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 "amdgpu.h"
36 #include "amdgpu_trace.h"
37 #include "amdgpu_gmc.h"
38 #include "amdgpu_gem.h"
39 #include "amdgpu_ras.h"
40 
41 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
42 				      struct drm_amdgpu_cs_chunk_fence *data,
43 				      uint32_t *offset)
44 {
45 	struct drm_gem_object *gobj;
46 	struct amdgpu_bo *bo;
47 	unsigned long size;
48 	int r;
49 
50 	gobj = drm_gem_object_lookup(p->filp, data->handle);
51 	if (gobj == NULL)
52 		return -EINVAL;
53 
54 	bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
55 	p->uf_entry.priority = 0;
56 	p->uf_entry.tv.bo = &bo->tbo;
57 	/* One for TTM and one for the CS job */
58 	p->uf_entry.tv.num_shared = 2;
59 
60 	drm_gem_object_put_unlocked(gobj);
61 
62 	size = amdgpu_bo_size(bo);
63 	if (size != PAGE_SIZE || (data->offset + 8) > size) {
64 		r = -EINVAL;
65 		goto error_unref;
66 	}
67 
68 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
69 		r = -EINVAL;
70 		goto error_unref;
71 	}
72 
73 	*offset = data->offset;
74 
75 	return 0;
76 
77 error_unref:
78 	amdgpu_bo_unref(&bo);
79 	return r;
80 }
81 
82 static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
83 				      struct drm_amdgpu_bo_list_in *data)
84 {
85 	int r;
86 	struct drm_amdgpu_bo_list_entry *info = NULL;
87 
88 	r = amdgpu_bo_create_list_entry_array(data, &info);
89 	if (r)
90 		return r;
91 
92 	r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
93 				  &p->bo_list);
94 	if (r)
95 		goto error_free;
96 
97 	kvfree(info);
98 	return 0;
99 
100 error_free:
101 	if (info)
102 		kvfree(info);
103 
104 	return r;
105 }
106 
107 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, union drm_amdgpu_cs *cs)
108 {
109 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
110 	struct amdgpu_vm *vm = &fpriv->vm;
111 	uint64_t *chunk_array_user;
112 	uint64_t *chunk_array;
113 	unsigned size, num_ibs = 0;
114 	uint32_t uf_offset = 0;
115 	int i;
116 	int ret;
117 
118 	if (cs->in.num_chunks == 0)
119 		return 0;
120 
121 	chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
122 	if (!chunk_array)
123 		return -ENOMEM;
124 
125 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
126 	if (!p->ctx) {
127 		ret = -EINVAL;
128 		goto free_chunk;
129 	}
130 
131 	mutex_lock(&p->ctx->lock);
132 
133 	/* skip guilty context job */
134 	if (atomic_read(&p->ctx->guilty) == 1) {
135 		ret = -ECANCELED;
136 		goto free_chunk;
137 	}
138 
139 	/* get chunks */
140 	chunk_array_user = u64_to_user_ptr(cs->in.chunks);
141 	if (copy_from_user(chunk_array, chunk_array_user,
142 			   sizeof(uint64_t)*cs->in.num_chunks)) {
143 		ret = -EFAULT;
144 		goto free_chunk;
145 	}
146 
147 	p->nchunks = cs->in.num_chunks;
148 	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
149 			    GFP_KERNEL);
150 	if (!p->chunks) {
151 		ret = -ENOMEM;
152 		goto free_chunk;
153 	}
154 
155 	for (i = 0; i < p->nchunks; i++) {
156 		struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
157 		struct drm_amdgpu_cs_chunk user_chunk;
158 		uint32_t __user *cdata;
159 
160 		chunk_ptr = u64_to_user_ptr(chunk_array[i]);
161 		if (copy_from_user(&user_chunk, chunk_ptr,
162 				       sizeof(struct drm_amdgpu_cs_chunk))) {
163 			ret = -EFAULT;
164 			i--;
165 			goto free_partial_kdata;
166 		}
167 		p->chunks[i].chunk_id = user_chunk.chunk_id;
168 		p->chunks[i].length_dw = user_chunk.length_dw;
169 
170 		size = p->chunks[i].length_dw;
171 		cdata = u64_to_user_ptr(user_chunk.chunk_data);
172 
173 		p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
174 		if (p->chunks[i].kdata == NULL) {
175 			ret = -ENOMEM;
176 			i--;
177 			goto free_partial_kdata;
178 		}
179 		size *= sizeof(uint32_t);
180 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
181 			ret = -EFAULT;
182 			goto free_partial_kdata;
183 		}
184 
185 		switch (p->chunks[i].chunk_id) {
186 		case AMDGPU_CHUNK_ID_IB:
187 			++num_ibs;
188 			break;
189 
190 		case AMDGPU_CHUNK_ID_FENCE:
191 			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
192 			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
193 				ret = -EINVAL;
194 				goto free_partial_kdata;
195 			}
196 
197 			ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
198 							 &uf_offset);
199 			if (ret)
200 				goto free_partial_kdata;
201 
202 			break;
203 
204 		case AMDGPU_CHUNK_ID_BO_HANDLES:
205 			size = sizeof(struct drm_amdgpu_bo_list_in);
206 			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
207 				ret = -EINVAL;
208 				goto free_partial_kdata;
209 			}
210 
211 			ret = amdgpu_cs_bo_handles_chunk(p, p->chunks[i].kdata);
212 			if (ret)
213 				goto free_partial_kdata;
214 
215 			break;
216 
217 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
218 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
219 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
220 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
221 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
222 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
223 			break;
224 
225 		default:
226 			ret = -EINVAL;
227 			goto free_partial_kdata;
228 		}
229 	}
230 
231 	ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
232 	if (ret)
233 		goto free_all_kdata;
234 
235 	if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
236 		ret = -ECANCELED;
237 		goto free_all_kdata;
238 	}
239 
240 	if (p->uf_entry.tv.bo)
241 		p->job->uf_addr = uf_offset;
242 	kfree(chunk_array);
243 
244 	/* Use this opportunity to fill in task info for the vm */
245 	amdgpu_vm_set_task_info(vm);
246 
247 	return 0;
248 
249 free_all_kdata:
250 	i = p->nchunks - 1;
251 free_partial_kdata:
252 	for (; i >= 0; i--)
253 		kvfree(p->chunks[i].kdata);
254 	kfree(p->chunks);
255 	p->chunks = NULL;
256 	p->nchunks = 0;
257 free_chunk:
258 	kfree(chunk_array);
259 
260 	return ret;
261 }
262 
263 /* Convert microseconds to bytes. */
264 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
265 {
266 	if (us <= 0 || !adev->mm_stats.log2_max_MBps)
267 		return 0;
268 
269 	/* Since accum_us is incremented by a million per second, just
270 	 * multiply it by the number of MB/s to get the number of bytes.
