xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c (revision fab183d632628381b466a41479489541ac0e29a0)
1 /*
2  * Copyright 2008 Jerome Glisse.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22  * DEALINGS IN THE SOFTWARE.
23  *
24  * Authors:
25  *    Jerome Glisse <glisse@freedesktop.org>
26  */
27 
28 #include <linux/file.h>
29 #include <linux/pagemap.h>
30 #include <linux/sync_file.h>
31 #include <linux/dma-buf.h>
32 
33 #include <drm/amdgpu_drm.h>
34 #include <drm/drm_syncobj.h>
35 #include <drm/ttm/ttm_tt.h>
36 
37 #include "amdgpu_cs.h"
38 #include "amdgpu.h"
39 #include "amdgpu_trace.h"
40 #include "amdgpu_gmc.h"
41 #include "amdgpu_gem.h"
42 #include "amdgpu_ras.h"
43 #include "amdgpu_hmm.h"
44 
45 /*
46  * Maximum IB length (dwords) for rings whose emit_ib packet format
47  * documents a 20-bit size field.
48  */
49 #define AMDGPU_GFX_SDMA_IB_PACKET_SIZE_MAX_DW	0xFFFFF
50 #define AMDGPU_MM_IB_PACKET_SIZE_MAX_DW	0x7FFFF0
51 
amdgpu_cs_ib_packet_size_max_dw(enum amdgpu_ring_type type)52 static u32 amdgpu_cs_ib_packet_size_max_dw(enum amdgpu_ring_type type)
53 {
54 	switch (type) {
55 	case AMDGPU_RING_TYPE_GFX:
56 	case AMDGPU_RING_TYPE_COMPUTE:
57 	case AMDGPU_RING_TYPE_SDMA:
58 	case AMDGPU_RING_TYPE_VPE:
59 		return AMDGPU_GFX_SDMA_IB_PACKET_SIZE_MAX_DW;
60 	default:
61 		return AMDGPU_MM_IB_PACKET_SIZE_MAX_DW;
62 	}
63 }
64 
amdgpu_cs_parser_init(struct amdgpu_cs_parser * p,struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_cs * cs)65 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p,
66 				 struct amdgpu_device *adev,
67 				 struct drm_file *filp,
68 				 union drm_amdgpu_cs *cs)
69 {
70 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
71 
72 	if (cs->in.num_chunks == 0)
73 		return -EINVAL;
74 
75 	memset(p, 0, sizeof(*p));
76 	p->adev = adev;
77 	p->filp = filp;
78 
79 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
80 	if (!p->ctx)
81 		return -EINVAL;
82 
83 	amdgpu_sync_create(&p->sync);
84 	drm_exec_init(&p->exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
85 		      DRM_EXEC_IGNORE_DUPLICATES, 0);
86 	return 0;
87 }
88 
amdgpu_cs_job_idx(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_ib * chunk_ib)89 static int amdgpu_cs_job_idx(struct amdgpu_cs_parser *p,
90 			     struct drm_amdgpu_cs_chunk_ib *chunk_ib)
91 {
92 	struct drm_sched_entity *entity;
93 	unsigned int i;
94 	int r;
95 
96 	r = amdgpu_ctx_get_entity(p->ctx, chunk_ib->ip_type,
97 				  chunk_ib->ip_instance,
98 				  chunk_ib->ring, &entity);
99 	if (r)
100 		return r;
101 
102 	/* Check if we can add this IB to some existing job */
103 	for (i = 0; i < p->gang_size; ++i)
104 		if (p->entities[i] == entity)
105 			return i;
106 
107 	/* If not increase the gang size if possible */
108 	if (i == AMDGPU_CS_GANG_SIZE)
109 		return -EINVAL;
110 
111 	p->entities[i] = entity;
112 	p->gang_size = i + 1;
113 	return i;
114 }
115 
amdgpu_cs_p1_ib(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_ib * chunk_ib,unsigned int * num_ibs)116 static int amdgpu_cs_p1_ib(struct amdgpu_cs_parser *p,
117 			   struct drm_amdgpu_cs_chunk_ib *chunk_ib,
118 			   unsigned int *num_ibs)
119 {
120 	int r;
121 
122 	r = amdgpu_cs_job_idx(p, chunk_ib);
123 	if (r < 0)
124 		return r;
125 
126 	if (num_ibs[r] >= amdgpu_ring_max_ibs(chunk_ib->ip_type))
127 		return -EINVAL;
128 
129 	++(num_ibs[r]);
130 	p->gang_leader_idx = r;
131 	return 0;
132 }
133 
amdgpu_cs_p1_user_fence(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_fence * data,uint32_t * offset)134 static int amdgpu_cs_p1_user_fence(struct amdgpu_cs_parser *p,
135 				   struct drm_amdgpu_cs_chunk_fence *data,
136 				   uint32_t *offset)
137 {
138 	struct drm_gem_object *gobj;
139 	unsigned long size;
140 
141 	gobj = drm_gem_object_lookup(p->filp, data->handle);
142 	if (gobj == NULL)
143 		return -EINVAL;
144 
145 	p->uf_bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
146 	drm_gem_object_put(gobj);
147 
148 	size = amdgpu_bo_size(p->uf_bo);
149 	if (size != PAGE_SIZE || data->offset > (size - 8))
150 		return -EINVAL;
151 
152 	if (amdgpu_ttm_tt_get_usermm(p->uf_bo->tbo.ttm))
153 		return -EINVAL;
154 
155 	*offset = data->offset;
156 	return 0;
157 }
158 
amdgpu_cs_p1_bo_handles(struct amdgpu_cs_parser * p,struct drm_amdgpu_bo_list_in * data)159 static int amdgpu_cs_p1_bo_handles(struct amdgpu_cs_parser *p,
160 				   struct drm_amdgpu_bo_list_in *data)
161 {
162 	struct drm_amdgpu_bo_list_entry *info;
163 	int r;
164 
165 	r = amdgpu_bo_create_list_entry_array(data, &info);
166 	if (r)
167 		return r;
168 
169 	r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
170 				  &p->bo_list);
171 	if (r)
172 		goto error_free;
173 
174 	kvfree(info);
175 	return 0;
176 
177 error_free:
178 	kvfree(info);
179 
180 	return r;
181 }
182 
183 /* Copy the data from userspace and go over it the first time */
amdgpu_cs_pass1(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)184 static int amdgpu_cs_pass1(struct amdgpu_cs_parser *p,
185 			   union drm_amdgpu_cs *cs)
186 {
187 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
188 	unsigned int num_ibs[AMDGPU_CS_GANG_SIZE] = { };
189 	struct amdgpu_vm *vm = &fpriv->vm;
190 	uint64_t *chunk_array;
191 	uint32_t uf_offset = 0;
192 	size_t size;
193 	int ret;
194 	int i;
195 
196 	chunk_array = memdup_array_user(u64_to_user_ptr(cs->in.chunks),
197 					cs->in.num_chunks,
198 					sizeof(uint64_t));
199 	if (IS_ERR(chunk_array))
200 		return PTR_ERR(chunk_array);
201 
202 	p->nchunks = cs->in.num_chunks;
203 	p->chunks = kvmalloc_objs(struct amdgpu_cs_chunk, p->nchunks);
204 	if (!p->chunks) {
205 		ret = -ENOMEM;
206 		goto free_chunk;
207 	}
208 
209 	for (i = 0; i < p->nchunks; i++) {
210 		struct drm_amdgpu_cs_chunk __user *chunk_ptr = NULL;
211 		struct drm_amdgpu_cs_chunk user_chunk;
212 
213 		chunk_ptr = u64_to_user_ptr(chunk_array[i]);
214 		if (copy_from_user(&user_chunk, chunk_ptr,
215 				       sizeof(struct drm_amdgpu_cs_chunk))) {
