xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/amd/amdgpu/amdgpu_cs.c (revision 0caae2224fa2e443b0194fe793325afc8e00f306)
1 /*	$NetBSD: amdgpu_cs.c,v 1.7 2021/12/19 12:02:39 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2008 Jerome Glisse.
5  * All Rights Reserved.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a
8  * copy of this software and associated documentation files (the "Software"),
9  * to deal in the Software without restriction, including without limitation
10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11  * and/or sell copies of the Software, and to permit persons to whom the
12  * Software is furnished to do so, subject to the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the next
15  * paragraph) shall be included in all copies or substantial portions of the
16  * Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
21  * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
22  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
23  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
24  * DEALINGS IN THE SOFTWARE.
25  *
26  * Authors:
27  *    Jerome Glisse <glisse@freedesktop.org>
28  */
29 
30 #include <sys/cdefs.h>
31 __KERNEL_RCSID(0, "$NetBSD: amdgpu_cs.c,v 1.7 2021/12/19 12:02:39 riastradh Exp $");
32 
33 #include <linux/file.h>
34 #include <linux/pagemap.h>
35 #include <linux/sync_file.h>
36 
37 #include <drm/amdgpu_drm.h>
38 #include <drm/drm_syncobj.h>
39 #include "amdgpu.h"
40 #include "amdgpu_trace.h"
41 #include "amdgpu_gmc.h"
42 #include "amdgpu_gem.h"
43 #include "amdgpu_ras.h"
44 
amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser * p,struct drm_amdgpu_cs_chunk_fence * data,uint32_t * offset)45 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
46 				      struct drm_amdgpu_cs_chunk_fence *data,
47 				      uint32_t *offset)
48 {
49 	struct drm_gem_object *gobj;
50 	struct amdgpu_bo *bo;
51 	unsigned long size;
52 	int r;
53 
54 	gobj = drm_gem_object_lookup(p->filp, data->handle);
55 	if (gobj == NULL)
56 		return -EINVAL;
57 
58 	bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
59 	p->uf_entry.priority = 0;
60 	p->uf_entry.tv.bo = &bo->tbo;
61 	/* One for TTM and one for the CS job */
62 	p->uf_entry.tv.num_shared = 2;
63 
64 	drm_gem_object_put_unlocked(gobj);
65 
66 	size = amdgpu_bo_size(bo);
67 	if (size != PAGE_SIZE || (data->offset + 8) > size) {
68 		r = -EINVAL;
69 		goto error_unref;
70 	}
71 
72 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
73 		r = -EINVAL;
74 		goto error_unref;
75 	}
76 
77 	*offset = data->offset;
78 
79 	return 0;
80 
81 error_unref:
82 	amdgpu_bo_unref(&bo);
83 	return r;
84 }
85 
amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser * p,struct drm_amdgpu_bo_list_in * data)86 static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
87 				      struct drm_amdgpu_bo_list_in *data)
88 {
89 	int r;
90 	struct drm_amdgpu_bo_list_entry *info = NULL;
91 
92 	r = amdgpu_bo_create_list_entry_array(data, &info);
93 	if (r)
94 		return r;
95 
96 	r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
97 				  &p->bo_list);
98 	if (r)
99 		goto error_free;
100 
101 	kvfree(info);
102 	return 0;
103 
104 error_free:
105 	if (info)
106 		kvfree(info);
107 
108 	return r;
109 }
110 
amdgpu_cs_parser_init(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)111 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, union drm_amdgpu_cs *cs)
112 {
113 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
114 	struct amdgpu_vm *vm = &fpriv->vm;
115 	uint64_t *chunk_array_user;
116 	uint64_t *chunk_array;
117 	unsigned size, num_ibs = 0;
118 	uint32_t uf_offset = 0;
119 	int i;
120 	int ret;
121 
122 	if (cs->in.num_chunks == 0)
123 		return 0;
124 
125 	chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
126 	if (!chunk_array)
127 		return -ENOMEM;
128 
129 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
130 	if (!p->ctx) {
131 		ret = -EINVAL;
132 		goto free_chunk;
