Helps avoiding deadlock during GPU reset.
Added mutex to amdgpu_ctx to preserve order of fences on a ring.
v2:
Put waiting logic in a function in a seperate function in amdgpu_ctx.c
Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Introduce amdgpu_ctx_priority_override(). A mechanism to override a
context's priority.
An override can be terminated by setting the override to
AMD_SCHED_PRIORITY_UNSET.
v2: change refcounted interface for a direct set
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Programming CP_HQD_QUEUE_PRIORITY enables a queue to take priority over
other queues on the same pipe. Multiple queues on a pipe are timesliced
so this gives us full precedence over other queues.
Programming CP_HQD_PIPE_PRIORITY changes the SPI_ARB_PRIORITY of the
wave as follows:
0x2: CS_H
0x1: CS_M
0x0: CS_L
The SPI block will then dispatch work according to the policy set by
SPI_ARB_PRIORITY. In the current policy CS_H is higher priority than
gfx.
In order to prevent getting stuck in loops of resources bouncing between
GFX and high priority compute and introducing further latency, we
statically reserve a portion of the pipe.
v2: fix srbm_select to ring->queue and use ring->funcs->type
v3: use AMD_SCHED_PRIORITY_* instead of AMDGPU_CTX_PRIORITY_*
v4: switch int to enum amd_sched_priority
v5: corresponding changes for srbm_lock
v6: change CU reservation to PIPE_PERCENT allocation
v7: use kiq instead of MMIO
v8: back to MMIO, and make the implementation sleep safe.
v9: corresponding changes for splitting HIGH into _HW/_SW
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>
SR-IOV need to reserve a piece of shared VRAM at the exact place
to exchange data betweem PF and VF. The start address and size of
the shared mem are passed to guest through VBIOS structure
VRAM_UsageByFirmware.
VRAM_UsageByFirmware is a general feature in VBIOS, it indicates
that VBIOS need to reserve a piece of memory on the VRAM.
Because the mem address is specified. Reserve it early in
amdgpu_ttm_init to make sure that it can monoplize the space.
Signed-off-by: Horace Chen <horace.chen@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
v2: drop hdp invalidate/flush.
v3: honor pgoff during prime mmap. Add a barrier after cpu access.
v4: drop begin/end_cpu_access() for now, revisit later.
Signed-off-by: Samuel Li <Samuel.Li@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When max_bytes is not 8 bytes aligned and bo size is larger than
max_bytes, the last 8 bytes in a ttm node may be left unchanged.
For example, on pre SDMA 4.0, max_bytes = 0x1fffff, and the bo size
is 0x200000, the problem will happen.
In order to fix the problem, we separately store the max nums of
PTEs/PDEs a single operation can set in amdgpu_vm_pte_funcs
structure, rather than inferring it from bytes limit of SDMA
constant fill, i.e. fill_max_bytes.
Together with the fix, we replace the hard code value "10" in
amdgpu_vm_bo_update_mapping() with the corresponding values from
structure amdgpu_vm_pte_funcs.
Signed-off-by: Yong Zhao <yong.zhao@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
GPU reset will require all hw doing hw_init thus
ucode_init_bo will be invoked again, which lead to
memory leak
skip the fw_buf allocation during sriov gpu reset to avoid
memory leak.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
otherwise a gpu hang will make application couldn't be killed
under timedout=0 mode
v2:
Fix memoryleak job/job->s_fence issue
unlock mn
remove the ERROR msg after waiting being interrupted
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
currently in_reset is only used in sriov gpu reset, and it
will be used for other non-gfx hw component later, like
PSP, so move it from gfx to adev and rename to in_sriov_reset
make more sense.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This is quite controversial because it adds another lock which is held during
page table updates, but I don't see much other option.
v2: allow multiple updates to be in flight at the same time
v3: simplify the patch, take the read side only once
v4: correctly fix rebase conflict
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Instead take the callback lock during the final parts of CS.
This should solve the last remaining locking order problems with BO reservations.
v2: rebase, make dummy functions static inline
v3: add one more missing inline and comments
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Move calling put_page into the unpopulate callback. Otherwise we mess up the pages
reference count when it is unbound multiple times.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add the IOCTL interface so that applications can allocate per VM BOs.
Still WIP since not all corner cases are tested yet, but this reduces average
CS overhead for 10K BOs from 21ms down to 48us.
v2: add some extra checks, remove the WIP tag
v3: rename new flag to AMDGPU_GEM_CREATE_VM_ALWAYS_VALID
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We need a larger gart for asics that do not support GPUVM on all
engines (e.g., MM) to make sure we have enough space for all
gtt buffers in physical mode. Change the default size based on
the asic type.
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>
While doing flr on VFs, there is possibility to lost the doorbell
writing for sdma, so enable poll mem for sdma, then sdma fw would
check the pollmem holding wptr.
Signed-off-by: Frank Min <Frank.Min@amd.com>
Signed-off-by: Xiangliang.Yu <Xiangliang.Yu@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Set a configurable SDMA phase quantum when enabling SDMA context
switching. The default value significantly reduces SDMA latency
in page table updates when user-mode SDMA queues have concurrent
activity, compared to the initial HW setting.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Andres Rodriguez <andres.rodriguez@amd.com>
Reviewed-by: Shaoyun Liu <shaoyun.liu@amd.com>
Acked-by: Chunming Zhou <david1.zhou@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The BO move throttling code is designed to allow VRAM to fill quickly if it
is relatively empty. However, this does not take into account situations
where the visible VRAM is smaller than total VRAM, and total VRAM may not
be close to full but the visible VRAM segment is under pressure. In such
situations, visible VRAM would experience unrestricted swapping and
performance would drop.
Add a separate counter specifically for moves involving visible VRAM, and
check it before moving BOs there.
v2: Only perform calculations for separate counter if visible VRAM is
smaller than total VRAM. (Michel Dänzer)
v3: [Michel Dänzer]
* Use BO's location rather than the AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
flag to determine whether to account a move for visible VRAM in most
cases.
* Use a single
if (adev->mc.visible_vram_size < adev->mc.real_vram_size) {
block in amdgpu_cs_get_threshold_for_moves.
Fixes: 95844d20ae (drm/amdgpu: throttle buffer migrations at CS using a fixed MBps limit (v2))
Signed-off-by: John Brooks <john@fastquake.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This allows us to write the mapped PTEs into
an IB instead of the table directly.
v2: fix build with debugfs enabled, remove unused assignment
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
