struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
struct drm_sched_entity *entity = p->entity;
enum drm_sched_priority priority;
- struct amdgpu_ring *ring;
struct amdgpu_bo_list_entry *e;
struct amdgpu_job *job;
uint64_t seq;
priority = job->base.s_priority;
drm_sched_entity_push_job(&job->base, entity);
- ring = to_amdgpu_ring(entity->rq->sched);
- amdgpu_ring_priority_get(ring, priority);
-
amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
return -EACCES;
}
+static enum gfx_pipe_priority amdgpu_ctx_sched_prio_to_compute_prio(enum drm_sched_priority prio)
+{
+ switch (prio) {
+ case DRM_SCHED_PRIORITY_HIGH_HW:
+ case DRM_SCHED_PRIORITY_KERNEL:
+ return AMDGPU_GFX_PIPE_PRIO_HIGH;
+ default:
+ return AMDGPU_GFX_PIPE_PRIO_NORMAL;
+ }
+}
+
static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, const u32 hw_ip, const u32 ring)
{
struct amdgpu_device *adev = ctx->adev;
struct amdgpu_ctx_entity *entity;
struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
unsigned num_scheds = 0;
+ enum gfx_pipe_priority hw_prio;
enum drm_sched_priority priority;
int r;
num_scheds = 1;
break;
case AMDGPU_HW_IP_COMPUTE:
- scheds = adev->gfx.compute_sched;
- num_scheds = adev->gfx.num_compute_sched;
+ hw_prio = amdgpu_ctx_sched_prio_to_compute_prio(priority);
+ scheds = adev->gfx.compute_prio_sched[hw_prio];
+ num_scheds = adev->gfx.num_compute_sched[hw_prio];
break;
case AMDGPU_HW_IP_DMA:
scheds = adev->sdma.sdma_sched;
mutex_destroy(&mgr->lock);
}
+
+static void amdgpu_ctx_init_compute_sched(struct amdgpu_device *adev)
+{
+ int num_compute_sched_normal = 0;
+ int num_compute_sched_high = AMDGPU_MAX_COMPUTE_RINGS - 1;
+ int i;
+
+ /* use one drm sched array, gfx.compute_sched to store both high and
+ * normal priority drm compute schedulers */
+ for (i = 0; i < adev->gfx.num_compute_rings; i++) {
+ if (!adev->gfx.compute_ring[i].has_high_prio)
+ adev->gfx.compute_sched[num_compute_sched_normal++] =
+ &adev->gfx.compute_ring[i].sched;
+ else
+ adev->gfx.compute_sched[num_compute_sched_high--] =
+ &adev->gfx.compute_ring[i].sched;
+ }
+
+ /* compute ring only has two priority for now */
+ i = AMDGPU_GFX_PIPE_PRIO_NORMAL;
+ adev->gfx.compute_prio_sched[i] = &adev->gfx.compute_sched[0];
+ adev->gfx.num_compute_sched[i] = num_compute_sched_normal;
+
+ i = AMDGPU_GFX_PIPE_PRIO_HIGH;
+ adev->gfx.compute_prio_sched[i] =
+ &adev->gfx.compute_sched[num_compute_sched_high - 1];
+ adev->gfx.num_compute_sched[i] =
+ adev->gfx.num_compute_rings - num_compute_sched_normal;
+}
+
void amdgpu_ctx_init_sched(struct amdgpu_device *adev)
{
int i, j;
+ amdgpu_ctx_init_compute_sched(adev);
for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
adev->gfx.gfx_sched[i] = &adev->gfx.gfx_ring[i].sched;
adev->gfx.num_gfx_sched++;
}
- for (i = 0; i < adev->gfx.num_compute_rings; i++) {
- adev->gfx.compute_sched[i] = &adev->gfx.compute_ring[i].sched;
- adev->gfx.num_compute_sched++;
- }
-
for (i = 0; i < adev->sdma.num_instances; i++) {
adev->sdma.sdma_sched[i] = &adev->sdma.instance[i].ring.sched;
adev->sdma.num_sdma_sched++;
return adev->gfx.mec.num_mec > 1;
