From: Rob Clark <robdclark@xxxxxxxxxxxx> Inspired by https://lore.kernel.org/dri-devel/20200604081224.863494-10-daniel.vetter@xxxxxxxx/ it seemed like a good idea to get rid of memory allocation in job_run() fence signaling path, and use lockdep annotations to yell at us about anything that could deadlock against shrinker/reclaim. Anything that can trigger reclaim, or block on any other thread that has triggered reclaim, can block the GPU shrinker from releasing memory if it is waiting the job to complete, causing deadlock. The first two patches decouple allocation from devfreq->lock, and teach lockdep that devfreq->lock can be acquired in paths that the shrinker indirectly depends on. The next three patches do the same for PM QoS. And the next two do a similar thing for interconnect. And then finally the last two patches enable the lockdep fence- signalling annotations. v2: Switch from embedding hw_fence in submit/job object to preallocating the hw_fence. Rework "fenced unpin" locking to drop obj lock from fence signaling path (ie. the part that was still WIP in the first iteration of the patchset). Adds the final patch to enable fence signaling annotations now that job_run() and job_free() are safe. The PM devfreq/QoS and interconnect patches are unchanged. v3: Mostly unchanged, but series is much smaller now that drm changes have landed, mostly consisting of the remaining devfreq/qos/ interconnect fixes. v4: Re-work PM / QoS patch based on Rafael's suggestion v5: Add a couple more drm/msm patches for issues I found as making my way to the bottom of the rabbit hole. In particular, I had to move power enable earlier, before enqueing to the scheduler, rather than after the scheduler waits for in-fences, which means we could be powering up slightly earlier than needed. If runpm had a separate prepare + enable similar to the clk framework, we wouldn't need this. Rob Clark (11): PM / devfreq: Drop unneed locking to appease lockdep PM / devfreq: Teach lockdep about locking order PM / QoS: Fix constraints alloc vs reclaim locking PM / QoS: Decouple request alloc from dev_pm_qos_mtx PM / QoS: Teach lockdep about dev_pm_qos_mtx locking order interconnect: Fix locking for runpm vs reclaim interconnect: Teach lockdep about icc_bw_lock order drm/msm/a6xx: Remove GMU lock from runpm paths drm/msm: Move runpm enable in submit path drm/sched: Add (optional) fence signaling annotation drm/msm: Enable fence signalling annotations drivers/base/power/qos.c | 98 +++++++++++++++++++------- drivers/devfreq/devfreq.c | 52 +++++++------- drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 15 +--- drivers/gpu/drm/msm/msm_gem_submit.c | 2 + drivers/gpu/drm/msm/msm_gpu.c | 2 - drivers/gpu/drm/msm/msm_ringbuffer.c | 1 + drivers/gpu/drm/scheduler/sched_main.c | 9 +++ drivers/interconnect/core.c | 18 ++++- include/drm/gpu_scheduler.h | 2 + 9 files changed, 130 insertions(+), 69 deletions(-) -- 2.41.0
