| Message ID | 20230314141904.1210824-1-tvrtko.ursulin@linux.intel.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | DRM scheduling cgroup controller | expand |
Hello, Tvrtko. On Tue, Mar 14, 2023 at 02:18:54PM +0000, Tvrtko Ursulin wrote: > DRM scheduling soft limits > ~~~~~~~~~~~~~~~~~~~~~~~~~~ > > Because of the heterogenous hardware and driver DRM capabilities, soft limits > are implemented as a loose co-operative (bi-directional) interface between the > controller and DRM core. > > The controller configures the GPU time allowed per group and periodically scans > the belonging tasks to detect the over budget condition, at which point it > invokes a callback notifying the DRM core of the condition. > > DRM core provides an API to query per process GPU utilization and 2nd API to > receive notification from the cgroup controller when the group enters or exits > the over budget condition. > > Individual DRM drivers which implement the interface are expected to act on this > in the best-effort manner only. There are no guarantees that the soft limits > will be respected. > > DRM scheduling soft limits interface files > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ In general, I'm in favor of your approach but can you please stop using the term "soft limit". That's a term with a specific historical meaning in cgroup, so it gets really confusing when you use the term for hierarchical weighted control. If you need a term to refer to how the weighted control is implemented by throttling cgroups at target rates, please just come up with a different term "usage threshold based control", "usage throttling based control" or whichever you may like. > drm.weight > Standard cgroup weight based control [1, 10000] used to configure the > relative distributing of GPU time between the sibling groups. > > This builds upon the per client GPU utilisation work which landed recently for a > few drivers. My thinking is that in principle, an intersect of drivers which > support both that and some sort of scheduling control, like priorities, could > also in theory support this controller. > > Another really interesting angle for this controller is that it mimics the same > control menthod used by the CPU scheduler. That is the proportional/weight based > GPU time budgeting. Which makes it easy to configure and does not need a new > mental model. FWIW, the hierarchical weighted distribution is also implemented by IO control. > However, as the introduction mentions, GPUs are much more heterogenous and > therefore the controller uses very "soft" wording as to what it promises. The > general statement is that it can define budgets, notify clients when they are > over them, and let individual drivers implement best effort handling of those > conditions. Maybe "best effort" is more suited than "soft"? ... > Roughly simultaneously we run the following two benchmarks in each session > respectively: > > 1) > ./GpuTest /test=pixmark_julia_fp32 /width=1920 /height=1080 /fullscreen /no_scorebox /benchmark /benchmark_duration_ms=60000 > > 2) > vblank_mode=0 bin/testfw_app --gl_api=desktop_core --width=1920 --height=1080 --fullscreen 1 --gfx=glfw -t gl_manhattan > > (The only reason for vsync off here is because I struggled to find an easily > runnable and demanding enough benchmark, or to run on a screen large enough to > make even a simpler ones demanding.) > > With this test we get 252fps from GpuTest and 96fps from GfxBenchmark. > > Premise here is that one of these GPU intensive benchmarks is intended to be ran > by the user with lower priority. Imagine kicking off some background compute > processing and continuing to use the UI for other tasks. Hence the user will now > re-run the test by first lowering the weight control of the first session (DRM > cgroup): > > 1) > echo 50 | sudo tee /sys/fs/cgroup/`cut -d':' -f3 /proc/self/cgroup`/drm.weight > ./GpuTest /test=pixmark_julia_fp32 /width=1920 /height=1080 /fullscreen /no_scorebox /benchmark /benchmark_duration_ms=60000 > > 2) > vblank_mode=0 bin/testfw_app --gl_api=desktop_core --width=1920 --height=1080 --fullscreen 1 --gfx=glfw -t gl_manhattan > > In this case we will see that GpuTest has recorded 208fps (~18% down) and > GfxBenchmark 114fps (18% up), demonstrating that even a very simple approach of > wiring up i915 to the DRM cgroup controller can enable external GPU scheduling > control. It's really nice to see it working pretty intuitively. > * For now (RFC) I haven't implemented the 2nd suggestion from Tejun of having > a shadow tree which would only contain groups with DRM clients. (Purpose > being less nodes to traverse in the scanning loop.) > > * Is the global state passing from can_attach to attach really okay? (I need > source and destination css.) Right now, it is and there are other places that depend on it. Obviously, it's not great and we probably want to add explicit context passed around instead in the future, but for now, it should be okay. While not fully polished, from cgroup POV, the series looks pretty good and I'd be happy to see it merged. Thanks.
