Message ID | 20230210130647.580135-1-tvrtko.ursulin@linux.intel.com (mailing list archive) |
---|---|
Headers | show |
Series | Waitboost drm syncobj waits | expand |
On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> wrote: > > From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > > In i915 we have this concept of "wait boosting" where we give a priority boost > for instance to fences which are actively waited upon from userspace. This has > it's pros and cons and can certainly be discussed at lenght. However fact is > some workloads really like it. > > Problem is that with the arrival of drm syncobj and a new userspace waiting > entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > this mini series really (really) quickly to see if some discussion can be had. > > It adds a concept of "wait count" to dma fence, which is incremented for every > explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > dma_fence_add_callback but from explicit/userspace wait paths). I was thinking about a similar thing, but in the context of dma_fence (or rather sync_file) fd poll()ing. How does the kernel differentiate between "housekeeping" poll()ers that don't want to trigger boost but simply know when to do cleanup, and waiters who are waiting with some urgency. I think we could use EPOLLPRI for this purpose. Not sure how that translates to waits via the syncobj. But I think we want to let userspace give some hint about urgent vs housekeeping waits. Also, on a related topic: https://lwn.net/Articles/868468/ BR, -R > Individual drivers can then inspect this via dma_fence_wait_count() and decide > to wait boost the waits on such fences. > > Again, quickly put together and smoke tested only - no guarantees whatsoever and > I will rely on interested parties to test and report if it even works or how > well. > > v2: > * Small fixups based on CI feedback: > * Handle decrement correctly for already signalled case while adding callback. > * Remove i915 assert which was making sure struct i915_request does not grow. > * Split out the i915 patch into three separate functional changes. > > Tvrtko Ursulin (5): > dma-fence: Track explicit waiters > drm/syncobj: Mark syncobj waits as external waiters > drm/i915: Waitboost external waits > drm/i915: Mark waits as explicit > drm/i915: Wait boost requests waited upon by others > > drivers/dma-buf/dma-fence.c | 102 ++++++++++++++++------ > drivers/gpu/drm/drm_syncobj.c | 6 +- > drivers/gpu/drm/i915/gt/intel_engine_pm.c | 1 - > drivers/gpu/drm/i915/i915_request.c | 13 ++- > include/linux/dma-fence.h | 14 +++ > 5 files changed, 101 insertions(+), 35 deletions(-) > > -- > 2.34.1 >
On Tue, Feb 14, 2023 at 11:14 AM Rob Clark <robdclark@gmail.com> wrote: > > On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > <tvrtko.ursulin@linux.intel.com> wrote: > > > > From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > > > > In i915 we have this concept of "wait boosting" where we give a priority boost > > for instance to fences which are actively waited upon from userspace. This has > > it's pros and cons and can certainly be discussed at lenght. However fact is > > some workloads really like it. > > > > Problem is that with the arrival of drm syncobj and a new userspace waiting > > entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > > this mini series really (really) quickly to see if some discussion can be had. > > > > It adds a concept of "wait count" to dma fence, which is incremented for every > > explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > > dma_fence_add_callback but from explicit/userspace wait paths). > > I was thinking about a similar thing, but in the context of dma_fence > (or rather sync_file) fd poll()ing. How does the kernel differentiate > between "housekeeping" poll()ers that don't want to trigger boost but > simply know when to do cleanup, and waiters who are waiting with some > urgency. I think we could use EPOLLPRI for this purpose. > > Not sure how that translates to waits via the syncobj. But I think we > want to let userspace give some hint about urgent vs housekeeping > waits. So probably the syncobj equiv of this would be to add something along the lines of DRM_SYNCOBJ_WAIT_FLAGS_WAIT_PRI BR, -R > Also, on a related topic: https://lwn.net/Articles/868468/ > > BR, > -R > > > Individual drivers can then inspect this via dma_fence_wait_count() and decide > > to wait boost the waits on such fences. > > > > Again, quickly put together and smoke tested only - no guarantees whatsoever and > > I will rely on interested parties to test and report if it even works or how > > well. > > > > v2: > > * Small fixups based on CI feedback: > > * Handle decrement correctly for already signalled case while adding callback. > > * Remove i915 assert which was making sure struct i915_request does not grow. > > * Split out the i915 patch into three separate functional changes. > > > > Tvrtko Ursulin (5): > > dma-fence: Track explicit waiters > > drm/syncobj: Mark syncobj waits as external waiters > > drm/i915: Waitboost external waits > > drm/i915: Mark waits as explicit > > drm/i915: Wait boost requests waited upon by others > > > > drivers/dma-buf/dma-fence.c | 102 ++++++++++++++++------ > > drivers/gpu/drm/drm_syncobj.c | 6 +- > > drivers/gpu/drm/i915/gt/intel_engine_pm.c | 1 - > > drivers/gpu/drm/i915/i915_request.c | 13 ++- > > include/linux/dma-fence.h | 14 +++ > > 5 files changed, 101 insertions(+), 35 deletions(-) > > > > -- > > 2.34.1 > >
On 14/02/2023 19:14, Rob Clark wrote: > On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > <tvrtko.ursulin@linux.intel.com> wrote: >> >> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> >> >> In i915 we have this concept of "wait boosting" where we give a priority boost >> for instance to fences which are actively waited upon from userspace. This has >> it's pros and cons and can certainly be discussed at lenght. However fact is >> some workloads really like it. >> >> Problem is that with the arrival of drm syncobj and a new userspace waiting >> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up >> this mini series really (really) quickly to see if some discussion can be had. >> >> It adds a concept of "wait count" to dma fence, which is incremented for every >> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like >> dma_fence_add_callback but from explicit/userspace wait paths). > > I was thinking about a similar thing, but in the context of dma_fence > (or rather sync_file) fd poll()ing. How does the kernel differentiate > between "housekeeping" poll()ers that don't want to trigger boost but > simply know when to do cleanup, and waiters who are waiting with some > urgency. I think we could use EPOLLPRI for this purpose. Sounds plausible to allow distinguishing the two. I wasn't aware one can set POLLPRI in pollfd.events but it appears it could be allowed: /* Event types that can be polled for. These bits may be set in `events' to indicate the interesting event types; they will appear in `revents' to indicate the status of the file descriptor. */ #define POLLIN 0x001 /* There is data to read. */ #define POLLPRI 0x002 /* There is urgent data to read. */ #define POLLOUT 0x004 /* Writing now will not block. */ > Not sure how that translates to waits via the syncobj. But I think we > want to let userspace give some hint about urgent vs housekeeping > waits. Probably DRM_SYNCOBJ_WAIT_FLAGS_<something>. Both look easy additions on top of my series. It would be just a matter of dma_fence_add_callback vs dma_fence_add_wait_callback based on flags, as that's how I called the "explicit userspace wait" one. It would require userspace changes to make use of it but that is probably okay, or even preferable, since it makes the thing less of a heuristic. What I don't know however is how feasible is to wire it up with say OpenCL, OpenGL or Vulkan, to allow application writers distinguish between house keeping vs performance sensitive waits. > Also, on a related topic: https://lwn.net/Articles/868468/ Right, I missed that one. One thing to mention is that my motivation here wasn't strictly waits relating to frame presentation but clvk workloads which constantly move between the CPU and GPU. Even outside the compute domain, I think this is a workload characteristic where waitboost in general helps. The concept of deadline could still be used I guess, just setting it for some artificially early value, when the actual time does not exist. But scanning that discussion seems the proposal got bogged down in interactions between mode setting and stuff? Regards, Tvrtko
On Thu, Feb 16, 2023 at 3:19 AM Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> wrote: > > > On 14/02/2023 19:14, Rob Clark wrote: > > On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > > <tvrtko.ursulin@linux.intel.com> wrote: > >> > >> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > >> > >> In i915 we have this concept of "wait boosting" where we give a priority boost > >> for instance to fences which are actively waited upon from userspace. This has > >> it's pros and cons and can certainly be discussed at lenght. However fact is > >> some workloads really like it. > >> > >> Problem is that with the arrival of drm syncobj and a new userspace waiting > >> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > >> this mini series really (really) quickly to see if some discussion can be had. > >> > >> It adds a concept of "wait count" to dma fence, which is incremented for every > >> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > >> dma_fence_add_callback but from explicit/userspace wait paths). > > > > I was thinking about a similar thing, but in the context of dma_fence > > (or rather sync_file) fd poll()ing. How does the kernel differentiate > > between "housekeeping" poll()ers that don't want to trigger boost but > > simply know when to do cleanup, and waiters who are waiting with some > > urgency. I think we could use EPOLLPRI for this purpose. > > Sounds plausible to allow distinguishing the two. > > I wasn't aware one can set POLLPRI in pollfd.events but it appears it could be allowed: > > /* Event types that can be polled for. These bits may be set in `events' > to indicate the interesting event types; they will appear in `revents' > to indicate the status of the file descriptor. */ > #define POLLIN 0x001 /* There is data to read. */ > #define POLLPRI 0x002 /* There is urgent data to read. */ > #define POLLOUT 0x004 /* Writing now will not block. */ > > > Not sure how that translates to waits via the syncobj. But I think we > > want to let userspace give some hint about urgent vs housekeeping > > waits. > > Probably DRM_SYNCOBJ_WAIT_FLAGS_<something>. > > Both look easy additions on top of my series. It would be just a matter of dma_fence_add_callback vs dma_fence_add_wait_callback based on flags, as that's how I called the "explicit userspace wait" one. > > It would require userspace changes to make use of it but that is probably okay, or even preferable, since it makes the thing less of a heuristic. What I don't know however is how feasible is to wire it up with say OpenCL, OpenGL or Vulkan, to allow application writers distinguish between house keeping vs performance sensitive waits. > I think to start with, we consider API level waits as POLLPRI/DRM_SYNCOBJ_WAIT_PRI until someone types up an extension to give the app control. I guess most housekeeping waits will be within the driver. (I could see the argument for making "PRI" the default and having a new flag for non-boosting waits.. but POLLPRI is also some sort of precedent) > > Also, on a related topic: https://lwn.net/Articles/868468/ > > Right, I missed that one. > > One thing to mention is that my motivation here wasn't strictly waits relating to frame presentation but clvk workloads which constantly move between the CPU and GPU. Even outside the compute domain, I think this is a workload characteristic where waitboost in general helps. The concept of deadline could still be used I guess, just setting it for some artificially early value, when the actual time does not exist. But scanning that discussion seems the proposal got bogged down in interactions between mode setting and stuff? > Yeah, it isn't _exactly_ the same thing but it is the same class of problem where GPU stalling on something else sends the freq in the wrong direction. Probably we could consider wait-boosting as simply an immediate deadline to unify the two things. BR, -R > Regards, > > Tvrtko
On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: > On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > <tvrtko.ursulin@linux.intel.com> wrote: > > > > From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > > > > In i915 we have this concept of "wait boosting" where we give a priority boost > > for instance to fences which are actively waited upon from userspace. This has > > it's pros and cons and can certainly be discussed at lenght. However fact is > > some workloads really like it. > > > > Problem is that with the arrival of drm syncobj and a new userspace waiting > > entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > > this mini series really (really) quickly to see if some discussion can be had. > > > > It adds a concept of "wait count" to dma fence, which is incremented for every > > explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > > dma_fence_add_callback but from explicit/userspace wait paths). > > I was thinking about a similar thing, but in the context of dma_fence > (or rather sync_file) fd poll()ing. How does the kernel differentiate > between "housekeeping" poll()ers that don't want to trigger boost but > simply know when to do cleanup, and waiters who are waiting with some > urgency. I think we could use EPOLLPRI for this purpose. > > Not sure how that translates to waits via the syncobj. But I think we > want to let userspace give some hint about urgent vs housekeeping > waits. Should the hint be on the waits, or should the hints be on the executed context? In the end we need some way to quickly ramp-up the frequency to avoid the execution bubbles. waitboost is trying to guess that, but in some cases it guess wrong and waste power. btw, this is something that other drivers might need: https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 Cc: Alex Deucher <alexander.deucher@amd.com> > > Also, on a related topic: https://lwn.net/Articles/868468/ > > BR, > -R > > > Individual drivers can then inspect this via dma_fence_wait_count() and decide > > to wait boost the waits on such fences. > > > > Again, quickly put together and smoke tested only - no guarantees whatsoever and > > I will rely on interested parties to test and report if it even works or how > > well. > > > > v2: > > * Small fixups based on CI feedback: > > * Handle decrement correctly for already signalled case while adding callback. > > * Remove i915 assert which was making sure struct i915_request does not grow. > > * Split out the i915 patch into three separate functional changes. > > > > Tvrtko Ursulin (5): > > dma-fence: Track explicit waiters > > drm/syncobj: Mark syncobj waits as external waiters > > drm/i915: Waitboost external waits > > drm/i915: Mark waits as explicit > > drm/i915: Wait boost requests waited upon by others > > > > drivers/dma-buf/dma-fence.c | 102 ++++++++++++++++------ > > drivers/gpu/drm/drm_syncobj.c | 6 +- > > drivers/gpu/drm/i915/gt/intel_engine_pm.c | 1 - > > drivers/gpu/drm/i915/i915_request.c | 13 ++- > > include/linux/dma-fence.h | 14 +++ > > 5 files changed, 101 insertions(+), 35 deletions(-) > > > > -- > > 2.34.1 > >
