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_cs.h"
36 #include "amdgpu.h"
37 #include "amdgpu_trace.h"
38 #include "amdgpu_gmc.h"
39 #include "amdgpu_gem.h"
40 #include "amdgpu_ras.h"
41 
amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_fence * data,uint32_t * offset)42 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
43 				      struct drm_amdgpu_cs_chunk_fence *data,
44 				      uint32_t *offset)
45 {
46 	struct drm_gem_object *gobj;
47 	struct amdgpu_bo *bo;
48 	unsigned long size;
49 	int r;
50 
51 	gobj = drm_gem_object_lookup(p->filp, data->handle);
52 	if (gobj == NULL)
53 		return -EINVAL;
54 
55 	bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
56 	p->uf_entry.priority = 0;
57 	p->uf_entry.tv.bo = &bo->tbo;
58 	/* One for TTM and two for the CS job */
59 	p->uf_entry.tv.num_shared = 3;
60 
61 	drm_gem_object_put(gobj);
62 
63 	size = amdgpu_bo_size(bo);
64 	if (size != PAGE_SIZE || (data->offset + 8) > size) {
65 		r = -EINVAL;
66 		goto error_unref;
67 	}
68 
69 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
70 		r = -EINVAL;
71 		goto error_unref;
72 	}
73 
74 	*offset = data->offset;
75 
76 	return 0;
77 
78 error_unref:
79 	amdgpu_bo_unref(&bo);
80 	return r;
81 }
82 
amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser * p,struct drm_amdgpu_bo_list_in * data)83 static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
84 				      struct drm_amdgpu_bo_list_in *data)
85 {
86 	int r;
87 	struct drm_amdgpu_bo_list_entry *info = NULL;
88 
89 	r = amdgpu_bo_create_list_entry_array(data, &info);
90 	if (r)
91 		return r;
92 
93 	r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
94 				  &p->bo_list);
95 	if (r)
96 		goto error_free;
97 
98 	kvfree(info);
99 	return 0;
100 
101 error_free:
102 	kvfree(info);
103 
104 	return r;
105 }
106 
amdgpu_cs_parser_init(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)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 -EINVAL;
120 
121 	chunk_array = kvmalloc_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 = kvmalloc_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 	kvfree(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 	kvfree(p->chunks);
255 	p->chunks = NULL;
256 	p->nchunks = 0;
257 free_chunk:
258 	kvfree(chunk_array);
259 
260 	return ret;
261 }
262 
263 /* Convert microseconds to bytes. */
us_to_bytes(struct amdgpu_device * adev,s64 us)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 
bytes_to_us(struct amdgpu_device * adev,u64 bytes)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  */
amdgpu_cs_get_threshold_for_moves(struct amdgpu_device * adev,u64 * max_bytes,u64 * max_vis_bytes)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 	/* Allow a maximum of 200 accumulated ms. This is basically per-IB
303 	 * throttling.
304 	 *
305 	 * It means that in order to get full max MBps, at least 5 IBs per
306 	 * second must be submitted and not more than 200ms apart from each
307 	 * other.
308 	 */
309 	const s64 us_upper_bound = 200000;
310 
311 	if (!adev->mm_stats.log2_max_MBps) {
312 		*max_bytes = 0;
313 		*max_vis_bytes = 0;
314 		return;
315 	}
316 
317 	total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
318 	used_vram = ttm_resource_manager_usage(&adev->mman.vram_mgr.manager);
319 	free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
320 
321 	spin_lock(&adev->mm_stats.lock);
322 
323 	/* Increase the amount of accumulated us. */
324 	time_us = ktime_to_us(ktime_get());
325 	increment_us = time_us - adev->mm_stats.last_update_us;
326 	adev->mm_stats.last_update_us = time_us;
327 	adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
328 				      us_upper_bound);
329 
330 	/* This prevents the short period of low performance when the VRAM
331 	 * usage is low and the driver is in debt or doesn't have enough
332 	 * accumulated us to fill VRAM quickly.
333 	 *
334 	 * The situation can occur in these cases:
335 	 * - a lot of VRAM is freed by userspace
336 	 * - the presence of a big buffer causes a lot of evictions
337 	 *   (solution: split buffers into smaller ones)
338 	 *
339 	 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
340 	 * accum_us to a positive number.
341 	 */
342 	if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
343 		s64 min_us;
344 
345 		/* Be more aggressive on dGPUs. Try to fill a portion of free
346 		 * VRAM now.
347 		 */
348 		if (!(adev->flags & AMD_IS_APU))
349 			min_us = bytes_to_us(adev, free_vram / 4);
350 		else
351 			min_us = 0; /* Reset accum_us on APUs. */
352 
353 		adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
354 	}
355 
356 	/* This is set to 0 if the driver is in debt to disallow (optional)
357 	 * buffer moves.
358 	 */
359 	*max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
360 
361 	/* Do the same for visible VRAM if half of it is free */
362 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
363 		u64 total_vis_vram = adev->gmc.visible_vram_size;
364 		u64 used_vis_vram =
365 		  amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr);
366 
367 		if (used_vis_vram < total_vis_vram) {
368 			u64 free_vis_vram = total_vis_vram - used_vis_vram;
369 			adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
370 							  increment_us, us_upper_bound);
371 
372 			if (free_vis_vram >= total_vis_vram / 2)
373 				adev->mm_stats.accum_us_vis =
374 					max(bytes_to_us(adev, free_vis_vram / 2),
375 					    adev->mm_stats.accum_us_vis);
376 		}
377 
378 		*max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
379 	} else {
380 		*max_vis_bytes = 0;
381 	}
382 
383 	spin_unlock(&adev->mm_stats.lock);
384 }
385 
386 /* Report how many bytes have really been moved for the last command
387  * submission. This can result in a debt that can stop buffer migrations
388  * temporarily.
