AMD-GFX Archive on lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH 0/4] drm/amdkfd: don't stall other threads during large userptr registration
@ 2026-11-18 15:10 William Palacek
  2026-11-18 15:10 ` [PATCH 1/4] drm/amdgpu: limit userptr sg segments to half of sg->length William Palacek
                   ` (3 more replies)
  0 siblings, 4 replies; 5+ messages in thread
From: William Palacek @ 2026-11-18 15:10 UTC (permalink / raw)
  To: amd-gfx
  Cc: William.Palacek, Harish.Kasiviswanathan, Felix.Kuehling,
	christian.koenig, philip.yang

Registering a large userptr BO (hipHostMalloc) holds mmap_lock, p->mutex
and process_info->lock while every page of the range is faulted in. For
32 GiB on MI300X that takes seconds, and every other thread of the
process that submits GPU work waits behind it for the whole time.

Patch 1 limits userptr sg segments to half of sg->length. Larger
physically contiguous segments can overflow the padded segment length in
iommu_dma_map_sg(), and the pin fails. This already happens without the
series; patch 3 makes the pages contiguous enough that it hits on a
freshly booted system, so it goes first.

Patch 2 drops p->mutex from AMDKFD_IOC_GET_CLOCK_COUNTERS, which other
threads call while submitting work.

Patch 3 drops mmap_lock between 2 MiB chunks while faulting userptr
pages in.

Patch 4 faults in and validates userptr pages without p->mutex and
process_info->lock, and adds the BO to the valid list only if no
invalidation happened meanwhile.

Tested with two 32 GiB hipHostMalloc/hipHostFree cycles next to 4
PyTorch streams on 2 GPUs, measuring the worst GPU submission latency
of the other threads (limit 1 s), on the amdgpu 7.2 DKMS driver:

  MI300X, IOMMU DMA-FQ     unpatched    0/3 (6.1-6.2 s; on other boots
                                         the 32 GiB pin itself fails)
                           series       3/3 (0.80-0.95 s)
  MI300X, IOMMU identity   unpatched    0/3 (9.6-10.1 s)
                           patches 2-3  0/3 (10.5-11.0 s)
                           patches 2-4  3/3 (0.81-0.93 s)
  MI350X, IOMMU identity   unpatched    0/4 (4.55-4.84 s)
                           series       7/7 (0.21-0.39 s)

kfdtest KFDMemoryTest.* and KFDEvictTest.* pass on both GPUs.

William Palacek (4):
  drm/amdgpu: limit userptr sg segments to half of sg->length
  drm/amdkfd: don't take p->mutex in AMDKFD_IOC_GET_CLOCK_COUNTERS
  drm/amdgpu: drop mmap_lock between chunks when faulting userptr pages
  drm/amdkfd: register userptr pages without holding KFD process locks

 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 53 +++++++++++--------
 drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c       | 14 ++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h       |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c       |  9 ++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h       |  7 +++
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c      | 10 +++-
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c          |  2 +-
 7 files changed, 66 insertions(+), 31 deletions(-)

-- 
2.43.0


^ permalink raw reply	[flat|nested] 5+ messages in thread

end of thread, other threads:[~2026-11-18 15:11 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-18 15:10 [PATCH 0/4] drm/amdkfd: don't stall other threads during large userptr registration William Palacek
2026-11-18 15:10 ` [PATCH 1/4] drm/amdgpu: limit userptr sg segments to half of sg->length William Palacek
2026-11-18 15:10 ` [PATCH 2/4] drm/amdkfd: don't take p->mutex in AMDKFD_IOC_GET_CLOCK_COUNTERS William Palacek
2026-11-18 15:10 ` [PATCH 3/4] drm/amdgpu: drop mmap_lock between chunks when faulting userptr pages William Palacek
2026-11-18 15:10 ` [PATCH 4/4] drm/amdkfd: register userptr pages without holding KFD process locks William Palacek

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox