forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2024-05-10 37f49e37ab4cb5d0bc4c60eb5c6d4dd57db767bb 
 kernel/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
....@@ -22,14 +22,33 @@
2222  * Authors: monk liu <monk.liu@amd.com>
2323  */
2424 
25
-#include <drm/drmP.h>
2625 #include <drm/drm_auth.h>
2726 #include "amdgpu.h"
2827 #include "amdgpu_sched.h"
28
+#include "amdgpu_ras.h"
29
+#include <linux/nospec.h>
30
+
31
+#define to_amdgpu_ctx_entity(e)	\
32
+	container_of((e), struct amdgpu_ctx_entity, entity)
33
+
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,
44
+};
2945 
3046 static int amdgpu_ctx_priority_permit(struct drm_file *filp,
3147 				      enum drm_sched_priority priority)
3248 {
49
+	if (priority < 0 || priority >= DRM_SCHED_PRIORITY_COUNT)
50
+		return -EINVAL;
51
+
3352 	/* NORMAL and below are accessible by everyone */
3453 	if (priority <= DRM_SCHED_PRIORITY_NORMAL)
3554 		return 0;
....@@ -43,36 +62,100 @@
4362 	return -EACCES;
4463 }
4564 
65
+static enum gfx_pipe_priority amdgpu_ctx_sched_prio_to_compute_prio(enum drm_sched_priority prio)
66
+{
67
+	switch (prio) {
68
+	case DRM_SCHED_PRIORITY_HIGH:
69
+	case DRM_SCHED_PRIORITY_KERNEL:
70
+		return AMDGPU_GFX_PIPE_PRIO_HIGH;
71
+	default:
72
+		return AMDGPU_GFX_PIPE_PRIO_NORMAL;
73
+	}
74
+}
75
+
76
+static unsigned int amdgpu_ctx_prio_sched_to_hw(struct amdgpu_device *adev,
77
+						 enum drm_sched_priority prio,
78
+						 u32 hw_ip)
79
+{
80
+	unsigned int hw_prio;
81
+
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;
88
+
89
+	return hw_prio;
90
+}
91
+
92
+static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
93
+				   const u32 ring)
94
+{
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;
99
+	unsigned int hw_prio;
100
+	enum drm_sched_priority priority;
101
+	int r;
102
+
103
+	entity = kcalloc(1, offsetof(typeof(*entity), fences[amdgpu_sched_jobs]),
104
+			 GFP_KERNEL);
105
+	if (!entity)
106
+		return  -ENOMEM;
107
+
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);
112
+
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;
116
+
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);
123
+		scheds = &sched;
124
+		num_scheds = 1;
125
+	}
126
+
127
+	r = drm_sched_entity_init(&entity->entity, priority, scheds, num_scheds,
128
+				  &ctx->guilty);
129
+	if (r)
130
+		goto error_free_entity;
131
+
132
+	ctx->entities[hw_ip][ring] = entity;
133
+	return 0;
134
+
135
+error_free_entity:
136
+	kfree(entity);
137
+
138
+	return r;
139
+}
140
+
46141 static int amdgpu_ctx_init(struct amdgpu_device *adev,
47142 			   enum drm_sched_priority priority,
48143 			   struct drm_file *filp,
49144 			   struct amdgpu_ctx *ctx)
50145 {
51
-	unsigned i, j;
52146 	int r;
53
-
54
-	if (priority < 0 || priority >= DRM_SCHED_PRIORITY_MAX)
55
-		return -EINVAL;
56147 
57148 	r = amdgpu_ctx_priority_permit(filp, priority);
58149 	if (r)
59150 		return r;
60151 
61152 	memset(ctx, 0, sizeof(*ctx));
153
+
62154 	ctx->adev = adev;
155
+
63156 	kref_init(&ctx->refcount);
64157 	spin_lock_init(&ctx->ring_lock);
65
-	ctx->fences = kcalloc(amdgpu_sched_jobs * AMDGPU_MAX_RINGS,
66
-			      sizeof(struct dma_fence*), GFP_KERNEL);
67
-	if (!ctx->fences)
68
-		return -ENOMEM;
69
-
70158 	mutex_init(&ctx->lock);
71
-
72
-	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
73
-		ctx->rings[i].sequence = 1;
74
-		ctx->rings[i].fences = &ctx->fences[amdgpu_sched_jobs * i];
75
-	}
76159 
77160 	ctx->reset_counter = atomic_read(&adev->gpu_reset_counter);
78161 	ctx->reset_counter_query = ctx->reset_counter;
....@@ -80,34 +163,21 @@
80163 	ctx->init_priority = priority;
81164 	ctx->override_priority = DRM_SCHED_PRIORITY_UNSET;
82165 
83
-	/* create context entity for each ring */
