Instead of hammering hard on the GPU try a soft recovery first.
v2: reorder code a bit
v3: increase timeout to 10ms, increment GPU reset counter
v4: squash in compile fix (Christian)
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Huang Rui <ray.huang@amd.com>
Define extra dword for jpeg ring. Jpeg ring will allocate extra dword to store
the patch commands for fixing the known issue.
v2: dropping extra_dw for rings other than jpeg.
Signed-off-by: Boyuan Zhang <boyuan.zhang@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Vega20 has dual-UVD. Need add multiple instances support for uvd.
Restruct uvd.inst, using uvd.inst[0] to replace uvd.inst->.
Repurpose amdgpu_ring::me for instance index, and initialize to 0.
There are no any logical changes here.
Signed-off-by: James Zhu <James.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This callback writes a value to a register and then reads
back another register and waits for a value in a single
operation.
Provide a helper function using two operations for engines
that don't support this opertion.
Reviewed-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This moves and renames the AMDGPU scheduler to a common location in DRM
in order to facilitate re-use by other drivers. This is mostly a straight
forward rename with no code changes.
One notable exception is the function to_drm_sched_fence(), which is no
longer a inline header function to avoid the need to export the
drm_sched_fence_ops_scheduled and drm_sched_fence_ops_finished structures.
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
KIQ ring submission is used for register accessing on SRIOV
VF that could happen both in irq enabled and irq disabled cases.
Inversion lock could happen on adev->ring_lru_list_lock, while
this operation is useless and just adds overhead in this use
case.
Signed-off-by: Pixel Ding <Pixel.Ding@amd.com>
Reviewed-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add an initial framework for changing the HW priorities of rings. The
framework allows requesting priority changes for the lifetime of an
amdgpu_job. After the job completes the priority will decay to the next
lowest priority for which a request is still valid.
A new ring function set_priority() can now be populated to take care of
the HW specific programming sequence for priority changes.
v2: set priority before emitting IB, and take a ref on amdgpu_job
v3: use AMD_SCHED_PRIORITY_* instead of AMDGPU_CTX_PRIORITY_*
v4: plug amdgpu_ring_restore_priority_cb into amdgpu_job_free_cb
v5: use atomic for tracking job priorities instead of last_job
v6: rename amdgpu_ring_priority_[get/put]() and align parameters
v7: replace spinlocks with mutexes for KIQ compatibility
v8: raise ring priority during cs_ioctl, instead of job_run
v9: priority_get() before push_job()
Reviewed-by: Christian König <christian.koenig@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch provides a guarantee that the first n queues allocated by
an application will be on different pipes. Where n is the number of
pipes available from the hardware.
This helps avoid ring aliasing which can result in work executing in
time-sliced mode instead of truly parallel mode.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
KIQ doesn't really use the GPU scheduler. The base
drivers generally use the KIQ ring directly rather than
submitting IBs. However, amdgpu_sched_hw_submission
(which defaults to 2) limits the number of outstanding
fences to 2. KFD uses the KIQ for TLB flushes and the
2 fence limit hurts performance when there are several KFD
processes running.
v2: move some expressions to one line
change KIQ sched_hw_submission to at least 16
v3: bump to 256
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If a ring is not initialized, it also should not be finished.
For example, in Vega10's SR-IOV environment, UVD's decode ring is not
initialized, but will be finnished in amdgpu_uvd_sw_fini, because UVD
driver put all the uvd decode ring's finish operation into
amdgpu_uvd_sw_fini function, while not uvd_vXXX_0_sw_fini. This will
lead to amdgpu module unloading failure.
Signed-off-by: Trigger Huang <trigger.huang@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
May waste a bit of memory, but simplifies the interface
significantly.
v2: convert internal accounting to use 256bit slots
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In review, Christian would like to keep the logic
inside amdgpu_vm.c with a cost of slightly slower.
The loop is still optimized out with this patch.
v2: remove the if statement. Now it is not slower.
Signed-off-by: Alex Xie <AlexBin.Xie@amd.com>
Reviewed-by: Christian König <christian.koeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Depending on usage patterns, the current LRU policy may create a
non-injective mapping between userspace ring ids and kernel rings.
This behaviour is undesired as apps that attempt to fill all HW blocks
would be unable to reach some of them.
This change forces the LRU policy to create bijective mappings only.
v2: compress ring_blacklist
v3: simplify amdgpu_ring_is_blacklisted() logic
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Use an LRU policy to map usermode rings to HW compute queues.
Most compute clients use one queue, and usually the first queue
available. This results in poor pipe/queue work distribution when
multiple compute apps are running. In most cases pipe 0 queue 0 is
the only queue that gets used.
In order to better distribute work across multiple HW queues, we adopt
a policy to map the usermode ring ids to the LRU HW queue.
This fixes a large majority of multi-app compute workloads sharing the
same HW queue, even though 7 other queues are available.
v2: use ring->funcs->type instead of ring->hw_ip
v3: remove amdgpu_queue_mapper_funcs
v4: change ring_lru_list_lock to spinlock, grab only once in lru_get()
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
On gfx9 hardware the value is not wrapped and is a 64-bit value. So
we reduce it modulo the ring size.
Signed-off-by: Tom St Denis <tom.stdenis@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
(v2) use buf_mask instead of computing on the fly
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
ring->buf_mask need be set prior to ring_clear_ring invoke
and fix ring_clear_ring as well which should use buf_mask
instead of ptr_mask
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Christian König
Boyuan Zhang
James Zhu
Alex Deucher
Tom St Denis
Lucas Stach
Pixel Ding
Andres Rodriguez
Trigger Huang
Monk Liu
Alex Xie
Ken Wang