[PATCH v3 0/5] drm/amdgpu: SVM VRAM migration via drm_pagemap

[Junhua Shen <Junhua.Shen@amd.com>]

This series adds VRAM migration support to the amdgpu SVM (Shared
Virtual Memory) subsystem, built on top of the drm_pagemap
infrastructure [1].

It enables transparent page migration between system RAM and device
VRAM using SDMA, driven by userspace SVM attribute hints (prefetch
location, access attributes) through the existing AMDGPU SVM ioctl
interface.

Limitations:

  - Single GPU only; multi-GPU migration is not addressed
  - No XNACK-on GPU fault-driven migration (XNACK-off ioctl
    path only)
  - No VRAM-to-VRAM (peer GPU) migration
  - No eviction fence / VRAM overcommit handling yet

Design highlights:
  - ZONE_DEVICE pages managed via devm_memremap_pages / drm_pagemap
  - SDMA-based migration with proper DMA fence synchronization
  - Migration decision layer that evaluates SVM attributes to
    determine when and where to migrate
  - Zero modifications to the KFD subsystem

Patch breakdown:
  1. Core VRAM migration infrastructure (ZONE_DEVICE, drm_pagemap_ops)
  2. SDMA migration callbacks (copy_to_ram / copy_to_dev)
  3. Migration decision layer (policy evaluation)
  4. SVM attribute extensions (prefetch force-trigger)
  5. Integration into SVM range map path + ZONE_DEVICE registration

Built on top of the drm_pagemap SVM series [1].

Changes since v2:
  - Moved amdgpu_pagemap entirely to amdgpu side, eliminating all KFD
    modifications
  - Split commits for better reviewability: separated infrastructure
    from SDMA callbacks, decision layer from integration
  - Merged ZONE_DEVICE registration hook into the integration patch

Changes since v1:
  - Dropped the eviction fence patch per Christian König's review
    (violates dma_fence contract)

[1] https://lore.kernel.org/all/20260317-drm-svm-v2-0-4bceef04e41e@amd.com/
v1: https://lore.kernel.org/all/20260410113146.146212-1-Junhua.Shen@amd.com/
v2: https://lore.kernel.org/all/20260413103031.181953-1-Junhua.Shen@amd.com/

Junhua Shen (5):
  drm/amdgpu: add VRAM migration infrastructure for drm_pagemap
  drm/amdgpu: implement drm_pagemap SDMA migration callbacks
  drm/amdgpu: introduce SVM range migration decision layer
  drm/amdgpu: add SVM attr prefetch/force-trigger functionality
  drm/amdgpu: integrate VRAM migration into SVM range map path

 drivers/gpu/drm/amd/amdgpu/Makefile           |   6 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu.h           |   8 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |   4 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_migrate.c   | 789 ++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_migrate.h   |  98 +++
 drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c     |   4 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_svm.c       |   4 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_svm_attr.c  |  34 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_svm_range.c | 136 +--
 drivers/gpu/drm/amd/amdgpu/amdgpu_svm_range.h |   5 +-
 .../drm/amd/amdgpu/amdgpu_svm_range_migrate.c | 140 ++++
 .../drm/amd/amdgpu/amdgpu_svm_range_migrate.h |  60 ++
 12 files changed, 1212 insertions(+), 76 deletions(-)
 create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_migrate.c
 create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_migrate.h
 create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_svm_range_migrate.c
 create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_svm_range_migrate.h

-- 
2.34.1