On Thu, Feb 16, 2023 at 10:20 AM Rodrigo Vivi <rodrigo.vivi@intel.com> wrote: > > On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: > > On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > > <tvrtko.ursulin@linux.intel.com> wrote: > > > > > > From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > > > > > > In i915 we have this concept of "wait boosting" where we give a priority boost > > > for instance to fences which are actively waited upon from userspace. This has > > > it's pros and cons and can certainly be discussed at lenght. However fact is > > > some workloads really like it. > > > > > > Problem is that with the arrival of drm syncobj and a new userspace waiting > > > entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > > > this mini series really (really) quickly to see if some discussion can be had. > > > > > > It adds a concept of "wait count" to dma fence, which is incremented for every > > > explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > > > dma_fence_add_callback but from explicit/userspace wait paths). > > > > I was thinking about a similar thing, but in the context of dma_fence > > (or rather sync_file) fd poll()ing. How does the kernel differentiate > > between "housekeeping" poll()ers that don't want to trigger boost but > > simply know when to do cleanup, and waiters who are waiting with some > > urgency. I think we could use EPOLLPRI for this purpose. > > > > Not sure how that translates to waits via the syncobj. But I think we > > want to let userspace give some hint about urgent vs housekeeping > > waits. > > Should the hint be on the waits, or should the hints be on the executed > context? I think it should be on the wait, because different waits may be for different purposes. Ideally this could be exposed at the app API level, but I guess first step is to expose it to userspace. BR, -R > In the end we need some way to quickly ramp-up the frequency to avoid > the execution bubbles. > > waitboost is trying to guess that, but in some cases it guess wrong > and waste power. > > btw, this is something that other drivers might need: > > https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 > Cc: Alex Deucher <alexander.deucher@amd.com> > > > > > Also, on a related topic: https://lwn.net/Articles/868468/ > > > > BR, > > -R > > > > > Individual drivers can then inspect this via dma_fence_wait_count() and decide > > > to wait boost the waits on such fences. > > > > > > Again, quickly put together and smoke tested only - no guarantees whatsoever and > > > I will rely on interested parties to test and report if it even works or how > > > well. > > > > > > v2: > > > * Small fixups based on CI feedback: > > > * Handle decrement correctly for already signalled case while adding callback. > > > * Remove i915 assert which was making sure struct i915_request does not grow. > > > * Split out the i915 patch into three separate functional changes. > > > > > > Tvrtko Ursulin (5): > > > dma-fence: Track explicit waiters > > > drm/syncobj: Mark syncobj waits as external waiters > > > drm/i915: Waitboost external waits > > > drm/i915: Mark waits as explicit > > > drm/i915: Wait boost requests waited upon by others > > > > > > drivers/dma-buf/dma-fence.c | 102 ++++++++++++++++------ > > > drivers/gpu/drm/drm_syncobj.c | 6 +- > > > drivers/gpu/drm/i915/gt/intel_engine_pm.c | 1 - > > > drivers/gpu/drm/i915/i915_request.c | 13 ++- > > > include/linux/dma-fence.h | 14 +++ > > > 5 files changed, 101 insertions(+), 35 deletions(-) > > > > > > -- > > > 2.34.1 > > >
On 16/02/2023 18:19, Rodrigo Vivi wrote: > On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: >> On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin >> <tvrtko.ursulin@linux.intel.com> wrote: >>> >>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> >>> >>> In i915 we have this concept of "wait boosting" where we give a priority boost >>> for instance to fences which are actively waited upon from userspace. This has >>> it's pros and cons and can certainly be discussed at lenght. However fact is >>> some workloads really like it. >>> >>> Problem is that with the arrival of drm syncobj and a new userspace waiting >>> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up >>> this mini series really (really) quickly to see if some discussion can be had. >>> >>> It adds a concept of "wait count" to dma fence, which is incremented for every >>> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like >>> dma_fence_add_callback but from explicit/userspace wait paths). >> >> I was thinking about a similar thing, but in the context of dma_fence >> (or rather sync_file) fd poll()ing. How does the kernel differentiate >> between "housekeeping" poll()ers that don't want to trigger boost but >> simply know when to do cleanup, and waiters who are waiting with some >> urgency. I think we could use EPOLLPRI for this purpose. >> >> Not sure how that translates to waits via the syncobj. But I think we >> want to let userspace give some hint about urgent vs housekeeping >> waits. > > Should the hint be on the waits, or should the hints be on the executed > context? > > In the end we need some way to quickly ramp-up the frequency to avoid > the execution bubbles. > > waitboost is trying to guess that, but in some cases it guess wrong > and waste power. Do we have a list of workloads which shows who benefits and who loses from the current implementation of waitboost? > btw, this is something that other drivers might need: > > https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 > Cc: Alex Deucher <alexander.deucher@amd.com> I have several issues with the context hint if it would directly influence frequency selection in the "more power" direction. First of all, assume a context hint would replace the waitboost. Which applications would need to set it to restore the lost performance and how would they set it? Then I don't even think userspace necessarily knows. Think of a layer like OpenCL. It doesn't really know in advance the profile of submissions vs waits. It depends on the CPU vs GPU speed, so hardware generation, and the actual size of the workload which can be influenced by the application (or user) and not the library. The approach also lends itself well for the "arms race" where every application can say "Me me me, I am the most important workload there is!". The last concern is for me shared with the proposal to expose deadlines or high priority waits as explicit uapi knobs. Both come under the "what application told us it will do" category vs what it actually does. So I think it is slightly weaker than basing decisions of waits. The current waitboost is a bit detached from that problem because when we waitboost for flips we _know_ it is an actual framebuffer in the flip chain. When we waitboost for waits we also know someone is waiting. We are not trusting userspace telling us this will be a buffer in the flip chain or that this is a context which will have a certain duty-cycle. But yes, even if the input is truthful, latter is still only a heuristics because nothing says all waits are important. AFAIU it just happened to work well in the past. I do understand I am effectively arguing for more heuristics, which may sound a bit against the common wisdom. This is because in general I think the logic to do the right thing, be it in the driver or in the firmware, can work best if it has a holistic view. Simply put it needs to have more inputs to the decisions it is making. That is what my series is proposing - adding a common signal of "someone in userspace is waiting". What happens with that signal needs not be defined (promised) in the uapi contract. Say you route it to SLPC logic. It doesn't need to do with it what legacy i915 is doing today. It just needs to do something which works best for majority of workloads. It can even ignore it if that works for it. Finally, back to the immediate problem is when people replace the OpenCL NEO driver with clvk that performance tanks. Because former does waits using i915 specific ioctls and so triggers waitboost, latter waits on syncobj so no waitboost and performance is bad. What short term solution can we come up with? Or we say to not use clvk? I also wonder if other Vulkan based stuff is perhaps missing those easy performance gains.. Perhaps strictly speaking Rob's and mine proposal are not mutually exclusive. Yes I could piggy back on his with an "immediate deadline for waits" idea, but they could also be separate concepts if we concluded "someone is waiting" signal is useful to have. Or it takes to long to upstream the full deadline idea. Regards, Tvrtko >> >> Also, on a related topic: https://lwn.net/Articles/868468/ >> >> BR, >> -R >> >>> Individual drivers can then inspect this via dma_fence_wait_count() and decide >>> to wait boost the waits on such fences. >>> >>> Again, quickly put together and smoke tested only - no guarantees whatsoever and >>> I will rely on interested parties to test and report if it even works or how >>> well. >>> >>> v2: >>> * Small fixups based on CI feedback: >>> * Handle decrement correctly for already signalled case while adding callback. >>> * Remove i915 assert which was making sure struct i915_request does not grow. >>> * Split out the i915 patch into three separate functional changes. >>> >>> Tvrtko Ursulin (5): >>> dma-fence: Track explicit waiters >>> drm/syncobj: Mark syncobj waits as external waiters >>> drm/i915: Waitboost external waits >>> drm/i915: Mark waits as explicit >>> drm/i915: Wait boost requests waited upon by others >>> >>> drivers/dma-buf/dma-fence.c | 102 ++++++++++++++++------ >>> drivers/gpu/drm/drm_syncobj.c | 6 +- >>> drivers/gpu/drm/i915/gt/intel_engine_pm.c | 1 - >>> drivers/gpu/drm/i915/i915_request.c | 13 ++- >>> include/linux/dma-fence.h | 14 +++ >>> 5 files changed, 101 insertions(+), 35 deletions(-) >>> >>> -- >>> 2.34.1 >>>
On Fri, Feb 17, 2023 at 4:56 AM Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> wrote: > > > On 16/02/2023 18:19, Rodrigo Vivi wrote: > > On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: > >> On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > >> <tvrtko.ursulin@linux.intel.com> wrote: > >>> > >>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > >>> > >>> In i915 we have this concept of "wait boosting" where we give a priority boost > >>> for instance to fences which are actively waited upon from userspace. This has > >>> it's pros and cons and can certainly be discussed at lenght. However fact is > >>> some workloads really like it. > >>> > >>> Problem is that with the arrival of drm syncobj and a new userspace waiting > >>> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > >>> this mini series really (really) quickly to see if some discussion can be had. > >>> > >>> It adds a concept of "wait count" to dma fence, which is incremented for every > >>> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > >>> dma_fence_add_callback but from explicit/userspace wait paths). > >> > >> I was thinking about a similar thing, but in the context of dma_fence > >> (or rather sync_file) fd poll()ing. How does the kernel differentiate > >> between "housekeeping" poll()ers that don't want to trigger boost but > >> simply know when to do cleanup, and waiters who are waiting with some > >> urgency. I think we could use EPOLLPRI for this purpose. > >> > >> Not sure how that translates to waits via the syncobj. But I think we > >> want to let userspace give some hint about urgent vs housekeeping > >> waits. > > > > Should the hint be on the waits, or should the hints be on the executed > > context? > > > > In the end we need some way to quickly ramp-up the frequency to avoid > > the execution bubbles. > > > > waitboost is trying to guess that, but in some cases it guess wrong > > and waste power. > > Do we have a list of workloads which shows who benefits and who loses > from the current implementation of waitboost? > > btw, this is something that other drivers might need: > > > > https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 > > Cc: Alex Deucher <alexander.deucher@amd.com> > > I have several issues with the context hint if it would directly > influence frequency selection in the "more power" direction. > > First of all, assume a context hint would replace the waitboost. Which > applications would need to set it to restore the lost performance and > how would they set it? > > Then I don't even think userspace necessarily knows. Think of a layer > like OpenCL. It doesn't really know in advance the profile of > submissions vs waits. It depends on the CPU vs GPU speed, so hardware > generation, and the actual size of the workload which can be influenced > by the application (or user) and not the library. > > The approach also lends itself well for the "arms race" where every > application can say "Me me me, I am the most important workload there is!". since there is discussion happening in two places: https://gitlab.freedesktop.org/drm/intel/-/issues/8014#note_1777433 What I think you might want is a ctx boost_mask which lets an app or driver disable certain boost signals/classes. Where fence waits is one class of boost, but hypothetical other signals like touchscreen (or other) input events could be another class of boost. A compute workload might be interested in fence wait boosts but could care less about input events. > The last concern is for me shared with the proposal to expose deadlines > or high priority waits as explicit uapi knobs. Both come under the "what > application told us it will do" category vs what it actually does. So I > think it is slightly weaker than basing decisions of waits. > > The current waitboost is a bit detached from that problem because when > we waitboost for flips we _know_ it is an actual framebuffer in the flip > chain. When we waitboost for waits we also know someone is waiting. We > are not trusting userspace telling us this will be a buffer in the flip > chain or that this is a context which will have a certain duty-cycle. > > But yes, even if the input is truthful, latter is still only a > heuristics because nothing says all waits are important. AFAIU it just > happened to work well in the past. > > I do understand I am effectively arguing for more heuristics, which may > sound a bit against the common wisdom. This is because in general I > think the logic to do the right thing, be it in the driver or in the > firmware, can work best if it has a holistic view. Simply put it needs > to have more inputs to the decisions it is making. > > That is what my series is proposing - adding a common signal of "someone > in userspace is waiting". What happens with that signal needs not be > defined (promised) in the uapi contract. > > Say you route it to SLPC logic. It doesn't need to do with it what > legacy i915 is doing today. It just needs to do something which works > best for majority of workloads. It can even ignore it if that works for it. > > Finally, back to the immediate problem is when people replace the OpenCL > NEO driver with clvk that performance tanks. Because former does waits > using i915 specific ioctls and so triggers waitboost, latter waits on > syncobj so no waitboost and performance is bad. What short term solution > can we come up with? Or we say to not use clvk? I also wonder if other > Vulkan based stuff is perhaps missing those easy performance gains.. > > Perhaps strictly speaking Rob's and mine proposal are not mutually > exclusive. Yes I could piggy back on his with an "immediate deadline for > waits" idea, but they could also be separate concepts if we concluded > "someone is waiting" signal is useful to have. Or it takes to long to > upstream the full deadline idea. Let me re-spin my series and add the syncobj wait flag and i915 bits adapted from your patches.. I think the basic idea of deadlines (which includes "I want it NOW" ;-)) isn't controversial, but the original idea got caught up in some bikeshed (what about compositors that wait on fences in userspace to decide which surfaces to update in the next frame), plus me getting busy and generally not having a good plan for how to leverage this from VM guests (which is becoming increasingly important for CrOS). I think I can build on some ongoing virtgpu fencing improvement work to solve the latter. But now that we have a 2nd use-case for this, it makes sense to respin. BR, -R > Regards, > > Tvrtko > > >> > >> Also, on a related topic: https://lwn.net/Articles/868468/ > >> > >> BR, > >> -R > >> > >>> Individual drivers can then inspect this via dma_fence_wait_count() and decide > >>> to wait boost the waits on such fences. > >>> > >>> Again, quickly put together and smoke tested only - no guarantees whatsoever and > >>> I will rely on interested parties to test and report if it even works or how > >>> well. > >>> > >>> v2: > >>> * Small fixups based on CI feedback: > >>> * Handle decrement correctly for already signalled case while adding callback. > >>> * Remove i915 assert which was making sure struct i915_request does not grow. > >>> * Split out the i915 patch into three separate functional changes. > >>> > >>> Tvrtko Ursulin (5): > >>> dma-fence: Track explicit waiters > >>> drm/syncobj: Mark syncobj waits as external waiters > >>> drm/i915: Waitboost external waits > >>> drm/i915: Mark waits as explicit > >>> drm/i915: Wait boost requests waited upon by others > >>> > >>> drivers/dma-buf/dma-fence.c | 102 ++++++++++++++++------ > >>> drivers/gpu/drm/drm_syncobj.c | 6 +- > >>> drivers/gpu/drm/i915/gt/intel_engine_pm.c | 1 - > >>> drivers/gpu/drm/i915/i915_request.c | 13 ++- > >>> include/linux/dma-fence.h | 14 +++ > >>> 5 files changed, 101 insertions(+), 35 deletions(-) > >>> > >>> -- > >>> 2.34.1 > >>>