271 	 */
272 	return us << adev->mm_stats.log2_max_MBps;
273 }
274 
275 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
276 {
277 	if (!adev->mm_stats.log2_max_MBps)
278 		return 0;
279 
280 	return bytes >> adev->mm_stats.log2_max_MBps;
281 }
282 
283 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
284  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
285  * which means it can go over the threshold once. If that happens, the driver
286  * will be in debt and no other buffer migrations can be done until that debt
287  * is repaid.
288  *
289  * This approach allows moving a buffer of any size (it's important to allow
290  * that).
291  *
292  * The currency is simply time in microseconds and it increases as the clock
293  * ticks. The accumulated microseconds (us) are converted to bytes and
294  * returned.
295  */
296 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
297 					      u64 *max_bytes,
298 					      u64 *max_vis_bytes)
299 {
300 	s64 time_us, increment_us;
301 	u64 free_vram, total_vram, used_vram;
302 
303 	/* Allow a maximum of 200 accumulated ms. This is basically per-IB
304 	 * throttling.
305 	 *
306 	 * It means that in order to get full max MBps, at least 5 IBs per
307 	 * second must be submitted and not more than 200ms apart from each
308 	 * other.
309 	 */
310 	const s64 us_upper_bound = 200000;
311 
312 	if (!adev->mm_stats.log2_max_MBps) {
313 		*max_bytes = 0;
314 		*max_vis_bytes = 0;
315 		return;
316 	}
317 
318 	total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
319 	used_vram = amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
320 	free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
321 
322 	spin_lock(&adev->mm_stats.lock);
323 
324 	/* Increase the amount of accumulated us. */
325 	time_us = ktime_to_us(ktime_get());
326 	increment_us = time_us - adev->mm_stats.last_update_us;
327 	adev->mm_stats.last_update_us = time_us;
328 	adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
329                                       us_upper_bound);
330 
331 	/* This prevents the short period of low performance when the VRAM
332 	 * usage is low and the driver is in debt or doesn't have enough
333 	 * accumulated us to fill VRAM quickly.
334 	 *
335 	 * The situation can occur in these cases:
336 	 * - a lot of VRAM is freed by userspace
337 	 * - the presence of a big buffer causes a lot of evictions
338 	 *   (solution: split buffers into smaller ones)
339 	 *
340 	 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
341 	 * accum_us to a positive number.
342 	 */
343 	if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
344 		s64 min_us;
345 
346 		/* Be more aggresive on dGPUs. Try to fill a portion of free
347 		 * VRAM now.
348 		 */
349 		if (!(adev->flags & AMD_IS_APU))
350 			min_us = bytes_to_us(adev, free_vram / 4);
351 		else
352 			min_us = 0; /* Reset accum_us on APUs. */
353 
354 		adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
355 	}
356 
357 	/* This is set to 0 if the driver is in debt to disallow (optional)
358 	 * buffer moves.
359 	 */
360 	*max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
361 
362 	/* Do the same for visible VRAM if half of it is free */
363 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
364 		u64 total_vis_vram = adev->gmc.visible_vram_size;
365 		u64 used_vis_vram =
366 			amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
367 
368 		if (used_vis_vram < total_vis_vram) {
369 			u64 free_vis_vram = total_vis_vram - used_vis_vram;
370 			adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
371 							  increment_us, us_upper_bound);
372 
373 			if (free_vis_vram >= total_vis_vram / 2)
374 				adev->mm_stats.accum_us_vis =
375 					max(bytes_to_us(adev, free_vis_vram / 2),
376 					    adev->mm_stats.accum_us_vis);
377 		}
378 
379 		*max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
380 	} else {
381 		*max_vis_bytes = 0;
382 	}
383 
384 	spin_unlock(&adev->mm_stats.lock);
385 }
386 
387 /* Report how many bytes have really been moved for the last command
388  * submission. This can result in a debt that can stop buffer migrations
389  * temporarily.
390  */
391 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
392 				  u64 num_vis_bytes)
393 {
394 	spin_lock(&adev->mm_stats.lock);
395 	adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
396 	adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
397 	spin_unlock(&adev->mm_stats.lock);
398 }
399 
400 static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
401 				 struct amdgpu_bo *bo)
402 {
403 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
404 	struct ttm_operation_ctx ctx = {
405 		.interruptible = true,
406 		.no_wait_gpu = false,
407 		.resv = bo->tbo.base.resv,
408 		.flags = 0
409 	};
410 	uint32_t domain;
411 	int r;
412 
413 	if (bo->pin_count)
414 		return 0;
415 
416 	/* Don't move this buffer if we have depleted our allowance
417 	 * to move it. Don't move anything if the threshold is zero.
418 	 */
419 	if (p->bytes_moved < p->bytes_moved_threshold &&
420 	    (!bo->tbo.base.dma_buf ||
421 	    list_empty(&bo->tbo.base.dma_buf->attachments))) {
422 		if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
423 		    (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
424 			/* And don't move a CPU_ACCESS_REQUIRED BO to limited
425 			 * visible VRAM if we've depleted our allowance to do
426 			 * that.
