory management locks. > > > > > > > > > > > > * Do not allow to attach them to a dma-resv object. > > > > > > > > > > > > * Introduce a new interface for adding callbacks making the > > > > > > > > > > > > helper adding > > > > > > > > > > > > a callback sign off on that it is aware that the dma-fence may not > > > > > > > > > > > > complete anytime soon. Typically this will be the > > > > > > > > > > > > scheduler chaining > > > > > > > > > > > > a new long-running fence on another one. > > > > > > > > > > > > > > > > > > > > > > Well that's pretty much what I tried before: > > > > > > > > > > > https://lwn.net/Articles/893704/ > > > > > > > > > > > > > > > > > > > > > > > > > I don't think this quite the same, this explictly enforces that we don't > > > > > > > break the dma-fence rules (in path of memory allocations, exported in > > > > > > > any way), essentially this just SW sync point reusing dma-fence the > > > > > > > infrastructure for signaling / callbacks. I believe your series tried to > > > > > > > export these fences to user space (admittedly I haven't fully read your > > > > > > > series). > > > > > > > > > > > > > > In this use case we essentially just want to flow control the ring via > > > > > > > the dma-scheduler + maintain a list of pending jobs so the TDR can be > > > > > > > used for cleanup if LR entity encounters an error. To me this seems > > > > > > > perfectly reasonable but I know dma-femce rules are akin to a holy war. > > > > > > > > > > > > > > If we return NULL in run_job, now we have to be able to sink all jobs > > > > > > > in the backend regardless on ring space, maintain a list of jobs pending > > > > > > > for cleanup after errors, and write a different cleanup path as now the > > > > > > > TDR doesn't work. Seems very, very silly to duplicate all of this code > > > > > > > when the DRM scheduler provides all of this for us. Also if we go this > > > > > > > route, now all drivers are going to invent ways to handle LR jobs /w the > > > > > > > DRM scheduler. > > > > > > > > > > > > > > This solution is pretty clear, mark the scheduler as LR, and don't > > > > > > > export any fences from the scheduler. If you try to export these fences > > > > > > > a blow up happens. > > > > > > > > > > > > The problem is if you mix things up. Like for resets you need all the > > > > > > schedulers on an engine/set-of-engines to quiescent or things get > > > > > > potentially hilarious. If you now have a scheduler in forever limbo, the > > > > > > dma_fence guarantees are right out the window. > > > > > > > > > > > > > > > > Right, a GT reset on Xe is: > > > > > > > > > > Stop all schedulers > > > > > Do a reset > > > > > Ban any schedulers which we think caused the GT reset > > > > > Resubmit all schedulers which we think were good > > > > > Restart all schedulers > > > > > > > > > > None of this flow depends on LR dma-fences, all of this uses the DRM > > > > > sched infrastructure and work very well compared to the i915. Rewriting > > > > > all this with a driver specific implementation is what we are trying to > > > > > avoid. > > > > > > > > > > Similarly if LR entity hangs on its own (not a GT reset, rather the > > > > > firmware does the reset for us) we use all the DRM scheduler > > > > > infrastructure to handle this. Again this works rather well... > > > > > > > > Yeah this is why I don't think duplicating everything that long-running > > > > jobs need makes any sense. iow I agree with you. > > > > > > > > > > Glad we agree. > > > > > > > > > But the issue you're having is fairly specific if it's just about > > > > > > ringspace. I think the dumbest fix is to just block in submit if you run > > > > > > out of per-ctx ringspace, and call it a day. This notion that somehow the > > > > > > > > > > How does that not break the dma-fence rules? A job can publish its > > > > > finished fence after ARM, if the finished fence fence waits on ring > > > > > space that may not free up in a reasonable amount of time we now have > > > > > broken the dma-dence rules. My understanding is any dma-fence must only > > > > > on other dma-fence, Christian seems to agree and NAK'd just blocking if > > > > > no space available [1]. IMO this series ensures we don't break dma-fence > > > > > rules by restricting how the finished fence can be used. > > > > > > > > Oh I meant in the submit ioctl, _before_ you even call > > > > drm_sched_job_arm(). It's ok to block in there indefinitely. > > > > > > > > > > Ok, but how do we determine if their is ring space, wait on xe_hw_fence > > > which is a dma-fence. We just move a wait from the scheduler to the exec > > > IOCTL and I realy fail to see the point of that. > > > > Fill in anything you need into the ring at ioctl time, but don't update > > the tail pointers? If there's no space, then EWOULDBLCK. > > > > Ok, I can maybe buy this approach and