Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On Wed, Aug 18, 2021 at 2:28 AM John Harrison wrote: > On 8/9/2021 23:36, Daniel Vetter wrote: > > On Mon, Aug 09, 2021 at 04:12:52PM -0700, John Harrison wrote: > >> On 8/6/2021 12:46, Daniel Vetter wrote: > >>> Seen this fly by and figured I dropped a few thoughts in here. At the > >>> likely cost of looking a bit out of whack :-) > >>> > >>> On Fri, Aug 6, 2021 at 8:01 PM John Harrison > >>> wrote: > On 8/2/2021 02:40, Tvrtko Ursulin wrote: > > On 30/07/2021 19:13, John Harrison wrote: > >> On 7/30/2021 02:49, Tvrtko Ursulin wrote: > >>> On 30/07/2021 01:13, John Harrison wrote: > On 7/28/2021 17:34, Matthew Brost wrote: > > If an engine associated with a context does not have a heartbeat, > > ban it > > immediately. This is needed for GuC submission as a idle pulse > > doesn't > > kick the context off the hardware where it then can check for a > > heartbeat and ban the context. > >>> Pulse, that is a request with I915_PRIORITY_BARRIER, does not > >>> preempt a running normal priority context? > >>> > >>> Why does it matter then whether or not heartbeats are enabled - when > >>> heartbeat just ends up sending the same engine pulse (eventually, > >>> with raising priority)? > >> The point is that the pulse is pointless. See the rest of my comments > >> below, specifically "the context will get resubmitted to the hardware > >> after the pulse completes". To re-iterate... > >> > >> Yes, it preempts the context. Yes, it does so whether heartbeats are > >> enabled or not. But so what? Who cares? You have preempted a context. > >> It is no longer running on the hardware. BUT IT IS STILL A VALID > >> CONTEXT. > > It is valid yes, and it even may be the current ABI so another > > question is whether it is okay to change that. > > > >> The backend scheduler will just resubmit it to the hardware as soon > >> as the pulse completes. The only reason this works at all is because > >> of the horrid hack in the execlist scheduler's back end > >> implementation (in __execlists_schedule_in): > >>if (unlikely(intel_context_is_closed(ce) && > >> !intel_engine_has_heartbeat(engine))) > >>intel_context_set_banned(ce); > > Right, is the above code then needed with this patch - when ban is > > immediately applied on the higher level? > > > >> The actual back end scheduler is saying "Is this a zombie context? Is > >> the heartbeat disabled? Then ban it". No other scheduler backend is > >> going to have knowledge of zombie context status or of the heartbeat > >> status. Nor are they going to call back into the higher levels of the > >> i915 driver to trigger a ban operation. Certainly a hardware > >> implemented scheduler is not going to be looking at private i915 > >> driver information to decide whether to submit a context or whether > >> to tell the OS to kill it off instead. > >> > >> For persistence to work with a hardware scheduler (or a non-Intel > >> specific scheduler such as the DRM one), the handling of zombie > >> contexts, banning, etc. *must* be done entirely in the front end. It > >> cannot rely on any backend hacks. That means you can't rely on any > >> fancy behaviour of pulses. > >> > >> If you want to ban a context then you must explicitly ban that > >> context. If you want to ban it at some later point then you need to > >> track it at the top level as a zombie and then explicitly ban that > >> zombie at whatever later point. > > I am still trying to understand it all. If I go by the commit message: > > > > """ > > This is needed for GuC submission as a idle pulse doesn't > > kick the context off the hardware where it then can check for a > > heartbeat and ban the context. > > """ > > > > That did not explain things for me. Sentence does not appear to make > > sense. Now, it seems "kick off the hardware" is meant as revoke and > > not just preempt. Which is fine, perhaps just needs to be written more > > explicitly. But the part of checking for heartbeat after idle pulse > > does not compute for me. It is the heartbeat which emits idle pulses, > > not idle pulse emitting heartbeats. > I am in agreement that the commit message is confusing and does not > explain either the problem or the solution. > > > > But anyway, I can buy the handling at the front end story completely. > > It makes sense. We just need to agree that a) it is okay to change the > > ABI and b) remove the backend check from execlists if it is not needed > > any longer. > > > > And if ABI change is okay then commit message needs to talk about it > > loudly and clearly. > I don't think we have a choice. The current ABI is
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On 8/9/2021 23:36, Daniel Vetter wrote: On Mon, Aug 09, 2021 at 04:12:52PM -0700, John Harrison wrote: On 8/6/2021 12:46, Daniel Vetter wrote: Seen this fly by and figured I dropped a few thoughts in here. At the likely cost of looking a bit out of whack :-) On Fri, Aug 6, 2021 at 8:01 PM John Harrison wrote: On 8/2/2021 02:40, Tvrtko Ursulin wrote: On 30/07/2021 19:13, John Harrison wrote: On 7/30/2021 02:49, Tvrtko Ursulin wrote: On 30/07/2021 01:13, John Harrison wrote: On 7/28/2021 17:34, Matthew Brost wrote: If an engine associated with a context does not have a heartbeat, ban it immediately. This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. Pulse, that is a request with I915_PRIORITY_BARRIER, does not preempt a running normal priority context? Why does it matter then whether or not heartbeats are enabled - when heartbeat just ends up sending the same engine pulse (eventually, with raising priority)? The point is that the pulse is pointless. See the rest of my comments below, specifically "the context will get resubmitted to the hardware after the pulse completes". To re-iterate... Yes, it preempts the context. Yes, it does so whether heartbeats are enabled or not. But so what? Who cares? You have preempted a context. It is no longer running on the hardware. BUT IT IS STILL A VALID CONTEXT. It is valid yes, and it even may be the current ABI so another question is whether it is okay to change that. The backend scheduler will just resubmit it to the hardware as soon as the pulse completes. The only reason this works at all is because of the horrid hack in the execlist scheduler's back end implementation (in __execlists_schedule_in): if (unlikely(intel_context_is_closed(ce) && !intel_engine_has_heartbeat(engine))) intel_context_set_banned(ce); Right, is the above code then needed with this patch - when ban