389  */
amdgpu_cs_report_moved_bytes(struct amdgpu_device * adev,u64 num_bytes,u64 num_vis_bytes)390 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
391 				  u64 num_vis_bytes)
392 {
393 	spin_lock(&adev->mm_stats.lock);
394 	adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
395 	adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
396 	spin_unlock(&adev->mm_stats.lock);
397 }
398 
amdgpu_cs_bo_validate(void * param,struct amdgpu_bo * bo)399 static int amdgpu_cs_bo_validate(void *param, struct amdgpu_bo *bo)
400 {
401 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
402 	struct amdgpu_cs_parser *p = param;
403 	struct ttm_operation_ctx ctx = {
404 		.interruptible = true,
405 		.no_wait_gpu = false,
406 		.resv = bo->tbo.base.resv
407 	};
408 	uint32_t domain;
409 	int r;
410 
411 	if (bo->tbo.pin_count)
412 		return 0;
413 
414 	/* Don't move this buffer if we have depleted our allowance
415 	 * to move it. Don't move anything if the threshold is zero.
416 	 */
417 	if (p->bytes_moved < p->bytes_moved_threshold &&
418 	    (!bo->tbo.base.dma_buf ||
419 	    list_empty(&bo->tbo.base.dma_buf->attachments))) {
420 		if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
421 		    (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
422 			/* And don't move a CPU_ACCESS_REQUIRED BO to limited
423 			 * visible VRAM if we've depleted our allowance to do
424 			 * that.
425 			 */
426 			if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
427 				domain = bo->preferred_domains;
428 			else
429 				domain = bo->allowed_domains;
430 		} else {
431 			domain = bo->preferred_domains;
432 		}
433 	} else {
434 		domain = bo->allowed_domains;
435 	}
436 
437 retry:
438 	amdgpu_bo_placement_from_domain(bo, domain);
439 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
440 
441 	p->bytes_moved += ctx.bytes_moved;
442 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
443 	    amdgpu_bo_in_cpu_visible_vram(bo))
444 		p->bytes_moved_vis += ctx.bytes_moved;
445 
446 	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
447 		domain = bo->allowed_domains;
448 		goto retry;
449 	}
450 
451 	return r;
452 }
453 
amdgpu_cs_list_validate(struct amdgpu_cs_parser * p,struct list_head * validated)454 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
455 			    struct list_head *validated)
456 {
457 	struct ttm_operation_ctx ctx = { true, false };
458 	struct amdgpu_bo_list_entry *lobj;
459 	int r;
460 
461 	list_for_each_entry(lobj, validated, tv.head) {
462 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
463 		struct mm_struct *usermm;
464 
465 		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
466 		if (usermm && usermm != current->mm)
467 			return -EPERM;
468 
469 		if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
470 		    lobj->user_invalidated && lobj->user_pages) {
471 			amdgpu_bo_placement_from_domain(bo,
472 							AMDGPU_GEM_DOMAIN_CPU);
473 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
474 			if (r)
475 				return r;
476 
477 			amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
478 						     lobj->user_pages);
479 		}
480 
481 		r = amdgpu_cs_bo_validate(p, bo);
482 		if (r)
483 			return r;
484 
485 		kvfree(lobj->user_pages);
486 		lobj->user_pages = NULL;
487 	}
488 	return 0;
489 }
490 
amdgpu_cs_parser_bos(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)491 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
492 				union drm_amdgpu_cs *cs)
493 {
494 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
495 	struct amdgpu_vm *vm = &fpriv->vm;
496 	struct amdgpu_bo_list_entry *e;
497 	struct list_head duplicates;
498 	struct amdgpu_bo *gds;
499 	struct amdgpu_bo *gws;
500 	struct amdgpu_bo *oa;
501 	int r;
502 
503 	INIT_LIST_HEAD(&p->validated);
504 
505 	/* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
506 	if (cs->in.bo_list_handle) {
507 		if (p->bo_list)
508 			return -EINVAL;
509 
510 		r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
511 				       &p->bo_list);
512 		if (r)
513 			return r;
514 	} else if (!p->bo_list) {
515 		/* Create a empty bo_list when no handle is provided */
516 		r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
517 					  &p->bo_list);
518 		if (r)
519 			return r;
520 	}
521 
522 	mutex_lock(&p->bo_list->bo_list_mutex);
523 
524 	/* One for TTM and one for the CS job */
525 	amdgpu_bo_list_for_each_entry(e, p->bo_list)
526 		e->tv.num_shared = 2;
527 
528 	amdgpu_bo_list_get_list(p->bo_list, &p->validated);
529 
530 	INIT_LIST_HEAD(&duplicates);
531 	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
532 
533 	if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
534 		list_add(&p->uf_entry.tv.head, &p->validated);
535 
536 	/* Get userptr backing pages. If pages are updated after registered
537 	 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
538 	 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
539 	 */
540 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
541 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
542 		bool userpage_invalidated = false;
543 		int i;
544 
545 		e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
546 					sizeof(struct page *),
547 					GFP_KERNEL | __GFP_ZERO);
548 		if (!e->user_pages) {
549 			DRM_ERROR("kvmalloc_array failure\n");
550 			r = -ENOMEM;
551 			goto out_free_user_pages;
552 		}
553 
554 		r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
555 		if (r) {
556 			kvfree(e->user_pages);
557 			e->user_pages = NULL;
558 			goto out_free_user_pages;
559 		}
560 
561 		for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
562 			if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
563 				userpage_invalidated = true;
564 				break;
565 			}
566 		}
567 		e->user_invalidated = userpage_invalidated;
568 	}
569 
570 	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
571 				   &duplicates);
572 	if (unlikely(r != 0)) {
573 		if (r != -ERESTARTSYS)
574 			DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
575 		goto out_free_user_pages;
576 	}
577 
578 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
579 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
580 
581 		e->bo_va = amdgpu_vm_bo_find(vm, bo);
582 	}
583 
584 	/* Move fence waiting after getting reservation lock of
585 	 * PD root. Then there is no need on a ctx mutex lock.