On 17/02/2023 14:55, Rob Clark wrote: > On Fri, Feb 17, 2023 at 4:56 AM Tvrtko Ursulin > <tvrtko.ursulin@linux.intel.com> wrote: >> >> >> On 16/02/2023 18:19, Rodrigo Vivi wrote: >>> On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: >>>> On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin >>>> <tvrtko.ursulin@linux.intel.com> wrote: >>>>> >>>>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> >>>>> >>>>> In i915 we have this concept of "wait boosting" where we give a priority boost >>>>> for instance to fences which are actively waited upon from userspace. This has >>>>> it's pros and cons and can certainly be discussed at lenght. However fact is >>>>> some workloads really like it. >>>>> >>>>> Problem is that with the arrival of drm syncobj and a new userspace waiting >>>>> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up >>>>> this mini series really (really) quickly to see if some discussion can be had. >>>>> >>>>> It adds a concept of "wait count" to dma fence, which is incremented for every >>>>> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like >>>>> dma_fence_add_callback but from explicit/userspace wait paths). >>>> >>>> I was thinking about a similar thing, but in the context of dma_fence >>>> (or rather sync_file) fd poll()ing. How does the kernel differentiate >>>> between "housekeeping" poll()ers that don't want to trigger boost but >>>> simply know when to do cleanup, and waiters who are waiting with some >>>> urgency. I think we could use EPOLLPRI for this purpose. >>>> >>>> Not sure how that translates to waits via the syncobj. But I think we >>>> want to let userspace give some hint about urgent vs housekeeping >>>> waits. >>> >>> Should the hint be on the waits, or should the hints be on the executed >>> context? >>> >>> In the end we need some way to quickly ramp-up the frequency to avoid >>> the execution bubbles. >>> >>> waitboost is trying to guess that, but in some cases it guess wrong >>> and waste power. >> >> Do we have a list of workloads which shows who benefits and who loses >> from the current implementation of waitboost? >>> btw, this is something that other drivers might need: >>> >>> https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 >>> Cc: Alex Deucher <alexander.deucher@amd.com> >> >> I have several issues with the context hint if it would directly >> influence frequency selection in the "more power" direction. >> >> First of all, assume a context hint would replace the waitboost. Which >> applications would need to set it to restore the lost performance and >> how would they set it? >> >> Then I don't even think userspace necessarily knows. Think of a layer >> like OpenCL. It doesn't really know in advance the profile of >> submissions vs waits. It depends on the CPU vs GPU speed, so hardware >> generation, and the actual size of the workload which can be influenced >> by the application (or user) and not the library. >> >> The approach also lends itself well for the "arms race" where every >> application can say "Me me me, I am the most important workload there is!". > > since there is discussion happening in two places: > > https://gitlab.freedesktop.org/drm/intel/-/issues/8014#note_1777433 > > What I think you might want is a ctx boost_mask which lets an app or > driver disable certain boost signals/classes. Where fence waits is > one class of boost, but hypothetical other signals like touchscreen > (or other) input events could be another class of boost. A compute > workload might be interested in fence wait boosts but could care less > about input events. I think it can only be apps which could have any chance knowing whether their use of a library is latency sensitive or not. Which means new library extensions and their adoption. So I have some strong reservation that route is feasible. Or we tie with priority which many drivers do. Normal and above gets the boosting and what lowered itself does not (aka SCHED_IDLE/SCHED_BATCH). Related note is that we lack any external control of our scheduling decisions so we really do suck compared to other scheduling domains like CPU and IO etc. >> The last concern is for me shared with the proposal to expose deadlines >> or high priority waits as explicit uapi knobs. Both come under the "what >> application told us it will do" category vs what it actually does. So I >> think it is slightly weaker than basing decisions of waits. >> >> The current waitboost is a bit detached from that problem because when >> we waitboost for flips we _know_ it is an actual framebuffer in the flip >> chain. When we waitboost for waits we also know someone is waiting. We >> are not trusting userspace telling us this will be a buffer in the flip >> chain or that this is a context which will have a certain duty-cycle. >> >> But yes, even if the input is truthful, latter is still only a >> heuristics because nothing says all waits are important. AFAIU it just >> happened to work well in the past. >> >> I do understand I am effectively arguing for more heuristics, which may >> sound a bit against the common wisdom. This is because in general I >> think the logic to do the right thing, be it in the driver or in the >> firmware, can work best if it has a holistic view. Simply put it needs >> to have more inputs to the decisions it is making. >> >> That is what my series is proposing - adding a common signal of "someone >> in userspace is waiting". What happens with that signal needs not be >> defined (promised) in the uapi contract. >> >> Say you route it to SLPC logic. It doesn't need to do with it what >> legacy i915 is doing today. It just needs to do something which works >> best for majority of workloads. It can even ignore it if that works for it. >> >> Finally, back to the immediate problem is when people replace the OpenCL >> NEO driver with clvk that performance tanks. Because former does waits >> using i915 specific ioctls and so triggers waitboost, latter waits on >> syncobj so no waitboost and performance is bad. What short term solution >> can we come up with? Or we say to not use clvk? I also wonder if other >> Vulkan based stuff is perhaps missing those easy performance gains.. >> >> Perhaps strictly speaking Rob's and mine proposal are not mutually >> exclusive. Yes I could piggy back on his with an "immediate deadline for >> waits" idea, but they could also be separate concepts if we concluded >> "someone is waiting" signal is useful to have. Or it takes to long to >> upstream the full deadline idea. > > Let me re-spin my series and add the syncobj wait flag and i915 bits I think wait flag is questionable unless it is inverted to imply waits which can be de-prioritized (again same parallel with SCHED_IDLE/BATCH). Having a flag which "makes things faster" IMO should require elevated privilege (to avoid the "arms race") at which point I fear it quickly becomes uninteresting. > adapted from your patches.. I think the basic idea of deadlines > (which includes "I want it NOW" ;-)) isn't controversial, but the > original idea got caught up in some bikeshed (what about compositors > that wait on fences in userspace to decide which surfaces to update in > the next frame), plus me getting busy and generally not having a good > plan for how to leverage this from VM guests (which is becoming > increasingly important for CrOS). I think I can build on some ongoing > virtgpu fencing improvement work to solve the latter. But now that we > have a 2nd use-case for this, it makes sense to respin. Sure, I was looking at the old version already. It is interesting. But also IMO needs quite a bit more work to approach achieving what is implied from the name of the feature. It would need proper deadline based sched job picking, and even then drm sched is mostly just a frontend. So once past runnable status and jobs handed over to backend, without further driver work it probably wouldn't be very effective past very lightly loaded systems. Then if we fast forward to a world where schedulers perhaps become fully deadline aware (we even had this for i915 few years back) then the question will be does equating waits with immediate deadlines still works. Maybe not too well because we wouldn't have the ability to distinguish between the "someone is waiting" signal from the otherwise propagated deadlines. Regards, Tvrtko
On Fri, Feb 17, 2023 at 8:03 AM Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> wrote: > > > On 17/02/2023 14:55, Rob Clark wrote: > > On Fri, Feb 17, 2023 at 4:56 AM Tvrtko Ursulin > > <tvrtko.ursulin@linux.intel.com> wrote: > >> > >> > >> On 16/02/2023 18:19, Rodrigo Vivi wrote: > >>> On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: > >>>> On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > >>>> <tvrtko.ursulin@linux.intel.com> wrote: > >>>>> > >>>>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > >>>>> > >>>>> In i915 we have this concept of "wait boosting" where we give a priority boost > >>>>> for instance to fences which are actively waited upon from userspace. This has > >>>>> it's pros and cons and can certainly be discussed at lenght. However fact is > >>>>> some workloads really like it. > >>>>> > >>>>> Problem is that with the arrival of drm syncobj and a new userspace waiting > >>>>> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > >>>>> this mini series really (really) quickly to see if some discussion can be had. > >>>>> > >>>>> It adds a concept of "wait count" to dma fence, which is incremented for every > >>>>> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > >>>>> dma_fence_add_callback but from explicit/userspace wait paths). > >>>> > >>>> I was thinking about a similar thing, but in the context of dma_fence > >>>> (or rather sync_file) fd poll()ing. How does the kernel differentiate > >>>> between "housekeeping" poll()ers that don't want to trigger boost but > >>>> simply know when to do cleanup, and waiters who are waiting with some > >>>> urgency. I think we could use EPOLLPRI for this purpose. > >>>> > >>>> Not sure how that translates to waits via the syncobj. But I think we > >>>> want to let userspace give some hint about urgent vs housekeeping > >>>> waits. > >>> > >>> Should the hint be on the waits, or should the hints be on the executed > >>> context? > >>> > >>> In the end we need some way to quickly ramp-up the frequency to avoid > >>> the execution bubbles. > >>> > >>> waitboost is trying to guess that, but in some cases it guess wrong > >>> and waste power. > >> > >> Do we have a list of workloads which shows who benefits and who loses > >> from the current implementation of waitboost? > >>> btw, this is something that other drivers might need: > >>> > >>> https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 > >>> Cc: Alex Deucher <alexander.deucher@amd.com> > >> > >> I have several issues with the context hint if it would directly > >> influence frequency selection in the "more power" direction. > >> > >> First of all, assume a context hint would replace the waitboost. Which > >> applications would need to set it to restore the lost performance and > >> how would they set it? > >> > >> Then I don't even think userspace necessarily knows. Think of a layer > >> like OpenCL. It doesn't really know in advance the profile of > >> submissions vs waits. It depends on the CPU vs GPU speed, so hardware > >> generation, and the actual size of the workload which can be influenced > >> by the application (or user) and not the library. > >> > >> The approach also lends itself well for the "arms race" where every > >> application can say "Me me me, I am the most important workload there is!". > > > > since there is discussion happening in two places: > > > > https://gitlab.freedesktop.org/drm/intel/-/issues/8014#note_1777433 > > > > What I think you might want is a ctx boost_mask which lets an app or > > driver disable certain boost signals/classes. Where fence waits is > > one class of boost, but hypothetical other signals like touchscreen > > (or other) input events could be another class of boost. A compute > > workload might be interested in fence wait boosts but could care less > > about input events. > > I think it can only be apps which could have any chance knowing whether > their use of a library is latency sensitive or not. Which means new > library extensions and their adoption. So I have some strong reservation > that route is feasible. > > Or we tie with priority which many drivers do. Normal and above gets the > boosting and what lowered itself does not (aka SCHED_IDLE/SCHED_BATCH). yeah, that sounds reasonable. > Related note is that we lack any external control of our scheduling > decisions so we really do suck compared to other scheduling domains like > CPU and IO etc. > > >> The last concern is for me shared with the proposal to expose deadlines > >> or high priority waits as explicit uapi knobs. Both come under the "what > >> application told us it will do" category vs what it actually does. So I > >> think it is slightly weaker than basing decisions of waits. > >> > >> The current waitboost is a bit detached from that problem because when > >> we waitboost for flips we _know_ it is an actual framebuffer in the flip > >> chain. When we waitboost for waits we also know someone is waiting. We > >> are not trusting userspace telling us this will be a buffer in the flip > >> chain or that this is a context which will have a certain duty-cycle. > >> > >> But yes, even if the input is truthful, latter is still only a > >> heuristics because nothing says all waits are important. AFAIU it just > >> happened to work well in the past. > >> > >> I do understand I am effectively arguing for more heuristics, which may > >> sound a bit against the common wisdom. This is because in general I > >> think the logic to do the right thing, be it in the driver or in the > >> firmware, can work best if it has a holistic view. Simply put it needs > >> to have more inputs to the decisions it is making. > >> > >> That is what my series is proposing - adding a common signal of "someone > >> in userspace is waiting". What happens with that signal needs not be > >> defined (promised) in the uapi contract. > >> > >> Say you route it to SLPC logic. It doesn't need to do with it what > >> legacy i915 is doing today. It just needs to do something which works > >> best for majority of workloads. It can even ignore it if that works for it. > >> > >> Finally, back to the immediate problem is when people replace the OpenCL > >> NEO driver with clvk that performance tanks. Because former does waits > >> using i915 specific ioctls and so triggers waitboost, latter waits on > >> syncobj so no waitboost and performance is bad. What short term solution > >> can we come up with? Or we say to not use clvk? I also wonder if other > >> Vulkan based stuff is perhaps missing those easy performance gains.. > >> > >> Perhaps strictly speaking Rob's and mine proposal are not mutually > >> exclusive. Yes I could piggy back on his with an "immediate deadline for > >> waits" idea, but they could also be separate concepts if we concluded > >> "someone is waiting" signal is useful to have. Or it takes to long to > >> upstream the full deadline idea. > > > > Let me re-spin my series and add the syncobj wait flag and i915 bits > > I think wait flag is questionable unless it is inverted to imply waits > which can be de-prioritized (again same parallel with SCHED_IDLE/BATCH). > Having a flag which "makes things faster" IMO should require elevated > privilege (to avoid the "arms race") at which point I fear it quickly > becomes uninteresting. I guess you could make the argument in either direction. Making the default behavior ramp up clocks could be a power regression. I also think the "arms race" scenario isn't really as much of a problem as you think. There aren't _that_ many things using the GPU at the same time (compared to # of things using CPU). And a lot of mobile games throttle framerate to avoid draining your battery too quickly (after all, if your battery is dead you can't keep buying loot boxes or whatever). > > adapted from your patches.. I think the basic idea of deadlines > > (which includes "I want it NOW" ;-)) isn't controversial, but the > > original idea got caught up in some bikeshed (what about compositors > > that wait on fences in userspace to decide which surfaces to update in > > the next frame), plus me getting busy and generally not having a good > > plan for how to leverage this from VM guests (which is becoming > > increasingly important for CrOS). I think I can build on some ongoing > > virtgpu fencing improvement work to solve the latter. But now that we > > have a 2nd use-case for this, it makes sense to respin. > > Sure, I was looking at the old version already. It is interesting. But > also IMO needs quite a bit more work to approach achieving what is > implied from the name of the feature. It would need proper deadline > based sched job picking, and even then drm sched is mostly just a > frontend. So once past runnable status and jobs handed over to backend, > without further driver work it probably wouldn't be very effective past > very lightly loaded systems. Yes, but all of that is not part of dma_fence ;-) A pretty common challenging usecase is still the single fullscreen game, where scheduling isn't the problem, but landing at an appropriate GPU freq absolutely is. (UI workloads are perhaps more interesting from a scheduler standpoint, but they generally aren't challenging from a load/freq standpoint.) Fwiw, the original motivation of the series was to implement something akin to i915 pageflip boosting without having to abandon the atomic helpers. (And, I guess it would also let i915 preserve that feature if it switched to atomic helpers.. I'm unsure if there are still other things blocking i915's migration.) > Then if we fast forward to a world where schedulers perhaps become fully > deadline aware (we even had this for i915 few years back) then the > question will be does equating waits with immediate deadlines still > works. Maybe not too well because we wouldn't have the ability to > distinguish between the "someone is waiting" signal from the otherwise > propagated deadlines. Is there any other way to handle a wait boost than expressing it as an ASAP deadline? BR, -R > > Regards, > > Tvrtko
On Fri, Feb 17, 2023 at 09:00:49AM -0800, Rob Clark wrote: > On Fri, Feb 17, 2023 at 8:03 AM Tvrtko Ursulin > <tvrtko.ursulin@linux.intel.com> wrote: > > > > > > On 17/02/2023 14:55, Rob Clark wrote: > > > On Fri, Feb 17, 2023 at 4:56 AM Tvrtko Ursulin > > > <tvrtko.ursulin@linux.intel.com> wrote: > > >> > > >> > > >> On 16/02/2023 18:19, Rodrigo Vivi wrote: > > >>> On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: > > >>>> On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > > >>>> <tvrtko.ursulin@linux.intel.com> wrote: > > >>>>> > > >>>>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > > >>>>> > > >>>>> In i915 we have this concept of "wait boosting" where we give a priority boost > > >>>>> for instance to fences which are actively waited upon from userspace. This has > > >>>>> it's pros and cons and can certainly be discussed at lenght. However fact is > > >>>>> some workloads really like it. > > >>>>> > > >>>>> Problem is that with the arrival of drm syncobj and a new userspace waiting > > >>>>> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > > >>>>> this mini series really (really) quickly to see if some discussion can be had. > > >>>>> > > >>>>> It adds a concept of "wait count" to dma fence, which is incremented for every > > >>>>> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > > >>>>> dma_fence_add_callback but from explicit/userspace wait paths). > > >>>> > > >>>> I was thinking about a similar thing, but in the context of dma_fence > > >>>> (or rather sync_file) fd poll()ing. How does the kernel differentiate > > >>>> between "housekeeping" poll()ers that don't want to trigger boost but > > >>>> simply know when to do cleanup, and waiters who are waiting with some > > >>>> urgency. I think we could use EPOLLPRI for this purpose. > > >>>> > > >>>> Not sure how that translates to waits via the syncobj. But I think we > > >>>> want to let userspace give some hint about urgent vs housekeeping > > >>>> waits. > > >>> > > >>> Should the hint be on the waits, or should the hints be on the executed > > >>> context? > > >>> > > >>> In the end we need some way to quickly ramp-up the frequency to avoid > > >>> the execution bubbles. > > >>> > > >>> waitboost is trying to guess that, but in some cases it guess wrong > > >>> and waste power. > > >> > > >> Do we have a list of workloads which shows who benefits and who loses > > >> from the current implementation of waitboost? > > >>> btw, this is something that other drivers might need: > > >>> > > >>> https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 > > >>> Cc: Alex Deucher <alexander.deucher@amd.com> > > >> > > >> I have several issues with the context hint if it would directly > > >> influence frequency selection in the "more power" direction. > > >> > > >> First of all, assume a context hint would replace the waitboost. Which > > >> applications would need to set it to restore the lost performance and > > >> how would they set it? > > >> > > >> Then I don't even think userspace necessarily knows. Think of a layer > > >> like OpenCL. It doesn't really know in advance the profile of > > >> submissions vs waits. It depends on the CPU vs GPU speed, so hardware > > >> generation, and the actual size of the workload which can be influenced > > >> by the application (or user) and not the library. > > >> > > >> The approach also lends itself well for the "arms race" where every > > >> application can say "Me me me, I am the most important workload there is!". > > > > > > since there is discussion happening in two places: > > > > > > https://gitlab.freedesktop.org/drm/intel/-/issues/8014#note_1777433 > > > > > > What I think you might want is a ctx boost_mask which lets an app or > > > driver disable certain boost signals/classes. Where fence waits is > > > one class of boost, but hypothetical other signals like touchscreen > > > (or other) input events could be another class of boost. A compute > > > workload might be interested in fence wait boosts but could care less > > > about input events. > > > > I think it can only be apps which could have any chance knowing whether > > their use of a library is latency sensitive or not. Which means new > > library extensions and their adoption. So I have some strong reservation > > that route is feasible. > > > > Or we tie with priority which many drivers do. Normal and above gets the > > boosting and what lowered itself does not (aka SCHED_IDLE/SCHED_BATCH). > > yeah, that sounds reasonable. > on that gitlab-issue discussion Emma Anholt was against using the priority to influence frequency since that should be more about latency. or we are talking about something different priority here? > > Related note is that we lack any external control of our scheduling > > decisions so we really do suck compared to other scheduling domains like > > CPU and IO etc. > > > > >> The last concern is for me shared with the proposal to expose deadlines > > >> or high priority waits as explicit uapi knobs. Both come under the "what > > >> application told us it will do" category vs what it actually does. So I > > >> think it is slightly weaker than basing decisions of waits. > > >> > > >> The current waitboost is a bit detached from that problem because when > > >> we waitboost for flips we _know_ it is an actual framebuffer in the flip > > >> chain. When we waitboost for waits we also know someone is waiting. We > > >> are not trusting userspace telling us this will be a buffer in the flip > > >> chain or that this is a context which will have a certain duty-cycle. > > >> > > >> But yes, even if the input is truthful, latter is still only a > > >> heuristics because nothing says all waits are important. AFAIU it just > > >> happened to work well in the past. > > >> > > >> I do understand I am effectively arguing for more heuristics, which may > > >> sound a bit against the common wisdom. This is because in general I > > >> think the logic to do the right thing, be it in the driver or in the > > >> firmware, can work best if it has a holistic view. Simply put it needs > > >> to have more inputs to the decisions it is making. > > >> > > >> That is what my series is proposing - adding a common signal of "someone > > >> in userspace is waiting". What happens with that signal needs not be > > >> defined (promised) in the uapi contract. > > >> > > >> Say you route it to SLPC logic. It doesn't need to do with it what > > >> legacy i915 is doing today. It just needs to do something which works > > >> best for majority of workloads. It can even ignore it if that works for it. > > >> > > >> Finally, back to the immediate problem is when people replace the OpenCL > > >> NEO driver with clvk that performance tanks. Because former does waits > > >> using i915 specific ioctls and so triggers waitboost, latter waits on > > >> syncobj so no waitboost and performance is bad. What short term solution > > >> can we come up with? Or we say to not use clvk? I also wonder if other > > >> Vulkan based stuff is perhaps missing those easy performance gains.. > > >> > > >> Perhaps strictly speaking Rob's and mine proposal are not mutually > > >> exclusive. Yes I could piggy back on his with an "immediate deadline for > > >> waits" idea, but they could also be separate concepts if we concluded > > >> "someone is waiting" signal is useful to have. Or it takes to long to > > >> upstream the full deadline idea. > > > > > > Let me re-spin my series and add the syncobj wait flag and i915 bits > > > > I think wait flag is questionable unless it is inverted to imply waits > > which can be de-prioritized (again same parallel with SCHED_IDLE/BATCH). > > Having a flag which "makes things faster" IMO should require elevated > > privilege (to avoid the "arms race") at which point I fear it quickly > > becomes uninteresting. > > I guess you could make the argument in either direction. Making the > default behavior ramp up clocks could be a power regression. yeap, exactly the media / video conference case. > > I also think the "arms race" scenario isn't really as much of a > problem as you think. There aren't _that_ many things using the GPU > at the same time (compared to # of things using CPU). And a lot of > mobile games throttle framerate to avoid draining your battery too > quickly (after all, if your battery is dead you can't keep buying loot > boxes or whatever). Very good point. And in the GPU case they rely a lot on the profiles. Which btw, seems to be the Radeon solution. They boost the freq if the high performance profile is selected and don't care about the execution bubbles if low or mid profiles are selected, or something like that. > > > > adapted from your patches.. I think the basic idea of deadlines > > > (which includes "I want it NOW" ;-)) isn't controversial, but the > > > original idea got caught up in some bikeshed (what about compositors > > > that wait on fences in userspace to decide which surfaces to update in > > > the next frame), plus me getting busy and generally not having a good > > > plan for how to leverage this from VM guests (which is becoming > > > increasingly important for CrOS). I think I can build on some ongoing > > > virtgpu fencing improvement work to solve the latter. But now that we > > > have a 2nd use-case for this, it makes sense to respin. > > > > Sure, I was looking at the old version already. It is interesting. But > > also IMO needs quite a bit more work to approach achieving what is > > implied from the name of the feature. It would need proper deadline > > based sched job picking, and even then drm sched is mostly just a > > frontend. So once past runnable status and jobs handed over to backend, > > without further driver work it probably wouldn't be very effective past > > very lightly loaded systems. > > Yes, but all of that is not part of dma_fence ;-) > > A pretty common challenging usecase is still the single fullscreen > game, where scheduling isn't the problem, but landing at an > appropriate GPU freq absolutely is. (UI workloads are perhaps more > interesting from a scheduler standpoint, but they generally aren't > challenging from a load/freq standpoint.) > > Fwiw, the original motivation of the series was to implement something > akin to i915 pageflip boosting without having to abandon the atomic > helpers. (And, I guess it would also let i915 preserve that feature > if it switched to atomic helpers.. I'm unsure if there are still other > things blocking i915's migration.) > > > Then if we fast forward to a world where schedulers perhaps become fully > > deadline aware (we even had this for i915 few years back) then the > > question will be does equating waits with immediate deadlines still > > works. Maybe not too well because we wouldn't have the ability to > > distinguish between the "someone is waiting" signal from the otherwise > > propagated deadlines. > > Is there any other way to handle a wait boost than expressing it as an > ASAP deadline? > > BR, > -R > > > > > Regards, > > > > Tvrtko