216 			ret = -EFAULT;
217 			i--;
218 			goto free_partial_kdata;
219 		}
220 		p->chunks[i].chunk_id = user_chunk.chunk_id;
221 		p->chunks[i].length_dw = user_chunk.length_dw;
222 
223 		size = p->chunks[i].length_dw;
224 
225 		p->chunks[i].kdata = vmemdup_array_user(u64_to_user_ptr(user_chunk.chunk_data),
226 							size,
227 							sizeof(uint32_t));
228 		if (IS_ERR(p->chunks[i].kdata)) {
229 			ret = PTR_ERR(p->chunks[i].kdata);
230 			i--;
231 			goto free_partial_kdata;
232 		}
233 		size *= sizeof(uint32_t);
234 
235 		/* Assume the worst on the following checks */
236 		ret = -EINVAL;
237 		switch (p->chunks[i].chunk_id) {
238 		case AMDGPU_CHUNK_ID_IB:
239 			if (size < sizeof(struct drm_amdgpu_cs_chunk_ib))
240 				goto free_partial_kdata;
241 
242 			ret = amdgpu_cs_p1_ib(p, p->chunks[i].kdata, num_ibs);
243 			if (ret)
244 				goto free_partial_kdata;
245 			break;
246 
247 		case AMDGPU_CHUNK_ID_FENCE:
248 			if (size < sizeof(struct drm_amdgpu_cs_chunk_fence))
249 				goto free_partial_kdata;
250 
251 			/* Only a single user fence is allowed to simplify handling. */
252 			if (p->uf_bo)
253 				goto free_partial_kdata;
254 
255 			ret = amdgpu_cs_p1_user_fence(p, p->chunks[i].kdata,
256 						      &uf_offset);
257 			if (ret)
258 				goto free_partial_kdata;
259 			break;
260 
261 		case AMDGPU_CHUNK_ID_BO_HANDLES:
262 			if (size < sizeof(struct drm_amdgpu_bo_list_in))
263 				goto free_partial_kdata;
264 
265 			/* Only a single BO list is allowed to simplify handling. */
266 			if (p->bo_list)
267 				goto free_partial_kdata;
268 
269 			ret = amdgpu_cs_p1_bo_handles(p, p->chunks[i].kdata);
270 			if (ret)
271 				goto free_partial_kdata;
272 			break;
273 
274 		case AMDGPU_CHUNK_ID_CP_GFX_SHADOW:
275 			if (size < sizeof(struct drm_amdgpu_cs_chunk_cp_gfx_shadow))
276 				goto free_partial_kdata;
277 			break;
278 
279 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
280 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
281 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
282 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
283 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
284 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
285 			break;
286 
287 		default:
288 			goto free_partial_kdata;
289 		}
290 	}
291 
292 	if (!p->gang_size || (amdgpu_sriov_vf(p->adev) && p->gang_size > 1)) {
293 		ret = -EINVAL;
294 		goto free_all_kdata;
295 	}
296 
297 	for (i = 0; i < p->gang_size; ++i) {
298 		ret = amdgpu_job_alloc(p->adev, vm, p->entities[i], vm,
299 				       num_ibs[i], &p->jobs[i],
300 				       p->filp->client_id);
301 		if (ret)
302 			goto free_all_kdata;
303 		switch (p->adev->enforce_isolation[fpriv->xcp_id]) {
304 		case AMDGPU_ENFORCE_ISOLATION_DISABLE:
305 		default:
306 			p->jobs[i]->enforce_isolation = false;
307 			p->jobs[i]->run_cleaner_shader = false;
308 			break;
309 		case AMDGPU_ENFORCE_ISOLATION_ENABLE:
310 			p->jobs[i]->enforce_isolation = true;
311 			p->jobs[i]->run_cleaner_shader = true;
312 			break;
313 		case AMDGPU_ENFORCE_ISOLATION_ENABLE_LEGACY:
314 			p->jobs[i]->enforce_isolation = true;
315 			p->jobs[i]->run_cleaner_shader = false;
316 			break;
317 		case AMDGPU_ENFORCE_ISOLATION_NO_CLEANER_SHADER:
318 			p->jobs[i]->enforce_isolation = true;
319 			p->jobs[i]->run_cleaner_shader = false;
320 			break;
321 		}
322 	}
323 	p->gang_leader = p->jobs[p->gang_leader_idx];
324 
325 	if (p->ctx->generation != p->gang_leader->generation) {
326 		ret = -ECANCELED;
327 		goto free_all_kdata;
328 	}
329 
330 	if (p->uf_bo)
331 		p->gang_leader->uf_addr = uf_offset;
332 	kvfree(chunk_array);
333 
334 	/* Use this opportunity to fill in task info for the vm */
335 	amdgpu_vm_set_task_info(vm);
336 
337 	return 0;
338 
339 free_all_kdata:
340 	i = p->nchunks - 1;
341 free_partial_kdata:
342 	for (; i >= 0; i--)
343 		kvfree(p->chunks[i].kdata);
344 	kvfree(p->chunks);
345 	p->chunks = NULL;
346 	p->nchunks = 0;
347 free_chunk:
348 	kvfree(chunk_array);
349 
350 	return ret;
351 }
352 
amdgpu_cs_p2_ib(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk,unsigned int * ce_preempt,unsigned int * de_preempt)353 static int amdgpu_cs_p2_ib(struct amdgpu_cs_parser *p,
354 			   struct amdgpu_cs_chunk *chunk,
355 			   unsigned int *ce_preempt,
356 			   unsigned int *de_preempt)
357 {
358 	struct drm_amdgpu_cs_chunk_ib *chunk_ib = chunk->kdata;
359 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
360 	struct amdgpu_vm *vm = &fpriv->vm;
361 	struct amdgpu_ring *ring;
362 	struct amdgpu_job *job;
363 	struct amdgpu_ib *ib;
364 	int r;
365 
366 	r = amdgpu_cs_job_idx(p, chunk_ib);
367 	if (r < 0)
368 		return r;
369 
370 	job = p->jobs[r];
371 	ring = amdgpu_job_ring(job);
372 
373 	/* submissions to kernel queues are disabled */
374 	if (ring->no_user_submission)
375 		return -EINVAL;
376 
377 	/* MM engine doesn't support user fences */
378 	if (p->uf_bo && ring->funcs->no_user_fence)
379 		return -EINVAL;
380 
381 	if (!p->adev->debug_enable_ce_cs &&
382 	    chunk_ib->flags & AMDGPU_IB_FLAG_CE) {
383 		dev_err_ratelimited(p->adev->dev, "CE CS is blocked, use debug=0x400 to override\n");
384 		return -EINVAL;
385 	}
386 
387 	if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
388 	    chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
389 		if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
390 			(*ce_preempt)++;
391 		else
392 			(*de_preempt)++;
393 
394 		/* Each GFX command submit allows only 1 IB max
395 		 * preemptible for CE & DE */
396 		if (*ce_preempt > 1 || *de_preempt > 1)
397 			return -EINVAL;
398 	}
399 
400 	if (chunk_ib->ib_bytes / 4 >
401 	    amdgpu_cs_ib_packet_size_max_dw(ring->funcs->type))
402 		return -EINVAL;
403 
404 	ib = &job->ibs[job->num_ibs++];
405 
406 	if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
407 		job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
408 
409 	r =  amdgpu_ib_get(p->adev, vm, ring->funcs->parse_cs ?
