To free fences, call_rcu() is used, which calls amdgpu_fence_free()
after a grace period. During teardown, there is no guarantee all
callbacks have finished, so amdgpu_fence_slab may be destroyed before
all fences have been freed. If we are lucky, this results in some slab
warnings, if not, we get a crash in one of rcu threads because callback
is called after amdgpu has already been unloaded.
Fix it with a rcu_barrier().
Fixes: b44135351a ("drm/amdgpu: RCU protected amdgpu_fence_release")
Acked-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Right now it's possible to trigger fence_drv.fences[] dereference after
the array has been freed. While the real problem is elsewhere, this still
results in confusing errors that depend on how the freed memory was
reused (I've seen "kernel tried to execute NX-protected page"), it's
better to clear them and get NULL dereference so that it's obvious what's
going wrong.
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This avoids problems with multiple GPUs. For example,
if the first GPU failed before amdgpu_fence_init() was
called, amdgpu_fence_slab_ref is still 0 and it will
get decremented in amdgpu_fence_driver_fini(). This
will lead to a crash during init of the second GPU since
amdgpu_fence_slab_ref is not 0.
v2: add functions for init/exit instead of
moving the variables into the driver.
Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The rcu_assign_pointer() ensures that the initialization of a structure
is carried out before storing a pointer to that structre. It is always
safe to use RCU_INIT_POINTER() to NULL a pointer, instead of
rcu_assign_pointer().
This results in slightly smaller/faster code.
The following semantic patch was used:
<smpl>
@@
@@
- rcu_assign_pointer
+ RCU_INIT_POINTER
(..., NULL)
</smpl>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Muhammad Falak R Wani <falakreyaz@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
since the rcu slot was initialized to be num_hw_submission,
if command submission doesn't use scheduler, this limitation
will be invalid like uvd test.
Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Amdgpu doesn't support using scratch registers for fences any more.
So we won't see values like 0xdeadbeef as fence value any more.
v2: reschedule timer even if no change detected
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Just wait for last fence instead of waiting for the sequence manually.
v2: don't use amdgpu_sched_jobs for the mask
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Just keep all HW fences in a RCU protected array as a
first step to replace the wait queue.
v2: update commit message, move fixes into separate patch.
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Fences must be freed RCU protected, otherwise the reservation_object_*_rcu()
functions can run into problems.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
The comment about the loop counter was never valid, even when you have
multiple threads this loop only runs as long as the sequence increases.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
No need to duplicate the functionality any more.
v2: fix handling if no fence is available.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v1)
amdgpu_fence_default_wait isn't needed any more the default wait does the same
thing and amdgpu_test_signaled is dead as well.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Embed the scheduler into the ring structure instead of allocating it.
Use the ring name directly instead of the id.
v2: rebased, whitespace cleanup
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Chunming Zhou<david1.zhou@amd.com>
Move the fence related stuff into amdgpu_fence.c
v2: rework commit message, cause this is actually not a bug
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou<david1.zhou@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
The fence in the array may be skipped if wait_all is false,
thus the related callback is not initialized with list head.
So removing this kind callback will cause NULL pointer reference.
Signed-off-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
