github-mirror/linux
1
0
Fork 0
mirror of https://github.com/torvalds/linux.git synced 2026-09-22 20:54:03 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity Graph

drm/amdkfd: enable rs64mem for kfd queue

Browse Source 
Enabled RS64mem for KFD queues by integrating
process and gang context index allocation in
the per KFD device process and queue creation.

Signed-off-by: Prike Liang <Prike.Liang@amd.com>
Reviewed-by: Michael Chen <michael.chen@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This commit is contained in:
Prike Liang 2026-07-28 15:00:03 +08:00 committed by Alex Deucher
parent 08bb824ae3
commit e32b68c6b1
4 changed files with 32 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
View File

@ -226,8 +226,10 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
/* MES unit for quantum is 100ns */
queue_input.process_quantum = KFD_MES_PROCESS_QUANTUM; /* Equivalent to 10ms. */
queue_input.process_context_addr = pdd->proc_ctx_gpu_addr;
queue_input.process_context_array_index = pdd->proc_ctx_array_index;
queue_input.gang_quantum = KFD_MES_GANG_QUANTUM; /* Equivalent to 1ms */
queue_input.gang_context_addr = q->gang_ctx_gpu_addr;
queue_input.gang_context_array_index = q->gang_ctx_array_index;
queue_input.inprocess_gang_priority = q->properties.priority;
queue_input.gang_global_priority_level =
AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
@ -303,6 +305,7 @@ static int remove_queue_mes_on_reset_option(struct device_queue_manager *dqm, st
queue_input.queue_type = convert_to_amdgpu_ring_type(q->properties.type);
queue_input.remove_queue_after_reset = flush_mes_queue;
queue_input.xcc_id = ffs(dqm->dev->xcc_mask) - 1;
queue_input.gang_context_array_index = q->gang_ctx_array_index;
amdgpu_mes_lock(&adev->mes);
r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);

3
drivers/gpu/drm/amd/amdkfd/kfd_priv.h
View File

@ -635,6 +635,7 @@ struct queue {
void *gang_ctx_bo;
uint64_t gang_ctx_gpu_addr;
void *gang_ctx_cpu_ptr;
uint32_t gang_ctx_array_index;
struct amdgpu_bo *wptr_bo_gart;
};
@ -871,6 +872,8 @@ struct kfd_process_device {
uint64_t proc_ctx_gpu_addr;
void *proc_ctx_cpu_ptr;
uint32_t proc_ctx_array_index;
/* Tracks queue reset status */
bool has_reset_queue;

5
drivers/gpu/drm/amd/amdkfd/kfd_process.c
View File

@ -1215,9 +1215,12 @@ static void kfd_process_destroy_pdds(struct kfd_process *p)
kfd_free_process_doorbells(pdd->dev->kfd, pdd);
if (pdd->dev->kfd->shared_resources.enable_mes &&
pdd->proc_ctx_cpu_ptr)
pdd->proc_ctx_cpu_ptr) {
amdgpu_mes_free_proc_ctx_index(&pdd->dev->adev->mes,
pdd->proc_ctx_array_index);
amdgpu_amdkfd_free_kernel_mem(pdd->dev->adev,
&pdd->proc_ctx_bo);
}
/*
* before destroying pdd, make sure to report availability
* for auto suspend

23
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
View File

@ -210,6 +210,7 @@ static void pqm_clean_queue_resource(struct process_queue_manager *pqm,
}
if (dev->kfd->shared_resources.enable_mes) {
amdgpu_mes_free_gang_ctx_index(&dev->adev->mes, pqn->q->gang_ctx_array_index);
amdgpu_amdkfd_free_kernel_mem(dev->adev, &pqn->q->gang_ctx_bo);
amdgpu_amdkfd_free_kernel_mem(dev->adev, (void **)&pqn->q->wptr_bo_gart);
}
@ -282,7 +283,15 @@ static int init_user_queue(struct process_queue_manager *pqm,
goto cleanup;
}
memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
/* Bind one MES gang context slot per queue (gang). */
if (dev->adev->mes.use_rs64mem) {
retval = amdgpu_mes_alloc_gang_ctx_index(&dev->adev->mes,
&(*q)->gang_ctx_array_index);
if (retval) {
pr_err("failed to allocate gang context index slot\n");
goto cleanup;
}
}
/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
* on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
*/
@ -304,6 +313,7 @@ static int init_user_queue(struct process_queue_manager *pqm,
return 0;
free_gang_ctx_bo:
amdgpu_mes_free_gang_ctx_index(&dev->adev->mes, (*q)->gang_ctx_array_index);
amdgpu_amdkfd_free_kernel_mem(dev->adev, &(*q)->gang_ctx_bo);
cleanup:
uninit_queue(*q);
@ -386,6 +396,17 @@ int pqm_create_queue(struct process_queue_manager *pqm,
goto err_allocate_pqn;
}
memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
/* Bind one MES process context slot to the whole process
* (per device); every queue of this process reuses it.
*/
if (dev->adev->mes.use_rs64mem) {
retval = amdgpu_mes_alloc_proc_ctx_index(&dev->adev->mes,
&pdd->proc_ctx_array_index);
if (retval) {
dev_err(dev->adev->dev, "failed to allocate process context index\n");
goto err_allocate_pqn;
}
}
}
pqn = kzalloc_obj(*pqn);