On Fri, Feb 17, 2023 at 12:45 PM Rodrigo Vivi <rodrigo.vivi@intel.com> wrote: > > On Fri, Feb 17, 2023 at 09:00:49AM -0800, Rob Clark wrote: > > On Fri, Feb 17, 2023 at 8:03 AM Tvrtko Ursulin > > <tvrtko.ursulin@linux.intel.com> wrote: > > > > > > > > > On 17/02/2023 14:55, Rob Clark wrote: > > > > On Fri, Feb 17, 2023 at 4:56 AM Tvrtko Ursulin > > > > <tvrtko.ursulin@linux.intel.com> wrote: > > > >> > > > >> > > > >> On 16/02/2023 18:19, Rodrigo Vivi wrote: > > > >>> On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: > > > >>>> On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > > > >>>> <tvrtko.ursulin@linux.intel.com> wrote: > > > >>>>> > > > >>>>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > > > >>>>> > > > >>>>> In i915 we have this concept of "wait boosting" where we give a priority boost > > > >>>>> for instance to fences which are actively waited upon from userspace. This has > > > >>>>> it's pros and cons and can certainly be discussed at lenght. However fact is > > > >>>>> some workloads really like it. > > > >>>>> > > > >>>>> Problem is that with the arrival of drm syncobj and a new userspace waiting > > > >>>>> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > > > >>>>> this mini series really (really) quickly to see if some discussion can be had. > > > >>>>> > > > >>>>> It adds a concept of "wait count" to dma fence, which is incremented for every > > > >>>>> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > > > >>>>> dma_fence_add_callback but from explicit/userspace wait paths). > > > >>>> > > > >>>> I was thinking about a similar thing, but in the context of dma_fence > > > >>>> (or rather sync_file) fd poll()ing. How does the kernel differentiate > > > >>>> between "housekeeping" poll()ers that don't want to trigger boost but > > > >>>> simply know when to do cleanup, and waiters who are waiting with some > > > >>>> urgency. I think we could use EPOLLPRI for this purpose. > > > >>>> > > > >>>> Not sure how that translates to waits via the syncobj. But I think we > > > >>>> want to let userspace give some hint about urgent vs housekeeping > > > >>>> waits. > > > >>> > > > >>> Should the hint be on the waits, or should the hints be on the executed > > > >>> context? > > > >>> > > > >>> In the end we need some way to quickly ramp-up the frequency to avoid > > > >>> the execution bubbles. > > > >>> > > > >>> waitboost is trying to guess that, but in some cases it guess wrong > > > >>> and waste power. > > > >> > > > >> Do we have a list of workloads which shows who benefits and who loses > > > >> from the current implementation of waitboost? > > > >>> btw, this is something that other drivers might need: > > > >>> > > > >>> https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 > > > >>> Cc: Alex Deucher <alexander.deucher@amd.com> > > > >> > > > >> I have several issues with the context hint if it would directly > > > >> influence frequency selection in the "more power" direction. > > > >> > > > >> First of all, assume a context hint would replace the waitboost. Which > > > >> applications would need to set it to restore the lost performance and > > > >> how would they set it? > > > >> > > > >> Then I don't even think userspace necessarily knows. Think of a layer > > > >> like OpenCL. It doesn't really know in advance the profile of > > > >> submissions vs waits. It depends on the CPU vs GPU speed, so hardware > > > >> generation, and the actual size of the workload which can be influenced > > > >> by the application (or user) and not the library. > > > >> > > > >> The approach also lends itself well for the "arms race" where every > > > >> application can say "Me me me, I am the most important workload there is!". > > > > > > > > since there is discussion happening in two places: > > > > > > > > https://gitlab.freedesktop.org/drm/intel/-/issues/8014#note_1777433 > > > > > > > > What I think you might want is a ctx boost_mask which lets an app or > > > > driver disable certain boost signals/classes. Where fence waits is > > > > one class of boost, but hypothetical other signals like touchscreen > > > > (or other) input events could be another class of boost. A compute > > > > workload might be interested in fence wait boosts but could care less > > > > about input events. > > > > > > I think it can only be apps which could have any chance knowing whether > > > their use of a library is latency sensitive or not. Which means new > > > library extensions and their adoption. So I have some strong reservation > > > that route is feasible. > > > > > > Or we tie with priority which many drivers do. Normal and above gets the > > > boosting and what lowered itself does not (aka SCHED_IDLE/SCHED_BATCH). > > > > yeah, that sounds reasonable. > > > > on that gitlab-issue discussion Emma Anholt was against using the priority > to influence frequency since that should be more about latency. > > or we are talking about something different priority here? I was thinking to only _not_ boost on the lowest priority, but boost on norm/high priority. But not something I feel too strongly about. Ie. deciding on policy doesn't affect or need to block getting the dma_fence and syncobj plumbing in place. BR, -R > > > Related note is that we lack any external control of our scheduling > > > decisions so we really do suck compared to other scheduling domains like > > > CPU and IO etc. > > > > > > >> The last concern is for me shared with the proposal to expose deadlines > > > >> or high priority waits as explicit uapi knobs. Both come under the "what > > > >> application told us it will do" category vs what it actually does. So I > > > >> think it is slightly weaker than basing decisions of waits. > > > >> > > > >> The current waitboost is a bit detached from that problem because when > > > >> we waitboost for flips we _know_ it is an actual framebuffer in the flip > > > >> chain. When we waitboost for waits we also know someone is waiting. We > > > >> are not trusting userspace telling us this will be a buffer in the flip > > > >> chain or that this is a context which will have a certain duty-cycle. > > > >> > > > >> But yes, even if the input is truthful, latter is still only a > > > >> heuristics because nothing says all waits are important. AFAIU it just > > > >> happened to work well in the past. > > > >> > > > >> I do understand I am effectively arguing for more heuristics, which may > > > >> sound a bit against the common wisdom. This is because in general I > > > >> think the logic to do the right thing, be it in the driver or in the > > > >> firmware, can work best if it has a holistic view. Simply put it needs > > > >> to have more inputs to the decisions it is making. > > > >> > > > >> That is what my series is proposing - adding a common signal of "someone > > > >> in userspace is waiting". What happens with that signal needs not be > > > >> defined (promised) in the uapi contract. > > > >> > > > >> Say you route it to SLPC logic. It doesn't need to do with it what > > > >> legacy i915 is doing today. It just needs to do something which works > > > >> best for majority of workloads. It can even ignore it if that works for it. > > > >> > > > >> Finally, back to the immediate problem is when people replace the OpenCL > > > >> NEO driver with clvk that performance tanks. Because former does waits > > > >> using i915 specific ioctls and so triggers waitboost, latter waits on > > > >> syncobj so no waitboost and performance is bad. What short term solution > > > >> can we come up with? Or we say to not use clvk? I also wonder if other > > > >> Vulkan based stuff is perhaps missing those easy performance gains.. > > > >> > > > >> Perhaps strictly speaking Rob's and mine proposal are not mutually > > > >> exclusive. Yes I could piggy back on his with an "immediate deadline for > > > >> waits" idea, but they could also be separate concepts if we concluded > > > >> "someone is waiting" signal is useful to have. Or it takes to long to > > > >> upstream the full deadline idea. > > > > > > > > Let me re-spin my series and add the syncobj wait flag and i915 bits > > > > > > I think wait flag is questionable unless it is inverted to imply waits > > > which can be de-prioritized (again same parallel with SCHED_IDLE/BATCH). > > > Having a flag which "makes things faster" IMO should require elevated > > > privilege (to avoid the "arms race") at which point I fear it quickly > > > becomes uninteresting. > > > > I guess you could make the argument in either direction. Making the > > default behavior ramp up clocks could be a power regression. > > yeap, exactly the media / video conference case. > > > > > I also think the "arms race" scenario isn't really as much of a > > problem as you think. There aren't _that_ many things using the GPU > > at the same time (compared to # of things using CPU). And a lot of > > mobile games throttle framerate to avoid draining your battery too > > quickly (after all, if your battery is dead you can't keep buying loot > > boxes or whatever). > > Very good point. > > And in the GPU case they rely a lot on the profiles. Which btw, seems > to be the Radeon solution. They boost the freq if the high performance > profile is selected and don't care about the execution bubbles if low > or mid profiles are selected, or something like that. > > > > > > > adapted from your patches.. I think the basic idea of deadlines > > > > (which includes "I want it NOW" ;-)) isn't controversial, but the > > > > original idea got caught up in some bikeshed (what about compositors > > > > that wait on fences in userspace to decide which surfaces to update in > > > > the next frame), plus me getting busy and generally not having a good > > > > plan for how to leverage this from VM guests (which is becoming > > > > increasingly important for CrOS). I think I can build on some ongoing > > > > virtgpu fencing improvement work to solve the latter. But now that we > > > > have a 2nd use-case for this, it makes sense to respin. > > > > > > Sure, I was looking at the old version already. It is interesting. But > > > also IMO needs quite a bit more work to approach achieving what is > > > implied from the name of the feature. It would need proper deadline > > > based sched job picking, and even then drm sched is mostly just a > > > frontend. So once past runnable status and jobs handed over to backend, > > > without further driver work it probably wouldn't be very effective past > > > very lightly loaded systems. > > > > Yes, but all of that is not part of dma_fence ;-) > > > > A pretty common challenging usecase is still the single fullscreen > > game, where scheduling isn't the problem, but landing at an > > appropriate GPU freq absolutely is. (UI workloads are perhaps more > > interesting from a scheduler standpoint, but they generally aren't > > challenging from a load/freq standpoint.) > > > > Fwiw, the original motivation of the series was to implement something > > akin to i915 pageflip boosting without having to abandon the atomic > > helpers. (And, I guess it would also let i915 preserve that feature > > if it switched to atomic helpers.. I'm unsure if there are still other > > things blocking i915's migration.) > > > > > Then if we fast forward to a world where schedulers perhaps become fully > > > deadline aware (we even had this for i915 few years back) then the > > > question will be does equating waits with immediate deadlines still > > > works. Maybe not too well because we wouldn't have the ability to > > > distinguish between the "someone is waiting" signal from the otherwise > > > propagated deadlines. > > > > Is there any other way to handle a wait boost than expressing it as an > > ASAP deadline? > > > > BR, > > -R > > > > > > > > Regards, > > > > > > Tvrtko
On 17/02/2023 20:45, Rodrigo Vivi wrote: > On Fri, Feb 17, 2023 at 09:00:49AM -0800, Rob Clark wrote: >> On Fri, Feb 17, 2023 at 8:03 AM Tvrtko Ursulin >> <tvrtko.ursulin@linux.intel.com> wrote: >>> >>> >>> On 17/02/2023 14:55, Rob Clark wrote: >>>> On Fri, Feb 17, 2023 at 4:56 AM Tvrtko Ursulin >>>> <tvrtko.ursulin@linux.intel.com> wrote: >>>>> >>>>> >>>>> On 16/02/2023 18:19, Rodrigo Vivi wrote: >>>>>> On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: >>>>>>> On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin >>>>>>> <tvrtko.ursulin@linux.intel.com> wrote: >>>>>>>> >>>>>>>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> >>>>>>>> >>>>>>>> In i915 we have this concept of "wait boosting" where we give a priority boost >>>>>>>> for instance to fences which are actively waited upon from userspace. This has >>>>>>>> it's pros and cons and can certainly be discussed at lenght. However fact is >>>>>>>> some workloads really like it. >>>>>>>> >>>>>>>> Problem is that with the arrival of drm syncobj and a new userspace waiting >>>>>>>> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up >>>>>>>> this mini series really (really) quickly to see if some discussion can be had. >>>>>>>> >>>>>>>> It adds a concept of "wait count" to dma fence, which is incremented for every >>>>>>>> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like >>>>>>>> dma_fence_add_callback but from explicit/userspace wait paths). >>>>>>> >>>>>>> I was thinking about a similar thing, but in the context of dma_fence >>>>>>> (or rather sync_file) fd poll()ing. How does the kernel differentiate >>>>>>> between "housekeeping" poll()ers that don't want to trigger boost but >>>>>>> simply know when to do cleanup, and waiters who are waiting with some >>>>>>> urgency. I think we could use EPOLLPRI for this purpose. >>>>>>> >>>>>>> Not sure how that translates to waits via the syncobj. But I think we >>>>>>> want to let userspace give some hint about urgent vs housekeeping >>>>>>> waits. >>>>>> >>>>>> Should the hint be on the waits, or should the hints be on the executed >>>>>> context? >>>>>> >>>>>> In the end we need some way to quickly ramp-up the frequency to avoid >>>>>> the execution bubbles. >>>>>> >>>>>> waitboost is trying to guess that, but in some cases it guess wrong >>>>>> and waste power. >>>>> >>>>> Do we have a list of workloads which shows who benefits and who loses >>>>> from the current implementation of waitboost? >>>>>> btw, this is something that other drivers might need: >>>>>> >>>>>> https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 >>>>>> Cc: Alex Deucher <alexander.deucher@amd.com> >>>>> >>>>> I have several issues with the context hint if it would directly >>>>> influence frequency selection in the "more power" direction. >>>>> >>>>> First of all, assume a context hint would replace the waitboost. Which >>>>> applications would need to set it to restore the lost performance and >>>>> how would they set it? >>>>> >>>>> Then I don't even think userspace necessarily knows. Think of a layer >>>>> like OpenCL. It doesn't really know in advance the profile of >>>>> submissions vs waits. It depends on the CPU vs GPU speed, so hardware >>>>> generation, and the actual size of the workload which can be influenced >>>>> by the application (or user) and not the library. >>>>> >>>>> The approach also lends itself well for the "arms race" where every >>>>> application can say "Me me me, I am the most important workload there is!". >>>> >>>> since there is discussion happening in two places: >>>> >>>> https://gitlab.freedesktop.org/drm/intel/-/issues/8014#note_1777433 >>>> >>>> What