From b249e18df151c9627af808321a8090c0b8d4d834 Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Tue, 22 Aug 2017 16:39:30 -0400 Subject: drm/amdgpu: set sched_hw_submission higher for KIQ (v3) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit KIQ doesn't really use the GPU scheduler. The base drivers generally use the KIQ ring directly rather than submitting IBs. However, amdgpu_sched_hw_submission (which defaults to 2) limits the number of outstanding fences to 2. KFD uses the KIQ for TLB flushes and the 2 fence limit hurts performance when there are several KFD processes running. v2: move some expressions to one line change KIQ sched_hw_submission to at least 16 v3: bump to 256 Reviewed-by: Christian König Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index 6c5646b48d1a..5ce65280b396 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -170,6 +170,16 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, unsigned irq_type) { int r; + int sched_hw_submission = amdgpu_sched_hw_submission; + + /* Set the hw submission limit higher for KIQ because + * it's used for a number of gfx/compute tasks by both + * KFD and KGD which may have outstanding fences and + * it doesn't really use the gpu scheduler anyway; + * KIQ tasks get submitted directly to the ring. + */ + if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ) + sched_hw_submission = max(sched_hw_submission, 256); if (ring->adev == NULL) { if (adev->num_rings >= AMDGPU_MAX_RINGS) @@ -178,8 +188,7 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, ring->adev = adev; ring->idx = adev->num_rings++; adev->rings[ring->idx] = ring; - r = amdgpu_fence_driver_init_ring(ring, - amdgpu_sched_hw_submission); + r = amdgpu_fence_driver_init_ring(ring, sched_hw_submission); if (r) return r; } @@ -218,8 +227,7 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, return r; } - ring->ring_size = roundup_pow_of_two(max_dw * 4 * - amdgpu_sched_hw_submission); + ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission); ring->buf_mask = (ring->ring_size / 4) - 1; ring->ptr_mask = ring->funcs->support_64bit_ptrs ? -- cgit v1.2.3 From 35161bbc135a748dd0a3c822030b3341cdefbd33 Mon Sep 17 00:00:00 2001 From: Andres Rodriguez Date: Tue, 26 Sep 2017 17:43:14 -0400 Subject: drm/amdgpu: map compute rings by least recently used pipe This patch provides a guarantee that the first n queues allocated by an application will be on different pipes. Where n is the number of pipes available from the hardware. This helps avoid ring aliasing which can result in work executing in time-sliced mode instead of truly parallel mode. Reviewed-by: Alex Deucher Signed-off-by: Andres Rodriguez Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c | 8 +++++--- drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 25 ++++++++++++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 5 +++-- 3 files changed, 28 insertions(+), 10 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c index befc09b68543..190e28cb827e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c @@ -121,7 +121,7 @@ static enum amdgpu_ring_type amdgpu_hw_ip_to_ring_type(int hw_ip) static int amdgpu_lru_map(struct amdgpu_device *adev, struct amdgpu_queue_mapper *mapper, - int user_ring, + int user_ring, bool lru_pipe_order, struct amdgpu_ring **out_ring) { int r, i, j; @@ -139,7 +139,7 @@ static int amdgpu_lru_map(struct amdgpu_device *adev, } r = amdgpu_ring_lru_get(adev, ring_type, ring_blacklist, - j, out_ring); + j, lru_pipe_order, out_ring); if (r) return r; @@ -284,8 +284,10 @@ int amdgpu_queue_mgr_map(struct amdgpu_device *adev, r = amdgpu_identity_map(adev, mapper, ring, out_ring); break; case AMDGPU_HW_IP_DMA: + r = amdgpu_lru_map(adev, mapper, ring, false, out_ring); + break; case AMDGPU_HW_IP_COMPUTE: - r = amdgpu_lru_map(adev, mapper, ring, out_ring); + r = amdgpu_lru_map(adev, mapper, ring, true, out_ring); break; default: *out_ring = NULL; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index 5ce65280b396..019932a7ea3a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -315,14 +315,16 @@ static bool amdgpu_ring_is_blacklisted(struct amdgpu_ring *ring, * @type: amdgpu_ring_type enum * @blacklist: blacklisted ring ids array * @num_blacklist: number of entries in @blacklist + * @lru_pipe_order: find a ring from the least recently used pipe * @ring: output ring * * Retrieve the amdgpu_ring structure for the least recently used ring of * a specific IP block (all asics). * Returns 0 on success, error on failure. */ -int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type, int *blacklist, - int num_blacklist, struct amdgpu_ring **ring) +int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type, + int *blacklist, int num_blacklist, + bool lru_pipe_order, struct amdgpu_ring **ring) { struct amdgpu_ring *entry; @@ -337,10 +339,23 @@ int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type, int *blacklist, if (amdgpu_ring_is_blacklisted(entry, blacklist, num_blacklist)) continue; - *ring = entry; - amdgpu_ring_lru_touch_locked(adev, *ring); - break; + if (!