is immediately applied on the higher level? The actual back end scheduler is saying "Is this a zombie context? Is the heartbeat disabled? Then ban it". No other scheduler backend is going to have knowledge of zombie context status or of the heartbeat status. Nor are they going to call back into the higher levels of the i915 driver to trigger a ban operation. Certainly a hardware implemented scheduler is not going to be looking at private i915 driver information to decide whether to submit a context or whether to tell the OS to kill it off instead. For persistence to work with a hardware scheduler (or a non-Intel specific scheduler such as the DRM one), the handling of zombie contexts, banning, etc. *must* be done entirely in the front end. It cannot rely on any backend hacks. That means you can't rely on any fancy behaviour of pulses. If you want to ban a context then you must explicitly ban that context. If you want to ban it at some later point then you need to track it at the top level as a zombie and then explicitly ban that zombie at whatever later point. I am still trying to understand it all. If I go by the commit message: """ This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. """ That did not explain things for me. Sentence does not appear to make sense. Now, it seems "kick off the hardware" is meant as revoke and not just preempt. Which is fine, perhaps just needs to be written more explicitly. But the part of checking for heartbeat after idle pulse does not compute for me. It is the heartbeat which emits idle pulses, not idle pulse emitting heartbeats. I am in agreement that the commit message is confusing and does not explain either the problem or the solution. But anyway, I can buy the handling at the front end story completely. It makes sense. We just need to agree that a) it is okay to change the ABI and b) remove the backend check from execlists if it is not needed any longer. And if ABI change is okay then commit message needs to talk about it loudly and clearly. I don't think we have a choice. The current ABI is not and cannot ever be compatible with any scheduler external to i915. It cannot be implemented with a hardware scheduler such as the GuC and it cannot be implemented with an external software scheduler such as the DRM one. So generally on linux we implement helper libraries, which means massive flexibility everywhere. https://blog.ffwll.ch/2016/12/midlayers-once-more-with-feeling.html So it shouldn't be an insurmountable problem to make this happen even with drm/scheduler, we can patch it up. Whether that's justified is another question. Helper libraries won't work with a hardware scheduler. Hm I guess I misunderstood then what exactly the hold-up is. This entire discussion feels at least a bit like "heartbeat is unchangeable and guc must fit", which is pretty much the midlayer
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On Mon, Aug 09, 2021 at 04:12:52PM -0700, John Harrison wrote: > On 8/6/2021 12:46, Daniel Vetter wrote: > > Seen this fly by and figured I dropped a few thoughts in here. At the > > likely cost of looking a bit out of whack :-) > > > > On Fri, Aug 6, 2021 at 8:01 PM John Harrison > > wrote: > > > On 8/2/2021 02:40, Tvrtko Ursulin wrote: > > > > On 30/07/2021 19:13, John Harrison wrote: > > > > > On 7/30/2021 02:49, Tvrtko Ursulin wrote: > > > > > > On 30/07/2021 01:13, John Harrison wrote: > > > > > > > On 7/28/2021 17:34, Matthew Brost wrote: > > > > > > > > If an engine associated with a context does not have a > > > > > > > > heartbeat, > > > > > > > > ban it > > > > > > > > immediately. This is needed for GuC submission as a idle pulse > > > > > > > > doesn't > > > > > > > > kick the context off the hardware where it then can check for a > > > > > > > > heartbeat and ban the context. > > > > > > Pulse, that is a request with I915_PRIORITY_BARRIER, does not > > > > > > preempt a running normal priority context? > > > > > > > > > > > > Why does it matter then whether or not heartbeats are enabled - when > > > > > > heartbeat just ends up sending the same engine pulse (eventually, > > > > > > with raising priority)? > > > > > The point is that the pulse is pointless. See the rest of my comments > > > > > below, specifically "the context will get resubmitted to the hardware > > > > > after the pulse completes". To re-iterate... > > > > > > > > > > Yes, it preempts the context. Yes, it does so whether heartbeats are > > > > > enabled or not. But so what? Who cares? You have preempted a context. > > > > > It is no longer running on the hardware. BUT IT IS STILL A VALID > > > > > CONTEXT. > > > > It is valid yes, and it even may be the current ABI so another > > > > question is whether it is okay to change that. > > > > > > > > > The backend scheduler will just resubmit it to the hardware as soon > > > > > as the pulse completes. The only reason this works at all is because > > > > > of the horrid hack in the execlist scheduler's back end > > > > > implementation (in __execlists_schedule_in): > > > > > if (unlikely(intel_context_is_closed(ce) && > > > > >!intel_engine_has_heartbeat(engine))) > > > > > intel_context_set_banned(ce); > > > > Right, is the above code then needed with this patch - when ban is > > > > immediately applied on the higher level? > > > > > > > > > The actual back end scheduler is saying "Is this a zombie context? Is > > > > > the heartbeat disabled? Then ban it". No other scheduler backend is > > > > > going to have knowledge of zombie context status or of the heartbeat > > > > > status. Nor are they going to call back into the higher levels of the > > > > > i915 driver to trigger a ban operation. Certainly a hardware > > > > > implemented scheduler is not going to be looking at private i915 > > > > > driver information to decide whether to submit a context or whether > > > > > to tell the OS to kill it off instead. > > > > > > > > > > For persistence to work with a hardware scheduler (or a non-Intel > > > > > specific scheduler such as the DRM one), the handling of zombie > > > > > contexts, banning, etc. *must* be done entirely in the front end. It > > > > > cannot rely on any backend hacks. That means you can't rely on any > > > > > fancy behaviour of pulses. > > > > > > > > > > If you want to ban a context then you must explicitly ban that > > > > > context. If you want to ban it at some later point then you need to > > > > > track it at the top level as a zombie and then explicitly ban that > > > > > zombie at whatever