This is a continuation of the previous discussion which was here: https://lore.kernel.org/dri-devel/[email protected]/
Work is now converted to a much simpler approach by adding real-time scheduling workqueues based on Tejun's feedback To re-cap, when an userspace thread submits GPU work, due how the DRM scheduler uses workqueues to feed the GPU, and regardless of the GPU rendering context priority, or the CPU scheduling priority of the userspace thread itself, the use of workqueues can add significant latency to the submit path. When CPU is busy with enough backround load this translates to severe latency spikes measured as time between userspace submitting work and GPU actually being given that work to execute. With the panthor workqueue upgraded to use WQ_HIGHPRI and varying the CPU priority of the submit thread, the test program from https://gitlab.freedesktop.org/panfrost/linux/-/work_items/49 reproduces these kind of latencies: . N RT M 27 28 32 95% 163 246 809 98% 924 991 1882 Legend: M = Median submit latency in us 95% = Percentile latency in us . = Userspace submit thread SCHED_OTHER N = -||= nice -1 RT = -||- FIFO 1 Upgrading the panthor workqueues so that the realtime GPU priority queue uses WQ_RT, submit latency becomes completely controlled with the median of 14us and 95 and 98-th percentiles at 23us and 25us respectively. Important to note is that VK_QUEUE_GLOBAL_PRIORITY_REALTIME already requires that userspace either has CAP_SYS_NICE, or the DRM master capability. The latter can be had by a single DRM client per GPU, usually the compositor since DRM master is required to drive the display hardware. It is not possible to have two DRM master clients simultaneously. On the overall this limits the number of userspace threads which can access the RT queue - it either has to be a compositor, or a rendering client which was explicitly granted CAP_SYS_NICE. v2: * See patch 1 changelog. v3: * See patch 1 changelog + apologies for v3 following so quickly after v2. I have found a race condition as I expanded the testing to a second platform. v4: * See patch 1 changelog. v5: * Rebased on top of tj/for-7.4 branch. * New patch in series simplifies unbound sysfs attribute handling. * WQ_RTPRI reworked to fold RT marker into attr->nice. v6: * Fixed accidental sysfs attribute rename (patch 1). * Various review feedback as per patch 2 changelog. v7: * Describe DRM master in the cover letter. * New patch in series to separate attribute sanitizing steps. * WQ_RT updated per the respective change log. Cc: Boris Brezillon <[email protected]> Cc: Chia-I Wu <[email protected]> Cc: Liviu Dudau <[email protected]> Cc: Matthew Brost <[email protected]> Cc: Steven Price <[email protected]> Cc: Tejun Heo <[email protected]> Tvrtko Ursulin (4): workqueue: Simplify unbound sysfs attribute registration workqueue: Separate sysfs mutable vs immutable wq attribute sanitizing workqueue: Add support for real-time workers drm/panthor: Create per queue priority workqueues Documentation/core-api/workqueue.rst | 8 + drivers/gpu/drm/panthor/panthor_sched.c | 37 +++- include/linux/workqueue.h | 5 +- kernel/workqueue.c | 237 +++++++++++++++--------- tools/workqueue/wq_dump.py | 9 +- 5 files changed, 197 insertions(+), 99 deletions(-) -- 2.56.0
