2 * Copyright 2015 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
25 #include <drm/drm_auth.h>
27 #include "amdgpu_sched.h"
28 #include "amdgpu_ras.h"
29 #include <linux/nospec.h>
31 #define to_amdgpu_ctx_entity(e) \
32 container_of((e), struct amdgpu_ctx_entity, entity)
34 const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = {
35 [AMDGPU_HW_IP_GFX] = 1,
36 [AMDGPU_HW_IP_COMPUTE] = 4,
37 [AMDGPU_HW_IP_DMA] = 2,
38 [AMDGPU_HW_IP_UVD] = 1,
39 [AMDGPU_HW_IP_VCE] = 1,
40 [AMDGPU_HW_IP_UVD_ENC] = 1,
41 [AMDGPU_HW_IP_VCN_DEC] = 1,
42 [AMDGPU_HW_IP_VCN_ENC] = 1,
43 [AMDGPU_HW_IP_VCN_JPEG] = 1,
46 static int amdgpu_ctx_priority_permit(struct drm_file *filp,
47 enum drm_sched_priority priority)
49 if (priority < 0 || priority >= DRM_SCHED_PRIORITY_COUNT)
52 /* NORMAL and below are accessible by everyone */
53 if (priority <= DRM_SCHED_PRIORITY_NORMAL)
56 if (capable(CAP_SYS_NICE))
59 if (drm_is_current_master(filp))
65 static enum gfx_pipe_priority amdgpu_ctx_sched_prio_to_compute_prio(enum drm_sched_priority prio)
68 case DRM_SCHED_PRIORITY_HIGH:
69 case DRM_SCHED_PRIORITY_KERNEL:
70 return AMDGPU_GFX_PIPE_PRIO_HIGH;
72 return AMDGPU_GFX_PIPE_PRIO_NORMAL;
76 static unsigned int amdgpu_ctx_prio_sched_to_hw(struct amdgpu_device *adev,
77 enum drm_sched_priority prio,
82 hw_prio = (hw_ip == AMDGPU_HW_IP_COMPUTE) ?
83 amdgpu_ctx_sched_prio_to_compute_prio(prio) :
84 AMDGPU_RING_PRIO_DEFAULT;
85 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
86 if (adev->gpu_sched[hw_ip][hw_prio].num_scheds == 0)
87 hw_prio = AMDGPU_RING_PRIO_DEFAULT;
92 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
95 struct amdgpu_device *adev = ctx->adev;
96 struct amdgpu_ctx_entity *entity;
97 struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
98 unsigned num_scheds = 0;
100 enum drm_sched_priority priority;
103 entity = kzalloc(struct_size(entity, fences, amdgpu_sched_jobs),
108 entity->sequence = 1;
109 priority = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
110 ctx->init_priority : ctx->override_priority;
111 hw_prio = amdgpu_ctx_prio_sched_to_hw(adev, priority, hw_ip);
113 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
114 scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
115 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
117 /* disable load balance if the hw engine retains context among dependent jobs */
118 if (hw_ip == AMDGPU_HW_IP_VCN_ENC ||
119 hw_ip == AMDGPU_HW_IP_VCN_DEC ||
120 hw_ip == AMDGPU_HW_IP_UVD_ENC ||
121 hw_ip == AMDGPU_HW_IP_UVD) {
122 sched = drm_sched_pick_best(scheds, num_scheds);
127 r = drm_sched_entity_init(&entity->entity, priority, scheds, num_scheds,
130 goto error_free_entity;
132 ctx->entities[hw_ip][ring] = entity;
141 static int amdgpu_ctx_init(struct amdgpu_device *adev,
142 enum drm_sched_priority priority,
143 struct drm_file *filp,
144 struct amdgpu_ctx *ctx)
148 r = amdgpu_ctx_priority_permit(filp, priority);
152 memset(ctx, 0, sizeof(*ctx));
156 kref_init(&ctx->refcount);
157 spin_lock_init(&ctx->ring_lock);
158 mutex_init(&ctx->lock);
160 ctx->reset_counter = atomic_read(&adev->gpu_reset_counter);
161 ctx->reset_counter_query = ctx->reset_counter;
162 ctx->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
163 ctx->init_priority = priority;
164 ctx->override_priority = DRM_SCHED_PRIORITY_UNSET;
169 static void amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity)
177 for (i = 0; i < amdgpu_sched_jobs; ++i)
178 dma_fence_put(entity->fences[i]);
183 static void amdgpu_ctx_fini(struct kref *ref)
185 struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount);
186 struct amdgpu_device *adev = ctx->adev;
192 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
193 for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) {
194 amdgpu_ctx_fini_entity(ctx->entities[i][j]);
195 ctx->entities[i][j] = NULL;
199 mutex_destroy(&ctx->lock);
203 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
204 u32 ring, struct drm_sched_entity **entity)
208 if (hw_ip >= AMDGPU_HW_IP_NUM) {
209 DRM_ERROR("unknown HW IP type: %d\n", hw_ip);
213 /* Right now all IPs have only one instance - multiple rings. */
215 DRM_DEBUG("invalid ip instance: %d\n", instance);
219 if (ring >= amdgpu_ctx_num_entities[hw_ip]) {
220 DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring);