later point. > > > > I am still trying to understand it all. If I go by the commit message: > > > > > > > > """ > > > > This is needed for GuC submission as a idle pulse doesn't > > > > kick the context off the hardware where it then can check for a > > > > heartbeat and ban the context. > > > > """ > > > > > > > > That did not explain things for me. Sentence does not appear to make > > > > sense. Now, it seems "kick off the hardware" is meant as revoke and > > > > not just preempt. Which is fine, perhaps just needs to be written more > > > > explicitly. But the part of checking for heartbeat after idle pulse > > > > does not compute for me. It is the heartbeat which emits idle pulses, > > > > not idle pulse emitting heartbeats. > > > I am in agreement that the commit message is confusing and does not > > > explain either the problem or the solution. > > > > > > > > > > > > > > But anyway, I can buy the handling at the front end story completely. > > > > It makes sense. We just need to agree that a) it is okay to change the > > > > ABI and b) remove the backend check from execlists if it is not needed > > > > any longer. > > > > > > > > And if ABI change is okay then commit message needs to talk about it > > > > loudly and clearly. > > > I don't think we have a choice. The current ABI is not and
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On 8/6/2021 12:46, Daniel Vetter wrote: Seen this fly by and figured I dropped a few thoughts in here. At the likely cost of looking a bit out of whack :-) On Fri, Aug 6, 2021 at 8:01 PM John Harrison wrote: On 8/2/2021 02:40, Tvrtko Ursulin wrote: On 30/07/2021 19:13, John Harrison wrote: On 7/30/2021 02:49, Tvrtko Ursulin wrote: On 30/07/2021 01:13, John Harrison wrote: On 7/28/2021 17:34, Matthew Brost wrote: If an engine associated with a context does not have a heartbeat, ban it immediately. This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. Pulse, that is a request with I915_PRIORITY_BARRIER, does not preempt a running normal priority context? Why does it matter then whether or not heartbeats are enabled - when heartbeat just ends up sending the same engine pulse (eventually, with raising priority)? The point is that the pulse is pointless. See the rest of my comments below, specifically "the context will get resubmitted to the hardware after the pulse completes". To re-iterate... Yes, it preempts the context. Yes, it does so whether heartbeats are enabled or not. But so what? Who cares? You have preempted a context. It is no longer running on the hardware. BUT IT IS STILL A VALID CONTEXT. It is valid yes, and it even may be the current ABI so another question is whether it is okay to change that. The backend scheduler will just resubmit it to the hardware as soon as the pulse completes. The only reason this works at all is because of the horrid hack in the execlist scheduler's back end implementation (in __execlists_schedule_in): if (unlikely(intel_context_is_closed(ce) && !intel_engine_has_heartbeat(engine))) intel_context_set_banned(ce); Right, is the above code then needed with this patch - when ban is immediately applied on the higher level? The actual back end scheduler is saying "Is this a zombie context? Is the heartbeat disabled? Then ban it". No other scheduler backend is going to have knowledge of zombie context status or of the heartbeat status. Nor are they going to call back into the higher levels of the i915 driver to trigger a ban operation. Certainly a hardware implemented scheduler is not going to be looking at private i915 driver information to decide whether to submit a context or whether to tell the OS to kill it off instead. For persistence to work with a hardware scheduler (or a non-Intel specific scheduler such as the DRM one), the handling of zombie contexts, banning, etc. *must* be done entirely in the front end. It cannot rely on any backend hacks. That means you can't rely on any fancy behaviour of pulses. If you want to ban a context then you must explicitly ban that context. If you want to ban it at some later point then you need to track it at the top level as a zombie and then explicitly ban that zombie at whatever later point. I am still trying to understand it all. If I go by the commit message: """ This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. """ That did not explain things for me. Sentence does not appear to make sense. Now, it seems "kick off the hardware" is meant as revoke and not just preempt. Which is fine, perhaps just needs to be written more explicitly. But the part of checking for heartbeat after idle pulse does not compute for me. It is the heartbeat which emits idle pulses, not idle pulse emitting heartbeats. I am in agreement that the commit message is confusing and does not explain either the problem or the solution. But anyway, I can buy the handling at the front end story completely. It makes sense. We just need to agree that a) it is okay to change the ABI and b) remove the backend check from execlists if it is not needed any longer. And if ABI change is okay then commit message needs to talk about it loudly and clearly. I don't think we have a choice. The current ABI is not and cannot ever be compatible with any scheduler external to i915. It cannot be implemented with a hardware scheduler such as the GuC and it cannot be implemented with an external software scheduler such as the DRM one. So generally on linux we implement helper libraries, which means massive flexibility everywhere. https://blog.ffwll.ch/2016/12/midlayers-once-more-with-feeling.html So it shouldn't be an insurmountable problem to make this happen even with drm/scheduler, we can patch it up. Whether that's justified is another question. Helper libraries won't work with a hardware scheduler. My view is that any implementation involving knowledge of the heartbeat is fundamentally broken. According to Daniel Vetter, the DRM ABI on this subject is that an actively executing context should persist until the DRM file handle is closed. That seems like a much more plausible and simple ABI
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