84
-	for (i = 0; i < adev->num_rings; i++) {
85
-		struct amdgpu_ring *ring = adev->rings[i];
86
-		struct drm_sched_rq *rq;
87
-
88
-		rq = &ring->sched.sched_rq[priority];
89
-
90
-		if (ring == &adev->gfx.kiq.ring)
91
-			continue;
92
-
93
-		r = drm_sched_entity_init(&ctx->rings[i].entity,
94
-					  &rq, 1, &ctx->guilty);
95
-		if (r)
96
-			goto failed;
97
-	}
98
-
99
-	r = amdgpu_queue_mgr_init(adev, &ctx->queue_mgr);
100
-	if (r)
101
-		goto failed;
102
-
103166 	return 0;
167
+}
104168 
105
-failed:
106
-	for (j = 0; j < i; j++)
107
-		drm_sched_entity_destroy(&ctx->rings[j].entity);
108
-	kfree(ctx->fences);
109
-	ctx->fences = NULL;
110
-	return r;
169
+static void amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity)
170
+{
171
+
172
+	int i;
173
+
174
+	if (!entity)
175
+		return;
176
+
177
+	for (i = 0; i < amdgpu_sched_jobs; ++i)
178
+		dma_fence_put(entity->fences[i]);
179
+
180
+	kfree(entity);
111181 }
112182 
113183 static void amdgpu_ctx_fini(struct kref *ref)
....@@ -119,17 +189,46 @@
119189 	if (!adev)
120190 		return;
121191 
122
-	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
123
-		for (j = 0; j < amdgpu_sched_jobs; ++j)
124
-			dma_fence_put(ctx->rings[i].fences[j]);
125
-	kfree(ctx->fences);
126
-	ctx->fences = NULL;
127
-
128
-	amdgpu_queue_mgr_fini(adev, &ctx->queue_mgr);
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;
196
+		}
197
+	}
129198 
130199 	mutex_destroy(&ctx->lock);
131
-
132200 	kfree(ctx);
201
+}
202
+
203
+int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
204
+			  u32 ring, struct drm_sched_entity **entity)
205
+{
206
+	int r;
207
+
208
+	if (hw_ip >= AMDGPU_HW_IP_NUM) {
209
+		DRM_ERROR("unknown HW IP type: %d\n", hw_ip);
210
+		return -EINVAL;
211
+	}
212
+
213
+	/* Right now all IPs have only one instance - multiple rings. */
214
+	if (instance != 0) {
215
+		DRM_DEBUG("invalid ip instance: %d\n", instance);
216
+		return -EINVAL;
217
+	}
218
+
219
+	if (ring >= amdgpu_ctx_num_entities[hw_ip]) {
220
+		DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring);
221
+		return -EINVAL;
222
+	}
223
+
224
+	if (ctx->entities[hw_ip][ring] == NULL) {
225
+		r = amdgpu_ctx_init_entity(ctx, hw_ip, ring);
226
+		if (r)
227
+			return r;
228
+	}
229
+
230
+	*entity = &ctx->entities[hw_ip][ring]->entity;
231
+	return 0;
133232 }
134233 
135234 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
....@@ -147,7 +246,7 @@
147246 		return -ENOMEM;
148247 
149248 	mutex_lock(&mgr->lock);
150
-	r = idr_alloc(&mgr->ctx_handles, ctx, 1, 0, GFP_KERNEL);
249
+	r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL);
151250 	if (r < 0) {
152251 		mutex_unlock(&mgr->lock);
153252 		kfree(ctx);
....@@ -168,16 +267,16 @@
168267 static void amdgpu_ctx_do_release(struct kref *ref)
169268 {
170269 	struct amdgpu_ctx *ctx;
171
-	u32 i;
270
+	u32 i, j;
172271 
173272 	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])
276
+				continue;
174277 
175
-	for (i = 0; i < ctx->adev->num_rings; i++) {
176
-
177
-		if (ctx->adev->rings[i] == &ctx->adev->gfx.kiq.ring)
178
-			continue;
179
-
180
-		drm_sched_entity_destroy(&ctx->rings[i].entity);
278
+			drm_sched_entity_destroy(&ctx->entities[i][j]->entity);
279
+		}
181280 	}
182281 
183282 	amdgpu_ctx_fini(ref);
....@@ -274,16 +373,15 @@
274373 	enum drm_sched_priority priority;
275374 
276375 	union drm_amdgpu_ctx *args = data;
277
-	struct amdgpu_device *adev = dev->dev_private;
376
+	struct amdgpu_device *adev = drm_to_adev(dev);
278377 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
279378 
280
-	r = 0;
281379 	id = args->in.ctx_id;
282
-	priority = amdgpu_to_sched_priority(args->in.priority);
380
+	r = amdgpu_to_sched_priority(args->in.priority, &priority);