586 	 */
587 	r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entity);
588 	if (unlikely(r != 0)) {
589 		if (r != -ERESTARTSYS)
590 			DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
591 		goto error_validate;
592 	}
593 
594 	amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
595 					  &p->bytes_moved_vis_threshold);
596 	p->bytes_moved = 0;
597 	p->bytes_moved_vis = 0;
598 
599 	r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
600 				      amdgpu_cs_bo_validate, p);
601 	if (r) {
602 		DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
603 		goto error_validate;
604 	}
605 
606 	r = amdgpu_cs_list_validate(p, &duplicates);
607 	if (r)
608 		goto error_validate;
609 
610 	r = amdgpu_cs_list_validate(p, &p->validated);
611 	if (r)
612 		goto error_validate;
613 
614 	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
615 				     p->bytes_moved_vis);
616 
617 	gds = p->bo_list->gds_obj;
618 	gws = p->bo_list->gws_obj;
619 	oa = p->bo_list->oa_obj;
620 
621 	if (gds) {
622 		p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
623 		p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
624 	}
625 	if (gws) {
626 		p->job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
627 		p->job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
628 	}
629 	if (oa) {
630 		p->job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
631 		p->job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
632 	}
633 
634 	if (!r && p->uf_entry.tv.bo) {
635 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
636 
637 		r = amdgpu_ttm_alloc_gart(&uf->tbo);
638 		p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
639 	}
640 
641 error_validate:
642 	if (r)
643 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
644 
645 out_free_user_pages:
646 	if (r) {
647 		amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
648 			struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
649 
650 			if (!e->user_pages)
651 				continue;
652 			amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
653 			kvfree(e->user_pages);
654 			e->user_pages = NULL;
655 		}
656 		mutex_unlock(&p->bo_list->bo_list_mutex);
657 	}
658 	return r;
659 }
660 
amdgpu_cs_sync_rings(struct amdgpu_cs_parser * p)661 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
662 {
663 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
664 	struct amdgpu_bo_list_entry *e;
665 	int r;
666 
667 	list_for_each_entry(e, &p->validated, tv.head) {
668 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
669 		struct dma_resv *resv = bo->tbo.base.resv;
670 		enum amdgpu_sync_mode sync_mode;
671 
672 		sync_mode = amdgpu_bo_explicit_sync(bo) ?
673 			AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
674 		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, sync_mode,
675 				     &fpriv->vm);
676 		if (r)
677 			return r;
678 	}
679 	return 0;
680 }
681 
682 /**
683  * amdgpu_cs_parser_fini() - clean parser states
684  * @parser:	parser structure holding parsing context.
685  * @error:	error number
686  * @backoff:	indicator to backoff the reservation
687  *
688  * If error is set then unvalidate buffer, otherwise just free memory
689  * used by parsing context.
690  **/
amdgpu_cs_parser_fini(struct amdgpu_cs_parser * parser,int error,bool backoff)691 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
692 				  bool backoff)
693 {
694 	unsigned i;
695 
696 	if (error && backoff) {
697 		ttm_eu_backoff_reservation(&parser->ticket,
698 					   &parser->validated);
699 		mutex_unlock(&parser->bo_list->bo_list_mutex);
700 	}
701 
702 	for (i = 0; i < parser->num_post_deps; i++) {
703 		drm_syncobj_put(parser->post_deps[i].syncobj);
704 		kfree(parser->post_deps[i].chain);
705 	}
706 	kfree(parser->post_deps);
707 
708 	dma_fence_put(parser->fence);
709 
710 	if (parser->ctx) {
711 		mutex_unlock(&parser->ctx->lock);
712 		amdgpu_ctx_put(parser->ctx);
713 	}
714 	if (parser->bo_list)
715 		amdgpu_bo_list_put(parser->bo_list);
716 
717 	for (i = 0; i < parser->nchunks; i++)
718 		kvfree(parser->chunks[i].kdata);
719 	kvfree(parser->chunks);
720 	if (parser->job)
721 		amdgpu_job_free(parser->job);
722 	if (parser->uf_entry.tv.bo) {
723 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
724 
725 		amdgpu_bo_unref(&uf);
726 	}
727 }
728 
amdgpu_cs_vm_handling(struct amdgpu_cs_parser * p)729 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
730 {
731 	struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
732 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
733 	struct amdgpu_device *adev = p->adev;
734 	struct amdgpu_vm *vm = &fpriv->vm;
735 	struct amdgpu_bo_list_entry *e;
736 	struct amdgpu_bo_va *bo_va;
737 	struct amdgpu_bo *bo;
738 	int r;
739 
740 	/* Only for UVD/VCE VM emulation */
741 	if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
742 		unsigned i, j;
743 
744 		for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
745 			struct drm_amdgpu_cs_chunk_ib *chunk_ib;
746 			struct amdgpu_bo_va_mapping *m;
747 			struct amdgpu_bo *aobj = NULL;
748 			struct amdgpu_cs_chunk *chunk;
749 			uint64_t offset, va_start;
750 			struct amdgpu_ib *ib;
751 			uint8_t *kptr;
752 
753 			chunk = &p->chunks[i];
754 			ib = &p->job->ibs[j];
755 			chunk_ib = chunk->kdata;
756 
757 			if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
758 				continue;
759 
760 			va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
761 			r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
762 			if (r) {
763 				DRM_ERROR("IB va_start is invalid\n");
764 				return r;
765 			}
766 
767 			if ((va_start + chunk_ib->ib_bytes) >
768 			    (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
769 				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
770 				return -EINVAL;
771 			}
772 
773 			/* the IB should be reserved at this point */
774 			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
775 			if (r) {
776 				return r;
777 			}
778 
779 			offset = m->start * AMDGPU_GPU_PAGE_SIZE;
780 			kptr += va_start - offset;
781 
782 			if (ring->funcs->parse_cs) {
783 				memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
784 				amdgpu_bo_kunmap(aobj);
785 
786 				r = amdgpu_ring_parse_cs(ring, p, p->job, ib);
787 				if (r)
788 					return r;
789 			} else {
790 				ib->ptr = (uint32_t *)kptr;
791 				r = amdgpu_ring_patch_cs_in_place(ring, p, p->job, ib);
792 				amdgpu_bo_kunmap(aobj);
793 				if (r)
794 					return r;
795 			}
796 
797 			j++;
798 		}
799 	}
800 
801 	if (!p->job->vm)
802 		return amdgpu_cs_sync_rings(p);
803 
804 
805 	r = amdgpu_vm_clear_freed(adev, vm, NULL);
806 	if (r)
807 		return r;
808 
809 	r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
810 	if (r)
811 		return r;
812 
813 	r = amdgpu_sync_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
814 	if (r)
815 		return r;
816 
817 	if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
818 		bo_va = fpriv->csa_va;
819 		BUG_ON(!bo_va);
820 		r = amdgpu_vm_bo_update(adev, bo_va, false);
821 		if (r)
822 			return r;
823 
824 		r = amdgpu_sync_fence(&p->job->sync, bo_va->last_pt_update);
825 		if (r)
826 			return r;
827 	}
828 
829 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
830 		/* ignore duplicates */
831 		bo = ttm_to_amdgpu_bo(e->tv.bo);
832 		if (!bo)
833 			continue;
834 
835 		bo_va = e->bo_va;
836 		if (bo_va == NULL)
837 			continue;
838 
839 		r = amdgpu_vm_bo_update(adev, bo_va, false);
840 		if (r)
841 			return r;
842 
843 		r = amdgpu_sync_fence(&p->job->sync, bo_va->last_pt_update);
844 		if (r)
845 			return r;
846 	}
847 
848 	r = amdgpu_vm_handle_moved(adev, vm);
849 	if (r)
850 		return r;
851 
852 	r = amdgpu_vm_update_pdes(adev, vm, false);
853 	if (r)
854 		return r;
855 
856 	r = amdgpu_sync_fence(&p->job->sync, vm->last_update);
857 	if (r)
858 		return r;
859 
860 	p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
861 
862 	if (amdgpu_vm_debug) {
863 		/* Invalidate all BOs to test for userspace bugs */
864 		amdgpu_bo_list_for_each_entry(e, p->bo_list) {
865 			struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
866 
867 			/* ignore duplicates */
868 			if (!bo)
869 				continue;
870 
871 			amdgpu_vm_bo_invalidate(adev, bo, false);
872 		}
873 	}
874 
875 	return amdgpu_cs_sync_rings(p);
876 }
877 
amdgpu_cs_ib_fill(struct amdgpu_device * adev,struct amdgpu_cs_parser * parser)878 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
879 			     struct amdgpu_cs_parser *parser)
880 {
881 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
882 	struct amdgpu_vm *vm = &fpriv->vm;
883 	int r, ce_preempt = 0, de_preempt = 0;
884 	struct amdgpu_ring *ring;
885 	int i, j;
886 
887 	for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
888 		struct amdgpu_cs_chunk *chunk;
889 		struct amdgpu_ib *ib;
890 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
891 		struct drm_sched_entity *entity;
892 
893 		chunk = &parser->chunks[i];
894 		ib = &parser->job->ibs[j];
895 		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
896 
897 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
898 			continue;
899 
900 		if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
901 		    (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
902 			if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
903 				if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
904 					ce_preempt++;
905 				else
906 					de_preempt++;
907 			}
908 
909 			/* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
910 			if (ce_preempt > 1 || de_preempt > 1)
911 				return -EINVAL;
912 		}
913 
914 		r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
915 					  chunk_ib->ip_instance, chunk_ib->ring,
916 					  &entity);
917 		if (r)
918 			return r;
919 
920 		if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
921 			parser->job->preamble_status |=
922 				AMDGPU_PREAMBLE_IB_PRESENT;
923 
924 		if (parser->entity && parser->entity != entity)
925 			return -EINVAL;
926 
927 		/* Return if there is no run queue associated with this entity.