From: Tvrtko Ursulin <tvrtko.ursulin@intel.com> This series contains a proposal for a DRM scheduling cgroup controller which implements a weight based hierarchical GPU usage budget based controller similar in concept to some of the existing controllers. Motivation mostly comes from my earlier proposal where I identified that GPU scheduling lags significantly behind what is available for CPU and IO. Whereas back then I was proposing to somehow tie this with process nice, feedback mostly was that people wanted cgroups. So here it is - in the world of heterogenous computing pipelines I think it is time to do something about this gap. Code is not finished but should survive some light experimenting with. I am sharing it early since the topic has been controversial in the past. I hope to demonstrate there are gains to be had in real world usage(*), today, and that the concepts the proposal relies are well enough established and stable. *) Specifically under ChromeOS which uses cgroups to control CPU bandwith for VMs based on the window focused status. It can be demonstrated how GPU scheduling control can easily be integrated into that setup. *) Another real world example later in the cover letter. There should be no conflict with this proposal and any efforts to implement memory usage based controller. Skeleton DRM cgroup controller is deliberatly purely a skeleton patch where any further functionality can be added with no real conflicts. [In fact, perhaps scheduling is even easier to deal with than memory accounting.] Structure of the series is as follows: 1-2) Improve client ownership tracking in DRM core. 3) Adds a skeleton DRM cgroup controller with no functionality. 4-7) Laying down some infrastructure to enable the controller. 8) The controller itself. 9-10) i915 support for the controller. The proposals defines a delegation of duties between the tree parties: cgroup controller, DRM core and individual drivers. Two way communication interfaces are then defined to enable the delegation to work. DRM scheduling soft limits ~~~~~~~~~~~~~~~~~~~~~~~~~~ Because of the heterogenous hardware and driver DRM capabilities, soft limits are implemented as a loose co-operative (bi-directional) interface between the controller and DRM core. The controller configures the GPU time allowed per group and periodically scans the belonging tasks to detect the over budget condition, at which point it invokes a callback notifying the DRM core of the condition. DRM core provides an API to query per process GPU utilization and 2nd API to receive notification from the cgroup controller when the group enters or exits the over budget condition. Individual DRM drivers which implement the interface are expected to act on this in the best-effort manner only. There are no guarantees that the soft limits will be respected. DRM scheduling soft limits interface files ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ drm.weight Standard cgroup weight based control [1, 10000] used to configure the relative distributing of GPU time between the sibling groups. This builds upon the per client GPU utilisation work which landed recently for a few drivers. My thinking is that in principle, an intersect of drivers which support both that and some sort of scheduling control, like priorities, could also in theory support this controller. Another really interesting angle for this controller is that it mimics the same control menthod used by the CPU scheduler. That is the proportional/weight based GPU time budgeting. Which makes it easy to configure and does not need a new mental model. However, as the introduction mentions, GPUs are much more heterogenous and therefore the controller uses very "soft" wording as to what it promises. The general statement is that it can define budgets, notify clients when they are over them, and let individual drivers implement best effort handling of those conditions. Delegation of duties in the implementation goes likes this: * DRM cgroup controller implements the control files, the scanning loop and tracks the DRM clients associated with each cgroup. It provides API DRM core needs to call to (de)register and migrate clients. * DRM core defines two call-backs which the core calls directly: First for querying GPU time by a client and second for notifying the client that it is over budget. It calls controller API for (de)registering clients and migrating then between tasks on file descriptor hand over. * Individual drivers implement the above mentiopned callbacks and register them with the DRM core. What I have demonstrated in practice is that when wired to i915, in a really primitive way where the over-budget condition simply lowers the scheduling priority, the concept can be almost equally effective as the static priority control. I say almost because the design where budget control depends on the periodic usage scanning has a fundamental delay, so responsiveness will depend on the scanning period, which may or may not be a problem for a particular use case. There are also interesting conversations to be had around mental models for what is GPU usage as a single number when faced with GPUs which have different execution engines. To an extent this is similar to the multi-core and cgroup CPU controller problems, but definitely goes further than that. I deliberately did not want to include any such