283381 
284382 	/* For backwards compatibility reasons, we need to accept
285383 	 * ioctls with garbage in the priority field */
286
-	if (priority == DRM_SCHED_PRIORITY_INVALID)
384
+	if (r == -EINVAL)
287385 		priority = DRM_SCHED_PRIORITY_NORMAL;
288386 
289387 	switch (args->in.op) {
....@@ -334,106 +432,129 @@
334432 	return 0;
335433 }
336434 
337
-int amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
338
-			      struct dma_fence *fence, uint64_t* handler)
435
+void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
436
+			  struct drm_sched_entity *entity,
437
+			  struct dma_fence *fence, uint64_t* handle)
339438 {
340
-	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
341
-	uint64_t seq = cring->sequence;
342
-	unsigned idx = 0;
439
+	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
440
+	uint64_t seq = centity->sequence;
343441 	struct dma_fence *other = NULL;
442
+	unsigned idx = 0;
344443 
345444 	idx = seq & (amdgpu_sched_jobs - 1);
346
-	other = cring->fences[idx];
445
+	other = centity->fences[idx];
347446 	if (other)
348447 		BUG_ON(!dma_fence_is_signaled(other));
349448 
350449 	dma_fence_get(fence);
351450 
352451 	spin_lock(&ctx->ring_lock);
353
-	cring->fences[idx] = fence;
354
-	cring->sequence++;
452
+	centity->fences[idx] = fence;
453
+	centity->sequence++;
355454 	spin_unlock(&ctx->ring_lock);
356455 
357456 	dma_fence_put(other);
358
-	if (handler)
359
-		*handler = seq;
360
-
361
-	return 0;
457
+	if (handle)
458
+		*handle = seq;
362459 }
363460 
364461 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
365
-				       struct amdgpu_ring *ring, uint64_t seq)
462
+				       struct drm_sched_entity *entity,
463
+				       uint64_t seq)
366464 {
367
-	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
465
+	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
368466 	struct dma_fence *fence;
369467 
370468 	spin_lock(&ctx->ring_lock);
371469 
372470 	if (seq == ~0ull)
373
-		seq = ctx->rings[ring->idx].sequence - 1;
471
+		seq = centity->sequence - 1;
374472 
375
-	if (seq >= cring->sequence) {
473
+	if (seq >= centity->sequence) {
376474 		spin_unlock(&ctx->ring_lock);
377475 		return ERR_PTR(-EINVAL);
378476 	}
379477 
380478 
381
-	if (seq + amdgpu_sched_jobs < cring->sequence) {
479
+	if (seq + amdgpu_sched_jobs < centity->sequence) {
382480 		spin_unlock(&ctx->ring_lock);
383481 		return NULL;
384482 	}
385483 
386
-	fence = dma_fence_get(cring->fences[seq & (amdgpu_sched_jobs - 1)]);
484
+	fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]);
387485 	spin_unlock(&ctx->ring_lock);
388486 
389487 	return fence;
390488 }
391489 
490
+static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
491
+					    struct amdgpu_ctx_entity *aentity,
492
+					    int hw_ip,
493
+					    enum drm_sched_priority priority)
494
+{
495
+	struct amdgpu_device *adev = ctx->adev;
496
+	unsigned int hw_prio;
497
+	struct drm_gpu_scheduler **scheds = NULL;
498
+	unsigned num_scheds;
499
+
500
+	/* set sw priority */
501
+	drm_sched_entity_set_priority(&aentity->entity, priority);
502
+
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,
511
+					      num_scheds);
512
+	}
513
+}
514
+
392515 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
393516 				  enum drm_sched_priority priority)
394517 {
395
-	int i;
396
-	struct amdgpu_device *adev = ctx->adev;
397
-	struct drm_sched_rq *rq;
398
-	struct drm_sched_entity *entity;
399
-	struct amdgpu_ring *ring;
400518 	enum drm_sched_priority ctx_prio;
519
+	unsigned i, j;
401520 
402521 	ctx->override_priority = priority;
403522 
404523 	ctx_prio = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
405524 			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])