427 			 */
428 			if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
429 				domain = bo->preferred_domains;
430 			else
431 				domain = bo->allowed_domains;
432 		} else {
433 			domain = bo->preferred_domains;
434 		}
435 	} else {
436 		domain = bo->allowed_domains;
437 	}
438 
439 retry:
440 	amdgpu_bo_placement_from_domain(bo, domain);
441 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
442 
443 	p->bytes_moved += ctx.bytes_moved;
444 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
445 	    amdgpu_bo_in_cpu_visible_vram(bo))
446 		p->bytes_moved_vis += ctx.bytes_moved;
447 
448 	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
449 		domain = bo->allowed_domains;
450 		goto retry;
451 	}
452 
453 	return r;
454 }
455 
456 static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
457 {
458 	struct amdgpu_cs_parser *p = param;
459 	int r;
460 
461 	r = amdgpu_cs_bo_validate(p, bo);
462 	if (r)
463 		return r;
464 
465 	if (bo->shadow)
466 		r = amdgpu_cs_bo_validate(p, bo->shadow);
467 
468 	return r;
469 }
470 
471 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
472 			    struct list_head *validated)
473 {
474 	struct ttm_operation_ctx ctx = { true, false };
475 	struct amdgpu_bo_list_entry *lobj;
476 	int r;
477 
478 	list_for_each_entry(lobj, validated, tv.head) {
479 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
480 		struct mm_struct *usermm;
481 
482 		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
483 		if (usermm && usermm != current->mm)
484 			return -EPERM;
485 
486 		if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
487 		    lobj->user_invalidated && lobj->user_pages) {
488 			amdgpu_bo_placement_from_domain(bo,
489 							AMDGPU_GEM_DOMAIN_CPU);
490 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
491 			if (r)
492 				return r;
493 
494 			amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
495 						     lobj->user_pages);
496 		}
497 
498 		r = amdgpu_cs_validate(p, bo);
499 		if (r)
500 			return r;
501 
502 		kvfree(lobj->user_pages);
503 		lobj->user_pages = NULL;
504 	}
505 	return 0;
506 }
507 
508 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
509 				union drm_amdgpu_cs *cs)
510 {
511 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
512 	struct amdgpu_vm *vm = &fpriv->vm;
513 	struct amdgpu_bo_list_entry *e;
514 	struct list_head duplicates;
515 	struct amdgpu_bo *gds;
516 	struct amdgpu_bo *gws;
517 	struct amdgpu_bo *oa;
518 	int r;
519 
520 	INIT_LIST_HEAD(&p->validated);
521 
522 	/* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
523 	if (cs->in.bo_list_handle) {
524 		if (p->bo_list)
525 			return -EINVAL;
526 
527 		r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
528 				       &p->bo_list);
529 		if (r)
530 			return r;
531 	} else if (!p->bo_list) {
532 		/* Create a empty bo_list when no handle is provided */
533 		r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
534 					  &p->bo_list);
535 		if (r)
536 			return r;
537 	}
538 
539 	/* One for TTM and one for the CS job */
540 	amdgpu_bo_list_for_each_entry(e, p->bo_list)
541 		e->tv.num_shared = 2;
542 
543 	amdgpu_bo_list_get_list(p->bo_list, &p->validated);
544 
545 	INIT_LIST_HEAD(&duplicates);
546 	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
547 
548 	if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
549 		list_add(&p->uf_entry.tv.head, &p->validated);
550 
551 	/* Get userptr backing pages. If pages are updated after registered
552 	 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
553 	 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
554 	 */
555 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
556 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
557 		bool userpage_invalidated = false;
558 		int i;
559 
560 		e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
561 					sizeof(struct page *),
562 					GFP_KERNEL | __GFP_ZERO);
563 		if (!e->user_pages) {
564 			DRM_ERROR("calloc failure\n");
565 			return -ENOMEM;
566 		}
567 
568 		r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
569 		if (r) {
570 			kvfree(e->user_pages);
571 			e->user_pages = NULL;
572 			return r;
573 		}
574 
575 		for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
576 			if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
577 				userpage_invalidated = true;
578 				break;
579 			}
580 		}
581 		e->user_invalidated = userpage_invalidated;
582 	}
583 
584 	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
585 				   &duplicates);
586 	if (unlikely(r != 0)) {
587 		if (r != -ERESTARTSYS)
588 			DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
589 		goto out;
590 	}
591 
592 	amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
593 					  &p->bytes_moved_vis_threshold);
594 	p->bytes_moved = 0;
595 	p->bytes_moved_vis = 0;
596 
597 	r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
598 				      amdgpu_cs_validate, p);
599 	if (r) {
600 		DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
601 		goto error_validate;
602 	}
603 
604 	r = amdgpu_cs_list_validate(p, &duplicates);
605 	if (r)
606 		goto error_validate;
607 
608 	r = amdgpu_cs_list_validate(p, &p->validated);
609 	if (r)
610 		goto error_validate;
611 
612 	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
613 				     p->bytes_moved_vis);
614 
615 	gds = p->bo_list->gds_obj;
616 	gws = p->bo_list->gws_obj;
617 	oa = p->bo_list->oa_obj;
618 
619 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
620 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
621 
622 		/* Make sure we use the exclusive slot for shared BOs */
623 		if (bo->prime_shared_count)
624 			e->tv.num_shared = 0;
625 		e->bo_va = amdgpu_vm_bo_find(vm, bo);
626 	}
627 
628 	if (gds) {
629 		p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
630 		p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
631 	}
632 	if (gws) {
633 		p->job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
634 		p->job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
635 	}
636 	if (oa) {
637 		p->job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
638 		p->job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
639 	}
640 
641 	if (!r && p->uf_entry.tv.bo) {
642 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
643 
644 		r = amdgpu_ttm_alloc_gart(&uf->tbo);
645 		p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
646 	}
647 
648 error_validate:
649 	if (r)
650 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
651 out:
652 	return r;
653 }
654 
655 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
656 {
657 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
658 	struct amdgpu_bo_list_entry *e;
659 	int r;
660 
661 	list_for_each_entry(e, &p->validated, tv.head) {
662 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
663 		struct dma_resv *resv = bo->tbo.base.resv;
664 		enum amdgpu_sync_mode sync_mode;
665 
666 		sync_mode = amdgpu_bo_explicit_sync(bo) ?