complications in the controller itself and left the individual drivers to handle it. For instance in the i915 over-budget callback it will not do anything unless client's GPU usage is on a physical engine which is oversubscribed. This enables multiple clients to be harmlessly over budget, as long as they are not competing for the same GPU resource. Example usage from within a Linux desktop ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Standard Linux distributions like Ubuntu already uses cgroups heavily for session management and that could easily be extended with the DRM controller. After logging into the system graphically we can enable the DRM controller throughout the cgroups hierarchy: echo +drm > /sys/fs/cgroup/cgroup.subtree_control echo +drm > /sys/fs/cgroup/user.slice/cgroup.subtree_control echo +drm > /sys/fs/cgroup/user.slice/user-$(id -u).slice/cgroup.subtree_control Next we will open two SSH sessions, just so separate cgroups are handily created by systemd for this experiment. Roughly simultaneously we run the following two benchmarks in each session respectively: 1) ./GpuTest /test=pixmark_julia_fp32 /width=1920 /height=1080 /fullscreen /no_scorebox /benchmark /benchmark_duration_ms=60000 2) vblank_mode=0 bin/testfw_app --gl_api=desktop_core --width=1920 --height=1080 --fullscreen 1 --gfx=glfw -t gl_manhattan (The only reason for vsync off here is because I struggled to find an easily runnable and demanding enough benchmark, or to run on a screen large enough to make even a simpler ones demanding.) With this test we get 252fps from GpuTest and 96fps from GfxBenchmark. Premise here is that one of these GPU intensive benchmarks is intended to be ran by the user with lower priority. Imagine kicking off some background compute processing and continuing to use the UI for other tasks. Hence the user will now re-run the test by first lowering the weight control of the first session (DRM cgroup): 1) echo 50 | sudo tee /sys/fs/cgroup/`cut -d':' -f3 /proc/self/cgroup`/drm.weight ./GpuTest /test=pixmark_julia_fp32 /width=1920 /height=1080 /fullscreen /no_scorebox /benchmark /benchmark_duration_ms=60000 2) vblank_mode=0 bin/testfw_app --gl_api=desktop_core --width=1920 --height=1080 --fullscreen 1 --gfx=glfw -t gl_manhattan In this case we will see that GpuTest has recorded 208fps (~18% down) and GfxBenchmark 114fps (18% up), demonstrating that even a very simple approach of wiring up i915 to the DRM cgroup controller can enable external GPU scheduling control. * Note here that default weight is 100, so setting 50 for the background session is asking the controller to give it half as much GPU bandwidth. v2: * Prefaced the series with some core DRM work as suggested by Christian. * Dropped the priority based controller for now. * Dropped the introspection cgroup controller file. * Implemented unused budget sharing/propagation. * Some small fixes/tweak as per review feedback and in general. v3: * Dropped one upstreamed patch. * Logging cleanup (use DRM macros where available). v4: * Dropped the struct pid tracking indirection in favour of tracking individual DRM clients directly in the controller. (Michal Koutný) * Added boot time param for configuring the scanning period. (Tejun Heo) * Improved spreading of unused budget to over budget clients, regardless of their location in the tree so that all unused budget can be utilized. * Made scanning more robust by not re-starting it on every client de- registration and removal. Instead new baseline GPU activity data is simply collected on those events and next scan invocation can proceed as scheduled. * Dropped the debugging aids from the series. TODOs/Opens: * For now (RFC) I haven't implemented the 2nd suggestion from Tejun of having a shadow tree which would only contain groups with DRM clients. (Purpose being less nodes to traverse in the scanning loop.) * Is the global state passing from can_attach to attach really okay? (I need source and destination css.) Tvrtko Ursulin (10): drm: Track clients by tgid and not tid drm: Update file owner during use cgroup: Add the DRM cgroup controller drm/cgroup: Track DRM clients per cgroup drm/cgroup: Add ability to query drm cgroup GPU time drm/cgroup: Add over budget signalling callback drm/cgroup: Only track clients which are providing drm_cgroup_ops cgroup/drm: Introduce weight based drm cgroup control drm/i915: Wire up with drm controller GPU time query drm/i915: Implement cgroup controller over budget throttling Documentation/admin-guide/cgroup-v2.rst | 31 + drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 6 +- drivers/gpu/drm/drm_auth.c | 3 +- drivers/gpu/drm/drm_debugfs.c | 12 +- drivers/gpu/drm/drm_file.c | 46 +- drivers/gpu/drm/drm_ioctl.c | 3 + .../gpu/drm/i915/gem/i915_gem_execbuffer.c | 38 +- drivers/gpu/drm/i915/i915_driver.c | 11 + drivers/gpu/drm/i915/i915_drm_client.c | 209 ++++++- drivers/gpu/drm/i915/i915_drm_client.h | 13 + drivers/gpu/drm/nouveau/nouveau_drm.c | 5 +- drivers/gpu/drm/vmwgfx/vmwgfx_gem.c | 6 +- include/drm/drm_drv.h | 36 ++ include/drm/drm_file.h | 19 +- include/linux/cgroup_drm.h | 29 + include/linux/cgroup_subsys.h | 4 + init/Kconfig | 7 + kernel/cgroup/Makefile | 1 + kernel/cgroup/drm.c | 568 ++++++++++++++++++ 19 files changed, 1020 insertions(+), 27 deletions(-) create mode 100644 include/linux/cgroup_drm.h create mode 100644 kernel/cgroup/drm.c