410 			   chunk_ib->ib_bytes : 0,
411 			   AMDGPU_IB_POOL_DELAYED, ib);
412 	if (r) {
413 		drm_err(adev_to_drm(p->adev), "Failed to get ib !\n");
414 		return r;
415 	}
416 
417 	ib->gpu_addr = chunk_ib->va_start;
418 	ib->length_dw = chunk_ib->ib_bytes / 4;
419 	ib->flags = chunk_ib->flags;
420 	return 0;
421 }
422 
amdgpu_cs_p2_dependencies(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)423 static int amdgpu_cs_p2_dependencies(struct amdgpu_cs_parser *p,
424 				     struct amdgpu_cs_chunk *chunk)
425 {
426 	struct drm_amdgpu_cs_chunk_dep *deps = chunk->kdata;
427 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
428 	unsigned int num_deps;
429 	int i, r;
430 
431 	num_deps = chunk->length_dw * 4 /
432 		sizeof(struct drm_amdgpu_cs_chunk_dep);
433 
434 	for (i = 0; i < num_deps; ++i) {
435 		struct amdgpu_ctx *ctx;
436 		struct drm_sched_entity *entity;
437 		struct dma_fence *fence;
438 
439 		ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
440 		if (ctx == NULL)
441 			return -EINVAL;
442 
443 		r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
444 					  deps[i].ip_instance,
445 					  deps[i].ring, &entity);
446 		if (r) {
447 			amdgpu_ctx_put(ctx);
448 			return r;
449 		}
450 
451 		fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
452 		amdgpu_ctx_put(ctx);
453 
454 		if (IS_ERR(fence))
455 			return PTR_ERR(fence);
456 		else if (!fence)
457 			continue;
458 
459 		if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
460 			struct drm_sched_fence *s_fence;
461 			struct dma_fence *old = fence;
462 
463 			s_fence = to_drm_sched_fence(fence);
464 			fence = dma_fence_get(&s_fence->scheduled);
465 			dma_fence_put(old);
466 		}
467 
468 		r = amdgpu_sync_fence(&p->sync, fence, GFP_KERNEL);
469 		dma_fence_put(fence);
470 		if (r)
471 			return r;
472 	}
473 	return 0;
474 }
475 
amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser * p,uint32_t handle,u64 point,u64 flags)476 static int amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser *p,
477 					 uint32_t handle, u64 point,
478 					 u64 flags)
479 {
480 	struct dma_fence *fence;
481 	int r;
482 
483 	r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
484 	if (r) {
485 		drm_err(adev_to_drm(p->adev), "syncobj %u failed to find fence @ %llu (%d)!\n",
486 			  handle, point, r);
487 		return r;
488 	}
489 
490 	r = amdgpu_sync_fence(&p->sync, fence, GFP_KERNEL);
491 	dma_fence_put(fence);
492 	return r;
493 }
494 
amdgpu_cs_p2_syncobj_in(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)495 static int amdgpu_cs_p2_syncobj_in(struct amdgpu_cs_parser *p,
496 				   struct amdgpu_cs_chunk *chunk)
497 {
498 	struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
499 	unsigned int num_deps;
500 	int i, r;
501 
502 	num_deps = chunk->length_dw * 4 /
503 		sizeof(struct drm_amdgpu_cs_chunk_sem);
504 	for (i = 0; i < num_deps; ++i) {
505 		r = amdgpu_syncobj_lookup_and_add(p, deps[i].handle, 0, 0);
506 		if (r)
507 			return r;
508 	}
509 
510 	return 0;
511 }
512 
amdgpu_cs_p2_syncobj_timeline_wait(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)513 static int amdgpu_cs_p2_syncobj_timeline_wait(struct amdgpu_cs_parser *p,
514 					      struct amdgpu_cs_chunk *chunk)
515 {
516 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
517 	unsigned int num_deps;
518 	int i, r;
519 
520 	num_deps = chunk->length_dw * 4 /
521 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
522 	for (i = 0; i < num_deps; ++i) {
523 		r = amdgpu_syncobj_lookup_and_add(p, syncobj_deps[i].handle,
524 						  syncobj_deps[i].point,
525 						  syncobj_deps[i].flags);
526 		if (r)
527 			return r;
528 	}
529 
530 	return 0;
531 }
532 
amdgpu_cs_p2_syncobj_out(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)533 static int amdgpu_cs_p2_syncobj_out(struct amdgpu_cs_parser *p,
534 				    struct amdgpu_cs_chunk *chunk)
535 {
536 	struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
537 	unsigned int num_deps;
538 	int i;
539 
540 	num_deps = chunk->length_dw * 4 /
541 		sizeof(struct drm_amdgpu_cs_chunk_sem);
542 
543 	if (p->post_deps)
544 		return -EINVAL;
545 
546 	p->post_deps = kmalloc_objs(*p->post_deps, num_deps);
547 	p->num_post_deps = 0;
548 
549 	if (!p->post_deps)
550 		return -ENOMEM;
551 
552 
553 	for (i = 0; i < num_deps; ++i) {
554 		p->post_deps[i].syncobj =
555 			drm_syncobj_find(p->filp, deps[i].handle);
556 		if (!p->post_deps[i].syncobj)
557 			return -EINVAL;
558 		p->post_deps[i].chain = NULL;
559 		p->post_deps[i].point = 0;
560 		p->num_post_deps++;
561 	}
562 
563 	return 0;
564 }
565 
amdgpu_cs_p2_syncobj_timeline_signal(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)566 static int amdgpu_cs_p2_syncobj_timeline_signal(struct amdgpu_cs_parser *p,
567 						struct amdgpu_cs_chunk *chunk)
568 {
569 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
570 	unsigned int num_deps;
571 	int i;
572 
573 	num_deps = chunk->length_dw * 4 /
574 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
575 
576 	if (p->post_deps)
577 		return -EINVAL;
578 
579 	p->post_deps = kmalloc_objs(*p->post_deps, num_deps);
580 	p->num_post_deps = 0;
581 
582 	if (!p->post_deps)
583 		return -ENOMEM;
584 
585 	for (i = 0; i < num_deps; ++i) {
586 		struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
587 
588 		dep->chain = NULL;
589 		if (syncobj_deps[i].point) {
590 			dep->chain = dma_fence_chain_alloc();
591 			if (!dep->chain)
592 				return -ENOMEM;
593 		}
594 
595 		dep->syncobj = drm_syncobj_find(p->filp,
596 						syncobj_deps[i].handle);
597 		if (!dep->syncobj) {
598 			dma_fence_chain_free(dep->chain);
599 			return -EINVAL;
600 		}
601 		dep->point = syncobj_deps[i].point;
602 		p->num_post_deps++;
603 	}
604 
605 	return 0;
606 }
607 
amdgpu_cs_p2_shadow(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)608 static int amdgpu_cs_p2_shadow(struct amdgpu_cs_parser *p,
609 			       struct amdgpu_cs_chunk *chunk)
610 {
611 	struct drm_amdgpu_cs_chunk_cp_gfx_shadow *shadow = chunk->kdata;
612 	int i;
613 
614 	if (shadow->flags & ~AMDGPU_CS_CHUNK_CP_GFX_SHADOW_FLAGS_INIT_SHADOW)
615 		return -EINVAL;
616 
617 	for (i = 0; i < p->gang_size; ++i) {
618 		p->jobs[i]->shadow_va = shadow->shadow_va;
619 		p->jobs[i]->csa_va = shadow->csa_va;
620 		p->jobs[i]->gds_va = shadow->gds_va;
621 		p->jobs[i]->init_shadow =
622 			shadow->flags & AMDGPU_CS_CHUNK_CP_GFX_SHADOW_FLAGS_INIT_SHADOW;
623 	}
624 
625 	return 0;
626 }
627 
amdgpu_cs_pass2(struct amdgpu_cs_parser * p)628 static int amdgpu_cs_pass2(struct amdgpu_cs_parser *p)
629 {
630 	unsigned int ce_preempt = 0, de_preempt = 0;
631 	int i, r;
632 
633 	for (i = 0; i < p->nchunks; ++i) {
634 		struct amdgpu_cs_chunk *chunk;
635 
636 		chunk = &p->chunks[i];
637 
638 		switch (chunk->chunk_id) {
639 		case AMDGPU_CHUNK_ID_IB:
640 			r = amdgpu_cs_p2_ib(p, chunk, &ce_preempt, &de_preempt);
641 			if (r)
642 				return r;
643 			break;
644 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
645 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
646 			r = amdgpu_cs_p2_dependencies(p, chunk);
647 			if (r)
648 				return r;
649 			break;
650 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
651 			r = amdgpu_cs_p2_syncobj_in(p, chunk);
652 			if (r)
653 				return r;
654 			break;
655 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
656 			r = amdgpu_cs_p2_syncobj_out(p, chunk);
657 			if (r)
658 				return r;
659 			break;
660 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
661 			r = amdgpu_cs_p2_syncobj_timeline_wait(p, chunk);
662 			if (r)
663 				return r;
664 			break;
665 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
666 			r = amdgpu_cs_p2_syncobj_timeline_signal(p, chunk);
667 			if (r)
668 				return r;
669 			break;
670 		case AMDGPU_CHUNK_ID_CP_GFX_SHADOW:
671 			r = amdgpu_cs_p2_shadow(p, chunk);
672 			if (r)
673 				return r;
674 			break;
675 		}
676 	}
677 
678 	return 0;
679 }
680 
681 /* Convert microseconds to bytes. */
us_to_bytes(struct amdgpu_device * adev,s64 us)682 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
683 {
684 	if (us <= 0 || !adev->mm_stats.log2_max_MBps)
685 		return 0;
686 
687 	/* Since accum_us is incremented by a million per second, just
688 	 * multiply it by the number of MB/s to get the number of bytes.