Seen this fly by and figured I dropped a few thoughts in here. At the likely cost of looking a bit out of whack :-) On Fri, Aug 6, 2021 at 8:01 PM John Harrison wrote: > On 8/2/2021 02:40, Tvrtko Ursulin wrote: > > On 30/07/2021 19:13, John Harrison wrote: > >> On 7/30/2021 02:49, Tvrtko Ursulin wrote: > >>> On 30/07/2021 01:13, John Harrison wrote: > On 7/28/2021 17:34, Matthew Brost wrote: > > If an engine associated with a context does not have a heartbeat, > > ban it > > immediately. This is needed for GuC submission as a idle pulse > > doesn't > > kick the context off the hardware where it then can check for a > > heartbeat and ban the context. > >>> > >>> Pulse, that is a request with I915_PRIORITY_BARRIER, does not > >>> preempt a running normal priority context? > >>> > >>> Why does it matter then whether or not heartbeats are enabled - when > >>> heartbeat just ends up sending the same engine pulse (eventually, > >>> with raising priority)? > >> The point is that the pulse is pointless. See the rest of my comments > >> below, specifically "the context will get resubmitted to the hardware > >> after the pulse completes". To re-iterate... > >> > >> Yes, it preempts the context. Yes, it does so whether heartbeats are > >> enabled or not. But so what? Who cares? You have preempted a context. > >> It is no longer running on the hardware. BUT IT IS STILL A VALID > >> CONTEXT. > > > > It is valid yes, and it even may be the current ABI so another > > question is whether it is okay to change that. > > > >> The backend scheduler will just resubmit it to the hardware as soon > >> as the pulse completes. The only reason this works at all is because > >> of the horrid hack in the execlist scheduler's back end > >> implementation (in __execlists_schedule_in): > >> if (unlikely(intel_context_is_closed(ce) && > >> !intel_engine_has_heartbeat(engine))) > >> intel_context_set_banned(ce); > > > > Right, is the above code then needed with this patch - when ban is > > immediately applied on the higher level? > > > >> The actual back end scheduler is saying "Is this a zombie context? Is > >> the heartbeat disabled? Then ban it". No other scheduler backend is > >> going to have knowledge of zombie context status or of the heartbeat > >> status. Nor are they going to call back into the higher levels of the > >> i915 driver to trigger a ban operation. Certainly a hardware > >> implemented scheduler is not going to be looking at private i915 > >> driver information to decide whether to submit a context or whether > >> to tell the OS to kill it off instead. > >> > >> For persistence to work with a hardware scheduler (or a non-Intel > >> specific scheduler such as the DRM one), the handling of zombie > >> contexts, banning, etc. *must* be done entirely in the front end. It > >> cannot rely on any backend hacks. That means you can't rely on any > >> fancy behaviour of pulses. > >> > >> If you want to ban a context then you must explicitly ban that > >> context. If you want to ban it at some later point then you need to > >> track it at the top level as a zombie and then explicitly ban that > >> zombie at whatever later point. > > > > I am still trying to understand it all. If I go by the commit message: > > > > """ > > This is needed for GuC submission as a idle pulse doesn't > > kick the context off the hardware where it then can check for a > > heartbeat and ban the context. > > """ > > > > That did not explain things for me. Sentence does not appear to make > > sense. Now, it seems "kick off the hardware" is meant as revoke and > > not just preempt. Which is fine, perhaps just needs to be written more > > explicitly. But the part of checking for heartbeat after idle pulse > > does not compute for me. It is the heartbeat which emits idle pulses, > > not idle pulse emitting heartbeats. > I am in agreement that the commit message is confusing and does not > explain either the problem or the solution. > > > > > > > > But anyway, I can buy the handling at the front end story completely. > > It makes sense. We just need to agree that a) it is okay to change the > > ABI and b) remove the backend check from execlists if it is not needed > > any longer. > > > > And if ABI change is okay then commit message needs to talk about it > > loudly and clearly. > I don't think we have a choice. The current ABI is not and cannot ever > be compatible with any scheduler external to i915. It cannot be > implemented with a hardware scheduler such as the GuC and it cannot be > implemented with an external software scheduler such as the DRM one. So generally on linux we implement helper libraries, which means massive flexibility everywhere. https://blog.ffwll.ch/2016/12/midlayers-once-more-with-feeling.html So it shouldn't be an insurmountable problem to make this happen even with drm/scheduler, we can patch it up. Whether that's justified is another question. >
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On 8/2/2021 02:40, Tvrtko Ursulin wrote: On 30/07/2021 19:13, John Harrison wrote: On 7/30/2021 02:49, Tvrtko Ursulin wrote: On 30/07/2021 01:13, John Harrison wrote: On 7/28/2021 17:34, Matthew Brost wrote: If an engine associated with a context does not have a heartbeat, ban it immediately. This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. Pulse, that is a request with I915_PRIORITY_BARRIER, does not preempt a running normal priority context? Why does it matter then whether or not heartbeats are enabled - when heartbeat just ends up sending the same engine pulse (eventually, with raising priority)? The point is that the pulse is pointless. See the rest of my comments below, specifically "the context will get resubmitted to the hardware after the pulse completes". To re-iterate... Yes, it preempts the context. Yes, it does so whether heartbeats are enabled or not. But so what? Who cares? You have preempted a context. It is no longer running on the hardware. BUT IT IS STILL A VALID CONTEXT. It is valid yes, and it even may be the current ABI so another question is whether it is okay to change that. The backend scheduler will just resubmit it to the hardware as soon as the pulse completes. The only reason this works at all is because of the horrid hack in the execlist scheduler's back end implementation (in __execlists_schedule_in): if (unlikely(intel_context_is_closed(ce) && !intel_engine_has_heartbeat(engine))) intel_context_set_banned(ce); Right, is the above code then needed with this patch - when ban is immediately applied on the higher level? The actual back end scheduler is saying "Is this a zombie context? Is the heartbeat disabled? Then ban it". No other scheduler backend is going to have knowledge of zombie context status or of the heartbeat status. Nor are they going to call back into the higher levels of the i915 driver to trigger a ban operation. Certainly a hardware implemented scheduler