this is fairly easy to do. I'm > going to do this for LR jobs only though (non-LR job will still flow > control on the ring via the scheduler + write ring in run_job). A bit of > duplicate code but I can live with this. > > > > > > > kernel is supposed to provide a bottomless queue of anything userspace > > > > > > submits simply doesn't hold up in reality (as much as userspace standards > > > > > > committees would like it to), and as long as it doesn't have a real-world > > > > > > perf impact it doesn't really matter why we end up blocking in the submit > > > > > > ioctl. It might also be a simple memory allocation that hits a snag in > > > > > > page reclaim. > > > > > > > > > > > > > > > > > And the reasons why it was rejected haven't changed. > > > > > > > > > > > > > > > > > > > > > > Regards, > > > > > > > > > > > Christian. > > > > > > > > > > > > > > > > > > > > > Yes, TBH this was mostly to get discussion going how we'd best > > > > > > > > > > tackle this problem while being able to reuse the scheduler for > > > > > > > > > > long-running workloads. > > > > > > > > > > > > > > > > > > > > I couldn't see any clear decision on your series, though, but one > > > > > > > > > > main difference I see is that this is intended for driver-internal > > > > > > > > > > use only. (I'm counting using the drm_scheduler as a helper for > > > > > > > > > > driver-private use). This is by no means a way to try tackle the > > > > > > > > > > indefinite fence problem. > > > > > > > > > > > > > > > > > > Well this was just my latest try to tackle this, but essentially the > > > > > > > > > problems are the same as with your approach: When we express such > > > > > > > > > operations as dma_fence there is always the change that we leak that > > > > > > > > > somewhere. > > > > > > > > > > > > > > > > > > My approach of adding a flag noting that this operation is dangerous and > > > > > > > > > can't be synced with something memory management depends on tried to > > > > > > > > > contain this as much as possible, but Daniel still pretty clearly > > > > > > > > > rejected it (for good reasons I think). > > > > > > > > > > > > > > > > > > > > > > > > > > > > > We could ofc invent a completely different data-type that abstracts > > > > > > > > > > the synchronization the scheduler needs in the long-running case, or > > > > > > > > > > each driver could hack something up, like sleeping in the > > > > > > > > > > prepare_job() or run_job() callback for throttling, but those waits > > > > > > > > > > should still be annotated in one way or annotated one way or another > > > > > > > > > > (and probably in a similar way across drivers) to make sure we don't > > > > > > > > > > do anything bad. > > > > > > > > > > > > > > > > > > > > So any suggestions as to what would be the better solution here > > > > > > > > > > would be appreciated. > > > > > > > > > > > > > > > > > > Mhm, do we really the the GPU scheduler for that? > > > > > > > > > > > > > > > > > > > > > > > I think we need to solve this within the DRM scheduler one way or > > > > > > > another. > > > > > > > > > > > > Yeah so if we conclude that the queue really must be bottomless then I > > > > > > agree drm-sched should help out sort out the mess. Because I'm guessing > > > > > > that every driver will have this issue. But that's a big if. > > > > > > > > > > > > I guess if we teach the drm scheduler that some jobs are fairly endless > > > > > > then maybe it wouldn't be too far-fetched to also teach it to wait for a > > > > > > previous one to finish (but not with the dma_fence that preempts, which we > > > > > > put into the dma_resv for memory management, but some other struct > > > > > > completion). The scheduler already has a concept of not stuffing too much > > > > > > stuff into the same queue after all, so this should fit? > > > > > > > > > > See above, exact same situation as spinning on flow controling the ring, > > > > > this IMO absolutely breaks the dma-fence rules. IMO the correct solution > > > > > is to have a DRM that doesn't export dma-fences, this is exactly what > > > > > this series does as if we try to, boom lockdep / warn on blow up. > > > > > > > > I dont think it's impossible to do this correctly, but definitely very, > > > > very hard. Which is why neither Christian nor me like the idea :-) > > > > > > > > Essentially you'd have to make sure that any indefinite way will still > > > > react to drm_sched_job, so that you're not holding up a gt reset or > > > > anything like that, but only ever hold up forward progress for this > > > > specific scheduler/drm_sched_entity. Which you can do as long (and again, > > > > another hugely tricky detail) you still obey the preempt-ctx dma_fence and > > > > manage to preempt the underlying long-running ctx even when the drm/sched > > > > is stuck waiting for an indefinite fence (like