689 	 */
690 	return us << adev->mm_stats.log2_max_MBps;
691 }
692 
bytes_to_us(struct amdgpu_device * adev,u64 bytes)693 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
694 {
695 	if (!adev->mm_stats.log2_max_MBps)
696 		return 0;
697 
698 	return bytes >> adev->mm_stats.log2_max_MBps;
699 }
700 
701 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
702  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
703  * which means it can go over the threshold once. If that happens, the driver
704  * will be in debt and no other buffer migrations can be done until that debt
705  * is repaid.
706  *
707  * This approach allows moving a buffer of any size (it's important to allow
708  * that).
709  *
710  * The currency is simply time in microseconds and it increases as the clock
711  * ticks. The accumulated microseconds (us) are converted to bytes and
712  * returned.
713  */
amdgpu_cs_get_threshold_for_moves(struct amdgpu_device * adev,u64 * max_bytes,u64 * max_vis_bytes)714 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
715 					      u64 *max_bytes,
716 					      u64 *max_vis_bytes)
717 {
718 	s64 time_us, increment_us;
719 	u64 free_vram, total_vram, used_vram;
720 	/* Allow a maximum of 200 accumulated ms. This is basically per-IB
721 	 * throttling.
722 	 *
723 	 * It means that in order to get full max MBps, at least 5 IBs per
724 	 * second must be submitted and not more than 200ms apart from each
725 	 * other.
726 	 */
727 	const s64 us_upper_bound = 200000;
728 
729 	if ((!adev->mm_stats.log2_max_MBps) || !ttm_resource_manager_used(&adev->mman.vram_mgr.manager)) {
730 		*max_bytes = 0;
731 		*max_vis_bytes = 0;
732 		return;
733 	}
734 
735 	total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
736 	used_vram = ttm_resource_manager_usage(&adev->mman.vram_mgr.manager);
737 	free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
738 
739 	spin_lock(&adev->mm_stats.lock);
740 
741 	/* Increase the amount of accumulated us. */
742 	time_us = ktime_to_us(ktime_get());
743 	increment_us = time_us - adev->mm_stats.last_update_us;
744 	adev->mm_stats.last_update_us = time_us;
745 	adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
746 				      us_upper_bound);
747 
748 	/* This prevents the short period of low performance when the VRAM
749 	 * usage is low and the driver is in debt or doesn't have enough
750 	 * accumulated us to fill VRAM quickly.
751 	 *
752 	 * The situation can occur in these cases:
753 	 * - a lot of VRAM is freed by userspace
754 	 * - the presence of a big buffer causes a lot of evictions
755 	 *   (solution: split buffers into smaller ones)
756 	 *
757 	 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
758 	 * accum_us to a positive number.
759 	 */
760 	if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
761 		s64 min_us;
762 
763 		/* Be more aggressive on dGPUs. Try to fill a portion of free
764 		 * VRAM now.
765 		 */
766 		if (!(adev->flags & AMD_IS_APU))
767 			min_us = bytes_to_us(adev, free_vram / 4);
768 		else
769 			min_us = 0; /* Reset accum_us on APUs. */
770 
771 		adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
772 	}
773 
774 	/* This is set to 0 if the driver is in debt to disallow (optional)
775 	 * buffer moves.
776 	 */
777 	*max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
778 
779 	/* Do the same for visible VRAM if half of it is free */
780 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
781 		u64 total_vis_vram = adev->gmc.visible_vram_size;
782 		u64 used_vis_vram =
783 		  amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr);
784 
785 		if (used_vis_vram < total_vis_vram) {
786 			u64 free_vis_vram = total_vis_vram - used_vis_vram;
787 
788 			adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
789 							  increment_us, us_upper_bound);
790 
791 			if (free_vis_vram >= total_vis_vram / 2)
792 				adev->mm_stats.accum_us_vis =
793 					max(bytes_to_us(adev, free_vis_vram / 2),
794 					    adev->mm_stats.accum_us_vis);
795 		}
796 
797 		*max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
798 	} else {
799 		*max_vis_bytes = 0;
800 	}
801 
802 	spin_unlock(&adev->mm_stats.lock);
803 }
804 
805 /* Report how many bytes have really been moved for the last command
806  * submission. This can result in a debt that can stop buffer migrations
807  * temporarily.
808  */
amdgpu_cs_report_moved_bytes(struct amdgpu_device * adev,u64 num_bytes,u64 num_vis_bytes)809 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
810 				  u64 num_vis_bytes)
811 {
812 	spin_lock(&adev->mm_stats.lock);
813 	adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
814 	adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
815 	spin_unlock(&adev->mm_stats.lock);
816 }
817 
amdgpu_cs_bo_validate(void * param,struct amdgpu_bo * bo)818 static int amdgpu_cs_bo_validate(void *param, struct amdgpu_bo *bo)
819 {
820 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
821 	struct amdgpu_cs_parser *p = param;
822 	struct ttm_operation_ctx ctx = {
823 		.interruptible = true,
824 		.no_wait_gpu = false,
825 		.resv = bo->tbo.base.resv
826 	};
827 	uint32_t domain;
828 	int r;
829 
830 	if (bo->tbo.pin_count)
831 		return 0;
832 
833 	/* Don't move this buffer if we have depleted our allowance
834 	 * to move it. Don't move anything if the threshold is zero.
835 	 */
836 	if (p->bytes_moved < p->bytes_moved_threshold &&
837 	    (!bo->tbo.base.dma_buf ||
838 	    list_empty(&bo->tbo.base.dma_buf->attachments))) {
839 		if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
840 		    (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
841 			/* And don't move a CPU_ACCESS_REQUIRED BO to limited
842 			 * visible VRAM if we've depleted our allowance to do
843 			 * that.