928 		 * Possibly because of disabled HW IP*/
929 		if (entity->rq == NULL)
930 			return -EINVAL;
931 
932 		parser->entity = entity;
933 
934 		ring = to_amdgpu_ring(entity->rq->sched);
935 		r =  amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
936 				   chunk_ib->ib_bytes : 0,
937 				   AMDGPU_IB_POOL_DELAYED, ib);
938 		if (r) {
939 			DRM_ERROR("Failed to get ib !\n");
940 			return r;
941 		}
942 
943 		ib->gpu_addr = chunk_ib->va_start;
944 		ib->length_dw = chunk_ib->ib_bytes / 4;
945 		ib->flags = chunk_ib->flags;
946 
947 		j++;
948 	}
949 
950 	/* MM engine doesn't support user fences */
951 	ring = to_amdgpu_ring(parser->entity->rq->sched);
952 	if (parser->job->uf_addr && ring->funcs->no_user_fence)
953 		return -EINVAL;
954 
955 	return 0;
956 }
957 
amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)958 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
959 				       struct amdgpu_cs_chunk *chunk)
960 {
961 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
962 	unsigned num_deps;
963 	int i, r;
964 	struct drm_amdgpu_cs_chunk_dep *deps;
965 
966 	deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
967 	num_deps = chunk->length_dw * 4 /
968 		sizeof(struct drm_amdgpu_cs_chunk_dep);
969 
970 	for (i = 0; i < num_deps; ++i) {
971 		struct amdgpu_ctx *ctx;
972 		struct drm_sched_entity *entity;
973 		struct dma_fence *fence;
974 
975 		ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
976 		if (ctx == NULL)
977 			return -EINVAL;
978 
979 		r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
980 					  deps[i].ip_instance,
981 					  deps[i].ring, &entity);
982 		if (r) {
983 			amdgpu_ctx_put(ctx);
984 			return r;
985 		}
986 
987 		fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
988 		amdgpu_ctx_put(ctx);
989 
990 		if (IS_ERR(fence))
991 			return PTR_ERR(fence);
992 		else if (!fence)
993 			continue;
994 
995 		if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
996 			struct drm_sched_fence *s_fence;
997 			struct dma_fence *old = fence;
998 
999 			s_fence = to_drm_sched_fence(fence);
1000 			fence = dma_fence_get(&s_fence->scheduled);
1001 			dma_fence_put(old);
1002 		}
1003 
1004 		r = amdgpu_sync_fence(&p->job->sync, fence);
1005 		dma_fence_put(fence);
1006 		if (r)
1007 			return r;
1008 	}
1009 	return 0;
1010 }
1011 
amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser * p,uint32_t handle,u64 point,u64 flags)1012 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1013 						 uint32_t handle, u64 point,
1014 						 u64 flags)
1015 {
1016 	struct dma_fence *fence;
1017 	int r;
1018 
1019 	r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
1020 	if (r) {
1021 		DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
1022 			  handle, point, r);
1023 		return r;
1024 	}
1025 
1026 	r = amdgpu_sync_fence(&p->job->sync, fence);
1027 	dma_fence_put(fence);
1028 
1029 	return r;
1030 }
1031 
amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)1032 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1033 					    struct amdgpu_cs_chunk *chunk)
1034 {
1035 	struct drm_amdgpu_cs_chunk_sem *deps;
1036 	unsigned num_deps;
1037 	int i, r;
1038 
1039 	deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1040 	num_deps = chunk->length_dw * 4 /
1041 		sizeof(struct drm_amdgpu_cs_chunk_sem);
1042 	for (i = 0; i < num_deps; ++i) {
1043 		r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
1044 							  0, 0);
1045 		if (r)
1046 			return r;
1047 	}
1048 
1049 	return 0;
1050 }
1051 
1052 
amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)1053 static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
1054 						     struct amdgpu_cs_chunk *chunk)
1055 {
1056 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1057 	unsigned num_deps;
1058 	int i, r;
1059 
1060 	syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1061 	num_deps = chunk->length_dw * 4 /
1062 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1063 	for (i = 0; i < num_deps; ++i) {
1064 		r = amdgpu_syncobj_lookup_and_add_to_sync(p,
1065 							  syncobj_deps[i].handle,
1066 							  syncobj_deps[i].point,
1067 							  syncobj_deps[i].flags);
1068 		if (r)
1069 			return r;
1070 	}
1071 
1072 	return 0;
1073 }
1074 
amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)1075 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1076 					     struct amdgpu_cs_chunk *chunk)
1077 {
1078 	struct drm_amdgpu_cs_chunk_sem *deps;
1079 	unsigned num_deps;
1080 	int i;
1081 
1082 	deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1083 	num_deps = chunk->length_dw * 4 /
1084 		sizeof(struct drm_amdgpu_cs_chunk_sem);
1085 
1086 	if (p->post_deps)
1087 		return -EINVAL;
1088 
1089 	p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1090 				     GFP_KERNEL);
1091 	p->num_post_deps = 0;
1092 
1093 	if (!p->post_deps)
1094 		return -ENOMEM;
1095 
1096 
1097 	for (i = 0; i < num_deps; ++i) {
1098 		p->post_deps[i].syncobj =
1099 			drm_syncobj_find(p->filp, deps[i].handle);
1100 		if (!p->post_deps[i].syncobj)
1101 			return -EINVAL;
1102 		p->post_deps[i].chain = NULL;
1103 		p->post_deps[i].point = 0;
1104 		p->num_post_deps++;
1105 	}
1106 
1107 	return 0;
1108 }
1109 
1110 
amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)1111 static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
1112 						      struct amdgpu_cs_chunk *chunk)
1113 {
1114 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1115 	unsigned num_deps;
1116 	int i;
1117 
1118 	syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1119 	num_deps = chunk->length_dw * 4 /
1120 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1121 
1122 	if (p->post_deps)
1123 		return -EINVAL;
1124 
1125 	p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1126 				     GFP_KERNEL);
1127 	p->num_post_deps = 0;
1128 
1129 	if (!p->post_deps)
1130 		return -ENOMEM;
1131 
1132 	for (i = 0; i < num_deps; ++i) {
1133 		struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
1134 
1135 		dep->chain = NULL;
1136 		if (syncobj_deps[i].point) {
1137 			dep->chain = dma_fence_chain_alloc();
1138 			if (!dep->chain)
1139 				return -ENOMEM;
1140 		}
1141 
1142 		dep->syncobj = drm_syncobj_find(p->filp,
1143 						syncobj_deps[i].handle);
1144 		if (!dep->syncobj) {
1145 			dma_fence_chain_free(dep->chain);
1146 			return -EINVAL;
1147 		}
1148 		dep->point = syncobj_deps[i].point;
1149 		p->num_post_deps++;
1150 	}
1151 
1152 	return 0;
1153 }
1154 