667 			AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
668 		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, sync_mode,
669 				     &fpriv->vm);
670 		if (r)
671 			return r;
672 	}
673 	return 0;
674 }
675 
676 /**
677  * cs_parser_fini() - clean parser states
678  * @parser:	parser structure holding parsing context.
679  * @error:	error number
680  *
681  * If error is set than unvalidate buffer, otherwise just free memory
682  * used by parsing context.
683  **/
684 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
685 				  bool backoff)
686 {
687 	unsigned i;
688 
689 	if (error && backoff)
690 		ttm_eu_backoff_reservation(&parser->ticket,
691 					   &parser->validated);
692 
693 	for (i = 0; i < parser->num_post_deps; i++) {
694 		drm_syncobj_put(parser->post_deps[i].syncobj);
695 		kfree(parser->post_deps[i].chain);
696 	}
697 	kfree(parser->post_deps);
698 
699 	dma_fence_put(parser->fence);
700 
701 	if (parser->ctx) {
702 		mutex_unlock(&parser->ctx->lock);
703 		amdgpu_ctx_put(parser->ctx);
704 	}
705 	if (parser->bo_list)
706 		amdgpu_bo_list_put(parser->bo_list);
707 
708 	for (i = 0; i < parser->nchunks; i++)
709 		kvfree(parser->chunks[i].kdata);
710 	kfree(parser->chunks);
711 	if (parser->job)
712 		amdgpu_job_free(parser->job);
713 	if (parser->uf_entry.tv.bo) {
714 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
715 
716 		amdgpu_bo_unref(&uf);
717 	}
718 }
719 
720 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
721 {
722 	struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
723 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
724 	struct amdgpu_device *adev = p->adev;
725 	struct amdgpu_vm *vm = &fpriv->vm;
726 	struct amdgpu_bo_list_entry *e;
727 	struct amdgpu_bo_va *bo_va;
728 	struct amdgpu_bo *bo;
729 	int r;
730 
731 	/* Only for UVD/VCE VM emulation */
732 	if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
733 		unsigned i, j;
734 
735 		for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
736 			struct drm_amdgpu_cs_chunk_ib *chunk_ib;
737 			struct amdgpu_bo_va_mapping *m;
738 			struct amdgpu_bo *aobj = NULL;
739 			struct amdgpu_cs_chunk *chunk;
740 			uint64_t offset, va_start;
741 			struct amdgpu_ib *ib;
742 			uint8_t *kptr;
743 
744 			chunk = &p->chunks[i];
745 			ib = &p->job->ibs[j];
746 			chunk_ib = chunk->kdata;
747 
748 			if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
749 				continue;
750 
751 			va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
752 			r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
753 			if (r) {
754 				DRM_ERROR("IB va_start is invalid\n");
755 				return r;
756 			}
757 
758 			if ((va_start + chunk_ib->ib_bytes) >
759 			    (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
760 				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
761 				return -EINVAL;
762 			}
763 
764 			/* the IB should be reserved at this point */
765 			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
766 			if (r) {
767 				return r;
768 			}
769 
770 			offset = m->start * AMDGPU_GPU_PAGE_SIZE;
771 			kptr += va_start - offset;
772 
773 			if (ring->funcs->parse_cs) {
774 				memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
775 				amdgpu_bo_kunmap(aobj);
776 
777 				r = amdgpu_ring_parse_cs(ring, p, j);
778 				if (r)
779 					return r;
780 			} else {
781 				ib->ptr = (uint32_t *)kptr;
782 				r = amdgpu_ring_patch_cs_in_place(ring, p, j);
783 				amdgpu_bo_kunmap(aobj);
784 				if (r)
785 					return r;
786 			}
787 
788 			j++;
789 		}
790 	}
791 
792 	if (!p->job->vm)
793 		return amdgpu_cs_sync_rings(p);
794 
795 
796 	r = amdgpu_vm_clear_freed(adev, vm, NULL);
797 	if (r)
798 		return r;
799 
800 	r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
801 	if (r)
802 		return r;
803 
804 	r = amdgpu_sync_vm_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
805 	if (r)
806 		return r;
807 
808 	if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
809 		bo_va = fpriv->csa_va;
810 		BUG_ON(!bo_va);
811 		r = amdgpu_vm_bo_update(adev, bo_va, false);
812 		if (r)
813 			return r;
814 
815 		r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
816 		if (r)
817 			return r;
818 	}
819 
820 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
821 		/* ignore duplicates */
822 		bo = ttm_to_amdgpu_bo(e->tv.bo);
823 		if (!bo)
824 			continue;
825 
826 		bo_va = e->bo_va;
827 		if (bo_va == NULL)
828 			continue;
829 
830 		r = amdgpu_vm_bo_update(adev, bo_va, false);
831 		if (r)
832 			return r;
833 
834 		r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
835 		if (r)
836 			return r;
837 	}
838 
839 	r = amdgpu_vm_handle_moved(adev, vm);
840 	if (r)
841 		return r;
842 
843 	r = amdgpu_vm_update_pdes(adev, vm, false);
844 	if (r)
845 		return r;
846 
847 	r = amdgpu_sync_vm_fence(&p->job->sync, vm->last_update);
848 	if (r)
849 		return r;
850 
851 	p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.base.bo);
852 
853 	if (amdgpu_vm_debug) {
854 		/* Invalidate all BOs to test for userspace bugs */
855 		amdgpu_bo_list_for_each_entry(e, p->bo_list) {
856 			struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
857 
858 			/* ignore duplicates */
859 			if (!bo)
860 				continue;
861 
862 			amdgpu_vm_bo_invalidate(adev, bo, false);
863 		}
864 	}
865 
866 	return amdgpu_cs_sync_rings(p);
867 }
868 
869 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
870 			     struct amdgpu_cs_parser *parser)
871 {
872 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
873 	struct amdgpu_vm *vm = &fpriv->vm;
874 	int r, ce_preempt = 0, de_preempt = 0;
875 	struct amdgpu_ring *ring;
876 	int i, j;
877 
878 	for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
879 		struct amdgpu_cs_chunk *chunk;
880 		struct amdgpu_ib *ib;
881 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
882 		struct drm_sched_entity *entity;
883 
884 		chunk = &parser->chunks[i];
885 		ib = &parser->job->ibs[j];
886 		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
887 
888 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
889 			continue;
890 
891 		if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
892 		    (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
893 			if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
894 				if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
895 					ce_preempt++;
896 				else
897 					de_preempt++;
898 			}
899 
900 			/* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
901 			if (ce_preempt > 1 || de_preempt > 1)
902 				return -EINVAL;
903 		}
904 
905 		r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
906 					  chunk_ib->ip_instance, chunk_ib->ring,
907 					  &entity);
908 		if (r)
909 			return r;
910 
911 		if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
912 			parser->job->preamble_status |=
913 				AMDGPU_PREAMBLE_IB_PRESENT;
914 
915 		if (parser->entity && parser->entity != entity)
916 			return -EINVAL;
917 
918 		/* Return if there is no run queue associated with this entity.