844 			 */
845 			if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
846 				domain = bo->preferred_domains;
847 			else
848 				domain = bo->allowed_domains;
849 		} else {
850 			domain = bo->preferred_domains;
851 		}
852 	} else {
853 		domain = bo->allowed_domains;
854 	}
855 
856 retry:
857 	amdgpu_bo_placement_from_domain(bo, domain);
858 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
859 
860 	p->bytes_moved += ctx.bytes_moved;
861 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
862 	    amdgpu_res_cpu_visible(adev, bo->tbo.resource))
863 		p->bytes_moved_vis += ctx.bytes_moved;
864 
865 	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
866 		domain = bo->allowed_domains;
867 		goto retry;
868 	}
869 
870 	return r;
871 }
872 
amdgpu_cs_parser_bos(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)873 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
874 				union drm_amdgpu_cs *cs)
875 {
876 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
877 	struct ttm_operation_ctx ctx = { true, false };
878 	struct amdgpu_vm *vm = &fpriv->vm;
879 	struct amdgpu_bo_list_entry *e;
880 	struct drm_gem_object *obj;
881 	unsigned int i;
882 	int r;
883 
884 	/* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
885 	if (cs->in.bo_list_handle) {
886 		if (p->bo_list)
887 			return -EINVAL;
888 
889 		r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
890 				       &p->bo_list);
891 		if (r)
892 			return r;
893 	} else if (!p->bo_list) {
894 		/* Create a empty bo_list when no handle is provided */
895 		r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
896 					  &p->bo_list);
897 		if (r)
898 			return r;
899 	}
900 
901 	mutex_lock(&p->bo_list->bo_list_mutex);
902 
903 	/* Get userptr backing pages. If pages are updated after registered
904 	 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
905 	 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
906 	 */
907 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
908 		bool userpage_invalidated = false;
909 		struct amdgpu_bo *bo = e->bo;
910 
911 		e->range = amdgpu_hmm_range_alloc(NULL);
912 		if (unlikely(!e->range)) {
913 			r = -ENOMEM;
914 			goto out_free_user_pages;
915 		}
916 
917 		r = amdgpu_ttm_tt_get_user_pages(bo, e->range);
918 		if (r)
919 			goto out_free_user_pages;
920 
921 		for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
922 			if (bo->tbo.ttm->pages[i] !=
923 				hmm_pfn_to_page(e->range->hmm_range.hmm_pfns[i])) {
924 				userpage_invalidated = true;
925 				break;
926 			}
927 		}
928 		e->user_invalidated = userpage_invalidated;
929 	}
930 
931 	drm_exec_until_all_locked(&p->exec) {
932 		r = amdgpu_vm_lock_pd(&fpriv->vm, &p->exec, 1 + p->gang_size);
933 		drm_exec_retry_on_contention(&p->exec);
934 		if (unlikely(r))
935 			goto out_free_user_pages;
936 
937 		amdgpu_bo_list_for_each_entry(e, p->bo_list) {
938 			r = drm_exec_prepare_obj(&p->exec, &e->bo->tbo.base,
939 						 TTM_NUM_MOVE_FENCES + p->gang_size);
940 			drm_exec_retry_on_contention(&p->exec);
941 			if (unlikely(r))
942 				goto out_free_user_pages;
943 
944 			e->bo_va = amdgpu_vm_bo_find(vm, e->bo);
945 		}
946 
947 		if (p->uf_bo) {
948 			r = drm_exec_prepare_obj(&p->exec, &p->uf_bo->tbo.base,
949 						 TTM_NUM_MOVE_FENCES + p->gang_size);
950 			drm_exec_retry_on_contention(&p->exec);
951 			if (unlikely(r))
952 				goto out_free_user_pages;
953 		}
954 	}
955 
956 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
957 		struct mm_struct *usermm;
958 
959 		usermm = amdgpu_ttm_tt_get_usermm(e->bo->tbo.ttm);
960 		if (usermm && usermm != current->mm) {
961 			r = -EPERM;
962 			goto out_free_user_pages;
963 		}
964 
965 		if (amdgpu_ttm_tt_is_userptr(e->bo->tbo.ttm) &&
966 		    e->user_invalidated) {
967 			amdgpu_bo_placement_from_domain(e->bo,
968 							AMDGPU_GEM_DOMAIN_CPU);
969 			r = ttm_bo_validate(&e->bo->tbo, &e->bo->placement,
970 					    &ctx);
971 			if (r)
972 				goto out_free_user_pages;
973 
974 			amdgpu_ttm_tt_set_user_pages(e->bo->tbo.ttm,
975 						     e->range);
976 		}
977 	}
978 
979 	amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
980 					  &p->bytes_moved_vis_threshold);
981 	p->bytes_moved = 0;
982 	p->bytes_moved_vis = 0;
983 
984 	r = amdgpu_vm_validate(p->adev, &fpriv->vm, NULL,
985 			       amdgpu_cs_bo_validate, p);
986 	if (r) {
987 		drm_err(adev_to_drm(p->adev), "amdgpu_vm_validate() failed.\n");
988 		goto out_free_user_pages;
989 	}
990 
991 	drm_exec_for_each_locked_object(&p->exec, obj) {
992 		r = amdgpu_cs_bo_validate(p, gem_to_amdgpu_bo(obj));
993 		if (unlikely(r))
994 			goto out_free_user_pages;
995 	}
996 
997 	if (p->uf_bo) {
998 		r = amdgpu_ttm_alloc_gart(&p->uf_bo->tbo);
999 		if (unlikely(r))
1000 			goto out_free_user_pages;
1001 
1002 		p->gang_leader->uf_addr += amdgpu_bo_gpu_offset(p->uf_bo);
1003 	}
1004 
1005 	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
1006 				     p->bytes_moved_vis);
1007 
1008 	for (i = 0; i < p->gang_size; ++i)
1009 		amdgpu_job_set_resources(p->jobs[i], p->bo_list->gds_obj,
1010 					 p->bo_list->gws_obj,
1011 					 p->bo_list->oa_obj);
1012 	return 0;
1013 
1014 out_free_user_pages:
1015 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1016 		amdgpu_hmm_range_free(e->range);
1017 		e->range = NULL;
1018 	}
1019 	mutex_unlock(&p->bo_list->bo_list_mutex);
1020 	return r;
1021 }
1022 
trace_amdgpu_cs_ibs(struct amdgpu_cs_parser * p)1023 static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *p)
1024 {
1025 	int i, j;
1026 
1027 	if (!trace_amdgpu_cs_enabled())
1028 		return;
1029 
1030 	for (i = 0; i < p->gang_size; ++i) {
1031 		struct amdgpu_job *job = p->jobs[i];
1032 
1033 		for (j = 0; j < job->num_ibs; ++j)
1034 			trace_amdgpu_cs(p, job, &job->ibs[j]);
1035 	}
1036 }
1037 
amdgpu_cs_patch_ibs(struct amdgpu_cs_parser * p,struct amdgpu_job * job)1038 static int amdgpu_cs_patch_ibs(struct amdgpu_cs_parser *p,
1039 			       struct amdgpu_job *job)
1040 {
1041 	struct amdgpu_ring *ring = amdgpu_job_ring(job);
1042 	struct amdgpu_device *adev = ring->adev;
1043 	unsigned int i;
1044 	int r;
1045 
1046 	/* Only for UVD/VCE VM emulation */
1047 	if (!ring->funcs->parse_cs && !ring->funcs->patch_cs_in_place)
1048 		return 0;
1049 
1050 	for (i = 0; i < job->num_ibs; ++i) {
1051 		struct amdgpu_ib *ib = &job->ibs[i];
1052 		struct amdgpu_bo_va_mapping *m;
1053 		struct amdgpu_bo *aobj;
1054 		uint64_t va_start;
1055 		uint8_t *kptr;
1056 
1057 		va_start = ib->gpu_addr & AMDGPU_GMC_HOLE_MASK;
1058 		r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
1059 		if (r) {
1060 			drm_err(adev_to_drm(p->adev), "IB va_start is invalid\n");
1061 			return r;
1062 		}
1063 
1064 		if ((va_start + ib->length_dw * 4) >
1065 		    (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
1066 			drm_err(adev_to_drm(p->adev), "IB va_start+ib_bytes is invalid\n");
1067 			return -EINVAL;
1068 		}
1069 
1070 		/* the IB should be reserved at this point */
1071 		r = amdgpu_bo_kmap(aobj, (void **)&kptr);
1072 		if (r)
1073 			return r;
1074 
1075 		kptr += va_start - (m->start * AMDGPU_GPU_PAGE_SIZE);
1076 
1077 		if (ring->funcs->parse_cs) {
1078 			memcpy(ib->ptr, kptr, ib->length_dw * 4);
1079 			amdgpu_bo_kunmap(aobj);
1080 
1081 			r = amdgpu_ring_parse_cs(ring, p, job, ib);
1082 			if (r)
1083 				return r;
1084 
1085 			if (ib->sa_bo)
1086 				ib->gpu_addr =  amdgpu_sa_bo_gpu_addr(ib->sa_bo);
1087 		} else {
1088 			ib->ptr = (uint32_t *)kptr;
1089 			r = amdgpu_ring_patch_cs_in_place(ring, p, job, ib);
1090 			amdgpu_bo_kunmap(aobj);
1091 			if (r)
1092 				return r;
1093 		}
1094 	}
1095 
1096 	return 0;
1097 }
1098 
amdgpu_cs_patch_jobs(struct amdgpu_cs_parser * p)1099 static int amdgpu_cs_patch_jobs(struct amdgpu_cs_parser *p)
1100 {
1101 	unsigned int i;
1102 	int r;
1103 
1104 	for (i = 0; i < p->gang_size; ++i) {
1105 		r = amdgpu_cs_patch_ibs(p, p->jobs[i]);
1106 		if (r)
1107 			return r;
1108 	}
1109 	return 0;
1110 }
1111 
amdgpu_cs_vm_handling(struct amdgpu_cs_parser * p)1112 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
1113 {
1114 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1115 	struct amdgpu_job *job = p->gang_leader;
1116 	struct amdgpu_device *adev = p->adev;
1117 	struct amdgpu_vm *vm = &fpriv->vm;
1118 	struct amdgpu_bo_list_entry *e;
1119 	struct amdgpu_bo_va *bo_va;
1120 	unsigned int i;
1121 	int r;
1122 
1123 	/*
1124 	 * We can't use gang submit on with reserved VMIDs when the VM changes
1125 	 * can't be invalidated by more than one engine at the same time.