amdgpu_cs_dependencies(struct amdgpu_device * adev,struct amdgpu_cs_parser * p)1155 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1156 				  struct amdgpu_cs_parser *p)
1157 {
1158 	int i, r;
1159 
1160 	/* TODO: Investigate why we still need the context lock */
1161 	mutex_unlock(&p->ctx->lock);
1162 
1163 	for (i = 0; i < p->nchunks; ++i) {
1164 		struct amdgpu_cs_chunk *chunk;
1165 
1166 		chunk = &p->chunks[i];
1167 
1168 		switch (chunk->chunk_id) {
1169 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
1170 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
1171 			r = amdgpu_cs_process_fence_dep(p, chunk);
1172 			if (r)
1173 				goto out;
1174 			break;
1175 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
1176 			r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1177 			if (r)
1178 				goto out;
1179 			break;
1180 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
1181 			r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1182 			if (r)
1183 				goto out;
1184 			break;
1185 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
1186 			r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
1187 			if (r)
1188 				goto out;
1189 			break;
1190 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
1191 			r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
1192 			if (r)
1193 				goto out;
1194 			break;
1195 		}
1196 	}
1197 
1198 out:
1199 	mutex_lock(&p->ctx->lock);
1200 	return r;
1201 }
1202 
amdgpu_cs_post_dependencies(struct amdgpu_cs_parser * p)1203 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1204 {
1205 	int i;
1206 
1207 	for (i = 0; i < p->num_post_deps; ++i) {
1208 		if (p->post_deps[i].chain && p->post_deps[i].point) {
1209 			drm_syncobj_add_point(p->post_deps[i].syncobj,
1210 					      p->post_deps[i].chain,
1211 					      p->fence, p->post_deps[i].point);
1212 			p->post_deps[i].chain = NULL;
1213 		} else {
1214 			drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1215 						  p->fence);
1216 		}
1217 	}
1218 }
1219 
amdgpu_cs_submit(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)1220 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1221 			    union drm_amdgpu_cs *cs)
1222 {
1223 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1224 	struct drm_sched_entity *entity = p->entity;
1225 	struct amdgpu_bo_list_entry *e;
1226 	struct amdgpu_job *job;
1227 	uint64_t seq;
1228 	int r;
1229 
1230 	job = p->job;
1231 	p->job = NULL;
1232 
1233 	r = drm_sched_job_init(&job->base, entity, &fpriv->vm);
1234 	if (r)
1235 		goto error_unlock;
1236 
1237 	drm_sched_job_arm(&job->base);
1238 
1239 	/* No memory allocation is allowed while holding the notifier lock.
1240 	 * The lock is held until amdgpu_cs_submit is finished and fence is
1241 	 * added to BOs.
1242 	 */
1243 	mutex_lock(&p->adev->notifier_lock);
1244 
1245 	/* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1246 	 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1247 	 */
1248 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1249 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1250 
1251 		r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1252 	}
1253 	if (r) {
1254 		r = -EAGAIN;
1255 		goto error_abort;
1256 	}
1257 
1258 	p->fence = dma_fence_get(&job->base.s_fence->finished);
1259 
1260 	seq = amdgpu_ctx_add_fence(p->ctx, entity, p->fence);
1261 	amdgpu_cs_post_dependencies(p);
1262 
1263 	if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1264 	    !p->ctx->preamble_presented) {
1265 		job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1266 		p->ctx->preamble_presented = true;
1267 	}
1268 
1269 	cs->out.handle = seq;
1270 	job->uf_sequence = seq;
1271 
1272 	amdgpu_job_free_resources(job);
1273 
1274 	trace_amdgpu_cs_ioctl(job);
1275 	amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1276 	drm_sched_entity_push_job(&job->base);
1277 
1278 	amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1279 
1280 	/* Make sure all BOs are remembered as writers */
1281 	amdgpu_bo_list_for_each_entry(e, p->bo_list)
1282 		e->tv.num_shared = 0;
1283 
1284 	ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1285 	mutex_unlock(&p->adev->notifier_lock);
1286 	mutex_unlock(&p->bo_list->bo_list_mutex);
1287 
1288 	return 0;
1289 
1290 error_abort:
1291 	drm_sched_job_cleanup(&job->base);
1292 	mutex_unlock(&p->adev->notifier_lock);
1293 
1294 error_unlock:
1295 	amdgpu_job_free(job);
1296 	return r;
1297 }
1298 
trace_amdgpu_cs_ibs(struct amdgpu_cs_parser * parser)1299 static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *parser)
1300 {
1301 	int i;
1302 
1303 	if (!trace_amdgpu_cs_enabled())
1304 		return;
1305 
1306 	for (i = 0; i < parser->job->num_ibs; i++)
1307 		trace_amdgpu_cs(parser, i);
1308 }
1309 
amdgpu_cs_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1310 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1311 {
1312 	struct amdgpu_device *adev = drm_to_adev(dev);
1313 	union drm_amdgpu_cs *cs = data;
1314 	struct amdgpu_cs_parser parser = {};
1315 	bool reserved_buffers = false;
1316 	int r;
1317 
1318 	if (amdgpu_ras_intr_triggered())
1319 		return -EHWPOISON;
1320 
1321 	if (!adev->accel_working)
1322 		return -EBUSY;
1323 
1324 	parser.adev = adev;
1325 	parser.filp = filp;
1326 
1327 	r = amdgpu_cs_parser_init(&parser, data);
1328 	if (r) {
1329 		if (printk_ratelimit())
1330 			DRM_ERROR("Failed to initialize parser %d!\n", r);
1331 		goto out;
1332 	}
1333 
1334 	r = amdgpu_cs_ib_fill(adev, &parser);
1335 	if (r)
1336 		goto out;
1337 
1338 	r = amdgpu_cs_dependencies(adev, &parser);
1339 	if (r) {
1340 		DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1341 		goto out;
1342 	}
1343 
1344 	r = amdgpu_cs_parser_bos(&parser, data);
1345 	if (r) {
1346 		if (r == -ENOMEM)
1347 			DRM_ERROR("Not enough memory for command submission!\n");
1348 		else if (r != -ERESTARTSYS && r != -EAGAIN)
1349 			DRM_ERROR("Failed to process the buffer list %d!\n", r);
1350 		goto out;
1351 	}
1352 
1353 	reserved_buffers = true;
1354 
1355 	trace_amdgpu_cs_ibs(&parser);
1356 
1357 	r = amdgpu_cs_vm_handling(&parser);
1358 	if (r)
1359 		goto out;
1360 
1361 	r = amdgpu_cs_submit(&parser, cs);
1362 
1363 out:
1364 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1365 
1366 	return r;
1367 }
1368 
1369 /**
1370  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1371  *
1372  * @dev: drm device
1373  * @data: data from userspace
1374  * @filp: file private
1375  *
1376  * Wait for the command submission identified by handle to finish.