waiting for ringspace or > > > > something like that). > > > > > > > > So I don't think it's impossible, but very far away from "a good idea" :-) > > > > > > > > Hence to proposal to bail out of this entire mess by throwing EWOULDBLCK > > > > back to userspace directly from the ioctl function, where you still can do > > > > that without breaking any dma_fence rules. Or if it's not a case that > > > > matters in practice, simply block in the ioctl handler instead of > > > > returning EWOULDBLCK. > > > > > > Returning EWOULDBLCK on a full ring is reasonsible I guess but again > > > without returning a fence in run job the TDR can't be used for clean up > > > on LR entities which will result in duplicate code open coded by each > > > driver. Same goes for checking ring full in exec. > > > > > > How about this: > > > - We mark xe_hw_fence as LR to ensure it can't be exported, return this > > > in run_job which gives flow control on the ring + the handy TDR > > > functionality > > > - When a scheduler is marked as LR, we do not generate finished fences > > > for jobs > > > - We heavily, heavily scrutinize any usage of the LR fence flag going > > > foward > > > - We document all of this very loudly > > > > > > Is this reasonable? > > > > I'm not seeing why it's needed? If you're worried about TDR duplication > > then I think we need something else. Because for long-running ctx we never > > have a timeout of the ctx itself (by definition). The only thing we time > > out on is the preempt, so I guess what could be done: > > - have the minimal scaffolding to support the preempt-ctx fence in > > drm_sched_entity > > - when the preempt ctx fence enables signalling a) callback to the driver > > to start the preempt (which should signal the fence) b) start a timer, > > which should catch if the preempt takes too long > > The GuC does this for us, no need. > > > - if the timeout first (importantly we only enable that when the > > preemption is trigger, not by default), kick of the normal drm/sched tdr > > flow. maybe needs some adjustements in case there's different handling > > needed for when a preemption times out compared to just a job timing out > > > > The GuC imforms us this and yea we kick the TDR. You might still need the kernel fallback when guc dies? But yeah not sure how much similarity is then left with the end-of-batch timed-out. > > I think this might make sense for sharing timeout handling needs for > > long-running context. What you proposed I don't really follow why it > > should exist, because that kind of timeout handling should not ever happen > > for long-running jobs. > > We use just the TDR a as single cleanup point for all entities. In the > case of a LR entity this occurs if the GuC issues a reset on the > entity (liekly preempt timeout), the entity takes a non-recoverable page > fail, or the entity to the root cause of a GT reset. The pending job > list here is handy, that why I wanted to return xe_hw_fence in run_job > to hold the job in the scheduler pending list. The doesn't TDR won't > fire if the pending list is empty. > > Based on what you are saying my new proposal: > > 1. Write ring in exec for LR jobs, return -EWOULDBLCK if no space in > ring > 2. Return NULL in run_job (or alternatively a signaled fence) > 3. Have specical cased cleanup flow for LR entites (not the TDR, rather > likely a different worker we kick owned by the xe_engine). > 4. Document this some that this how drivers are expected to do LR > workloads plus DRM scheduler > > 1 & 3 are pretty clear duplicates of code but I can live with that if > I can get Ack on the plan + move on. The coding will not be all that > difficult either, I am just being difficult. In the is probably a 100ish > lines of code. For 1 I guess you could also write the ring stuff for end-of-batch code from the ioctl (and then just block for the ring to advance enough if needed). That is how we had i915-gem operate before i915-scheduler happened. For 3 I'm not sure there's really that much to share, the end-of-batch and preempt-ctx dma_fence are just rather different. I'm not too clear on why LR needs a special cleanup flow, isn't it roughly the same: - stop drm/sched from pushing in new jobs - preempt the ctx to kill it With xe I don't think we ever want to let existing jobs complete, neither for legacy nor for lr ctx. > What do you think Daniel, seem like a reasonable plan? Yeah my comments above are just minor side questions, I think this is roughly what we want. Plus/minus polish details of how much code sharing between legacy/lr ctx in xe and for lr with other drivers makes sense or not. I think the in-fences handling (because that ties into memory management when eviction and restoring a vm) is the big part which ideally has as much shared as possible. Which I think this achieves. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
