[v2,00/23] drm/msm+PM+icc: Make job_run() reclaim-safe

Message

Rob Clark March 20, 2023, 2:43 p.m. UTC

From: Rob Clark <[e-mail]>

Inspired by https://lore.kernel.org/dri-devel/[e-mail]/
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 patch pre-allocates the hw_fence, splitting allocation and
initialization, to avoid allocation in the job_run() path.  The next
eight decouple the obj lock from job_run(), as the obj lock is required
to pin/unpin backing pages (ie. holding an obj lock in job_run() could
deadlock the shrinker by blocking forward progress towards pinned buffers
becoming idle).  Followed by two so that we could idr_preload() in order
to avoid memory allocations under locks indirectly connected to the
shrinker path.

Next are three paths to decouple initialization (where allocations are
needed) from GPU runpm and devfreq, to avoid allocations in the fence
signaling path.  Followed by various PM devfreq/qos and interconnect
locking fixes to decouple initialization (allocation) from runtime.

And finally, the last patch is a modified version of danvet's patch to
add lockdep annotations to gpu scheduler, but does so conditionally so
that drivers can opt-in.

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.

Rob Clark (23):
  drm/msm: Pre-allocate hw_fence
  drm/msm: Move submit bo flags update from obj lock
  drm/msm/gem: Tidy up VMA API
  drm/msm: Decouple vma tracking from obj lock
  drm/msm/gem: Simplify vmap vs LRU tracking
  drm/gem: Export drm_gem_lru_move_tail_locked()
  drm/msm/gem: Move update_lru()
  drm/msm/gem: Protect pin_count/madv by LRU lock
  drm/msm/gem: Avoid obj lock in job_run()
  drm/msm: Switch idr_lock to spinlock
  drm/msm: Use idr_preload()
  drm/msm/gpu: Move fw loading out of hw_init() path
  drm/msm/gpu: Move BO allocation out of hw_init
  drm/msm/a6xx: Move ioremap out of hw_init path
  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
  soc: qcom: smd-rpm: Use GFP_ATOMIC in write path
  interconnect: Fix locking for runpm vs reclaim
  interconnect: Teach lockdep about icc_bw_lock order
  drm/sched: Add (optional) fence signaling annotation

 drivers/base/power/qos.c                   |  83 +++++++++---
 drivers/devfreq/devfreq.c                  |  52 ++++----
 drivers/gpu/drm/drm_gem.c                  |  11 +-
 drivers/gpu/drm/msm/adreno/a5xx_gpu.c      |  48 ++++---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.c      |  18 ++-
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c      |  46 ++++---
 drivers/gpu/drm/msm/adreno/adreno_device.c |   6 +
 drivers/gpu/drm/msm/adreno/adreno_gpu.c    |   9 +-
 drivers/gpu/drm/msm/msm_drv.c              |   6 +-
 drivers/gpu/drm/msm/msm_fence.c            |  12 +-
 drivers/gpu/drm/msm/msm_fence.h            |   3 +-
 drivers/gpu/drm/msm/msm_gem.c              | 145 ++++++++++++++-------
 drivers/gpu/drm/msm/msm_gem.h              |  29 +++--
 drivers/gpu/drm/msm/msm_gem_submit.c       |  27 ++--
 drivers/gpu/drm/msm/msm_gem_vma.c          |  91 ++++++++++---
 drivers/gpu/drm/msm/msm_gpu.h              |   8 +-
 drivers/gpu/drm/msm/msm_ringbuffer.c       |   9 +-
 drivers/gpu/drm/msm/msm_submitqueue.c      |   2 +-
 drivers/gpu/drm/scheduler/sched_main.c     |   9 ++
 drivers/interconnect/core.c                |  18 ++-
 drivers/soc/qcom/smd-rpm.c                 |   2 +-
 include/drm/drm_gem.h                      |   1 +
 include/drm/gpu_scheduler.h                |   2 +
 23 files changed, 416 insertions(+), 221 deletions(-)

Comments

Bjorn Andersson April 7, 2023, 5:41 p.m. UTC | #1

On Mon, 20 Mar 2023 07:43:22 -0700, Rob Clark wrote:
> From: Rob Clark <[e-mail]>
> 
> Inspired by https://lore.kernel.org/dri-devel/[e-mail]/
> 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.
> 
> [...]

Applied, thanks!

[20/23] soc: qcom: smd-rpm: Use GFP_ATOMIC in write path
        commit: 5808c532ca0a983d643319caca44f2bcb148298f

Best regards,