133 	}
134 
135 	mutex_lock(&p->ctx->lock);
136 
137 	/* skip guilty context job */
138 	if (atomic_read(&p->ctx->guilty) == 1) {
139 		ret = -ECANCELED;
140 		goto free_chunk;
141 	}
142 
143 	/* get chunks */
144 	chunk_array_user = u64_to_user_ptr(cs->in.chunks);
145 	if (copy_from_user(chunk_array, chunk_array_user,
146 			   sizeof(uint64_t)*cs->in.num_chunks)) {
147 		ret = -EFAULT;
148 		goto free_chunk;
149 	}
150 
151 	p->nchunks = cs->in.num_chunks;
152 	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
153 			    GFP_KERNEL);
154 	if (!p->chunks) {
155 		ret = -ENOMEM;
156 		goto free_chunk;
157 	}
158 
159 	for (i = 0; i < p->nchunks; i++) {
160 		struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
161 		struct drm_amdgpu_cs_chunk user_chunk;
162 		uint32_t __user *cdata;
163 
164 		chunk_ptr = u64_to_user_ptr(chunk_array[i]);
165 		if (copy_from_user(&user_chunk, chunk_ptr,
166 				       sizeof(struct drm_amdgpu_cs_chunk))) {
167 			ret = -EFAULT;
168 			i--;
169 			goto free_partial_kdata;
170 		}
171 		p->chunks[i].chunk_id = user_chunk.chunk_id;
172 		p->chunks[i].length_dw = user_chunk.length_dw;
173 
174 		size = p->chunks[i].length_dw;
175 		cdata = u64_to_user_ptr(user_chunk.chunk_data);
176 
177 		p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
178 		if (p->chunks[i].kdata == NULL) {
179 			ret = -ENOMEM;
180 			i--;
181 			goto free_partial_kdata;
182 		}
183 		size *= sizeof(uint32_t);
184 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
185 			ret = -EFAULT;
186 			goto free_partial_kdata;
187 		}
188 
189 		switch (p->chunks[i].chunk_id) {
190 		case AMDGPU_CHUNK_ID_IB:
191 			++num_ibs;
192 			break;
193 
194 		case AMDGPU_CHUNK_ID_FENCE:
195 			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
196 			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
197 				ret = -EINVAL;
198 				goto free_partial_kdata;
199 			}
200 
201 			ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
202 							 &uf_offset);
203 			if (ret)
204 				goto free_partial_kdata;
205 
206 			break;
207 
208 		case AMDGPU_CHUNK_ID_BO_HANDLES:
209 			size = sizeof(struct drm_amdgpu_bo_list_in);
210 			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
211 				ret = -EINVAL;
212 				goto free_partial_kdata;
213 			}
214 
215 			ret = amdgpu_cs_bo_handles_chunk(p, p->chunks[i].kdata);
216 			if (ret)
217 				goto free_partial_kdata;
218 
219 			break;
220 
221 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
222 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
223 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
224 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
225 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
226 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
227 			break;
228 
229 		default:
230 			ret = -EINVAL;
231 			goto free_partial_kdata;
232 		}
233 	}
234 
235 	ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
236 	if (ret)
237 		goto free_all_kdata;
238 
239 	if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
240 		ret = -ECANCELED;
241 		goto free_all_kdata;
242 	}
243 
244 	if (p->uf_entry.tv.bo)
245 		p->job->uf_addr = uf_offset;
246 	kfree(chunk_array);
247 
248 	/* Use this opportunity to fill in task info for the vm */
249 	amdgpu_vm_set_task_info(vm);
250 
251 	return 0;
252 
253 free_all_kdata:
254 	i = p->nchunks - 1;
255 free_partial_kdata:
256 	for (; i >= 0; i--)
257 		kvfree(p->chunks[i].kdata);
258 	kfree(p->chunks);
259 	p->chunks = NULL;
260 	p->nchunks = 0;
261 free_chunk:
262 	kfree(chunk_array);
263 
264 	return ret;
265 }
266 
267 /* Convert microseconds to bytes. */
us_to_bytes(struct amdgpu_device * adev,s64 us)268 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
269 {
270 	if (us <= 0 || !adev->mm_stats.log2_max_MBps)
271 		return 0;
272 
273 	/* Since accum_us is incremented by a million per second, just
274 	 * multiply it by the number of MB/s to get the number of bytes.