1126 	 */
1127 	if (p->gang_size > 1 && !adev->vm_manager.concurrent_flush) {
1128 		for (i = 0; i < p->gang_size; ++i) {
1129 			struct drm_sched_entity *entity = p->entities[i];
1130 			struct drm_gpu_scheduler *sched = entity->rq->sched;
1131 			struct amdgpu_ring *ring = to_amdgpu_ring(sched);
1132 
1133 			if (amdgpu_vmid_uses_reserved(vm, ring->vm_hub))
1134 				return -EINVAL;
1135 		}
1136 	}
1137 
1138 	if (!amdgpu_vm_ready(vm))
1139 		return -EINVAL;
1140 
1141 	r = amdgpu_vm_clear_freed(adev, vm, NULL);
1142 	if (r)
1143 		return r;
1144 
1145 	r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
1146 	if (r)
1147 		return r;
1148 
1149 	r = amdgpu_sync_fence(&p->sync, fpriv->prt_va->last_pt_update,
1150 			      GFP_KERNEL);
1151 	if (r)
1152 		return r;
1153 
1154 	if (fpriv->csa_va) {
1155 		bo_va = fpriv->csa_va;
1156 		BUG_ON(!bo_va);
1157 		r = amdgpu_vm_bo_update(adev, bo_va, false);
1158 		if (r)
1159 			return r;
1160 
1161 		r = amdgpu_sync_fence(&p->sync, bo_va->last_pt_update,
1162 				      GFP_KERNEL);
1163 		if (r)
1164 			return r;
1165 	}
1166 
1167 	/* FIXME: In theory this loop shouldn't be needed any more when
1168 	 * amdgpu_vm_handle_moved handles all moved BOs that are reserved
1169 	 * with p->ticket. But removing it caused test regressions, so I'm
1170 	 * leaving it here for now.
1171 	 */
1172 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1173 		bo_va = e->bo_va;
1174 		if (bo_va == NULL)
1175 			continue;
1176 
1177 		r = amdgpu_vm_bo_update(adev, bo_va, false);
1178 		if (r)
1179 			return r;
1180 
1181 		r = amdgpu_sync_fence(&p->sync, bo_va->last_pt_update,
1182 				      GFP_KERNEL);
1183 		if (r)
1184 			return r;
1185 	}
1186 
1187 	r = amdgpu_vm_handle_moved(adev, vm, drm_exec_ticket(&p->exec));
1188 	if (r)
1189 		return r;
1190 
1191 	r = amdgpu_vm_update_pdes(adev, vm, false);
1192 	if (r)
1193 		return r;
1194 
1195 	r = amdgpu_sync_fence(&p->sync, vm->last_update, GFP_KERNEL);
1196 	if (r)
1197 		return r;
1198 
1199 	for (i = 0; i < p->gang_size; ++i) {
1200 		job = p->jobs[i];
1201 
1202 		if (!job->vm)
1203 			continue;
1204 
1205 		job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
1206 	}
1207 
1208 	return 0;
1209 }
1210 
amdgpu_cs_sync_rings(struct amdgpu_cs_parser * p)1211 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
1212 {
1213 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1214 	struct drm_gpu_scheduler *sched;
1215 	struct drm_gem_object *obj;
1216 	struct dma_fence *fence;
1217 	unsigned int i;
1218 	int r;
1219 
1220 	r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entities[p->gang_leader_idx]);
1221 	if (r) {
1222 		if (r != -ERESTARTSYS)
1223 			drm_err(adev_to_drm(p->adev), "amdgpu_ctx_wait_prev_fence failed.\n");
1224 		return r;
1225 	}
1226 
1227 	drm_exec_for_each_locked_object(&p->exec, obj) {
1228 		struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
1229 
1230 		struct dma_resv *resv = bo->tbo.base.resv;
1231 		enum amdgpu_sync_mode sync_mode;
1232 
1233 		sync_mode = amdgpu_bo_explicit_sync(bo) ?
1234 			AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
1235 		r = amdgpu_sync_resv(p->adev, &p->sync, resv, sync_mode,
1236 				     &fpriv->vm);
1237 		if (r)
1238 			return r;
1239 	}
1240 
1241 	for (i = 0; i < p->gang_size; ++i) {
1242 		r = amdgpu_sync_push_to_job(&p->sync, p->jobs[i]);
1243 		if (r)
1244 			return r;
1245 	}
1246 
1247 	sched = p->gang_leader->base.entity->rq->sched;
1248 	while ((fence = amdgpu_sync_get_fence(&p->sync))) {
1249 		struct drm_sched_fence *s_fence = to_drm_sched_fence(fence);
1250 
1251 		/*
1252 		 * When we have an dependency it might be necessary to insert a
1253 		 * pipeline sync to make sure that all caches etc are flushed and the
1254 		 * next job actually sees the results from the previous one
1255 		 * before we start executing on the same scheduler ring.
1256 		 */
1257 		if (!s_fence || s_fence->sched != sched) {
1258 			dma_fence_put(fence);
1259 			continue;
1260 		}
1261 
1262 		r = amdgpu_sync_fence(&p->gang_leader->explicit_sync, fence,
1263 				      GFP_KERNEL);
1264 		dma_fence_put(fence);
1265 		if (r)
1266 			return r;
1267 	}
1268 	return 0;
1269 }
1270 
amdgpu_cs_post_dependencies(struct amdgpu_cs_parser * p)1271 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1272 {
1273 	int i;
1274 
1275 	for (i = 0; i < p->num_post_deps; ++i) {
1276 		if (p->post_deps[i].chain && p->post_deps[i].point) {
1277 			drm_syncobj_add_point(p->post_deps[i].syncobj,
1278 					      p->post_deps[i].chain,
1279 					      p->fence, p->post_deps[i].point);
1280 			p->post_deps[i].chain = NULL;
1281 		} else {
1282 			drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1283 						  p->fence);
1284 		}
1285 	}
1286 }
1287 
amdgpu_cs_submit(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)1288 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1289 			    union drm_amdgpu_cs *cs)
1290 {
1291 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1292 	struct amdgpu_job *leader = p->gang_leader;
1293 	struct amdgpu_vm *vm = &fpriv->vm;
1294 	struct amdgpu_bo_list_entry *e;
1295 	struct drm_gem_object *gobj;
1296 	unsigned int i;
1297 	uint64_t seq;
1298 	int r;
1299 
1300 	for (i = 0; i < p->gang_size; ++i)
1301 		drm_sched_job_arm(&p->jobs[i]->base);
1302 
1303 	for (i = 0; i < p->gang_size; ++i) {
1304 		struct dma_fence *fence;
1305 
1306 		if (p->jobs[i] == leader)
1307 			continue;
1308 
1309 		fence = &p->jobs[i]->base.s_fence->scheduled;
1310 		dma_fence_get(fence);
1311 		r = drm_sched_job_add_dependency(&leader->base, fence);
1312 		if (r) {
1313 			dma_fence_put(fence);
1314 			return r;
1315 		}
1316 	}
1317 
1318 	if (p->gang_size > 1) {
1319 		for (i = 0; i < p->gang_size; ++i)
1320 			amdgpu_job_set_gang_leader(p->jobs[i], leader);
1321 	}
1322 
1323 	/* No memory allocation is allowed while holding the notifier lock.
1324 	 * The lock is held until amdgpu_cs_submit is finished and fence is
1325 	 * added to BOs.