919 		 * Possibly because of disabled HW IP*/
920 		if (entity->rq == NULL)
921 			return -EINVAL;
922 
923 		parser->entity = entity;
924 
925 		ring = to_amdgpu_ring(entity->rq->sched);
926 		r =  amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
927 				   chunk_ib->ib_bytes : 0, ib);
928 		if (r) {
929 			DRM_ERROR("Failed to get ib !\n");
930 			return r;
931 		}
932 
933 		ib->gpu_addr = chunk_ib->va_start;
934 		ib->length_dw = chunk_ib->ib_bytes / 4;
935 		ib->flags = chunk_ib->flags;
936 
937 		j++;
938 	}
939 
940 	/* MM engine doesn't support user fences */
941 	ring = to_amdgpu_ring(parser->entity->rq->sched);
942 	if (parser->job->uf_addr && ring->funcs->no_user_fence)
943 		return -EINVAL;
944 
945 	return amdgpu_ctx_wait_prev_fence(parser->ctx, parser->entity);
946 }
947 
948 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
949 				       struct amdgpu_cs_chunk *chunk)
950 {
951 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
952 	unsigned num_deps;
953 	int i, r;
954 	struct drm_amdgpu_cs_chunk_dep *deps;
955 
956 	deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
957 	num_deps = chunk->length_dw * 4 /
958 		sizeof(struct drm_amdgpu_cs_chunk_dep);
959 
960 	for (i = 0; i < num_deps; ++i) {
961 		struct amdgpu_ctx *ctx;
962 		struct drm_sched_entity *entity;
963 		struct dma_fence *fence;
964 
965 		ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
966 		if (ctx == NULL)
967 			return -EINVAL;
968 
969 		r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
970 					  deps[i].ip_instance,
971 					  deps[i].ring, &entity);
972 		if (r) {
973 			amdgpu_ctx_put(ctx);
974 			return r;
975 		}
976 
977 		fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
978 		amdgpu_ctx_put(ctx);
979 
980 		if (IS_ERR(fence))
981 			return PTR_ERR(fence);
982 		else if (!fence)
983 			continue;
984 
985 		if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
986 			struct drm_sched_fence *s_fence;
987 			struct dma_fence *old = fence;
988 
989 			s_fence = to_drm_sched_fence(fence);
990 			fence = dma_fence_get(&s_fence->scheduled);
991 			dma_fence_put(old);
992 		}
993 
994 		r = amdgpu_sync_fence(&p->job->sync, fence, true);
995 		dma_fence_put(fence);
996 		if (r)
997 			return r;
998 	}
999 	return 0;
1000 }
1001 
1002 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1003 						 uint32_t handle, u64 point,
1004 						 u64 flags)
1005 {
1006 	struct dma_fence *fence;
1007 	int r;
1008 
1009 	r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
1010 	if (r) {
1011 		DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
1012 			  handle, point, r);
1013 		return r;
1014 	}
1015 
1016 	r = amdgpu_sync_fence(&p->job->sync, fence, true);
1017 	dma_fence_put(fence);
1018 
1019 	return r;
1020 }
1021 
1022 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1023 					    struct amdgpu_cs_chunk *chunk)
1024 {
1025 	struct drm_amdgpu_cs_chunk_sem *deps;
1026 	unsigned num_deps;
1027 	int i, r;
1028 
1029 	deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1030 	num_deps = chunk->length_dw * 4 /
1031 		sizeof(struct drm_amdgpu_cs_chunk_sem);
1032 	for (i = 0; i < num_deps; ++i) {
1033 		r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
1034 							  0, 0);
1035 		if (r)
1036 			return r;
1037 	}
1038 
1039 	return 0;
1040 }
1041 
1042 
1043 static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
1044 						     struct amdgpu_cs_chunk *chunk)
1045 {
1046 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1047 	unsigned num_deps;
1048 	int i, r;
1049 
1050 	syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1051 	num_deps = chunk->length_dw * 4 /
1052 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1053 	for (i = 0; i < num_deps; ++i) {
1054 		r = amdgpu_syncobj_lookup_and_add_to_sync(p,
1055 							  syncobj_deps[i].handle,
1056 							  syncobj_deps[i].point,
1057 							  syncobj_deps[i].flags);
1058 		if (r)
1059 			return r;
1060 	}
1061 
1062 	return 0;
1063 }
1064 
1065 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1066 					     struct amdgpu_cs_chunk *chunk)
1067 {
1068 	struct drm_amdgpu_cs_chunk_sem *deps;
1069 	unsigned num_deps;
1070 	int i;
1071 
1072 	deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1073 	num_deps = chunk->length_dw * 4 /