I think you might want is a ctx boost_mask which lets an app or >>>> driver disable certain boost signals/classes. Where fence waits is >>>> one class of boost, but hypothetical other signals like touchscreen >>>> (or other) input events could be another class of boost. A compute >>>> workload might be interested in fence wait boosts but could care less >>>> about input events. >>> >>> I think it can only be apps which could have any chance knowing whether >>> their use of a library is latency sensitive or not. Which means new >>> library extensions and their adoption. So I have some strong reservation >>> that route is feasible. >>> >>> Or we tie with priority which many drivers do. Normal and above gets the >>> boosting and what lowered itself does not (aka SCHED_IDLE/SCHED_BATCH). >> >> yeah, that sounds reasonable. >> > > on that gitlab-issue discussion Emma Anholt was against using the priority > to influence frequency since that should be more about latency. > > or we are talking about something different priority here? As Rob already explained - I was suggesting skipping waitboost for contexts which explicitly made themselves low priority. I don't see a controversial angle there. >>> Related note is that we lack any external control of our scheduling >>> decisions so we really do suck compared to other scheduling domains like >>> CPU and IO etc. >>> >>>>> The last concern is for me shared with the proposal to expose deadlines >>>>> or high priority waits as explicit uapi knobs. Both come under the "what >>>>> application told us it will do" category vs what it actually does. So I >>>>> think it is slightly weaker than basing decisions of waits. >>>>> >>>>> The current waitboost is a bit detached from that problem because when >>>>> we waitboost for flips we _know_ it is an actual framebuffer in the flip >>>>> chain. When we waitboost for waits we also know someone is waiting. We >>>>> are not trusting userspace telling us this will be a buffer in the flip >>>>> chain or that this is a context which will have a certain duty-cycle. >>>>> >>>>> But yes, even if the input is truthful, latter is still only a >>>>> heuristics because nothing says all waits are important. AFAIU it just >>>>> happened to work well in the past. >>>>> >>>>> I do understand I am effectively arguing for more heuristics, which may >>>>> sound a bit against the common wisdom. This is because in general I >>>>> think the logic to do the right thing, be it in the driver or in the >>>>> firmware, can work best if it has a holistic view. Simply put it needs >>>>> to have more inputs to the decisions it is making. >>>>> >>>>> That is what my series is proposing - adding a common signal of "someone >>>>> in userspace is waiting". What happens with that signal needs not be >>>>> defined (promised) in the uapi contract. >>>>> >>>>> Say you route it to SLPC logic. It doesn't need to do with it what >>>>> legacy i915 is doing today. It just needs to do something which works >>>>> best for majority of workloads. It can even ignore it if that works for it. >>>>> >>>>> Finally, back to the immediate problem is when people replace the OpenCL >>>>> NEO driver with clvk that performance tanks. Because former does waits >>>>> using i915 specific ioctls and so triggers waitboost, latter waits on >>>>> syncobj so no waitboost and performance is bad. What short term solution >>>>> can we come up with? Or we say to not use clvk? I also wonder if other >>>>> Vulkan based stuff is perhaps missing those easy performance gains.. >>>>> >>>>> Perhaps strictly speaking Rob's and mine proposal are not mutually >>>>> exclusive. Yes I could piggy back on his with an "immediate deadline for >>>>> waits" idea, but they could also be separate concepts if we concluded >>>>> "someone is waiting" signal is useful to have. Or it takes to long to >>>>> upstream the full deadline idea. >>>> >>>> Let me re-spin my series and add the syncobj wait flag and i915 bits >>> >>> I think wait flag is questionable unless it is inverted to imply waits >>> which can be de-prioritized (again same parallel with SCHED_IDLE/BATCH). >>> Having a flag which "makes things faster" IMO should require elevated >>> privilege (to avoid the "arms race") at which point I fear it quickly >>> becomes uninteresting. >> >> I guess you could make the argument in either direction. Making the >> default behavior ramp up clocks could be a power regression. > > yeap, exactly the media / video conference case. Yeah I agree. And as not all media use cases are the same, as are not all compute contexts someone somewhere will need to run a series of workloads for power and performance numbers. Ideally that someone would be the entity for which it makes sense to look at all use cases, from server room to client, 3d, media and compute for both. If we could get the capability to run this in some automated fashion, akin to CI, we would even have a chance to keep making good decisions in the future. Or we do some one off testing for this instance, but we still need a range of workloads and parts to do it properly.. >> I also think the "arms race" scenario isn't really as much of a >> problem as you think. There aren't _that_ many things using the GPU >> at the same time (compared to # of things using CPU). And a lot of >> mobile games throttle framerate to avoid draining your battery too >> quickly (after all, if your battery is dead you can't keep buying loot >> boxes or whatever). > > Very good point. On this one I still disagree from the point of view that it does not make it good uapi if we allow everyone to select themselves for priority handling (one flavour or the other). > And in the GPU case they rely a lot on the profiles. Which btw, seems > to be the Radeon solution. They boost the freq if the high performance > profile is selected and don't care about the execution bubbles if low > or mid profiles are selected, or something like that. Profile as something which controls the waitboost globally? What would be the mechanism for communicating it to the driver? Also, how would that reconcile the fact waitboost harms some workloads but helps others? If the latter not only improves the performance but also efficiency then assuming "battery" profile must mean "waitboost off" would be leaving battery life on the table. Conversely, if the "on a/c - max performance", would be global "waitboost on", then it could even be possible it wouldn't always be truly best performance if it causes thermal throttling. Regards, Tvrtko
On 17/02/2023 17:00, Rob Clark wrote: > On Fri, Feb 17, 2023 at 8:03 AM Tvrtko Ursulin > <tvrtko.ursulin@linux.intel.com> wrote: [snip] >>> adapted from your patches.. I think the basic idea of deadlines >>> (which includes "I want it NOW" ;-)) isn't controversial, but the >>> original idea got caught up in some bikeshed (what about compositors >>> that wait on fences in userspace to decide which surfaces to update in >>> the next frame), plus me getting busy and generally not having a good >>> plan for how to leverage this from VM guests (which is becoming >>> increasingly important for CrOS). I think I can build on some ongoing >>> virtgpu fencing improvement work to solve the latter. But now that we >>> have a 2nd use-case for this, it makes sense to respin. >> >> Sure, I was looking at the old version already. It is interesting. But >> also IMO needs quite a bit more work to approach achieving what is >> implied from the name of the feature. It would need proper deadline >> based sched job picking, and even then drm sched is mostly just a >> frontend. So once past runnable status and jobs handed over to backend, >> without further driver work it probably wouldn't be very effective past >> very lightly loaded systems. > > Yes, but all of that is not part of dma_fence ;-) :) Okay. Having said that, do we need a step back to think about whether adding deadline to dma-fences is not making them something too much different to what they were? Going from purely synchronisation primitive more towards scheduling paradigms. Just to brainstorm if there will not be any unintended consequences. I should mention this in your RFC thread actually. > A pretty common challenging usecase is still the single fullscreen > game, where scheduling isn't the problem, but landing at an > appropriate GPU freq absolutely is. (UI workloads are perhaps more > interesting from a scheduler standpoint, but they generally aren't > challenging from a load/freq standpoint.) Challenging as in picking the right operating point? Might be latency impacted (and so user perceived UI smoothness) due missing waitboost for anything syncobj related. I don't know if anything to measure that exists currently though. Assuming it is measurable then the question would be is it perceivable. > Fwiw, the original motivation of the series was to implement something > akin to i915 pageflip boosting without having to abandon the atomic > helpers. (And, I guess it would also let i915 preserve that feature > if it switched to atomic helpers.. I'm unsure if there are still other > things blocking i915's migration.) Question for display folks I guess. >> Then if we fast forward to a world where schedulers perhaps become fully >> deadline aware (we even had this for i915 few years back) then the >> question will be does equating waits with immediate deadlines still >> works. Maybe not too well because we wouldn't have the ability to >> distinguish between the "someone is waiting" signal from the otherwise >> propagated deadlines. > > Is there any other way to handle a wait boost than expressing it as an > ASAP deadline? A leading question or just a question? Nothing springs to my mind at the moment. Regards, Tvrtko
On Mon, Feb 20, 2023 at 4:22 AM Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> wrote: > > > On 17/02/2023 17:00, Rob Clark wrote: > > On Fri, Feb 17, 2023 at 8:03 AM Tvrtko Ursulin > > <tvrtko.ursulin@linux.intel.com> wrote: > > [snip] > > >>> adapted from your patches.. I think the basic idea of deadlines > >>> (which includes "I want it NOW" ;-)) isn't controversial, but the > >>> original idea got caught up in some bikeshed (what about compositors > >>> that wait on fences in userspace to decide which surfaces to update in > >>> the next frame), plus me getting busy and generally not having a good > >>> plan for how to leverage this from VM guests (which is becoming > >>> increasingly important for CrOS). I think I can build on some ongoing > >>> virtgpu fencing improvement work to solve the latter. But now that we > >>> have a 2nd use-case for this, it makes sense to respin. > >> > >> Sure, I was looking at the old version already. It is interesting. But > >> also IMO needs quite a bit more work to approach achieving what is > >> implied from the name of the feature. It would need proper deadline > >> based sched job picking, and even then drm sched is mostly just a > >> frontend. So once past runnable status and jobs handed over to backend, > >> without further driver work it probably wouldn't be very effective past > >> very lightly loaded systems. > > > > Yes, but all of that is not part of dma_fence ;-) > > :) Okay. > > Having said that, do we need a step back to think about whether adding > deadline to dma-fences is not making them something too much different > to what they were? Going from purely synchronisation primitive more > towards scheduling paradigms. Just to brainstorm if there will not be > any unintended consequences. I should mention this in your RFC thread > actually. Perhaps "deadline" isn't quite the right name, but I haven't thought of anything better. It is really a hint to the fence signaller about how soon it is interested in a result so the driver can factor that into freq scaling decisions. Maybe "goal" or some other term would be better? I guess that can factor into scheduling decisions as well.. but we already have priority for that. My main interest is freq mgmt. (Thankfully we don't have performance and efficiency cores to worry about, like CPUs ;-)) > > A pretty common challenging usecase is still the single fullscreen > > game, where scheduling isn't the problem, but landing at an > > appropriate GPU freq absolutely is. (UI workloads are perhaps more > > interesting from a scheduler standpoint, but they generally aren't > > challenging from a load/freq standpoint.) > > Challenging as in picking the right operating point? Might be latency > impacted (and so user perceived UI smoothness) due missing waitboost for > anything syncobj related. I don't know if anything to measure that > exists currently though. Assuming it is measurable then the question > would be is it perceivable. > > Fwiw, the original motivation of the series was to implement something > > akin to i915 pageflip boosting without having to abandon the atomic > > helpers. (And, I guess it would also let i915 preserve that feature > > if it switched to atomic helpers.. I'm unsure if there are still other > > things blocking i915's migration.) > > Question for display folks I guess. > > >> Then if we fast forward to a world where schedulers perhaps become fully > >> deadline aware (we even had this for i915 few years back) then the > >> question will be does equating waits with immediate deadlines still > >> works. Maybe not too well because we wouldn't have the ability to > >> distinguish between the "someone is waiting" signal from the otherwise > >> propagated deadlines. > > > > Is there any other way to handle a wait boost than expressing it as an > > ASAP deadline? > > A leading question or just a question? Nothing springs to my mind at the > moment. Just a question. The immediate deadline is the only thing that makes sense to me, but that could be because I'm looking at it from the perspective of also trying to handle the case where missing vblank reduces utilization and provides the wrong signal to gpufreq.. i915 already has a way to handle this internally, but it involves bypassing the atomic helpers, which isn't a thing I want to encourage other drivers to do. And completely doesn't work for situations where the gpu and display are separate devices. BR, -R > Regards, > > Tvrtko