}
+bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
+ int queue)
+{
+ /* Policy: make queue 0 of each pipe as high priority compute queue */
+ return (queue == 0);
+
+}
+
void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
{
int i, queue, pipe, mec;
#define AMDGPU_MAX_GFX_QUEUES KGD_MAX_QUEUES
#define AMDGPU_MAX_COMPUTE_QUEUES KGD_MAX_QUEUES
+enum gfx_pipe_priority {
+ AMDGPU_GFX_PIPE_PRIO_NORMAL = 1,
+ AMDGPU_GFX_PIPE_PRIO_HIGH,
+ AMDGPU_GFX_PIPE_PRIO_MAX
+};
+
+#define AMDGPU_GFX_QUEUE_PRIORITY_MINIMUM 0
+#define AMDGPU_GFX_QUEUE_PRIORITY_MAXIMUM 15
+
struct amdgpu_mec {
struct amdgpu_bo *hpd_eop_obj;
u64 hpd_eop_gpu_addr;
uint32_t num_gfx_sched;
unsigned num_gfx_rings;
struct amdgpu_ring compute_ring[AMDGPU_MAX_COMPUTE_RINGS];
+ struct drm_gpu_scheduler **compute_prio_sched[AMDGPU_GFX_PIPE_PRIO_MAX];
struct drm_gpu_scheduler *compute_sched[AMDGPU_MAX_COMPUTE_RINGS];
- uint32_t num_compute_sched;
+ uint32_t num_compute_sched[AMDGPU_GFX_PIPE_PRIO_MAX];
unsigned num_compute_rings;
struct amdgpu_irq_src eop_irq;
struct amdgpu_irq_src priv_reg_irq;
int *mec, int *pipe, int *queue);
bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev, int mec,
int pipe, int queue);
+bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
+ int queue);
int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev, int me,
int pipe, int queue);
void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
{
- struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
struct amdgpu_job *job = to_amdgpu_job(s_job);
drm_sched_job_cleanup(s_job);
- amdgpu_ring_priority_put(ring, s_job->s_priority);
dma_fence_put(job->fence);
amdgpu_sync_free(&job->sync);
amdgpu_sync_free(&job->sched_sync);
void *owner, struct dma_fence **f)
{
enum drm_sched_priority priority;
- struct amdgpu_ring *ring;
int r;
if (!f)
priority = job->base.s_priority;
drm_sched_entity_push_job(&job->base, entity);
- ring = to_amdgpu_ring(entity->rq->sched);
- amdgpu_ring_priority_get(ring, priority);
-
return 0;
}
struct mutex priority_mutex;
/* protected by priority_mutex */
int priority;
+ bool has_high_prio;
#if defined(CONFIG_DEBUG_FS)
struct dentry *ent;
return r;
}
+static void gfx_v10_0_compute_mqd_set_priority(struct amdgpu_ring *ring, struct v10_compute_mqd *mqd)
+{
+ struct amdgpu_device *adev = ring->adev;
+
+ if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE) {
+ if (amdgpu_gfx_is_high_priority_compute_queue(adev, ring->queue)) {
+ mqd->cp_hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_HIGH;
+ ring->has_high_prio = true;
+ mqd->cp_hqd_queue_priority =
+ AMDGPU_GFX_QUEUE_PRIORITY_MAXIMUM;
+ } else {
+ ring->has_high_prio = false;
+ }
+ }
+}
+
static int gfx_v10_0_compute_mqd_init(struct amdgpu_ring *ring)
{
struct amdgpu_device *adev = ring->adev;
tmp = REG_SET_FIELD(tmp, CP_HQD_IB_CONTROL, MIN_IB_AVAIL_SIZE, 3);
mqd->cp_hqd_ib_control = tmp;
+ /* set static priority for a compute queue/ring */
+ gfx_v10_0_compute_mqd_set_priority(ring, mqd);
+
/* map_queues packet doesn't need activate the queue,
* so only kiq need set this field.