1074 		sizeof(struct drm_amdgpu_cs_chunk_sem);
1075 
1076 	if (p->post_deps)
1077 		return -EINVAL;
1078 
1079 	p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1080 				     GFP_KERNEL);
1081 	p->num_post_deps = 0;
1082 
1083 	if (!p->post_deps)
1084 		return -ENOMEM;
1085 
1086 
1087 	for (i = 0; i < num_deps; ++i) {
1088 		p->post_deps[i].syncobj =
1089 			drm_syncobj_find(p->filp, deps[i].handle);
1090 		if (!p->post_deps[i].syncobj)
1091 			return -EINVAL;
1092 		p->post_deps[i].chain = NULL;
1093 		p->post_deps[i].point = 0;
1094 		p->num_post_deps++;
1095 	}
1096 
1097 	return 0;
1098 }
1099 
1100 
1101 static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
1102 						      struct amdgpu_cs_chunk *chunk)
1103 {
1104 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1105 	unsigned num_deps;
1106 	int i;
1107 
1108 	syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1109 	num_deps = chunk->length_dw * 4 /
1110 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1111 
1112 	if (p->post_deps)
1113 		return -EINVAL;
1114 
1115 	p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1116 				     GFP_KERNEL);
1117 	p->num_post_deps = 0;
1118 
1119 	if (!p->post_deps)
1120 		return -ENOMEM;
1121 
1122 	for (i = 0; i < num_deps; ++i) {
1123 		struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
1124 
1125 		dep->chain = NULL;
1126 		if (syncobj_deps[i].point) {
1127 			dep->chain = kmalloc(sizeof(*dep->chain), GFP_KERNEL);
1128 			if (!dep->chain)
1129 				return -ENOMEM;
1130 		}
1131 
1132 		dep->syncobj = drm_syncobj_find(p->filp,
1133 						syncobj_deps[i].handle);
1134 		if (!dep->syncobj) {
1135 			kfree(dep->chain);
1136 			return -EINVAL;
1137 		}
1138 		dep->point = syncobj_deps[i].point;
1139 		p->num_post_deps++;
1140 	}
1141 
1142 	return 0;
1143 }
1144 
1145 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1146 				  struct amdgpu_cs_parser *p)
1147 {
1148 	int i, r;
1149 
1150 	for (i = 0; i < p->nchunks; ++i) {
1151 		struct amdgpu_cs_chunk *chunk;
1152 
1153 		chunk = &p->chunks[i];
1154 
1155 		switch (chunk->chunk_id) {
1156 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
1157 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
1158 			r = amdgpu_cs_process_fence_dep(p, chunk);
1159 			if (r)
1160 				return r;
1161 			break;
1162 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
1163 			r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1164 			if (r)
1165 				return r;
1166 			break;
1167 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
1168 			r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1169 			if (r)
1170 				return r;
1171 			break;
1172 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
1173 			r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
1174 			if (r)
1175 				return r;
1176 			break;
1177 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
1178 			r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
1179 			if (r)
1180 				return r;
1181 			break;
1182 		}
1183 	}
1184 
1185 	return 0;
1186 }
1187 
1188 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1189 {
1190 	int i;
1191 
1192 	for (i = 0; i < p->num_post_deps; ++i) {
1193 		if (p->post_deps[i].chain && p->post_deps[i].point) {
1194 			drm_syncobj_add_point(p->post_deps[i].syncobj,
1195 					      p->post_deps[i].chain,
1196 					      p->fence, p->post_deps[i].point);
1197 			p->post_deps[i].chain = NULL;
1198 		} else {
1199 			drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1200 						  p->fence);
1201 		}
1202 	}
1203 }
1204 
1205 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1206 			    union drm_amdgpu_cs *cs)
1207 {
1208 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1209 	struct drm_sched_entity *entity = p->entity;
1210 	enum drm_sched_priority priority;
1211 	struct amdgpu_bo_list_entry *e;
1212 	struct amdgpu_job *job;
1213 	uint64_t seq;
1214 	int r;
1215 
1216 	job = p->job;
1217 	p->job = NULL;
1218 
1219 	r = drm_sched_job_init(&job->base, entity, &fpriv->vm);
1220 	if (r)
1221 		goto error_unlock;
1222 
1223 	/* No memory allocation is allowed while holding the notifier lock.
1224 	 * The lock is held until amdgpu_cs_submit is finished and fence is
1225 	 * added to BOs.