1326 	 */
1327 	mutex_lock(&p->adev->notifier_lock);
1328 
1329 	/* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1330 	 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1331 	 */
1332 	r = 0;
1333 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1334 		r |= !amdgpu_hmm_range_valid(e->range);
1335 		amdgpu_hmm_range_free(e->range);
1336 		e->range = NULL;
1337 	}
1338 
1339 	if (r || !list_empty(&vm->individual.needs_update)) {
1340 		r = -EAGAIN;
1341 		mutex_unlock(&p->adev->notifier_lock);
1342 		return r;
1343 	}
1344 
1345 	p->fence = dma_fence_get(&leader->base.s_fence->finished);
1346 	drm_exec_for_each_locked_object(&p->exec, gobj) {
1347 
1348 		ttm_bo_move_to_lru_tail_unlocked(&gem_to_amdgpu_bo(gobj)->tbo);
1349 
1350 		/* Everybody except for the gang leader uses READ */
1351 		for (i = 0; i < p->gang_size; ++i) {
1352 			if (p->jobs[i] == leader)
1353 				continue;
1354 
1355 			dma_resv_add_fence(gobj->resv,
1356 					   &p->jobs[i]->base.s_fence->finished,
1357 					   DMA_RESV_USAGE_READ);
1358 		}
1359 
1360 		/* The gang leader as remembered as writer */
1361 		dma_resv_add_fence(gobj->resv, p->fence, DMA_RESV_USAGE_WRITE);
1362 	}
1363 
1364 	seq = amdgpu_ctx_add_fence(p->ctx, p->entities[p->gang_leader_idx],
1365 				   p->fence);
1366 	amdgpu_cs_post_dependencies(p);
1367 
1368 	if ((leader->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1369 	    !p->ctx->preamble_presented) {
1370 		leader->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1371 		p->ctx->preamble_presented = true;
1372 	}
1373 
1374 	cs->out.handle = seq;
1375 	leader->uf_sequence = seq;
1376 
1377 	amdgpu_vm_bo_trace_cs(&fpriv->vm, drm_exec_ticket(&p->exec));
1378 	for (i = 0; i < p->gang_size; ++i) {
1379 		amdgpu_job_free_resources(p->jobs[i]);
1380 		trace_amdgpu_cs_ioctl(p->jobs[i]);
1381 		drm_sched_entity_push_job(&p->jobs[i]->base);
1382 		p->jobs[i] = NULL;
1383 	}
1384 
1385 	amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1386 
1387 	mutex_unlock(&p->adev->notifier_lock);
1388 	mutex_unlock(&p->bo_list->bo_list_mutex);
1389 	return 0;
1390 }
1391 
1392 /* Cleanup the parser structure */
amdgpu_cs_parser_fini(struct amdgpu_cs_parser * parser)1393 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser)
1394 {
1395 	struct amdgpu_device *adev = parser->adev;
1396 	struct amdgpu_bo_list_entry *e;
1397 	unsigned int i;
1398 
1399 	amdgpu_sync_free(&parser->sync);
1400 	drm_exec_fini(&parser->exec);
1401 
1402 	for (i = 0; i < parser->num_post_deps; i++) {
1403 		drm_syncobj_put(parser->post_deps[i].syncobj);
1404 		kfree(parser->post_deps[i].chain);
1405 	}
1406 	kfree(parser->post_deps);
1407 
1408 	dma_fence_put(parser->fence);
1409 
1410 	if (parser->ctx)
1411 		amdgpu_ctx_put(parser->ctx);
1412 	if (parser->bo_list) {
1413 		if (adev->debug_vm) {
1414 			/* Invalidate all BOs to test for userspace bugs */
1415 			amdgpu_bo_list_for_each_entry(e, parser->bo_list) {
1416 				struct amdgpu_bo *bo = e->bo;
1417 
1418 				/* ignore duplicates */
1419 				if (!bo)
1420 					continue;
1421 
1422 				amdgpu_vm_bo_invalidate(bo, false);
1423 			}
1424 		}
1425 		amdgpu_bo_list_put(parser->bo_list);
1426 	}
1427 
1428 	for (i = 0; i < parser->nchunks; i++)
1429 		kvfree(parser->chunks[i].kdata);
1430 	kvfree(parser->chunks);
1431 	for (i = 0; i < parser->gang_size; ++i) {
1432 		if (parser->jobs[i])
1433 			amdgpu_job_free(parser->jobs[i]);
1434 	}
1435 	amdgpu_bo_unref(&parser->uf_bo);
1436 }
1437 
amdgpu_cs_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1438 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1439 {
1440 	struct amdgpu_device *adev = drm_to_adev(dev);
1441 	struct amdgpu_cs_parser parser;
1442 	int r;
1443 
1444 	if (amdgpu_ras_intr_triggered())
1445 		return -EHWPOISON;
1446 
1447 	if (!adev->accel_working)
1448 		return -EBUSY;
1449 
1450 	r = amdgpu_cs_parser_init(&parser, adev, filp, data);
1451 	if (r) {
1452 		drm_err_ratelimited(dev, "Failed to initialize parser %d!\n", r);
1453 		return r;
1454 	}
1455 
1456 	r = amdgpu_cs_pass1(&parser, data);
1457 	if (r)
1458 		goto error_fini;
1459 
1460 	r = amdgpu_cs_pass2(&parser);
1461 	if (r)
1462 		goto error_fini;
1463 
1464 	r = amdgpu_cs_parser_bos(&parser, data);
1465 	if (r) {
1466 		if (r == -ENOMEM)
1467 			drm_err(dev, "Not enough memory for command submission!\n");
1468 		else if (r != -ERESTARTSYS && r != -EAGAIN)
1469 			drm_dbg(dev, "Failed to process the buffer list %d!\n", r);
1470 		goto error_fini;
1471 	}
1472 
1473 	r = amdgpu_cs_patch_jobs(&parser);
1474 	if (r)
1475 		goto error_backoff;
1476 
1477 	r = amdgpu_cs_vm_handling(&parser);
1478 	if (r)
1479 		goto error_backoff;
1480 
1481 	r = amdgpu_cs_sync_rings(&parser);
1482 	if (r)
1483 		goto error_backoff;
1484 
1485 	trace_amdgpu_cs_ibs(&parser);
1486 
1487 	r = amdgpu_cs_submit(&parser, data);
1488 	if (r)
1489 		goto error_backoff;
1490 
1491 	amdgpu_cs_parser_fini(&parser);
1492 	return 0;
1493 
1494 error_backoff:
1495 	mutex_unlock(&parser.bo_list->bo_list_mutex);
1496 
1497 error_fini:
1498 	amdgpu_cs_parser_fini(&parser);
1499 	return r;
1500 }
1501 
1502 /**
1503  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1504  *
1505  * @dev: drm device
1506  * @data: data from userspace
1507  * @filp: file private
1508  *
1509  * Wait for the command submission identified by handle to finish.