*ring) { + *ring = entry; + + /* We are done for ring LRU */ + if (!lru_pipe_order) + break; + } + + /* Move all rings on the same pipe to the end of the list */ + if (entry->pipe == (*ring)->pipe) + amdgpu_ring_lru_touch_locked(adev, entry); } + + /* Move the ring we found to the end of the list */ + if (*ring) + amdgpu_ring_lru_touch_locked(adev, *ring); + spin_unlock(&adev->ring_lru_list_lock); if (!*ring) { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h index 322d25299a00..491bd5512dcc 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h @@ -201,8 +201,9 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, unsigned ring_size, struct amdgpu_irq_src *irq_src, unsigned irq_type); void amdgpu_ring_fini(struct amdgpu_ring *ring); -int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type, int *blacklist, - int num_blacklist, struct amdgpu_ring **ring); +int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type, + int *blacklist, int num_blacklist, + bool lru_pipe_order, struct amdgpu_ring **ring); void amdgpu_ring_lru_touch(struct amdgpu_device *adev, struct amdgpu_ring *ring); static inline void amdgpu_ring_clear_ring(struct amdgpu_ring *ring) { -- cgit v1.2.3 From b2ff0e8ac4ce1fb647ae40feb4cf26bc9301e0c9 Mon Sep 17 00:00:00 2001 From: Andres Rodriguez Date: Mon, 20 Feb 2017 17:53:19 -0500 Subject: drm/amdgpu: add framework for HW specific priority settings v9 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Add an initial framework for changing the HW priorities of rings. The framework allows requesting priority changes for the lifetime of an amdgpu_job. After the job completes the priority will decay to the next lowest priority for which a request is still valid. A new ring function set_priority() can now be populated to take care of the HW specific programming sequence for priority changes. v2: set priority before emitting IB, and take a ref on amdgpu_job v3: use AMD_SCHED_PRIORITY_* instead of AMDGPU_CTX_PRIORITY_* v4: plug amdgpu_ring_restore_priority_cb into amdgpu_job_free_cb v5: use atomic for tracking job priorities instead of last_job v6: rename amdgpu_ring_priority_[get/put]() and align parameters v7: replace spinlocks with mutexes for KIQ compatibility v8: raise ring priority during cs_ioctl, instead of job_run v9: priority_get() before push_job() Reviewed-by: Christian König Acked-by: Christian König Signed-off-by: Andres Rodriguez Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 76 ++++++++++++++++++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 15 ++++++ drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 7 +++ 5 files changed, 103 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 38027a00f8ab..fe7dd44ac9fe 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -1177,6 +1177,8 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, job->uf_sequence = seq; amdgpu_job_free_resources(job); + amdgpu_ring_priority_get(job->ring, + amd_sched_get_job_priority(&job->base)); trace_amdgpu_cs_ioctl(job); amd_sched_entity_push_job(&job->base); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c index 4510627ae83e..83d13431cbdd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c @@ -103,6 +103,7 @@ static void amdgpu_job_free_cb(struct amd_sched_job *s_job) { struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base); + amdgpu_ring_priority_put(job->ring, amd_sched_get_job_priority(s_job)); dma_fence_put(job->fence); amdgpu_sync_free(&job->sync); amdgpu_sync_free(&job->dep_sync); @@ -139,6 +140,8 @@ int amdgpu_job_submit(struct amdgpu_job *job, struct amdgpu_ring *ring, job->fence_ctx = entity->fence_context; *f = dma_fence_get(&job->base.s_fence->finished); amdgpu_job_free_resources(job); + amdgpu_ring_priority_get(job->ring, + amd_sched_get_job_priority(&job->base)); amd_sched_entity_push_job(&job->base); return 0; @@ -203,6 +206,7 @@ static struct dma_fence *amdgpu_job_run(struct amd_sched_job *sched_job) /* if gpu reset, hw fence will be replaced here */ dma_fence_put(job->fence); job->fence = dma_fence_get(fence); + amdgpu_job_free_resources(job); return fence; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index 019932a7ea3a..e5ece1fae149 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -154,6 +154,75 @@ void amdgpu_ring_undo(struct amdgpu_ring *ring) ring->funcs->end_use(ring); } +/** + * amdgpu_ring_priority_put - restore a ring's priority + * + * @ring: amdgpu_ring structure holding the information + * @priority: target priority + * + * Release a request for executing at @priority + */ +void amdgpu_ring_priority_put(struct amdgpu_ring *ring, + enum amd_sched_priority priority) +{ + int i; + + if (!ring->funcs->set_priority) + return; + + if (atomic_dec_return(&ring->num_jobs[priority]) > 0) + return; + + /* no need to restore if the job is already at the lowest priority */ + if (priority == AMD_SCHED_PRIORITY_NORMAL) + return; + + mutex_lock(&ring->priority_mutex); + /* something higher prio is executing, no