1226 	 */
1227 	mutex_lock(&p->adev->notifier_lock);
1228 
1229 	/* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1230 	 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1231 	 */
1232 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1233 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1234 
1235 		r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1236 	}
1237 	if (r) {
1238 		r = -EAGAIN;
1239 		goto error_abort;
1240 	}
1241 
1242 	p->fence = dma_fence_get(&job->base.s_fence->finished);
1243 
1244 	amdgpu_ctx_add_fence(p->ctx, entity, p->fence, &seq);
1245 	amdgpu_cs_post_dependencies(p);
1246 
1247 	if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1248 	    !p->ctx->preamble_presented) {
1249 		job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1250 		p->ctx->preamble_presented = true;
1251 	}
1252 
1253 	cs->out.handle = seq;
1254 	job->uf_sequence = seq;
1255 
1256 	amdgpu_job_free_resources(job);
1257 
1258 	trace_amdgpu_cs_ioctl(job);
1259 	amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1260 	priority = job->base.s_priority;
1261 	drm_sched_entity_push_job(&job->base, entity);
1262 
1263 	amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1264 
1265 	ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1266 	mutex_unlock(&p->adev->notifier_lock);
1267 
1268 	return 0;
1269 
1270 error_abort:
1271 	drm_sched_job_cleanup(&job->base);
1272 	mutex_unlock(&p->adev->notifier_lock);
1273 
1274 error_unlock:
1275 	amdgpu_job_free(job);
1276 	return r;
1277 }
1278 
1279 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1280 {
1281 	struct amdgpu_device *adev = dev->dev_private;
1282 	union drm_amdgpu_cs *cs = data;
1283 	struct amdgpu_cs_parser parser = {};
1284 	bool reserved_buffers = false;
1285 	int i, r;
1286 
1287 	if (amdgpu_ras_intr_triggered())
1288 		return -EHWPOISON;
1289 
1290 	if (!adev->accel_working)
1291 		return -EBUSY;
1292 
1293 	parser.adev = adev;
1294 	parser.filp = filp;
1295 
1296 	r = amdgpu_cs_parser_init(&parser, data);
1297 	if (r) {
1298 		DRM_ERROR("Failed to initialize parser %d!\n", r);
1299 		goto out;
1300 	}
1301 
1302 	r = amdgpu_cs_ib_fill(adev, &parser);
1303 	if (r)
1304 		goto out;
1305 
1306 	r = amdgpu_cs_dependencies(adev, &parser);
1307 	if (r) {
1308 		DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1309 		goto out;
1310 	}
1311 
1312 	r = amdgpu_cs_parser_bos(&parser, data);
1313 	if (r) {
1314 		if (r == -ENOMEM)
1315 			DRM_ERROR("Not enough memory for command submission!\n");
1316 		else if (r != -ERESTARTSYS && r != -EAGAIN)
1317 			DRM_ERROR("Failed to process the buffer list %d!\n", r);
1318 		goto out;
1319 	}
1320 
1321 	reserved_buffers = true;
1322 
1323 	for (i = 0; i < parser.job->num_ibs; i++)
1324 		trace_amdgpu_cs(&parser, i);
1325 
1326 	r = amdgpu_cs_vm_handling(&parser);
1327 	if (r)
1328 		goto out;
1329 
1330 	r = amdgpu_cs_submit(&parser, cs);
1331 
1332 out:
1333 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1334 
1335 	return r;
1336 }
1337 
1338 /**
1339  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1340  *
1341  * @dev: drm device
1342  * @data: data from userspace
1343  * @filp: file private
1344  *
1345  * Wait for the command submission identified by handle to finish.
1346  */
1347 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1348 			 struct drm_file *filp)
1349 {
1350 	union drm_amdgpu_wait_cs *wait = data;
1351 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1352 	struct drm_sched_entity *entity;
1353 	struct amdgpu_ctx *ctx;
1354 	struct dma_fence *fence;
1355 	long r;
1356 
1357 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1358 	if (ctx == NULL)
1359 		return -EINVAL;
1360 
1361 	r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1362 				  wait->in.ring, &entity);
1363 	if (r) {
1364 		amdgpu_ctx_put(ctx);
1365 		return r;
1366 	}
1367 
1368 	fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1369 	if (IS_ERR(fence))
1370 		r = PTR_ERR(fence);
1371 	else if (fence) {
1372 		r = dma_fence_wait_timeout(fence, true, timeout);
1373 		if (r > 0 && fence->error)
1374 			r = fence->error;
1375 		dma_fence_put(fence);
1376 	} else
1377 		r = 1;
1378 
1379 	amdgpu_ctx_put(ctx);
1380 	if (r < 0)
1381 		return r;
1382 
1383 	memset(wait, 0, sizeof(*wait));
1384 	wait->out.status = (r == 0);
1385 
1386 	return 0;
1387 }
1388 
1389 /**
1390  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1391  *
1392  * @adev: amdgpu device
1393  * @filp: file private
1394  * @user: drm_amdgpu_fence copied from user space
1395  */
1396 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1397 					     struct drm_file *filp,
1398 					     struct drm_amdgpu_fence *user)
1399 {
1400 	struct drm_sched_entity *entity;
1401 	struct amdgpu_ctx *ctx;
1402 	struct dma_fence *fence;
1403 	int r;
1404 
1405 	ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1406 	if (ctx == NULL)
1407 		return ERR_PTR(-EINVAL);
1408 
1409 	r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1410 				  user->ring, &entity);
1411 	if (r) {
1412 		amdgpu_ctx_put(ctx);
1413 		return ERR_PTR(r);
1414 	}
1415 
1416 	fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1417 	amdgpu_ctx_put(ctx);
1418 
1419 	return fence;
1420 }
1421 
1422 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1423 				    struct drm_file *filp)
1424 {
1425 	struct amdgpu_device *adev = dev->dev_private;
1426 	union drm_amdgpu_fence_to_handle *info = data;
1427 	struct dma_fence *fence;
1428 	struct drm_syncobj *syncobj;
1429 	struct sync_file *sync_file;
1430 	int fd, r;
1431 
1432 	fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1433 	if (IS_ERR(fence))
1434 		return PTR_ERR(fence);
1435 
1436 	if (!fence)
1437 		fence = dma_fence_get_stub();
1438 
1439 	switch (info->in.what) {
1440 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1441 		r = drm_syncobj_create(&syncobj, 0, fence);
1442 		dma_fence_put(fence);
1443 		if (r)
1444 			return r;
1445 		r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1446 		drm_syncobj_put(syncobj);
1447 		return r;
1448 
1449 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1450 		r = drm_syncobj_create(&syncobj, 0, fence);
1451 		dma_fence_put(fence);
1452 		if (r)
1453 			return r;