On Mon, Feb 20, 2023 at 3:33 AM Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> wrote: > > > On 17/02/2023 20:45, Rodrigo Vivi wrote: > > On Fri, Feb 17, 2023 at 09:00:49AM -0800, Rob Clark wrote: > >> On Fri, Feb 17, 2023 at 8:03 AM Tvrtko Ursulin > >> <tvrtko.ursulin@linux.intel.com> wrote: > >>> > >>> > >>> On 17/02/2023 14:55, Rob Clark wrote: > >>>> On Fri, Feb 17, 2023 at 4:56 AM Tvrtko Ursulin > >>>> <tvrtko.ursulin@linux.intel.com> wrote: > >>>>> > >>>>> > >>>>> On 16/02/2023 18:19, Rodrigo Vivi wrote: > >>>>>> On Tue, Feb 14, 2023 at 11:14:00AM -0800, Rob Clark wrote: > >>>>>>> On Fri, Feb 10, 2023 at 5:07 AM Tvrtko Ursulin > >>>>>>> <tvrtko.ursulin@linux.intel.com> wrote: > >>>>>>>> > >>>>>>>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > >>>>>>>> > >>>>>>>> In i915 we have this concept of "wait boosting" where we give a priority boost > >>>>>>>> for instance to fences which are actively waited upon from userspace. This has > >>>>>>>> it's pros and cons and can certainly be discussed at lenght. However fact is > >>>>>>>> some workloads really like it. > >>>>>>>> > >>>>>>>> Problem is that with the arrival of drm syncobj and a new userspace waiting > >>>>>>>> entry point it added, the waitboost mechanism was bypassed. Hence I cooked up > >>>>>>>> this mini series really (really) quickly to see if some discussion can be had. > >>>>>>>> > >>>>>>>> It adds a concept of "wait count" to dma fence, which is incremented for every > >>>>>>>> explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like > >>>>>>>> dma_fence_add_callback but from explicit/userspace wait paths). > >>>>>>> > >>>>>>> I was thinking about a similar thing, but in the context of dma_fence > >>>>>>> (or rather sync_file) fd poll()ing. How does the kernel differentiate > >>>>>>> between "housekeeping" poll()ers that don't want to trigger boost but > >>>>>>> simply know when to do cleanup, and waiters who are waiting with some > >>>>>>> urgency. I think we could use EPOLLPRI for this purpose. > >>>>>>> > >>>>>>> Not sure how that translates to waits via the syncobj. But I think we > >>>>>>> want to let userspace give some hint about urgent vs housekeeping > >>>>>>> waits. > >>>>>> > >>>>>> Should the hint be on the waits, or should the hints be on the executed > >>>>>> context? > >>>>>> > >>>>>> In the end we need some way to quickly ramp-up the frequency to avoid > >>>>>> the execution bubbles. > >>>>>> > >>>>>> waitboost is trying to guess that, but in some cases it guess wrong > >>>>>> and waste power. > >>>>> > >>>>> Do we have a list of workloads which shows who benefits and who loses > >>>>> from the current implementation of waitboost? > >>>>>> btw, this is something that other drivers might need: > >>>>>> > >>>>>> https://gitlab.freedesktop.org/drm/amd/-/issues/1500#note_825883 > >>>>>> Cc: Alex Deucher <alexander.deucher@amd.com> > >>>>> > >>>>> I have several issues with the context hint if it would directly > >>>>> influence frequency selection in the "more power" direction. > >>>>> > >>>>> First of all, assume a context hint would replace the waitboost. Which > >>>>> applications would need to set it to restore the lost performance and > >>>>> how would they set it? > >>>>> > >>>>> Then I don't even think userspace necessarily knows. Think of a layer > >>>>> like OpenCL. It doesn't really know in advance the profile of > >>>>> submissions vs waits. It depends on the CPU vs GPU speed, so hardware > >>>>> generation, and the actual size of the workload which can be influenced > >>>>> by the application (or user) and not the library. > >>>>> > >>>>> The approach also lends itself well for the "arms race" where every > >>>>> application can say "Me me me, I am the most important workload there is!". > >>>> > >>>> since there is discussion happening in two places: > >>>> > >>>> https://gitlab.freedesktop.org/drm/intel/-/issues/8014#note_1777433 > >>>> > >>>> What I think you might want is a ctx boost_mask which lets an app or > >>>> driver disable certain boost signals/classes. Where fence waits is > >>>> one class of boost, but hypothetical other signals like touchscreen > >>>> (or other) input events could be another class of boost. A compute > >>>> workload might be interested in fence wait boosts but could care less > >>>> about input events. > >>> > >>> I think it can only be apps which could have any chance knowing whether > >>> their use of a library is latency sensitive or not. Which means new > >>> library extensions and their adoption. So I have some strong reservation > >>> that route is feasible. > >>> > >>> Or we tie with priority which many drivers do. Normal and above gets the > >>> boosting and what lowered itself does not (aka SCHED_IDLE/SCHED_BATCH). > >> > >> yeah, that sounds reasonable. > >> > > > > on that gitlab-issue discussion Emma Anholt was against using the priority > > to influence frequency since that should be more about latency. > > > > or we are talking about something different priority here? > > As Rob already explained - I was suggesting skipping waitboost for > contexts which explicitly made themselves low priority. I don't see a > controversial angle there. > > >>> Related note is that we lack any external control of our scheduling > >>> decisions so we really do suck compared to other scheduling domains like > >>> CPU and IO etc. > >>> > >>>>> The last concern is for me shared with the proposal to expose deadlines > >>>>> or high priority waits as explicit uapi knobs. Both come under the "what > >>>>> application told us it will do" category vs what it actually does. So I > >>>>> think it is slightly weaker than basing decisions of waits. > >>>>> > >>>>> The current waitboost is a bit detached from that problem because when > >>>>> we waitboost for flips we _know_ it is an actual framebuffer in the flip > >>>>> chain. When we waitboost for waits we also know someone is waiting. We > >>>>> are not trusting userspace telling us this will be a buffer in the flip > >>>>> chain or that this is a context which will have a certain duty-cycle. > >>>>> > >>>>> But yes, even if the input is truthful, latter is still only a > >>>>> heuristics because nothing says all waits are important. AFAIU it just > >>>>> happened to work well in the past. > >>>>> > >>>>> I do understand I am effectively arguing for more heuristics, which may > >>>>> sound a bit against the common wisdom. This is because in general I > >>>>> think the logic to do the right thing, be it in the driver or in the > >>>>> firmware, can work best if it has a holistic view. Simply put it needs > >>>>> to have more inputs to the decisions it is making. > >>>>> > >>>>> That is what my series is proposing - adding a common signal of "someone > >>>>> in userspace is waiting". What happens with that signal needs not be > >>>>> defined (promised) in the uapi contract. > >>>>> > >>>>> Say you route it to SLPC logic. It doesn't need to do with it what > >>>>> legacy i915 is doing today. It just needs to do something which works > >>>>> best for majority of workloads. It can even ignore it if that works for it. > >>>>> > >>>>> Finally, back to the immediate problem is when people replace the OpenCL > >>>>> NEO driver with clvk that performance tanks. Because former does waits > >>>>> using i915 specific ioctls and so triggers waitboost, latter waits on > >>>>> syncobj so no waitboost and performance is bad. What short term solution > >>>>> can we come up with? Or we say to not use clvk? I also wonder if other > >>>>> Vulkan based stuff is perhaps missing those easy performance gains.. > >>>>> > >>>>> Perhaps strictly speaking Rob's and mine proposal are not mutually > >>>>> exclusive. Yes I could piggy back on his with an "immediate deadline for > >>>>> waits" idea, but they could also be separate concepts if we concluded > >>>>> "someone is waiting" signal is useful to have. Or it takes to long to > >>>>> upstream the full deadline idea. > >>>> > >>>> Let me re-spin my series and add the syncobj wait flag and i915 bits > >>> > >>> I think wait flag is questionable unless it is inverted to imply waits > >>> which can be de-prioritized (again same parallel with SCHED_IDLE/BATCH). > >>> Having a flag which "makes things faster" IMO should require elevated > >>> privilege (to avoid the "arms race") at which point I fear it quickly > >>> becomes uninteresting. > >> > >> I guess you could make the argument in either direction. Making the > >> default behavior ramp up clocks could be a power regression. > > > > yeap, exactly the media / video conference case. > > Yeah I agree. And as not all media use cases are the same, as are not > all compute contexts someone somewhere will need to run a series of > workloads for power and performance numbers. Ideally that someone would > be the entity for which it makes sense to look at all use cases, from > server room to client, 3d, media and compute for both. If we could get > the capability to run this in some automated fashion, akin to CI, we > would even have a chance to keep making good decisions in the future. > > Or we do some one off testing for this instance, but we still need a > range of workloads and parts to do it properly.. > > >> I also think the "arms race" scenario isn't really as much of a > >> problem as you think. There aren't _that_ many things using the GPU > >> at the same time (compared to # of things using CPU). And a lot of > >> mobile games throttle framerate to avoid draining your battery too > >> quickly (after all, if your battery is dead you can't keep buying loot > >> boxes or whatever). > > > > Very good point. > > On this one I still disagree from the point of view that it does not > make it good uapi if we allow everyone to select themselves for priority > handling (one flavour or the other). There is plenty of precedent for userspace giving hints to the kernel about scheduling and freq mgmt. Like schedutil uclamp stuff. Although I think that is all based on cgroups. In the fence/syncobj case, I think we need per-wait hints.. because for a single process the driver will be doing both housekeeping waits and potentially urgent waits. There may also be some room for some cgroup or similar knobs to control things like what max priority an app can ask for, and whether or how aggressively the kernel responds to the "deadline" hints. So as far as "arms race", I don't think I'd change anything about my "fence deadline" proposal.. but that it might just be one piece of the overall puzzle. BR, -R > > And in the GPU case they rely a lot on the profiles. Which btw, seems > > to be the Radeon solution. They boost the freq if the high performance > > profile is selected and don't care about the execution bubbles if low > > or mid profiles are selected, or something like that. > > Profile as something which controls the waitboost globally? What would > be the mechanism for communicating it to the driver? > > Also, how would that reconcile the fact waitboost harms some workloads > but helps others? If the latter not only improves the performance but > also efficiency then assuming "battery" profile must mean "waitboost > off" would be leaving battery life on the table. Conversely, if the "on > a/c - max performance", would be global "waitboost on", then it could > even be possible it wouldn't always be truly best performance if it > causes thermal throttling. > > Regards, > > Tvrtko
On 20/02/2023 15:45, Rob Clark wrote: > On Mon, Feb 20, 2023 at 4:22 AM Tvrtko Ursulin > <tvrtko.ursulin@linux.intel.com> wrote: >> >> >> On 17/02/2023 17:00, Rob Clark wrote: >>> On Fri, Feb 17, 2023 at 8:03 AM Tvrtko Ursulin >>> <tvrtko.ursulin@linux.intel.com> wrote: >> >> [snip] >> >>>>> adapted from your patches.. I think the basic idea of deadlines >>>>> (which includes "I want it NOW" ;-)) isn't controversial, but the >>>>> original idea got caught up in some bikeshed (what about compositors >>>>> that wait on fences in userspace to decide which surfaces to update in >>>>> the next frame), plus me getting busy and generally not having a good >>>>> plan for how to leverage this from VM guests (which is becoming >>>>> increasingly important for CrOS). I think I can build on some ongoing >>>>> virtgpu fencing improvement work to solve the latter. But now that we >>>>> have a 2nd use-case for this, it makes sense to respin. >>>> >>>> Sure, I was looking at the old version already. It is interesting. But >>>> also IMO needs quite a bit more work to approach achieving what is >>>> implied from the name of the feature. It would need proper deadline >>>> based sched job picking, and even then drm sched is mostly just a >>>> frontend. So once past runnable status and jobs handed over to backend, >>>> without further driver work it probably wouldn't be very effective past >>>> very lightly loaded systems. >>> >>> Yes, but all of that is not part of dma_fence ;-) >> >> :) Okay. >> >> Having said that, do we need a step back to think about whether adding >> deadline to dma-fences is not making them something too much different >> to what they were? Going from purely synchronisation primitive more >> towards scheduling paradigms. Just to brainstorm if there will not be >> any unintended consequences. I should mention this in your RFC thread >> actually. > > Perhaps "deadline" isn't quite the right name, but I haven't thought > of anything better. It is really a hint to the fence signaller about > how soon it is interested in a result so the driver can factor that > into freq scaling decisions. Maybe "goal" or some other term would be > better? Don't know, no strong opinion on the name at the moment. For me it was more about the change of what type of side channel data is getting attached to dma-fence and whether it changes what the primitive is for. > I guess that can factor into scheduling decisions as well.. but we > already have priority for that. My main interest is freq mgmt. > > (Thankfully we don't have performance and efficiency cores to worry > about, like CPUs ;-)) > >>> A pretty common challenging usecase is still the single fullscreen >>> game, where scheduling isn't the problem, but landing at an >>> appropriate GPU freq absolutely is. (UI workloads are perhaps more >>> interesting from a scheduler standpoint, but they generally aren't >>> challenging from a load/freq standpoint.) >> >> Challenging as in picking the right operating point? Might be latency >> impacted (and so user perceived UI smoothness) due missing waitboost for >> anything syncobj related. I don't know if anything to measure that >> exists currently though. Assuming it is measurable then the question >> would be is it perceivable. >>> Fwiw, the original motivation of the series was to implement something >>> akin to i915 pageflip boosting without having to abandon the atomic >>> helpers. (And, I guess it would also let i915 preserve that feature >>> if it switched to atomic helpers.. I'm unsure if there are still other >>> things blocking i915's migration.) >> >> Question for display folks I guess. >> >>>> Then if we fast forward to a world where schedulers perhaps become fully >>>> deadline aware (we even had this for i915 few years back) then the >>>> question will be does equating waits with immediate deadlines still >>>> works. Maybe not too well because we wouldn't have the ability to >>>> distinguish between the "someone is waiting" signal from the otherwise >>>> propagated deadlines. >>> >>> Is there any other way to handle a wait boost than expressing it as an >>> ASAP deadline? >> >> A leading question or just a question? Nothing springs to my mind at the >> moment. > > Just a question. The immediate deadline is the only thing that makes > sense to me, but that could be because I'm looking at it from the > perspective of also trying to handle the case where missing vblank > reduces utilization and provides the wrong signal to gpufreq.. i915 > already has a way to handle this internally, but it involves bypassing > the atomic helpers, which isn't a thing I want to encourage other > drivers to do. And completely doesn't work for situations where the > gpu and display are separate devices. Right, there is yet another angle to discuss with Daniel here who AFAIR was a bit against i915 priority inheritance going past a single device instance. In which case DRI_PRIME=1 would lose the ability to boost frame buffer dependency chains. Opens up the question of deadline inheritance across different drivers too. Or perhaps Daniel would be okay with this working if implemented at the dma-fence layer. Regards, Tvrtko