528
+				continue;
406529 
407
-	for (i = 0; i < adev->num_rings; i++) {
408
-		ring = adev->rings[i];
409
-		entity = &ctx->rings[i].entity;
410
-		rq = &ring->sched.sched_rq[ctx_prio];
411
-
412
-		if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
413
-			continue;
414
-
415
-		drm_sched_entity_set_rq(entity, rq);
530
+			amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j],
531
+						       i, ctx_prio);
532
+		}
416533 	}
417534 }
418535 
419
-int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx, unsigned ring_id)
536
+int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
537
+			       struct drm_sched_entity *entity)
420538 {
421
-	struct amdgpu_ctx_ring *cring = &ctx->rings[ring_id];
422
-	unsigned idx = cring->sequence & (amdgpu_sched_jobs - 1);
423
-	struct dma_fence *other = cring->fences[idx];
539
+	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
540
+	struct dma_fence *other;
541
+	unsigned idx;
542
+	long r;
424543 
425
-	if (other) {
426
-		signed long r;
427
-		r = dma_fence_wait(other, true);
428
-		if (r < 0) {
429
-			if (r != -ERESTARTSYS)
430
-				DRM_ERROR("Error (%ld) waiting for fence!\n", r);
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);
431548 
432
-			return r;
433
-		}
434
-	}
549
+	if (!other)
550
+		return 0;
435551 
436
-	return 0;
552
+	r = dma_fence_wait(other, true);
553
+	if (r < 0 && r != -ERESTARTSYS)
554
+		DRM_ERROR("Error (%ld) waiting for fence!\n", r);
555
+
556
+	dma_fence_put(other);
557
+	return r;
437558 }
438559 
439560 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)
....@@ -442,57 +563,56 @@
442563 	idr_init(&mgr->ctx_handles);
443564 }
444565 
445
-void amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr)
566
+long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
446567 {
447568 	struct amdgpu_ctx *ctx;
448569 	struct idr *idp;
449
-	uint32_t id, i;
450
-	long max_wait = MAX_WAIT_SCHED_ENTITY_Q_EMPTY;
570
+	uint32_t id, i, j;
451571 
452572 	idp = &mgr->ctx_handles;
453573 
454574 	mutex_lock(&mgr->lock);
455575 	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;
456579 
457
-		if (!ctx->adev) {
458
-			mutex_unlock(&mgr->lock);
459
-			return;
460
-		}
580
+				if (!ctx->entities[i][j])
581
+					continue;
461582 
462
-		for (i = 0; i < ctx->adev->num_rings; i++) {
463
-
464
-			if (ctx->adev->rings[i] == &ctx->adev->gfx.kiq.ring)
465
-				continue;
466
-
467
-			max_wait = drm_sched_entity_flush(&ctx->rings[i].entity,
468
-							  max_wait);
583
+				entity = &ctx->entities[i][j]->entity;
584
+				timeout = drm_sched_entity_flush(entity, timeout);
585
+			}
469586 		}
470587 	}
471588 	mutex_unlock(&mgr->lock);
589
+	return timeout;
472590 }
473591 
474592 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
475593 {
476594 	struct amdgpu_ctx *ctx;
477595 	struct idr *idp;
478
-	uint32_t id, i;
596
+	uint32_t id, i, j;
479597 
480598 	idp = &mgr->ctx_handles;
481599 
482600 	idr_for_each_entry(idp, ctx, id) {
601
+		if (kref_read(&ctx->refcount) != 1) {
602
+			DRM_ERROR("ctx %p is still alive\n", ctx);
603
+			continue;
604
+		}
483605 
484
-		if (!ctx->adev)
485
-			return;
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;
486609 
487
-		for (i = 0; i < ctx->adev->num_rings; i++) {
610
+				if (!ctx->entities[i][j])
611
+					continue;
488612 
489
-			if (ctx->adev->rings[i] == &ctx->adev->gfx.kiq.ring)
490
-				continue;
491
-
492
-			if (kref_read(&ctx->refcount) == 1)
493
-				drm_sched_entity_fini(&ctx->rings[i].entity);
494
-			else
495
-				DRM_ERROR("ctx %p is still alive\n", ctx);
613
+				entity = &ctx->entities[i][j]->entity;
614
+				drm_sched_entity_fini(entity);
615
+			}
496616 		}
497617 	}
498618 }