224 if (ctx->entities[hw_ip][ring] == NULL) {
225 r = amdgpu_ctx_init_entity(ctx, hw_ip, ring);
230 *entity = &ctx->entities[hw_ip][ring]->entity;
234 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
235 struct amdgpu_fpriv *fpriv,
236 struct drm_file *filp,
237 enum drm_sched_priority priority,
240 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
241 struct amdgpu_ctx *ctx;
244 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
248 mutex_lock(&mgr->lock);
249 r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL);
251 mutex_unlock(&mgr->lock);
257 r = amdgpu_ctx_init(adev, priority, filp, ctx);
259 idr_remove(&mgr->ctx_handles, *id);
263 mutex_unlock(&mgr->lock);
267 static void amdgpu_ctx_do_release(struct kref *ref)
269 struct amdgpu_ctx *ctx;
272 ctx = container_of(ref, struct amdgpu_ctx, refcount);
273 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
274 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
275 if (!ctx->entities[i][j])
278 drm_sched_entity_destroy(&ctx->entities[i][j]->entity);
282 amdgpu_ctx_fini(ref);
285 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id)
287 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
288 struct amdgpu_ctx *ctx;
290 mutex_lock(&mgr->lock);
291 ctx = idr_remove(&mgr->ctx_handles, id);
293 kref_put(&ctx->refcount, amdgpu_ctx_do_release);
294 mutex_unlock(&mgr->lock);
295 return ctx ? 0 : -EINVAL;
298 static int amdgpu_ctx_query(struct amdgpu_device *adev,
299 struct amdgpu_fpriv *fpriv, uint32_t id,
300 union drm_amdgpu_ctx_out *out)
302 struct amdgpu_ctx *ctx;
303 struct amdgpu_ctx_mgr *mgr;
304 unsigned reset_counter;
309 mgr = &fpriv->ctx_mgr;
310 mutex_lock(&mgr->lock);
311 ctx = idr_find(&mgr->ctx_handles, id);
313 mutex_unlock(&mgr->lock);
317 /* TODO: these two are always zero */
318 out->state.flags = 0x0;
319 out->state.hangs = 0x0;
321 /* determine if a GPU reset has occured since the last call */
322 reset_counter = atomic_read(&adev->gpu_reset_counter);
323 /* TODO: this should ideally return NO, GUILTY, or INNOCENT. */
324 if (ctx->reset_counter_query == reset_counter)
325 out->state.reset_status = AMDGPU_CTX_NO_RESET;
327 out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET;
328 ctx->reset_counter_query = reset_counter;
330 mutex_unlock(&mgr->lock);
334 static int amdgpu_ctx_query2(struct amdgpu_device *adev,
335 struct amdgpu_fpriv *fpriv, uint32_t id,
336 union drm_amdgpu_ctx_out *out)
338 struct amdgpu_ctx *ctx;
339 struct amdgpu_ctx_mgr *mgr;
344 mgr = &fpriv->ctx_mgr;
345 mutex_lock(&mgr->lock);
346 ctx = idr_find(&mgr->ctx_handles, id);
348 mutex_unlock(&mgr->lock);
352 out->state.flags = 0x0;
353 out->state.hangs = 0x0;
355 if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter))
356 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET;
358 if (ctx->vram_lost_counter != atomic_read(&adev->vram_lost_counter))
359 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST;
361 if (atomic_read(&ctx->guilty))
362 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY;
364 mutex_unlock(&mgr->lock);
368 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
369 struct drm_file *filp)
373 enum drm_sched_priority priority;
375 union drm_amdgpu_ctx *args = data;
376 struct amdgpu_device *adev = drm_to_adev(dev);
377 struct amdgpu_fpriv *fpriv = filp->driver_priv;
379 id = args->in.ctx_id;
380 r = amdgpu_to_sched_priority(args->in.priority, &priority);
382 /* For backwards compatibility reasons, we need to accept
383 * ioctls with garbage in the priority field */
385 priority = DRM_SCHED_PRIORITY_NORMAL;
387 switch (args->in.op) {
388 case AMDGPU_CTX_OP_ALLOC_CTX:
389 r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id);
390 args->out.alloc.ctx_id = id;
392 case AMDGPU_CTX_OP_FREE_CTX:
393 r = amdgpu_ctx_free(fpriv, id);
395 case AMDGPU_CTX_OP_QUERY_STATE:
396 r = amdgpu_ctx_query(adev, fpriv, id, &args->out);
398 case AMDGPU_CTX_OP_QUERY_STATE2:
399 r = amdgpu_ctx_query2(adev, fpriv, id, &args->out);
408 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id)
410 struct amdgpu_ctx *ctx;
411 struct amdgpu_ctx_mgr *mgr;
416 mgr = &fpriv->ctx_mgr;
418 mutex_lock(&mgr->lock);
419 ctx = idr_find(&mgr->ctx_handles, id);
421 kref_get(&ctx->refcount);
422 mutex_unlock(&mgr->lock);
426 int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
431 kref_put(&ctx->refcount, amdgpu_ctx_do_release);