is not going to be looking at private i915 driver information to decide whether to submit a context or whether to tell the OS to kill it off instead. For persistence to work with a hardware scheduler (or a non-Intel specific scheduler such as the DRM one), the handling of zombie contexts, banning, etc. *must* be done entirely in the front end. It cannot rely on any backend hacks. That means you can't rely on any fancy behaviour of pulses. If you want to ban a context then you must explicitly ban that context. If you want to ban it at some later point then you need to track it at the top level as a zombie and then explicitly ban that zombie at whatever later point. I am still trying to understand it all. If I go by the commit message: """ This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. """ That did not explain things for me. Sentence does not appear to make sense. Now, it seems "kick off the hardware" is meant as revoke and not just preempt. Which is fine, perhaps just needs to be written more explicitly. But the part of checking for heartbeat after idle pulse does not compute for me. It is the heartbeat which emits idle pulses, not idle pulse emitting heartbeats. I am in agreement that the commit message is confusing and does not explain either the problem or the solution. But anyway, I can buy the handling at the front end story completely. It makes sense. We just need to agree that a) it is okay to change the ABI and b) remove the backend check from execlists if it is not needed any longer. And if ABI change is okay then commit message needs to talk about it loudly and clearly. I don't think we have a choice. The current ABI is not and cannot ever be compatible with any scheduler external to i915. It cannot be implemented with a hardware scheduler such as the GuC and it cannot be implemented with an external software scheduler such as the DRM one. My view is that any implementation involving knowledge of the heartbeat is fundamentally broken. According to Daniel Vetter, the DRM ABI on this subject is that an actively executing context should persist until the DRM file handle is closed. That seems like a much more plausible and simple ABI than one that says 'if the heartbeat is running then a context will persist forever, if the heartbeat is not running then it will be killed immediately, if the heart was running but then stops running then the context will be killed on the next context switch, ...'. And if I understand it correctly, the current ABI allows a badly written user app to cause a denial of service by leaving contexts permanently running an infinite loop on the hardware even after the app has been killed! How can that ever be considered a good idea?
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On 30/07/2021 19:13, John Harrison wrote: On 7/30/2021 02:49, Tvrtko Ursulin wrote: On 30/07/2021 01:13, John Harrison wrote: On 7/28/2021 17:34, Matthew Brost wrote: If an engine associated with a context does not have a heartbeat, ban it immediately. This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. Pulse, that is a request with I915_PRIORITY_BARRIER, does not preempt a running normal priority context? Why does it matter then whether or not heartbeats are enabled - when heartbeat just ends up sending the same engine pulse (eventually, with raising priority)? The point is that the pulse is pointless. See the rest of my comments below, specifically "the context will get resubmitted to the hardware after the pulse completes". To re-iterate... Yes, it preempts the context. Yes, it does so whether heartbeats are enabled or not. But so what? Who cares? You have preempted a context. It is no longer running on the hardware. BUT IT IS STILL A VALID CONTEXT. It is valid yes, and it even may be the current ABI so another question is whether it is okay to change that. The backend scheduler will just resubmit it to the hardware as soon as the pulse completes. The only reason this works at all is because of the horrid hack in the execlist scheduler's back end implementation (in __execlists_schedule_in): if (unlikely(intel_context_is_closed(ce) && !intel_engine_has_heartbeat(engine))) intel_context_set_banned(ce); Right, is the above code then needed with this patch - when ban is immediately applied on the higher level? The actual back end scheduler is saying "Is this a zombie context? Is the heartbeat disabled? Then ban it". No other scheduler backend is going to have knowledge of zombie context status or of the heartbeat status. Nor are they going to call back into the higher levels of the i915 driver to trigger a ban operation. Certainly a hardware implemented scheduler is not going to be looking at private i915 driver information to decide whether to submit a context or whether to tell the OS to kill it off instead. For persistence to work with a hardware scheduler (or a non-Intel specific scheduler such as the DRM one), the handling of zombie contexts, banning, etc. *must* be done entirely in the front end. It cannot rely on any backend hacks. That means you can't rely on any fancy behaviour of pulses. If you want to ban a context then you must explicitly ban that context. If you want to ban it at some later point then you need to track it at the top level as a zombie and then explicitly ban that zombie at whatever later point. I am still trying to understand it all. If I go by the commit message: """ This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. """ That did not explain things for me. Sentence does not appear to make sense. Now, it seems "kick off the hardware" is meant as revoke and not just preempt. Which is fine, perhaps just needs to be written more explicitly. But the part of checking for heartbeat after idle pulse does not compute for me. It is the heartbeat which emits idle pulses, not idle pulse emitting heartbeats. But anyway, I can buy the handling at the front end story completely. It makes sense. We just need to agree that a) it is okay to change the ABI and b) remove the backend check from execlists if it is not needed any longer. And if ABI change is okay then commit message needs to talk about it loudly and clearly. Or perhaps there is no ABI change? I am not really clear how does setting banned status propagate to the GuC backend. I mean at which point does i915 ends up passing that info to the firmware? Regards, Tvrtko It's worse than this. If the engine in question is an individual physical engine then sending a pulse (with sufficiently high priority) will pre-empt the engine and kick the context off. However, the GuC Why it is different for physical vs virtual, aren't both just schedulable contexts with different engine masks for what GuC is concerned? Oh, is it a matter of needing to send pulses to all engines which comprise a virtual one? It isn't different. It is totally broken for both. It is potentially more broken for virtual engines because of the question of which engine to pulse. But as stated above, the pulse is pointless anyway so the which engine question doesn't even matter. John. scheduler does not have hacks in it to check the state of the heartbeat or whether a context is actually a zombie or not. Thus, the context will get resubmitted to the hardware after the pulse completes and effectively nothing will have happened. I would assume that the DRM scheduler which we are meant to be switching to for execlist as well as
