amdgpu_init() does not undo its earlier steps when KFD init or PCI
driver registration fails. v1 [1] added that unwinding. While checking
it against amdgpu_exit(), two more things came up, so this is now a
series:

1. amdgpu_unregister_atpx_handler() clears the vga_switcheroo handler
   even when amdgpu did not register it, e.g. when apple-gmux holds the
   slot. amdgpu_exit() has always done this; v1 would also have done it
   on a failed module init. Patch 1 fixes it and goes first so that
   patch 3 does not add a new path to it.

2. amdgpu_exit() does not undo amdgpu_init() in reverse order. For the
   ATPX handler and the ACPI info that is only historical, so patch 2
   swaps them. For KFD and the PCI driver it is required, so patch 2
   keeps that order and adds a comment explaining why.

Patch 3 is v1 rebased on top: the error path of amdgpu_init() now
undoes the steps in reverse order, the same order amdgpu_exit() uses
after patch 2.

Changes since v1:
- New patches 1 and 2.
- Patch 3: commit message updated to refer to the order in
  amdgpu_exit(); no code change.
- Patches 1 and 3 carry Cc: stable. Patch 3 depends on patch 1 and
  should not be backported without it.

Only build-tested, on the ROCm amdgpu tree with W=1, LLVM=1 and
CONFIG_VGA_SWITCHEROO=y; I have no hardware with ATPX or apple-gmux to
run it on.

[1] https://lists.freedesktop.org/archives/amd-gfx/2026-November/157013.html

Jiaming Fan (3):
  drm/amdgpu: only unregister the ATPX handler if we registered it
  drm/amdgpu: undo module init in reverse order in module exit
  drm/amdgpu: unwind module init on failure

 .../gpu/drm/amd/amdgpu/amdgpu_atpx_handler.c  | 13 +++++++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c       | 24 +++++++++++++++----
 2 files changed, 32 insertions(+), 5 deletions(-)

-- 
2.56.0


Reply via email to