drm/amdkfd: Fix user queue validation on Gfx7/8
Subsystems (3)
PATCH 1/1 drm/amdkfd: Fix user queue validation on Gfx7/8
To workaround queue full h/w issue on Gfx7/8, when application create AQL queues, the ring buffer bo allocate size is queue_size/2 and mapped to GPU twice using 2 attachments with same ring_bo backing memory.
For this case, user queue buffer validation should use queue_size/2 to verify ring_bo allocation and mapping size.
--- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_queue.c +11 -1
@@ -233,6 +233,7 @@ void kfd_queue_buffer_put(struct amdgpu_bo **bo)int kfd_queue_acquire_buffers (struct kfd_process_device *pdd,struct queue_properties *properties) {struct kfd_topology_device *topo_dev; +u64 expected_queue_size;struct amdgpu_vm *vm;u32 total_cwsr_size;int err;
@@ -241,6 +242,15 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_propeif (!topo_dev)return -EINVAL ; +/* AQL queues on GFX7 and GFX8 appear twice their actual size */ +if (properties->type ==KFD_QUEUE_TYPE_COMPUTE && + properties->format ==KFD_QUEUE_FORMAT_AQL && + topo_dev->node_props.gfx_target_version >=70000 && + topo_dev->node_props.gfx_target_version <90000 ) + expected_queue_size = properties->queue_size /2 ; +else + expected_queue_size = properties->queue_size; + vm =drm_priv_to_vm (pdd->drm_priv); err =amdgpu_bo_reserve (vm->root.bo, false);if (err)
@@ -255,7 +265,7 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope goto out_err_unreserve; err =kfd_queue_buffer_get (vm, (void *)properties->queue_address, - &properties->ring_bo, properties->queue_size); + &properties->ring_bo, expected_queue_size);if (err) goto out_err_unreserve;
Signature · 1 line
2.47.1
ping...
On 2025-01-29 19:04, Philip Yang wrote:
To workaround queue full h/w issue on Gfx7/8, when application create AQL queues, the ring buffer bo allocate size is queue_size/2 and mapped to GPU twice using 2 attachments with same ring_bo backing memory. For this case, user queue buffer validation should use queue_size/2 to verify ring_bo allocation and mapping size. Fixes: 68e599db7a54 ("drm/amdkfd: Validate user queue buffers") Suggested-by: Tomáš Trnka <[email protected]> Signed-off-by: Philip Yang <[email protected]> --- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_queue.c index ecccd7adbab4..62c635e9d1aa 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_queue.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_queue.c @@ -233,6 +233,7 @@ void kfd_queue_buffer_put(struct amdgpu_bo **bo) int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_properties *properties) { struct kfd_topology_device *topo_dev; + u64 expected_queue_size; struct amdgpu_vm *vm; u32 total_cwsr_size; int err; @@ -241,6 +242,15 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope if (!topo_dev) return -EINVAL; + /* AQL queues on GFX7 and GFX8 appear twice their actual size */ + if (properties->type == KFD_QUEUE_TYPE_COMPUTE && + properties->format == KFD_QUEUE_FORMAT_AQL && + topo_dev->node_props.gfx_target_version >= 70000 && + topo_dev->node_props.gfx_target_version < 90000) + expected_queue_size = properties->queue_size / 2; + else + expected_queue_size = properties->queue_size; + vm = drm_priv_to_vm(pdd->drm_priv); err = amdgpu_bo_reserve(vm->root.bo, false); if (err) @@ -255,7 +265,7 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope goto out_err_unreserve; err = kfd_queue_buffer_get(vm, (void *)properties->queue_address, - &properties->ring_bo, properties->queue_size); + &properties->ring_bo, expected_queue_size); if (err) goto out_err_unreserve;
Quoted 6 lines · Philip Yang
To workaround queue full h/w issue on Gfx7/8, when application create AQL queues, the ring buffer bo allocate size is queue_size/2 and mapped to GPU twice using 2 attachments with same
ring_bobacking memory.
Quoted 6 lines
For this case, user queue buffer validation should use queue_size/2 to verify
ring_boallocation and mapping size.Fixes 68e599db7a54 ("drm/amdkfd: Validate user queue buffers")Suggested-by Tomáš Trnka <[email protected]>Signed-off-by Philip Yang <[email protected]>
Quoted 44 lines
--- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-)drivers/gpu/drm/amd/amdkfd/kfd_queue.c +11
-2@@ -233,6 +233,7 @@ void kfd_queue_buffer_put(struct amdgpu_bo **bo)int kfd_queue_acquire_buffers (struct kfd_process_device *pdd,struct queue_properties *properties) {struct kfd_topology_device *topo_dev; +u64 expected_queue_size;struct amdgpu_vm *vm;u32 total_cwsr_size;int err;@@ -241,6 +242,15 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_propeif (!topo_dev)return -EINVAL ; +/* AQL queues on GFX7 and GFX8 appear twice their actual size */ +if (properties->type ==KFD_QUEUE_TYPE_COMPUTE && + properties->format ==KFD_QUEUE_FORMAT_AQL && + topo_dev->node_props.gfx_target_version >=70000 && + topo_dev->node_props.gfx_target_version <90000 ) + expected_queue_size = properties->queue_size /2 ; +else + expected_queue_size = properties->queue_size; + vm =drm_priv_to_vm (pdd->drm_priv); err =amdgpu_bo_reserve (vm->root.bo, false);if (err)@@ -255,7 +265,7 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope goto out_err_unreserve; err =kfd_queue_buffer_get (vm, (void *)properties->queue_address, - &properties->ring_bo, properties->queue_size); + &properties->ring_bo, expected_queue_size);if (err) goto out_err_unreserve; --2.47.1