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On Fri, Jul 30, 2021 at 10:49:01AM +0100, Tvrtko Ursulin wrote: > > On 30/07/2021 01:13, John Harrison wrote: > > On 7/28/2021 17:34, Matthew Brost wrote: > > > If an engine associated with a context does not have a heartbeat, ban it > > > immediately. This is needed for GuC submission as a idle pulse doesn't > > > kick the context off the hardware where it then can check for a > > > heartbeat and ban the context. > > Pulse, that is a request with I915_PRIORITY_BARRIER, does not preempt a > running normal priority context? > Yes, in both execlists and GuC submission the contexts gets preempted. With execlists the i915 see the preempt CSB while with GuC submission the GUC sees it. > Why does it matter then whether or not heartbeats are enabled - when > heartbeat just ends up sending the same engine pulse (eventually, with > raising priority)? > With execlists when the request gets resubmitted, there is check if the context is closed and the heartbeat is disabled. If this is true, the context gets banned. See __execlists_schedule_in. With the Guc sense it owns the CSB / resubmission, the heartbeat / closed check doesn't exist to ban the context. > > It's worse than this. If the engine in question is an individual > > physical engine then sending a pulse (with sufficiently high priority) > > will pre-empt the engine and kick the context off. However, the GuC > > Why it is different for physical vs virtual, aren't both just schedulable > contexts with different engine masks for what GuC is concerned? Oh, is it a > matter of needing to send pulses to all engines which comprise a virtual > one? Yes. The whole idle pulse thing is kinda junk. It really makes an assumption that the backend is execlists. We likely have a bit more work here. > > > scheduler does not have hacks in it to check the state of the heartbeat > > or whether a context is actually a zombie or not. Thus, the context will > > get resubmitted to the hardware after the pulse completes and > > effectively nothing will have happened. > > > > I would assume that the DRM scheduler which we are meant to be switching > > to for execlist as well as GuC submission is also unlikely to have hacks > > for zombie contexts and tests for whether the i915 specific heartbeat > > has been disabled since the context became a zombie. So when that switch > > happens, this test will also fail in execlist mode as well as GuC mode. > > > > The choices I see here are to simply remove persistence completely (it > > is a basically a bug that became UAPI because it wasn't caught soon > > enough!) or to implement it in a way that does not require hacks in the > > back end scheduler. Apparently, the DRM scheduler is expected to allow > > zombie contexts to persist until the DRM file handle is closed. So > > presumably we will have to go with option two. > > > > That means flagging a context as being a zombie when it is closed but > > still active. The driver would then add it to a zombie list owned by the > > DRM client object. When that client object is closed, i915 would go > > through the list and genuinely kill all the contexts. No back end > > scheduler hacks required and no intimate knowledge of the i915 heartbeat > > mechanism required either. > > > > John. > > > > > > > > > > This patch also updates intel_engine_has_heartbeat to be a vfunc as we > > > now need to call this function on execlists virtual engines too. > > > > > > Signed-off-by: Matthew Brost > > > --- > > > drivers/gpu/drm/i915/gem/i915_gem_context.c | 5 +++-- > > > drivers/gpu/drm/i915/gt/intel_context_types.h | 2 ++ > > > drivers/gpu/drm/i915/gt/intel_engine.h | 21 ++- > > > .../drm/i915/gt/intel_execlists_submission.c | 14 + > > > .../gpu/drm/i915/gt/uc/intel_guc_submission.c | 6 +- > > > .../gpu/drm/i915/gt/uc/intel_guc_submission.h | 2 -- > > > 6 files changed, 26 insertions(+), 24 deletions(-) > > > > > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c > > > b/drivers/gpu/drm/i915/gem/i915_gem_context.c > > > index 9c3672bac0e2..b8e01c5ba9e5 100644 > > > --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c > > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c > > > @@ -1090,8 +1090,9 @@ static void kill_engines(struct > > > i915_gem_engines *engines, bool ban) > > > */ > > > for_each_gem_engine(ce, engines, it) { > > > struct intel_engine_cs *engine; > > > + bool local_ban = ban || !intel_engine_has_heartbeat(ce->engine); > > In any case (pending me understanding what's really going on there), why > would this check not be in kill_context with currently does this: > > bool ban = (!i915_gem_context_is_persistent(ctx) || > !ctx->i915->params.enable_hangcheck); > ... This gem_context level check, while the other check is per intel_context. We don't have the intel_context here. > kill_engines(pos, ban); > > So whether to ban decision would