1377  */
amdgpu_cs_wait_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1378 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1379 			 struct drm_file *filp)
1380 {
1381 	union drm_amdgpu_wait_cs *wait = data;
1382 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1383 	struct drm_sched_entity *entity;
1384 	struct amdgpu_ctx *ctx;
1385 	struct dma_fence *fence;
1386 	long r;
1387 
1388 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1389 	if (ctx == NULL)
1390 		return -EINVAL;
1391 
1392 	r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1393 				  wait->in.ring, &entity);
1394 	if (r) {
1395 		amdgpu_ctx_put(ctx);
1396 		return r;
1397 	}
1398 
1399 	fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1400 	if (IS_ERR(fence))
1401 		r = PTR_ERR(fence);
1402 	else if (fence) {
1403 		r = dma_fence_wait_timeout(fence, true, timeout);
1404 		if (r > 0 && fence->error)
1405 			r = fence->error;
1406 		dma_fence_put(fence);
1407 	} else
1408 		r = 1;
1409 
1410 	amdgpu_ctx_put(ctx);
1411 	if (r < 0)
1412 		return r;
1413 
1414 	memset(wait, 0, sizeof(*wait));
1415 	wait->out.status = (r == 0);
1416 
1417 	return 0;
1418 }
1419 
1420 /**
1421  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1422  *
1423  * @adev: amdgpu device
1424  * @filp: file private
1425  * @user: drm_amdgpu_fence copied from user space
1426  */
amdgpu_cs_get_fence(struct amdgpu_device * adev,struct drm_file * filp,struct drm_amdgpu_fence * user)1427 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1428 					     struct drm_file *filp,
1429 					     struct drm_amdgpu_fence *user)
1430 {
1431 	struct drm_sched_entity *entity;
1432 	struct amdgpu_ctx *ctx;
1433 	struct dma_fence *fence;
1434 	int r;
1435 
1436 	ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1437 	if (ctx == NULL)
1438 		return ERR_PTR(-EINVAL);
1439 
1440 	r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1441 				  user->ring, &entity);
1442 	if (r) {
1443 		amdgpu_ctx_put(ctx);
1444 		return ERR_PTR(r);
1445 	}
1446 
1447 	fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1448 	amdgpu_ctx_put(ctx);
1449 
1450 	return fence;
1451 }
1452 
amdgpu_cs_fence_to_handle_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1453 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1454 				    struct drm_file *filp)
1455 {
1456 	struct amdgpu_device *adev = drm_to_adev(dev);
1457 	union drm_amdgpu_fence_to_handle *info = data;
1458 	struct dma_fence *fence;
1459 	struct drm_syncobj *syncobj;
1460 	struct sync_file *sync_file;
1461 	int fd, r;
1462 
1463 	fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1464 	if (IS_ERR(fence))
1465 		return PTR_ERR(fence);
1466 
1467 	if (!fence)
1468 		fence = dma_fence_get_stub();
1469 
1470 	switch (info->in.what) {
1471 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1472 		r = drm_syncobj_create(&syncobj, 0, fence);
1473 		dma_fence_put(fence);
1474 		if (r)
1475 			return r;
1476 		r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1477 		drm_syncobj_put(syncobj);
1478 		return r;
1479 
1480 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1481 		r = drm_syncobj_create(&syncobj, 0, fence);
1482 		dma_fence_put(fence);
1483 		if (r)
1484 			return r;
1485 		r = drm_syncobj_get_fd(syncobj, (int *)&info->out.handle);
1486 		drm_syncobj_put(syncobj);
1487 		return r;
1488 
1489 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1490 		fd = get_unused_fd_flags(O_CLOEXEC);
1491 		if (fd < 0) {
1492 			dma_fence_put(fence);
1493 			return fd;
1494 		}
1495 
1496 		sync_file = sync_file_create(fence);
1497 		dma_fence_put(fence);
1498 		if (!sync_file) {
1499 			put_unused_fd(fd);
1500 			return -ENOMEM;
1501 		}
1502 
1503 		fd_install(fd, sync_file->file);
1504 		info->out.handle = fd;
1505 		return 0;
1506 
1507 	default:
1508 		dma_fence_put(fence);
1509 		return -EINVAL;
1510 	}
1511 }
1512 
1513 /**
1514  * amdgpu_cs_wait_all_fences - wait on all fences to signal
1515  *
1516  * @adev: amdgpu device
1517  * @filp: file private
1518  * @wait: wait parameters
1519  * @fences: array of drm_amdgpu_fence
1520  */
amdgpu_cs_wait_all_fences(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1521 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1522 				     struct drm_file *filp,
1523 				     union drm_amdgpu_wait_fences *wait,
1524 				     struct drm_amdgpu_fence *fences)
1525 {
1526 	uint32_t fence_count = wait->in.fence_count;
1527 	unsigned int i;
1528 	long r = 1;
1529 
1530 	for (i = 0; i < fence_count; i++) {
1531 		struct dma_fence *fence;
1532 		unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1533 
1534 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1535 		if (IS_ERR(fence))
1536 			return PTR_ERR(fence);
1537 		else if (!fence)
1538 			continue;
1539 
1540 		r = dma_fence_wait_timeout(fence, true, timeout);
1541 		dma_fence_put(fence);
1542 		if (r < 0)
1543 			return r;
1544 
1545 		if (r == 0)
1546 			break;
1547 
