Pinning of VRAM is for peer devices that don't support dynamic attachment
and move notifiers. But it requires that all such peer devices are able to
access VRAM via PCIe P2P. Any device without P2P access requires migration
to GTT, which fails if the memory is already pinned for another peer
device.
Sharing between GPUs should not require pinning in VRAM. However, if
DMABUF_MOVE_NOTIFY is disabled in the kernel build, even DMABufs shared
between GPUs must be pinned, which can lead to failures and functional
regressions on systems where some peer GPUs are not P2P accessible.
Disable VRAM pinning if move notifiers are disabled in the kernel build
to fix regressions when sharing BOs between GPUs.
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
Tested-by: Hao (Claire) Zhou <hao.zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When determining the domains for pinning DMABufs, filter allowed_domains
and fail with a warning if VRAM is forbidden and GTT is not an allowed
domain.
Fixes: f5e7fabd1f ("drm/amdgpu: allow pinning DMA-bufs into VRAM if all importers can do P2P")
Suggested-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Update the minimum firmware version for the Cleaner Shader in the
gfx_v10_0_sw_init function.
This change adjusts the minimum required firmware version for the MEC
firmware from 152 to 151, allowing for broader compatibility with
GFX10.1 GPUs.
Fixes: 25961bad92 ("drm/amdgpu/gfx10: Add cleaner shader for GFX10.1.10")
Cc: Christian König <christian.koenig@amd.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Srinivasan Shanmugam <srinivasan.shanmugam@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch removes the deprecated SDMA reset callback mechanism, which was previously used to register pre-reset and post-reset callbacks for SDMA engine resets.
The callback mechanism has been replaced with a more direct and efficient approach using `stop_queue` and `start_queue` functions in the ring's function table.
The SDMA reset callback mechanism allowed KFD and AMDGPU to register pre-reset and post-reset functions for handling SDMA engine resets.
However, this approach added unnecessary complexity and was no longer needed after the introduction of the `stop_queue` and `start_queue` functions in the ring's function table.
1. **Remove Callback Mechanism**:
- Removed the `amdgpu_sdma_register_on_reset_callbacks` function and its associated data structures (`sdma_on_reset_funcs`).
- Removed the callback registration logic from the SDMA v4.4.2 initialization code.
2. **Clean Up Related Code**:
- Removed the `sdma_v4_4_2_set_engine_reset_funcs` function, which was used to register the callbacks.
- Removed the `sdma_v4_4_2_engine_reset_funcs` structure, which contained the pre-reset and post-reset callback functions.
Signed-off-by: Jesse Zhang <jesse.zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch refactors the SDMA v5.2 queue reset and stop logic to improve
code readability, maintainability, and performance. The key changes include:
1. **Generalized `sdma_v5_2_gfx_stop` Function**:
- Added an `inst_mask` parameter to allow stopping specific SDMA instances
instead of all instances. This is useful for resetting individual queues.
2. **Simplified `sdma_v5_2_reset_queue` Function**:
- Removed redundant loops and checks by directly using the `ring->me` field
to identify the SDMA instance.
- Reused the `sdma_v5_2_gfx_stop` function to stop the queue, reducing code
duplication.
v1: The general coding style is to declare variables like "i" or "r" last. E.g. longest lines first and short lasts. (Chritian)
Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch refactors the SDMA v5.0 queue reset and stop logic to improve
code readability, maintainability, and performance. The key changes include:
1. **Generalized `sdma_v5_0_gfx_stop` Function**:
- Added an `inst_mask` parameter to allow stopping specific SDMA instances
instead of all instances. This is useful for resetting individual queues.
2. **Simplified `sdma_v5_0_reset_queue` Function**:
- Removed redundant loops and checks by directly using the `ring->me` field
to identify the SDMA instance.
- Reused the `sdma_v5_0_gfx_stop` function to stop the queue, reducing code
duplication.
v1: The general coding style is to declare variables like "i" or "r" last. E.g. longest lines first and short lasts. (Chritian)
Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch introduces a new function `amdgpu_sdma_soft_reset` to handle SDMA soft resets directly,
rather than relying on the DPM interface.
1. **New `amdgpu_sdma_soft_reset` Function**:
- Implements a soft reset for SDMA engines by directly writing to the hardware registers.
- Handles SDMA versions 4.x and 5.x separately:
- For SDMA 4.x, the existing `amdgpu_dpm_reset_sdma` function is used for backward compatibility.
- For SDMA 5.x, the driver directly manipulates the `GRBM_SOFT_RESET` register to reset the specified SDMA instance.