435 void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
436 struct drm_sched_entity *entity,
437 struct dma_fence *fence, uint64_t *handle)
439 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
440 uint64_t seq = centity->sequence;
441 struct dma_fence *other = NULL;
444 idx = seq & (amdgpu_sched_jobs - 1);
445 other = centity->fences[idx];
447 BUG_ON(!dma_fence_is_signaled(other));
449 dma_fence_get(fence);
451 spin_lock(&ctx->ring_lock);
452 centity->fences[idx] = fence;
454 spin_unlock(&ctx->ring_lock);
456 dma_fence_put(other);
461 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
462 struct drm_sched_entity *entity,
465 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
466 struct dma_fence *fence;
468 spin_lock(&ctx->ring_lock);
471 seq = centity->sequence - 1;
473 if (seq >= centity->sequence) {
474 spin_unlock(&ctx->ring_lock);
475 return ERR_PTR(-EINVAL);
479 if (seq + amdgpu_sched_jobs < centity->sequence) {
480 spin_unlock(&ctx->ring_lock);
484 fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]);
485 spin_unlock(&ctx->ring_lock);
490 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
491 struct amdgpu_ctx_entity *aentity,
493 enum drm_sched_priority priority)
495 struct amdgpu_device *adev = ctx->adev;
496 unsigned int hw_prio;
497 struct drm_gpu_scheduler **scheds = NULL;
500 /* set sw priority */
501 drm_sched_entity_set_priority(&aentity->entity, priority);
503 /* set hw priority */
504 if (hw_ip == AMDGPU_HW_IP_COMPUTE) {
505 hw_prio = amdgpu_ctx_prio_sched_to_hw(adev, priority,
506 AMDGPU_HW_IP_COMPUTE);
507 hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX);
508 scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
509 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
510 drm_sched_entity_modify_sched(&aentity->entity, scheds,
515 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
516 enum drm_sched_priority priority)
518 enum drm_sched_priority ctx_prio;
521 ctx->override_priority = priority;
523 ctx_prio = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
524 ctx->init_priority : ctx->override_priority;
525 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
526 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
527 if (!ctx->entities[i][j])
530 amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j],
536 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
537 struct drm_sched_entity *entity)
539 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
540 struct dma_fence *other;
544 spin_lock(&ctx->ring_lock);
545 idx = centity->sequence & (amdgpu_sched_jobs - 1);
546 other = dma_fence_get(centity->fences[idx]);
547 spin_unlock(&ctx->ring_lock);
552 r = dma_fence_wait(other, true);
553 if (r < 0 && r != -ERESTARTSYS)
554 DRM_ERROR("Error (%ld) waiting for fence!\n", r);
556 dma_fence_put(other);
560 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)
562 mutex_init(&mgr->lock);
563 idr_init(&mgr->ctx_handles);
566 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
568 struct amdgpu_ctx *ctx;
572 idp = &mgr->ctx_handles;
574 mutex_lock(&mgr->lock);
575 idr_for_each_entry(idp, ctx, id) {
576 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
577 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
578 struct drm_sched_entity *entity;
580 if (!ctx->entities[i][j])
583 entity = &ctx->entities[i][j]->entity;
584 timeout = drm_sched_entity_flush(entity, timeout);
588 mutex_unlock(&mgr->lock);
592 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
594 struct amdgpu_ctx *ctx;
598 idp = &mgr->ctx_handles;
600 idr_for_each_entry(idp, ctx, id) {
601 if (kref_read(&ctx->refcount) != 1) {
602 DRM_ERROR("ctx %p is still alive\n", ctx);
606 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
607 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
608 struct drm_sched_entity *entity;
610 if (!ctx->entities[i][j])
613 entity = &ctx->entities[i][j]->entity;
614 drm_sched_entity_fini(entity);
620 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
622 struct amdgpu_ctx *ctx;
626 amdgpu_ctx_mgr_entity_fini(mgr);
628 idp = &mgr->ctx_handles;
630 idr_for_each_entry(idp, ctx, id) {
631 if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1)
632 DRM_ERROR("ctx %p is still alive\n", ctx);
635 idr_destroy(&mgr->ctx_handles);
636 mutex_destroy(&mgr->lock);