1548 		if (fence->error)
1549 			return fence->error;
1550 	}
1551 
1552 	memset(wait, 0, sizeof(*wait));
1553 	wait->out.status = (r > 0);
1554 
1555 	return 0;
1556 }
1557 
1558 /**
1559  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1560  *
1561  * @adev: amdgpu device
1562  * @filp: file private
1563  * @wait: wait parameters
1564  * @fences: array of drm_amdgpu_fence
1565  */
amdgpu_cs_wait_any_fence(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1566 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1567 				    struct drm_file *filp,
1568 				    union drm_amdgpu_wait_fences *wait,
1569 				    struct drm_amdgpu_fence *fences)
1570 {
1571 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1572 	uint32_t fence_count = wait->in.fence_count;
1573 	uint32_t first = ~0;
1574 	struct dma_fence **array;
1575 	unsigned int i;
1576 	long r;
1577 
1578 	/* Prepare the fence array */
1579 	array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1580 
1581 	if (array == NULL)
1582 		return -ENOMEM;
1583 
1584 	for (i = 0; i < fence_count; i++) {
1585 		struct dma_fence *fence;
1586 
1587 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1588 		if (IS_ERR(fence)) {
1589 			r = PTR_ERR(fence);
1590 			goto err_free_fence_array;
1591 		} else if (fence) {
1592 			array[i] = fence;
1593 		} else { /* NULL, the fence has been already signaled */
1594 			r = 1;
1595 			first = i;
1596 			goto out;
1597 		}
1598 	}
1599 
1600 	r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1601 				       &first);
1602 	if (r < 0)
1603 		goto err_free_fence_array;
1604 
1605 out:
1606 	memset(wait, 0, sizeof(*wait));
1607 	wait->out.status = (r > 0);
1608 	wait->out.first_signaled = first;
1609 
1610 	if (first < fence_count && array[first])
1611 		r = array[first]->error;
1612 	else
1613 		r = 0;
1614 
1615 err_free_fence_array:
1616 	for (i = 0; i < fence_count; i++)
1617 		dma_fence_put(array[i]);
1618 	kfree(array);
1619 
1620 	return r;
1621 }
1622 
1623 /**
1624  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1625  *
1626  * @dev: drm device
1627  * @data: data from userspace
1628  * @filp: file private
1629  */
amdgpu_cs_wait_fences_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1630 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1631 				struct drm_file *filp)
1632 {
1633 	struct amdgpu_device *adev = drm_to_adev(dev);
1634 	union drm_amdgpu_wait_fences *wait = data;
1635 	uint32_t fence_count = wait->in.fence_count;
1636 	struct drm_amdgpu_fence *fences_user;
1637 	struct drm_amdgpu_fence *fences;
1638 	int r;
1639 
1640 	/* Get the fences from userspace */
1641 	fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1642 			GFP_KERNEL);
1643 	if (fences == NULL)
1644 		return -ENOMEM;
1645 
1646 	fences_user = u64_to_user_ptr(wait->in.fences);
1647 	if (copy_from_user(fences, fences_user,
1648 		sizeof(struct drm_amdgpu_fence) * fence_count)) {
1649 		r = -EFAULT;
1650 		goto err_free_fences;
1651 	}
1652 
1653 	if (wait->in.wait_all)
1654 		r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1655 	else
1656 		r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1657 
1658 err_free_fences:
1659 	kfree(fences);
1660 
1661 	return r;
1662 }
1663 
1664 /**
1665  * amdgpu_cs_find_mapping - find bo_va for VM address
1666  *
1667  * @parser: command submission parser context
1668  * @addr: VM address
1669  * @bo: resulting BO of the mapping found
1670  * @map: Placeholder to return found BO mapping
1671  *
1672  * Search the buffer objects in the command submission context for a certain
1673  * virtual memory address. Returns allocation structure when found, NULL
1674  * otherwise.
1675  */
amdgpu_cs_find_mapping(struct amdgpu_cs_parser * parser,uint64_t addr,struct amdgpu_bo ** bo,struct amdgpu_bo_va_mapping ** map)1676 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1677 			   uint64_t addr, struct amdgpu_bo **bo,
1678 			   struct amdgpu_bo_va_mapping **map)
1679 {
1680 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1681 	struct ttm_operation_ctx ctx = { false, false };
1682 	struct amdgpu_vm *vm = &fpriv->vm;
1683 	struct amdgpu_bo_va_mapping *mapping;
1684 	int r;
1685 
1686 	addr /= AMDGPU_GPU_PAGE_SIZE;
1687 
1688 	mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1689 	if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1690 		return -EINVAL;
1691 
1692 	*bo = mapping->bo_va->base.bo;
1693 	*map = mapping;
1694 
1695 	/* Double check that the BO is reserved by this CS */
1696 	if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1697 		return -EINVAL;
1698 
1699 	if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1700 		(*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1701 		amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1702 		r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1703 		if (r)
1704 			return r;
1705 	}
1706 
1707 	return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1708 }
1709