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On 7/30/2021 02:49, Tvrtko Ursulin wrote: On 30/07/2021 01:13, John Harrison wrote: On 7/28/2021 17:34, Matthew Brost wrote: If an engine associated with a context does not have a heartbeat, ban it immediately. This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. Pulse, that is a request with I915_PRIORITY_BARRIER, does not preempt a running normal priority context? Why does it matter then whether or not heartbeats are enabled - when heartbeat just ends up sending the same engine pulse (eventually, with raising priority)? The point is that the pulse is pointless. See the rest of my comments below, specifically "the context will get resubmitted to the hardware after the pulse completes". To re-iterate... Yes, it preempts the context. Yes, it does so whether heartbeats are enabled or not. But so what? Who cares? You have preempted a context. It is no longer running on the hardware. BUT IT IS STILL A VALID CONTEXT. The backend scheduler will just resubmit it to the hardware as soon as the pulse completes. The only reason this works at all is because of the horrid hack in the execlist scheduler's back end implementation (in __execlists_schedule_in): if (unlikely(intel_context_is_closed(ce) && !intel_engine_has_heartbeat(engine))) intel_context_set_banned(ce); The actual back end scheduler is saying "Is this a zombie context? Is the heartbeat disabled? Then ban it". No other scheduler backend is going to have knowledge of zombie context status or of the heartbeat status. Nor are they going to call back into the higher levels of the i915 driver to trigger a ban operation. Certainly a hardware implemented scheduler is not going to be looking at private i915 driver information to decide whether to submit a context or whether to tell the OS to kill it off instead. For persistence to work with a hardware scheduler (or a non-Intel specific scheduler such as the DRM one), the handling of zombie contexts, banning, etc. *must* be done entirely in the front end. It cannot rely on any backend hacks. That means you can't rely on any fancy behaviour of pulses. If you want to ban a context then you must explicitly ban that context. If you want to ban it at some later point then you need to track it at the top level as a zombie and then explicitly ban that zombie at whatever later point. It's worse than this. If the engine in question is an individual physical engine then sending a pulse (with sufficiently high priority) will pre-empt the engine and kick the context off. However, the GuC Why it is different for physical vs virtual, aren't both just schedulable contexts with different engine masks for what GuC is concerned? Oh, is it a matter of needing to send pulses to all engines which comprise a virtual one? It isn't different. It is totally broken for both. It is potentially more broken for virtual engines because of the question of which engine to pulse. But as stated above, the pulse is pointless anyway so the which engine question doesn't even matter. John. scheduler does not have hacks in it to check the state of the heartbeat or whether a context is actually a zombie or not. Thus, the context will get resubmitted to the hardware after the pulse completes and effectively nothing will have happened. I would assume that the DRM scheduler which we are meant to be switching to for execlist as well as GuC submission is also unlikely to have hacks for zombie contexts and tests for whether the i915 specific heartbeat has been disabled since the context became a zombie. So when that switch happens, this test will also fail in execlist mode as well as GuC mode. The choices I see here are to simply remove persistence completely (it is a basically a bug that became UAPI because it wasn't caught soon enough!) or to implement it in a way that does not require hacks in the back end scheduler. Apparently, the DRM scheduler is expected to allow zombie contexts to persist until the DRM file handle is closed. So presumably we will have to go with option two. That means flagging a context as being a zombie when it is closed but still active. The driver would then add it to a zombie list owned by the DRM client object. When that client object is closed, i915 would go through the list and genuinely kill all the contexts. No back end scheduler hacks required and no intimate knowledge of the i915 heartbeat mechanism required either. John. This patch also updates intel_engine_has_heartbeat to be a vfunc as we now need to call this function on execlists virtual engines too. Signed-off-by: Matthew Brost --- drivers/gpu/drm/i915/gem/i915_gem_context.c | 5 +++-- drivers/gpu/drm/i915/gt/intel_context_types.h | 2 ++ drivers/gpu/drm/i915/gt/intel_engine.h | 21 ++-
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On 30/07/2021 01:13, John Harrison wrote: On 7/28/2021 17:34, Matthew Brost wrote: If an engine associated with a context does not have a heartbeat, ban it immediately. This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. Pulse, that is a request with I915_PRIORITY_BARRIER, does not preempt a running normal priority context? Why does it matter then whether or not heartbeats are enabled - when heartbeat just ends up sending the same engine pulse (eventually, with raising priority)? It's worse than this. If the engine in question is an individual physical engine then sending a pulse (with sufficiently high priority) will pre-empt the engine and kick the context off. However, the GuC Why it is different for physical vs virtual, aren't both just schedulable contexts with different engine masks for what GuC is concerned? Oh, is it a matter of needing to send pulses to all engines which comprise a virtual one? scheduler does not have hacks in it to check the state of the heartbeat or whether a context is actually a zombie or not. Thus, the context will get resubmitted to the hardware after the pulse completes and effectively nothing will have happened. I would assume that the DRM scheduler which we are meant to be switching to for execlist as well as GuC submission is also unlikely to have hacks for zombie contexts and tests for whether the i915 specific heartbeat has been disabled since the context became a zombie. So when that switch happens, this test will also fail in execlist mode as well as GuC mode. The choices I see here are to simply remove persistence completely (it is a basically a bug that became UAPI because it wasn't caught soon enough!) or to implement it in a way that does not require hacks in the back end scheduler. Apparently, the DRM scheduler is expected to allow zombie contexts to persist until the DRM file handle is closed. So presumably we will have to go with option two. That means flagging a context as being a zombie when it is closed but still active. The driver would then add it to a zombie list owned by the DRM client object. When that client object is closed, i915 would go through the list and genuinely kill all the contexts. No back end scheduler hacks required and no intimate knowledge of the i915 heartbeat mechanism required either. John. This patch also updates intel_engine_has_heartbeat to be a vfunc as we now need to call this function on execlists