On 20/02/2023 15:52, Rob Clark wrote: > On Mon, Feb 20, 2023 at 3:33 AM Tvrtko Ursulin > <tvrtko.ursulin@linux.intel.com> wrote: >> >> >> On 17/02/2023 20:45, Rodrigo Vivi wrote: [snip] >> Yeah I agree. And as not all media use cases are the same, as are not >> all compute contexts someone somewhere will need to run a series of >> workloads for power and performance numbers. Ideally that someone would >> be the entity for which it makes sense to look at all use cases, from >> server room to client, 3d, media and compute for both. If we could get >> the capability to run this in some automated fashion, akin to CI, we >> would even have a chance to keep making good decisions in the future. >> >> Or we do some one off testing for this instance, but we still need a >> range of workloads and parts to do it properly.. >> >>>> I also think the "arms race" scenario isn't really as much of a >>>> problem as you think. There aren't _that_ many things using the GPU >>>> at the same time (compared to # of things using CPU). And a lot of >>>> mobile games throttle framerate to avoid draining your battery too >>>> quickly (after all, if your battery is dead you can't keep buying loot >>>> boxes or whatever). >>> >>> Very good point. >> >> On this one I still disagree from the point of view that it does not >> make it good uapi if we allow everyone to select themselves for priority >> handling (one flavour or the other). > > There is plenty of precedent for userspace giving hints to the kernel > about scheduling and freq mgmt. Like schedutil uclamp stuff. > Although I think that is all based on cgroups. I knew about SCHED_DEADLINE and that it requires CAP_SYS_NICE, but I did not know about uclamp. Quick experiment with uclampset suggests it indeed does not require elevated privilege. If that is indeed so, it is good enough for me as a precedent. It appears to work using sched_setscheduler so maybe could define something similar in i915/xe, per context or per client, not sure. Maybe it would start as a primitive implementation but the uapi would not preclude making it smart(er) afterwards. Or passing along to GuC to do it's thing with it. > In the fence/syncobj case, I think we need per-wait hints.. because > for a single process the driver will be doing both housekeeping waits > and potentially urgent waits. There may also be some room for some > cgroup or similar knobs to control things like what max priority an > app can ask for, and whether or how aggressively the kernel responds > to the "deadline" hints. So as far as "arms race", I don't think I'd Per wait hints are okay I guess even with "I am important" in their name if sched_setscheduler allows raising uclamp.min just like that. In which case cgroup limits to mimick cpu uclamp also make sense. > change anything about my "fence deadline" proposal.. but that it might > just be one piece of the overall puzzle. That SCHED_DEADLINE requires CAP_SYS_NICE does not worry you? Regards, Tvrtko
On 20/02/2023 16:44, Tvrtko Ursulin wrote: > > On 20/02/2023 15:52, Rob Clark wrote: >> On Mon, Feb 20, 2023 at 3:33 AM Tvrtko Ursulin >> <tvrtko.ursulin@linux.intel.com> wrote: >>> >>> >>> On 17/02/2023 20:45, Rodrigo Vivi wrote: > > [snip] > >>> Yeah I agree. And as not all media use cases are the same, as are not >>> all compute contexts someone somewhere will need to run a series of >>> workloads for power and performance numbers. Ideally that someone would >>> be the entity for which it makes sense to look at all use cases, from >>> server room to client, 3d, media and compute for both. If we could get >>> the capability to run this in some automated fashion, akin to CI, we >>> would even have a chance to keep making good decisions in the future. >>> >>> Or we do some one off testing for this instance, but we still need a >>> range of workloads and parts to do it properly.. >>> >>>>> I also think the "arms race" scenario isn't really as much of a >>>>> problem as you think. There aren't _that_ many things using the GPU >>>>> at the same time (compared to # of things using CPU). And a lot of >>>>> mobile games throttle framerate to avoid draining your battery too >>>>> quickly (after all, if your battery is dead you can't keep buying loot >>>>> boxes or whatever). >>>> >>>> Very good point. >>> >>> On this one I still disagree from the point of view that it does not >>> make it good uapi if we allow everyone to select themselves for priority >>> handling (one flavour or the other). >> >> There is plenty of precedent for userspace giving hints to the kernel >> about scheduling and freq mgmt. Like schedutil uclamp stuff. >> Although I think that is all based on cgroups. > > I knew about SCHED_DEADLINE and that it requires CAP_SYS_NICE, but I did > not know about uclamp. Quick experiment with uclampset suggests it > indeed does not require elevated privilege. If that is indeed so, it is > good enough for me as a precedent. > > It appears to work using sched_setscheduler so maybe could define > something similar in i915/xe, per context or per client, not sure. > > Maybe it would start as a primitive implementation but the uapi would > not preclude making it smart(er) afterwards. Or passing along to GuC to > do it's thing with it. Hmmm having said that, how would we fix clvk performance using that? We would either need the library to do a new step when creating contexts, or allow external control so outside entity can do it. And then the question is based on what it decides to do it? Is it possible to know which, for instance, Chrome tab will be (or is) using clvk so that tab management code does it? Regards, Tvrtko >> In the fence/syncobj case, I think we need per-wait hints.. because >> for a single process the driver will be doing both housekeeping waits >> and potentially urgent waits. There may also be some room for some >> cgroup or similar knobs to control things like what max priority an >> app can ask for, and whether or how aggressively the kernel responds >> to the "deadline" hints. So as far as "arms race", I don't think I'd > > Per wait hints are okay I guess even with "I am important" in their name > if sched_setscheduler allows raising uclamp.min just like that. In which > case cgroup limits to mimick cpu uclamp also make sense. > >> change anything about my "fence deadline" proposal.. but that it might >> just be one piece of the overall puzzle. > > That SCHED_DEADLINE requires CAP_SYS_NICE does not worry you? > > Regards, > > Tvrtko
On Mon, Feb 20, 2023 at 8:44 AM Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> wrote: > > > On 20/02/2023 15:52, Rob Clark wrote: > > On Mon, Feb 20, 2023 at 3:33 AM Tvrtko Ursulin > > <tvrtko.ursulin@linux.intel.com> wrote: > >> > >> > >> On 17/02/2023 20:45, Rodrigo Vivi wrote: > > [snip] > > >> Yeah I agree. And as not all media use cases are the same, as are not > >> all compute contexts someone somewhere will need to run a series of > >> workloads for power and performance numbers. Ideally that someone would > >> be the entity for which it makes sense to look at all use cases, from > >> server room to client, 3d, media and compute for both. If we could get > >> the capability to run this in some automated fashion, akin to CI, we > >> would even have a chance to keep making good decisions in the future. > >> > >> Or we do some one off testing for this instance, but we still need a > >> range of workloads and parts to do it properly.. > >> > >>>> I also think the "arms race" scenario isn't really as much of a > >>>> problem as you think. There aren't _that_ many things using the GPU > >>>> at the same time (compared to # of things using CPU). And a lot of > >>>> mobile games throttle framerate to avoid draining your battery too > >>>> quickly (after all, if your battery is dead you can't keep buying loot > >>>> boxes or whatever). > >>> > >>> Very good point. > >> > >> On this one I still disagree from the point of view that it does not > >> make it good uapi if we allow everyone to select themselves for priority > >> handling (one flavour or the other). > > > > There is plenty of precedent for userspace giving hints to the kernel > > about scheduling and freq mgmt. Like schedutil uclamp stuff. > > Although I think that is all based on cgroups. > > I knew about SCHED_DEADLINE and that it requires CAP_SYS_NICE, but I did > not know about uclamp. Quick experiment with uclampset suggests it > indeed does not require elevated privilege. If that is indeed so, it is > good enough for me as a precedent. > > It appears to work using sched_setscheduler so maybe could define > something similar in i915/xe, per context or per client, not sure. > > Maybe it would start as a primitive implementation but the uapi would > not preclude making it smart(er) afterwards. Or passing along to GuC to > do it's thing with it. > > > In the fence/syncobj case, I think we need per-wait hints.. because > > for a single process the driver will be doing both housekeeping waits > > and potentially urgent waits. There may also be some room for some > > cgroup or similar knobs to control things like what max priority an > > app can ask for, and whether or how aggressively the kernel responds > > to the "deadline" hints. So as far as "arms race", I don't think I'd > > Per wait hints are okay I guess even with "I am important" in their name > if sched_setscheduler allows raising uclamp.min just like that. In which > case cgroup limits to mimick cpu uclamp also make sense. > > > change anything about my "fence deadline" proposal.. but that it might > > just be one piece of the overall puzzle. > > That SCHED_DEADLINE requires CAP_SYS_NICE does not worry you? This gets to why the name "fence deadline" is perhaps not the best.. it really isn't meant to be analogous to SCHED_DEADLINE, but rather just a hint to the driver about what userspace is doing. Maybe we just document it more strongly as a hint? BR, -R > Regards, > > Tvrtko
On Mon, Feb 20, 2023 at 8:51 AM Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> wrote: > > > On 20/02/2023 16:44, Tvrtko Ursulin wrote: > > > > On 20/02/2023 15:52, Rob Clark wrote: > >> On Mon, Feb 20, 2023 at 3:33 AM Tvrtko Ursulin > >> <tvrtko.ursulin@linux.intel.com> wrote: > >>> > >>> > >>> On 17/02/2023 20:45, Rodrigo Vivi wrote: > > > > [snip] > > > >>> Yeah I agree. And as not all media use cases are the same, as are not > >>> all compute contexts someone somewhere will need to run a series of > >>> workloads for power and performance numbers. Ideally that someone would > >>> be the entity for which it makes sense to look at all use cases, from > >>> server room to client, 3d, media and compute for both. If we could get > >>> the capability to run this in some automated fashion, akin to CI, we > >>> would even have a chance to keep making good decisions in the future. > >>> > >>> Or we do some one off testing for this instance, but we still need a > >>> range of workloads and parts to do it properly.. > >>> > >>>>> I also think the "arms race" scenario isn't really as much of a > >>>>> problem as you think. There aren't _that_ many things using the GPU > >>>>> at the same time (compared to # of things using CPU). And a lot of > >>>>> mobile games throttle framerate to avoid draining your battery too > >>>>> quickly (after all, if your battery is dead you can't keep buying loot > >>>>> boxes or whatever). > >>>> > >>>> Very good point. > >>> > >>> On this one I still disagree from the point of view that it does not > >>> make it good uapi if we allow everyone to select themselves for priority > >>> handling (one flavour or the other). > >> > >> There is plenty of precedent for userspace giving hints to the kernel > >> about scheduling and freq mgmt. Like schedutil uclamp stuff. > >> Although I think that is all based on cgroups. > > > > I knew about SCHED_DEADLINE and that it requires CAP_SYS_NICE, but I did > > not know about uclamp. Quick experiment with uclampset suggests it > > indeed does not require elevated privilege. If that is indeed so, it is > > good enough for me as a precedent. > > > > It appears to work using sched_setscheduler so maybe could define > > something similar in i915/xe, per context or per client, not sure. > > > > Maybe it would start as a primitive implementation but the uapi would > > not preclude making it smart(er) afterwards. Or passing along to GuC to > > do it's thing with it. > > Hmmm having said that, how would we fix clvk performance using that? We > would either need the library to do a new step when creating contexts, > or allow external control so outside entity can do it. And then the > question is based on what it decides to do it? Is it possible to know > which, for instance, Chrome tab will be (or is) using clvk so that tab > management code does it? I am not sure.. the clvk usage is, I think, not actually in chrome itself, but something camera related? Presumably we could build some cgroup knobs to control how the driver reacts to the "deadline" hints (ie. ignore them completely, or impose some upper limit on how much freq boost will be applied, etc). I think this sort of control of how the driver responds to hints probably fits best with cgroups, as that is how we are already implementing similar tuning for cpufreq/sched. (Ie. foreground app or tab gets moved to a different cgroup.) But admittedly I haven't looked too closely at how cgroups work on the kernel side. BR, -R > Regards, > > Tvrtko > > >> In the fence/syncobj case, I think we need per-wait hints.. because > >> for a single process the driver will be doing both housekeeping waits > >> and potentially urgent waits. There may also be some room for some > >> cgroup or similar knobs to control things like what max priority an > >> app can ask for, and whether or how aggressively the kernel responds > >> to the "deadline" hints. So as far as "arms race", I don't think I'd > > > > Per wait hints are okay I guess even with "I am important" in their name > > if sched_setscheduler allows raising uclamp.min just like that. In which > > case cgroup limits to mimick cpu uclamp also make sense. > > > >> change anything about my "fence deadline" proposal.. but that it might > >> just be one piece of the overall puzzle. > > > > That SCHED_DEADLINE requires CAP_SYS_NICE does not worry you? > > > > Regards, > > > > Tvrtko
From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> In i915 we have this concept of "wait boosting" where we give a priority boost for instance to fences which are actively waited upon from userspace. This has it's pros and cons and can certainly be discussed at lenght. However fact is some workloads really like it. Problem is that with the arrival of drm syncobj and a new userspace waiting entry point it added, the waitboost mechanism was bypassed. Hence I cooked up this mini series really (really) quickly to see if some discussion can be had. It adds a concept of "wait count" to dma fence, which is incremented for every explicit dma_fence_enable_sw_signaling and dma_fence_add_wait_callback (like dma_fence_add_callback but from explicit/userspace wait paths). Individual drivers can then inspect this via dma_fence_wait_count() and decide to wait boost the waits on such fences. Again, quickly put together and smoke tested only - no guarantees whatsoever and I will rely on interested parties to test and report if it even works or how well. v2: * Small fixups based on CI feedback: * Handle decrement correctly for already signalled case while adding callback. * Remove i915 assert which was making sure struct i915_request does not grow. * Split out the i915 patch into three separate functional changes. Tvrtko Ursulin (5): dma-fence: Track explicit waiters drm/syncobj: Mark syncobj waits as external waiters drm/i915: Waitboost external waits drm/i915: Mark waits as explicit drm/i915: Wait boost requests waited upon by others drivers/dma-buf/dma-fence.c | 102 ++++++++++++++++------ drivers/gpu/drm/drm_syncobj.c | 6 +- drivers/gpu/drm/i915/gt/intel_engine_pm.c | 1 - drivers/gpu/drm/i915/i915_request.c | 13 ++- include/linux/dma-fence.h | 14 +++ 5 files changed, 101 insertions(+), 35 deletions(-)