1510  */
amdgpu_cs_wait_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1511 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1512 			 struct drm_file *filp)
1513 {
1514 	union drm_amdgpu_wait_cs *wait = data;
1515 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1516 	struct drm_sched_entity *entity;
1517 	struct amdgpu_ctx *ctx;
1518 	struct dma_fence *fence;
1519 	long r;
1520 
1521 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1522 	if (ctx == NULL)
1523 		return -EINVAL;
1524 
1525 	r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1526 				  wait->in.ring, &entity);
1527 	if (r) {
1528 		amdgpu_ctx_put(ctx);
1529 		return r;
1530 	}
1531 
1532 	fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1533 	if (IS_ERR(fence))
1534 		r = PTR_ERR(fence);
1535 	else if (fence) {
1536 		r = dma_fence_wait_timeout(fence, true, timeout);
1537 		if (r > 0 && fence->error)
1538 			r = fence->error;
1539 		dma_fence_put(fence);
1540 	} else
1541 		r = 1;
1542 
1543 	amdgpu_ctx_put(ctx);
1544 	if (r < 0)
1545 		return r;
1546 
1547 	memset(wait, 0, sizeof(*wait));
1548 	wait->out.status = (r == 0);
1549 
1550 	return 0;
1551 }
1552 
1553 /**
1554  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1555  *
1556  * @adev: amdgpu device
1557  * @filp: file private
1558  * @user: drm_amdgpu_fence copied from user space
1559  */
amdgpu_cs_get_fence(struct amdgpu_device * adev,struct drm_file * filp,struct drm_amdgpu_fence * user)1560 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1561 					     struct drm_file *filp,
1562 					     struct drm_amdgpu_fence *user)
1563 {
1564 	struct drm_sched_entity *entity;
1565 	struct amdgpu_ctx *ctx;
1566 	struct dma_fence *fence;
1567 	int r;
1568 
1569 	ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1570 	if (ctx == NULL)
1571 		return ERR_PTR(-EINVAL);
1572 
1573 	r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1574 				  user->ring, &entity);
1575 	if (r) {
1576 		amdgpu_ctx_put(ctx);
1577 		return ERR_PTR(r);
1578 	}
1579 
1580 	fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1581 	amdgpu_ctx_put(ctx);
1582 
1583 	return fence;
1584 }
1585 
amdgpu_cs_fence_to_handle_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1586 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1587 				    struct drm_file *filp)
1588 {
1589 	struct amdgpu_device *adev = drm_to_adev(dev);
1590 	union drm_amdgpu_fence_to_handle *info = data;
1591 	struct dma_fence *fence;
1592 	struct drm_syncobj *syncobj;
1593 	struct sync_file *sync_file;
1594 	int fd, r;
1595 
1596 	fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1597 	if (IS_ERR(fence))
1598 		return PTR_ERR(fence);
1599 
1600 	if (!fence)
1601 		fence = dma_fence_get_stub();
1602 
1603 	switch (info->in.what) {
1604 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1605 		r = drm_syncobj_create(&syncobj, 0, fence);
1606 		dma_fence_put(fence);
1607 		if (r)
1608 			return r;
1609 		r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1610 		drm_syncobj_put(syncobj);
1611 		return r;
1612 
1613 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1614 		r = drm_syncobj_create(&syncobj, 0, fence);
1615 		dma_fence_put(fence);
1616 		if (r)
1617 			return r;
1618 		r = drm_syncobj_get_fd(syncobj, (int *)&info->out.handle);
1619 		drm_syncobj_put(syncobj);
1620 		return r;
1621 
1622 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1623 		fd = get_unused_fd_flags(O_CLOEXEC);
1624 		if (fd < 0) {
1625 			dma_fence_put(fence);
1626 			return fd;
1627 		}
1628 
1629 		sync_file = sync_file_create(fence);
1630 		dma_fence_put(fence);
1631 		if (!sync_file) {
1632 			put_unused_fd(fd);
1633 			return -ENOMEM;
1634 		}
1635 
1636 		fd_install(fd, sync_file->file);
1637 		info->out.handle = fd;
1638 		return 0;
1639 
1640 	default:
1641 		dma_fence_put(fence);
1642 		return -EINVAL;
1643 	}
1644 }
1645 
1646 /**
1647  * amdgpu_cs_wait_all_fences - wait on all fences to signal
1648  *
1649  * @adev: amdgpu device
1650  * @filp: file private
1651  * @wait: wait parameters
1652  * @fences: array of drm_amdgpu_fence
1653  */
amdgpu_cs_wait_all_fences(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1654 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1655 				     struct drm_file *filp,
1656 				     union drm_amdgpu_wait_fences *wait,
1657 				     struct drm_amdgpu_fence *fences)
1658 {
1659 	uint32_t fence_count = wait->in.fence_count;
1660 	unsigned int i;
1661 	long r = 1;
1662 
1663 	for (i = 0; i < fence_count; i++) {
1664 		struct dma_fence *fence;
1665 		unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1666 
1667 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1668 		if (IS_ERR(fence))
1669 			return PTR_ERR(fence);
1670 		else if (!fence)
1671 			continue;
1672 
1673 		r = dma_fence_wait_timeout(fence, true, timeout);
1674 		if (r > 0 && fence->error)
1675 			r = fence->error;
1676 
1677 		dma_fence_put(fence);
1678 		if (r < 0)
1679 			return r;
1680 
1681 		if (r == 0)
1682 			break;
1683 	}
1684 
1685 	memset(wait, 0, sizeof(*wait));
1686 	wait->out.status = (r > 0);
1687 
1688 	return 0;
1689 }
1690 
1691 /**
1692  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1693  *
1694  * @adev: amdgpu device
1695  * @filp: file private
1696  * @wait: wait parameters
1697  * @fences: array of drm_amdgpu_fence
1698  */
amdgpu_cs_wait_any_fence(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1699 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1700 				    struct drm_file *filp,
1701 				    union drm_amdgpu_wait_fences *wait,
1702 				    struct drm_amdgpu_fence *fences)
1703 {
1704 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1705 	uint32_t fence_count = wait->in.fence_count;
1706 	uint32_t first = ~0;
1707 	struct dma_fence **array;
1708 	unsigned int i;
1709 	long r;
1710 
1711 	/* Prepare the fence array */
1712 	array = kzalloc_objs(struct dma_fence *, fence_count);
1713 
1714 	if (array == NULL)
1715 		return -ENOMEM;
1716 
1717 	for (i = 0; i < fence_count; i++) {
1718 		struct dma_fence *fence;
1719 
1720 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1721 		if (IS_ERR(fence)) {
1722 			r = PTR_ERR(fence);
1723 			goto err_free_fence_array;
1724 		} else if (fence) {
1725 			array[i] = fence;
1726 		} else { /* NULL, the fence has been already signaled */
1727 			r = 1;
1728 			first = i;
1729 			goto out;
1730 		}
1731 	}
1732 
1733 	r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1734 				       &first);
1735 	if (r < 0)
1736 		goto err_free_fence_array;
1737 
1738 out:
1739 	memset(wait, 0, sizeof(*wait));
1740 	wait->out.status = (r > 0);
1741 	wait->out.first_signaled = first;
1742 
1743 	if (first < fence_count && array[first])
1744 		r = array[first]->error;
1745 	else
1746 		r = 0;
1747 
1748 err_free_fence_array:
1749 	for (i = 0; i < fence_count; i++)
1750 		dma_fence_put(array[i]);
1751 	kfree(array);
1752 
1753 	return r;
1754 }
1755 
1756 /**
1757  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1758  *
1759  * @dev: drm device
1760  * @data: data from userspace
1761  * @filp: file private
1762  */
amdgpu_cs_wait_fences_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1763 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1764 				struct drm_file *filp)
1765 {
1766 	struct amdgpu_device *adev = drm_to_adev(dev);
1767 	union drm_amdgpu_wait_fences *wait = data;
1768 	struct drm_amdgpu_fence *fences;
1769 	int r;
1770 
1771 	/*
1772 	 * fence_count must be non-zero; dma_fence_wait_any_timeout()
1773 	 * does not accept an empty fence array.
1774 	 */
1775 	if (!wait->in.fence_count)
1776 		return -EINVAL;
1777 
1778 	/* Get the fences from userspace */
1779 	fences = memdup_array_user(u64_to_user_ptr(wait->in.fences),
1780 				   wait->in.fence_count,
1781 				   sizeof(struct drm_amdgpu_fence));
1782 	if (IS_ERR(fences))
1783 		return PTR_ERR(fences);
1784 
1785 	if (wait->in.wait_all)
1786 		r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1787 	else
1788 		r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1789 
1790 	kfree(fences);
1791 
1792 	return r;
1793 }
1794 
1795 /**
1796  * amdgpu_cs_find_mapping - find bo_va for VM address
1797  *
1798  * @parser: command submission parser context
1799  * @addr: VM address
1800  * @bo: resulting BO of the mapping found
1801  * @map: Placeholder to return found BO mapping
1802  *
1803  * Search the buffer objects in the command submission context for a certain
1804  * virtual memory address. Returns allocation structure when found, NULL
1805  * otherwise.
1806  */
amdgpu_cs_find_mapping(struct amdgpu_cs_parser * parser,uint64_t addr,struct amdgpu_bo ** bo,struct amdgpu_bo_va_mapping ** map)1807 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1808 			   uint64_t addr, struct amdgpu_bo **bo,
1809 			   struct amdgpu_bo_va_mapping **map)
1810 {
1811 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1812 	struct ttm_operation_ctx ctx = { false, false };
1813 	struct amdgpu_vm *vm = &fpriv->vm;
1814 	struct amdgpu_bo_va_mapping *mapping;
1815 	int i, r;
1816 
1817 	addr /= AMDGPU_GPU_PAGE_SIZE;
1818 
1819 	mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1820 	if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1821 		return -EINVAL;
1822 
1823 	*bo = mapping->bo_va->base.bo;
1824 	*map = mapping;
1825 
1826 	/* Double check that the BO is reserved by this CS */
1827 	if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != drm_exec_ticket(&parser->exec))
1828 		return -EINVAL;
1829 
1830 	/* Make sure VRAM is allocated contigiously */
1831 	(*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1832 	if ((*bo)->tbo.resource->mem_type == TTM_PL_VRAM &&
1833 	    !((*bo)->tbo.resource->placement & TTM_PL_FLAG_CONTIGUOUS)) {
1834 
1835 		amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1836 		for (i = 0; i < (*bo)->placement.num_placement; i++)
1837 			(*bo)->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
1838 		r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1839 		if (r)
1840 			return r;
1841 	}
1842 
1843 	return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1844 }
1845