1454 		r = drm_syncobj_get_fd(syncobj, (int*)&info->out.handle);
1455 		drm_syncobj_put(syncobj);
1456 		return r;
1457 
1458 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1459 		fd = get_unused_fd_flags(O_CLOEXEC);
1460 		if (fd < 0) {
1461 			dma_fence_put(fence);
1462 			return fd;
1463 		}
1464 
1465 		sync_file = sync_file_create(fence);
1466 		dma_fence_put(fence);
1467 		if (!sync_file) {
1468 			put_unused_fd(fd);
1469 			return -ENOMEM;
1470 		}
1471 
1472 		fd_install(fd, sync_file->file);
1473 		info->out.handle = fd;
1474 		return 0;
1475 
1476 	default:
1477 		return -EINVAL;
1478 	}
1479 }
1480 
1481 /**
1482  * amdgpu_cs_wait_all_fence - wait on all fences to signal
1483  *
1484  * @adev: amdgpu device
1485  * @filp: file private
1486  * @wait: wait parameters
1487  * @fences: array of drm_amdgpu_fence
1488  */
1489 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1490 				     struct drm_file *filp,
1491 				     union drm_amdgpu_wait_fences *wait,
1492 				     struct drm_amdgpu_fence *fences)
1493 {
1494 	uint32_t fence_count = wait->in.fence_count;
1495 	unsigned int i;
1496 	long r = 1;
1497 
1498 	for (i = 0; i < fence_count; i++) {
1499 		struct dma_fence *fence;
1500 		unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1501 
1502 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1503 		if (IS_ERR(fence))
1504 			return PTR_ERR(fence);
1505 		else if (!fence)
1506 			continue;
1507 
1508 		r = dma_fence_wait_timeout(fence, true, timeout);
1509 		dma_fence_put(fence);
1510 		if (r < 0)
1511 			return r;
1512 
1513 		if (r == 0)
1514 			break;
1515 
1516 		if (fence->error)
1517 			return fence->error;
1518 	}
1519 
1520 	memset(wait, 0, sizeof(*wait));
1521 	wait->out.status = (r > 0);
1522 
1523 	return 0;
1524 }
1525 
1526 /**
1527  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1528  *
1529  * @adev: amdgpu device
1530  * @filp: file private
1531  * @wait: wait parameters
1532  * @fences: array of drm_amdgpu_fence
1533  */
1534 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1535 				    struct drm_file *filp,
1536 				    union drm_amdgpu_wait_fences *wait,
1537 				    struct drm_amdgpu_fence *fences)
1538 {
1539 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1540 	uint32_t fence_count = wait->in.fence_count;
1541 	uint32_t first = ~0;
1542 	struct dma_fence **array;
1543 	unsigned int i;
1544 	long r;
1545 
1546 	/* Prepare the fence array */
1547 	array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1548 
1549 	if (array == NULL)
1550 		return -ENOMEM;
1551 
1552 	for (i = 0; i < fence_count; i++) {
1553 		struct dma_fence *fence;
1554 
1555 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1556 		if (IS_ERR(fence)) {
1557 			r = PTR_ERR(fence);
1558 			goto err_free_fence_array;
1559 		} else if (fence) {
1560 			array[i] = fence;
1561 		} else { /* NULL, the fence has been already signaled */
1562 			r = 1;
1563 			first = i;
1564 			goto out;
1565 		}
1566 	}
1567 
1568 	r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1569 				       &first);
1570 	if (r < 0)
1571 		goto err_free_fence_array;
1572 
1573 out:
1574 	memset(wait, 0, sizeof(*wait));
1575 	wait->out.status = (r > 0);
1576 	wait->out.first_signaled = first;
1577 
1578 	if (first < fence_count && array[first])
1579 		r = array[first]->error;
1580 	else
1581 		r = 0;
1582 
1583 err_free_fence_array:
1584 	for (i = 0; i < fence_count; i++)
1585 		dma_fence_put(array[i]);
1586 	kfree(array);
1587 
1588 	return r;
1589 }
1590 
1591 /**
1592  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1593  *
1594  * @dev: drm device
1595  * @data: data from userspace
1596  * @filp: file private
1597  */
1598 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1599 				struct drm_file *filp)
1600 {
1601 	struct amdgpu_device *adev = dev->dev_private;
1602 	union drm_amdgpu_wait_fences *wait = data;
1603 	uint32_t fence_count = wait->in.fence_count;
1604 	struct drm_amdgpu_fence *fences_user;
1605 	struct drm_amdgpu_fence *fences;
1606 	int r;
1607 
1608 	/* Get the fences from userspace */
1609 	fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1610 			GFP_KERNEL);
1611 	if (fences == NULL)
1612 		return -ENOMEM;
1613 
1614 	fences_user = u64_to_user_ptr(wait->in.fences);
1615 	if (copy_from_user(fences, fences_user,
1616 		sizeof(struct drm_amdgpu_fence) * fence_count)) {
1617 		r = -EFAULT;
1618 		goto err_free_fences;
1619 	}
1620 
1621 	if (wait->in.wait_all)
1622 		r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1623 	else
1624 		r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1625 
1626 err_free_fences:
1627 	kfree(fences);
1628 
1629 	return r;
1630 }
1631 
1632 /**
1633  * amdgpu_cs_find_bo_va - find bo_va for VM address
1634  *
1635  * @parser: command submission parser context
1636  * @addr: VM address
1637  * @bo: resulting BO of the mapping found
1638  *
1639  * Search the buffer objects in the command submission context for a certain
1640  * virtual memory address. Returns allocation structure when found, NULL
1641  * otherwise.
1642  */
1643 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1644 			   uint64_t addr, struct amdgpu_bo **bo,
1645 			   struct amdgpu_bo_va_mapping **map)
1646 {
1647 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1648 	struct ttm_operation_ctx ctx = { false, false };
1649 	struct amdgpu_vm *vm = &fpriv->vm;
1650 	struct amdgpu_bo_va_mapping *mapping;
1651 	int r;
1652 
1653 	addr /= AMDGPU_GPU_PAGE_SIZE;
1654 
1655 	mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1656 	if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1657 		return -EINVAL;
1658 
1659 	*bo = mapping->bo_va->base.bo;
1660 	*map = mapping;
1661 
1662 	/* Double check that the BO is reserved by this CS */
1663 	if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1664 		return -EINVAL;
1665 
1666 	if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1667 		(*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1668 		amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1669 		r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1670 		if (r)
1671 			return r;
1672 	}
1673 
1674 	return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1675 }
1676