*/
return r;
}
+static void gfx_v8_0_mqd_set_priority(struct amdgpu_ring *ring, struct vi_mqd *mqd)
+{
+ struct amdgpu_device *adev = ring->adev;
+
+ if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE) {
+ if (amdgpu_gfx_is_high_priority_compute_queue(adev, ring->queue)) {
+ mqd->cp_hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_HIGH;
+ ring->has_high_prio = true;
+ mqd->cp_hqd_queue_priority =
+ AMDGPU_GFX_QUEUE_PRIORITY_MAXIMUM;
+ } else {
+ ring->has_high_prio = false;
+ }
+ }
+}
+
static int gfx_v8_0_mqd_init(struct amdgpu_ring *ring)
{
struct amdgpu_device *adev = ring->adev;
/* defaults */
mqd->cp_hqd_eop_rptr = RREG32(mmCP_HQD_EOP_RPTR);
mqd->cp_hqd_eop_wptr = RREG32(mmCP_HQD_EOP_WPTR);
- mqd->cp_hqd_pipe_priority = RREG32(mmCP_HQD_PIPE_PRIORITY);
- mqd->cp_hqd_queue_priority = RREG32(mmCP_HQD_QUEUE_PRIORITY);
- mqd->cp_hqd_quantum = RREG32(mmCP_HQD_QUANTUM);
mqd->cp_hqd_ctx_save_base_addr_lo = RREG32(mmCP_HQD_CTX_SAVE_BASE_ADDR_LO);
mqd->cp_hqd_ctx_save_base_addr_hi = RREG32(mmCP_HQD_CTX_SAVE_BASE_ADDR_HI);
mqd->cp_hqd_cntl_stack_offset = RREG32(mmCP_HQD_CNTL_STACK_OFFSET);
mqd->cp_hqd_eop_wptr_mem = RREG32(mmCP_HQD_EOP_WPTR_MEM);
mqd->cp_hqd_eop_dones = RREG32(mmCP_HQD_EOP_DONES);
+ /* set static priority for a queue/ring */
+ gfx_v8_0_mqd_set_priority(ring, mqd);
+ mqd->cp_hqd_quantum = RREG32(mmCP_HQD_QUANTUM);
+
/* map_queues packet doesn't need activate the queue,
* so only kiq need set this field.
*/
WREG32_SOC15_RLC(GC, 0, mmRLC_CP_SCHEDULERS, tmp);
}
+static void gfx_v9_0_mqd_set_priority(struct amdgpu_ring *ring, struct v9_mqd *mqd)
+{
+ struct amdgpu_device *adev = ring->adev;
+
+ if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE) {
+ if (amdgpu_gfx_is_high_priority_compute_queue(adev, ring->queue)) {
+ mqd->cp_hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_HIGH;
+ ring->has_high_prio = true;
+ mqd->cp_hqd_queue_priority =
+ AMDGPU_GFX_QUEUE_PRIORITY_MAXIMUM;
+ } else {
+ ring->has_high_prio = false;
+ }
+ }
+}
+
static int gfx_v9_0_mqd_init(struct amdgpu_ring *ring)
{
struct amdgpu_device *adev = ring->adev;
tmp = REG_SET_FIELD(tmp, CP_HQD_IB_CONTROL, MIN_IB_AVAIL_SIZE, 3);
mqd->cp_hqd_ib_control = tmp;
+ /* set static priority for a queue/ring */
+ gfx_v9_0_mqd_set_priority(ring, mqd);
+ mqd->cp_hqd_quantum = RREG32(mmCP_HQD_QUANTUM);
+
/* map_queues packet doesn't need activate the queue,
* so only kiq need set this field.
*/