2. **Integration into `amdgpu_sdma_reset_engine`**:
- The `amdgpu_sdma_soft_reset` function is called during the SDMA reset process, replacing the previous call to `amdgpu_dpm_reset_sdma`.
v2: r should default to an error (Alex)
Suggested-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Jesse Zhang <jesse.zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Enable users to create queues at different priority levels.
The highest level is restricted to drm master.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Reviewed-by: Jesse.Zhang <Jesse.zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Convert driver priority levels to MES11 priority levels.
At the moment they are the same, but they may not always
be.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Reviewed-by: Jesse.Zhang <Jesse.zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Convert driver priority levels to MES11 priority levels.
At the moment they are the same, but they may not always
be.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Reviewed-by: Jesse.Zhang <Jesse.zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Enforce isolation serializes access to the GFX IP. User
queues are isolated in the MES scheduler, but we still
need to serialize between kernel queues and user queues.
For enforce isolation, group KGD user queues with KFD user
queues.
v2: split out variable renaming, add config guards
v3: use new function names
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Since they will be used for both KFD and KGD user queues,
rename them from kfd to userq. No intended functional
change.
Acked-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This will be used to stop/start user queue scheduling for
example when switching between kernel and user queues when
enforce isolation is enabled.
v2: use idx
v3: only stop compute/gfx queues
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Replace disable_kq parameter with user_queue parameter.
The parameter has the following logic:
-1 = auto (ASIC specific default)
0 = user queues disabled
1 = user queues enabled and kernel queues enabled (if supported)
2 = user queues enabled and kernel queues disabled
The default behavior (-1) is currently the same as 0 for current
ASICs. To enable user queues (in addition to kernel queues) set
user_queue=1. To enable user queues and disable kernel queues
(to make all resources available to user queues), set user_queue=2.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We need to take a reference to the interrupts to make
sure they stay enabled even if the kernel queues have
disabled them.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We need to take a reference to the interrupts to make
sure they stay enabled even if the kernel queues have
disabled them.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Regardless of whether we disable kernel queues, we need
to take an extra reference to the pipe interrupts for
user queues to make sure they stay enabled in case we
disable them for kernel queues.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Regardless of whether we disable kernel queues, we need
to take an extra reference to the pipe interrupts for
user queues to make sure they stay enabled in case we
disable them for kernel queues.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
1) Checkpatch complains if we print an error message for kzalloc()
failure. The kzalloc() failure already has it's own error messages
built in. Also this allocation is small enough that it is guaranteed
to succeed.
2) Return directly instead of doing a goto free_fence_drv. The
"fence_drv" is already NULL so no cleanup is necessary.
Reviewed-by: Arvind Yadav <arvind.yadav@amd.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The goto frees "fence_drv" so this is a double free bug. There is no
need to call amdgpu_seq64_free(adev, fence_drv->va) since the seq64
allocation failed so change the goto to goto free_fence_drv. Also
propagate the error code from amdgpu_seq64_alloc() instead of hard coding
it to -ENOMEM.
Fixes: e7cf21fbb2 ("drm/amdgpu: Few optimization and fixes for userq fence driver")
Reviewed-by: Arvind Yadav <arvind.yadav@amd.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Pull it out of the MES code and into the generic code.
It's not MES specific and needs to be applied to all user
queues regardless of the backend.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Reviewed-by: Shaoyun.liu <Shaoyun.liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch introduces new function pointers in the amdgpu_sdma structure
to handle queue stop, start and soft reset operations. These will replace
the older callback mechanism.
The new functions are:
- stop_kernel_queue: Stops a specific SDMA queue
- start_kernel_queue: Starts/Restores a specific SDMA queue
- soft_reset_kernel_queue: Performs soft reset on a specific SDMA queue
v2: Update stop_queue/start_queue function paramters to use ring pointer instead of device/instance(Chritian)
v3: move stop_queue/start_queue to struct amdgpu_sdma_instance and rename them. (Alex)
v4: rework the ordering a bit (Alex)
Suggested-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If userq creation fails, we need to properly unwind and free the
user queue fence driver.
v2: free idr as well (Sunil)
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Move some userq fence handling code into amdgpu_userq_fence.c.
This matches the other code in that file.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Split out the queue map from the mqd create call and split
out the queue unmap from the mqd destroy call. This splits
the queue setup and teardown with the actual enablement
in the firmware.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Rename to map and umap to better align with what is happening
at the firmware level and remove the extra level of indirection
in the MES userq code.
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
There are a few prechecks made before HDP flush like a flush is not
required on APU bare metal. Using hdp callback directly bypasses those
checks. Use amdgpu_device_flush_hdp which takes care of prechecks.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 1d9bff4cf8)