need to decay */ + if (ring->priority > priority) + goto out_unlock; + + /* decay priority to the next level with a job available */ + for (i = priority; i >= AMD_SCHED_PRIORITY_MIN; i--) { + if (i == AMD_SCHED_PRIORITY_NORMAL + || atomic_read(&ring->num_jobs[i])) { + ring->priority = i; + ring->funcs->set_priority(ring, i); + break; + } + } + +out_unlock: + mutex_unlock(&ring->priority_mutex); +} + +/** + * amdgpu_ring_priority_get - change the ring's priority + * + * @ring: amdgpu_ring structure holding the information + * @priority: target priority + * + * Request a ring's priority to be raised to @priority (refcounted). + */ +void amdgpu_ring_priority_get(struct amdgpu_ring *ring, + enum amd_sched_priority priority) +{ + if (!ring->funcs->set_priority) + return; + + atomic_inc(&ring->num_jobs[priority]); + + mutex_lock(&ring->priority_mutex); + if (priority <= ring->priority) + goto out_unlock; + + ring->priority = priority; + ring->funcs->set_priority(ring, priority); + +out_unlock: + mutex_unlock(&ring->priority_mutex); +} + /** * amdgpu_ring_init - init driver ring struct. * @@ -169,7 +238,7 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, unsigned max_dw, struct amdgpu_irq_src *irq_src, unsigned irq_type) { - int r; + int r, i; int sched_hw_submission = amdgpu_sched_hw_submission; /* Set the hw submission limit higher for KIQ because @@ -247,9 +316,14 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, } ring->max_dw = max_dw; + ring->priority = AMD_SCHED_PRIORITY_NORMAL; + mutex_init(&ring->priority_mutex); INIT_LIST_HEAD(&ring->lru_list); amdgpu_ring_lru_touch(adev, ring); + for (i = 0; i < AMD_SCHED_PRIORITY_MAX; ++i) + atomic_set(&ring->num_jobs[i], 0); + if (amdgpu_debugfs_ring_init(adev, ring)) { DRM_ERROR("Failed to register debugfs file for rings !\n"); } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h index 491bd5512dcc..0d9ce141404c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h @@ -24,6 +24,7 @@ #ifndef __AMDGPU_RING_H__ #define __AMDGPU_RING_H__ +#include #include "gpu_scheduler.h" /* max number of rings */ @@ -56,6 +57,7 @@ struct amdgpu_device; struct amdgpu_ring; struct amdgpu_ib; struct amdgpu_cs_parser; +struct amdgpu_job; /* * Fences. @@ -147,6 +149,9 @@ struct amdgpu_ring_funcs { void (*emit_rreg)(struct amdgpu_ring *ring, uint32_t reg); void (*emit_wreg)(struct amdgpu_ring *ring, uint32_t reg, uint32_t val); void (*emit_tmz)(struct amdgpu_ring *ring, bool start); + /* priority functions */ + void (*set_priority) (struct amdgpu_ring *ring, + enum amd_sched_priority priority); }; struct amdgpu_ring { @@ -187,6 +192,12 @@ struct amdgpu_ring { volatile u32 *cond_exe_cpu_addr; unsigned vm_inv_eng; bool has_compute_vm_bug; + + atomic_t num_jobs[AMD_SCHED_PRIORITY_MAX]; + struct mutex priority_mutex; + /* protected by priority_mutex */ + int priority; + #if defined(CONFIG_DEBUG_FS) struct dentry *ent; #endif @@ -197,6 +208,10 @@ void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count); void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib); void amdgpu_ring_commit(struct amdgpu_ring *ring); void amdgpu_ring_undo(struct amdgpu_ring *ring); +void amdgpu_ring_priority_get(struct amdgpu_ring *ring, + enum amd_sched_priority priority); +void amdgpu_ring_priority_put(struct amdgpu_ring *ring, + enum amd_sched_priority priority); int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, unsigned ring_size, struct amdgpu_irq_src *irq_src, unsigned irq_type); diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h index 38e622ce06de..dbcaa2e1c5c7 100644 --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h @@ -170,4 +170,11 @@ void amd_sched_job_recovery(struct amd_gpu_scheduler *sched); bool amd_sched_dependency_optimized(struct dma_fence* fence, struct amd_sched_entity *entity); void amd_sched_job_kickout(struct amd_sched_job *s_job); + +static inline enum amd_sched_priority +amd_sched_get_job_priority(struct amd_sched_job *job) +{ + return (job->s_entity->rq - job->sched->sched_rq); +} + #endif -- cgit v1.2.3 From dce1e131dd4dc68099ff1b70aa03cd2d0acf8639 Mon Sep 17 00:00:00 2001 From: Pixel Ding Date: Wed, 8 Nov 2017 10:20:01 +0800 Subject: drm/amdgpu: bypass lru touch for KIQ ring submission MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit KIQ ring submission is used for register accessing on SRIOV VF that could happen both in irq enabled and irq disabled cases. Inversion lock could happen on adev->ring_lru_list_lock, while this operation is useless and just adds overhead in this use case. Signed-off-by: Pixel Ding Reviewed-by: Monk Liu Reviewed-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index e5ece1fae149..a98fbbb4739f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -136,7 +136,8 @@ void amdgpu_ring_commit(struct amdgpu_ring *ring) if (ring->funcs->end_use) ring->funcs->end_use(ring); - amdgpu_ring_lru_touch(ring->adev, ring); + if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) + amdgpu_ring_lru_touch(ring->adev, ring); } /** -- cgit v1.2.3