virtual engines too. Signed-off-by: Matthew Brost --- drivers/gpu/drm/i915/gem/i915_gem_context.c | 5 +++-- drivers/gpu/drm/i915/gt/intel_context_types.h | 2 ++ drivers/gpu/drm/i915/gt/intel_engine.h | 21 ++- .../drm/i915/gt/intel_execlists_submission.c | 14 + .../gpu/drm/i915/gt/uc/intel_guc_submission.c | 6 +- .../gpu/drm/i915/gt/uc/intel_guc_submission.h | 2 -- 6 files changed, 26 insertions(+), 24 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c index 9c3672bac0e2..b8e01c5ba9e5 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c @@ -1090,8 +1090,9 @@ static void kill_engines(struct i915_gem_engines *engines, bool ban) */ for_each_gem_engine(ce, engines, it) { struct intel_engine_cs *engine; + bool local_ban = ban || !intel_engine_has_heartbeat(ce->engine); In any case (pending me understanding what's really going on there), why would this check not be in kill_context with currently does this: bool ban = (!i915_gem_context_is_persistent(ctx) || !ctx->i915->params.enable_hangcheck); ... kill_engines(pos, ban); So whether to ban decision would be consolidated to one place. In fact, decision on whether to allow persistent is tied to enable_hangcheck, which also drives hearbeat emission. So perhaps one part of the correct fix is to extend the above (kill_context) ban criteria to include hearbeat values anyway. Otherwise isn't it a simple miss that this check fails to account to hearbeat disablement via sysfs? Regards, Tvrtko - if (ban && intel_context_ban(ce, NULL)) + if (local_ban && intel_context_ban(ce, NULL)) continue; /* @@ -1104,7 +1105,7 @@ static void kill_engines(struct i915_gem_engines *engines, bool ban) engine = active_engine(ce); /* First attempt to gracefully cancel the context */ - if (engine && !__cancel_engine(engine) && ban) + if (engine && !__cancel_engine(engine) && local_ban) /* * If we are unable to send a preemptive pulse to bump * the context from the GPU, we have to resort to a full diff --git a/drivers/gpu/drm/i915/gt/intel_context_types.h
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Check if engine has heartbeat when closing a context
On 7/28/2021 17:34, Matthew Brost wrote: If an engine associated with a context does not have a heartbeat, ban it immediately. This is needed for GuC submission as a idle pulse doesn't kick the context off the hardware where it then can check for a heartbeat and ban the context. It's worse than this. If the engine in question is an individual physical engine then sending a pulse (with sufficiently high priority) will pre-empt the engine and kick the context off. However, the GuC scheduler does not have hacks in it to check the state of the heartbeat or whether a context is actually a zombie or not. Thus, the context will get resubmitted to the hardware after the pulse completes and effectively nothing will have happened. I would assume that the DRM scheduler which we are meant to be switching to for execlist as well as GuC submission is also unlikely to have hacks for zombie contexts and tests for whether the i915 specific heartbeat has been disabled since the context became a zombie. So when that switch happens, this test will also fail in execlist mode as well as GuC mode. The choices I see here are to simply remove persistence completely (it is a basically a bug that became UAPI because it wasn't caught soon enough!) or to implement it in a way that does not require hacks in the back end scheduler. Apparently, the DRM scheduler is expected to allow zombie contexts to persist until the DRM file handle is closed. So presumably we will have to go with option two. That means flagging a context as being a zombie when it is closed but still active. The driver would then add it to a zombie list owned by the DRM client object. When that client object is closed, i915 would go through the list and genuinely kill all the contexts. No back end scheduler hacks required and no intimate knowledge of the i915 heartbeat mechanism required either. John. This patch also updates intel_engine_has_heartbeat to be a vfunc as we now need to call this function on execlists virtual engines too. Signed-off-by: Matthew Brost --- drivers/gpu/drm/i915/gem/i915_gem_context.c | 5 +++-- drivers/gpu/drm/i915/gt/intel_context_types.h | 2 ++ drivers/gpu/drm/i915/gt/intel_engine.h| 21 ++- .../drm/i915/gt/intel_execlists_submission.c | 14 + .../gpu/drm/i915/gt/uc/intel_guc_submission.c | 6 +- .../gpu/drm/i915/gt/uc/intel_guc_submission.h | 2 -- 6 files changed, 26 insertions(+), 24 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c index 9c3672bac0e2..b8e01c5ba9e5 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c @@ -1090,8 +1090,9 @@ static void kill_engines(struct i915_gem_engines *engines, bool ban) */ for_each_gem_engine(ce, engines, it) { struct intel_engine_cs *engine; + bool local_ban = ban || !intel_engine_has_heartbeat(ce->engine); - if (ban && intel_context_ban(ce, NULL)) + if (local_ban && intel_context_ban(ce, NULL)) continue; /* @@ -1104,7 +1105,7 @@ static void kill_engines(struct i915_gem_engines *engines, bool ban) engine = active_engine(ce); /* First attempt to gracefully cancel the context */ - if (engine && !__cancel_engine(engine) && ban) + if (engine && !__cancel_engine(engine) && local_ban) /* * If we are unable to send a preemptive pulse to bump * the context from the GPU, we have to resort to a full diff --git a/drivers/gpu/drm/i915/gt/intel_context_types.h b/drivers/gpu/drm/i915/gt/intel_context_types.h index e54351a170e2..65f2eb2a78e4 100644 --- a/drivers/gpu/drm/i915/gt/intel_context_types.h +++ b/drivers/gpu/drm/i915/gt/intel_context_types.h @@ -55,6 +55,8 @@ struct intel_context_ops { void (*reset)(struct intel_context *ce); void (*destroy)(struct kref *kref); + bool (*has_heartbeat)(const struct intel_engine_cs *engine); + /* virtual engine/context interface */ struct intel_context *(*create_virtual)(struct intel_engine_cs **engine, unsigned int count); diff --git a/drivers/gpu/drm/i915/gt/intel_engine.h b/drivers/gpu/drm/i915/gt/intel_engine.h index c2a5640ae055..1b11a808acc4 100644 --- a/drivers/gpu/drm/i915/gt/intel_engine.h +++ b/drivers/gpu/drm/i915/gt/intel_engine.h @@ -283,28 +283,11 @@ struct intel_context * intel_engine_create_virtual(struct intel_engine_cs **siblings, unsigned int count); -static inline bool -intel_virtual_engine_has_heartbeat(const struct intel_engine_cs *engine) -{ - /* -* For non-GuC submission we expect the back-end to look at the -* heartbeat status of the actual physical engine that the work -