275 	 */
276 	return us << adev->mm_stats.log2_max_MBps;
277 }
278 
bytes_to_us(struct amdgpu_device * adev,u64 bytes)279 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
280 {
281 	if (!adev->mm_stats.log2_max_MBps)
282 		return 0;
283 
284 	return bytes >> adev->mm_stats.log2_max_MBps;
285 }
286 
287 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
288  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
289  * which means it can go over the threshold once. If that happens, the driver
290  * will be in debt and no other buffer migrations can be done until that debt
291  * is repaid.
292  *
293  * This approach allows moving a buffer of any size (it's important to allow
294  * that).
295  *
296  * The currency is simply time in microseconds and it increases as the clock
297  * ticks. The accumulated microseconds (us) are converted to bytes and
298  * returned.
299  */
amdgpu_cs_get_threshold_for_moves(struct amdgpu_device * adev,u64 * max_bytes,u64 * max_vis_bytes)300 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
301 					      u64 *max_bytes,
302 					      u64 *max_vis_bytes)
303 {
304 	s64 time_us, increment_us;
305 	u64 free_vram, total_vram, used_vram;
306 
307 	/* Allow a maximum of 200 accumulated ms. This is basically per-IB
308 	 * throttling.
309 	 *
310 	 * It means that in order to get full max MBps, at least 5 IBs per
311 	 * second must be submitted and not more than 200ms apart from each
312 	 * other.
313 	 */
314 	const s64 us_upper_bound = 200000;
315 
316 	if (!adev->mm_stats.log2_max_MBps) {
317 		*max_bytes = 0;
318 		*max_vis_bytes = 0;
319 		return;
320 	}
321 
322 	total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
323 	used_vram = amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
324 	free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
325 
326 	spin_lock(&adev->mm_stats.lock);
327 
328 	/* Increase the amount of accumulated us. */
329 	time_us = ktime_to_us(ktime_get());
330 	increment_us = time_us - adev->mm_stats.last_update_us;
331 	adev->mm_stats.last_update_us = time_us;
332 	adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
333                                       us_upper_bound);
334 
335 	/* This prevents the short period of low performance when the VRAM
336 	 * usage is low and the driver is in debt or doesn't have enough
337 	 * accumulated us to fill VRAM quickly.
338 	 *
339 	 * The situation can occur in these cases:
340 	 * - a lot of VRAM is freed by userspace
341 	 * - the presence of a big buffer causes a lot of evictions
342 	 *   (solution: split buffers into smaller ones)
343 	 *
344 	 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
345 	 * accum_us to a positive number.
346 	 */
347 	if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
348 		s64 min_us;
349 
350 		/* Be more aggresive on dGPUs. Try to fill a portion of free
351 		 * VRAM now.
352 		 */
353 		if (!(adev->flags & AMD_IS_APU))
354 			min_us = bytes_to_us(adev, free_vram / 4);
355 		else
356 			min_us = 0; /* Reset accum_us on APUs. */
357 
358 		adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
359 	}
360 
361 	/* This is set to 0 if the driver is in debt to disallow (optional)
362 	 * buffer moves.
363 	 */
364 	*max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
365 
366 	/* Do the same for visible VRAM if half of it is free */
367 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
368 		u64 total_vis_vram = adev->gmc.visible_vram_size;
369 		u64 used_vis_vram =
370 			amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
371 
372 		if (used_vis_vram < total_vis_vram) {
373 			u64 free_vis_vram = total_vis_vram - used_vis_vram;
374 			adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
375 							  increment_us, us_upper_bound);
376 
377 			if (free_vis_vram >= total_vis_vram / 2)
378 				adev->mm_stats.accum_us_vis =
379 					max(bytes_to_us(adev, free_vis_vram / 2),
380 					    adev->mm_stats.accum_us_vis);
381 		}
382 
383 		*max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
384 	} else {
385 		*max_vis_bytes = 0;
386 	}
387 
388 	spin_unlock(&adev->mm_stats.lock);
389 }
390 
391 /* Report how many bytes have really been moved for the last command
392  * submission. This can result in a debt that can stop buffer migrations
393  * temporarily.
394  */
amdgpu_cs_report_moved_bytes(struct amdgpu_device * adev,u64 num_bytes,u64 num_vis_bytes)395 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
396 				  u64 num_vis_bytes)
397 {
398 	spin_lock(&adev->mm_stats.lock);
399 	adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
400 	adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
401 	spin_unlock(&adev->mm_stats.lock);
402 }
403 
amdgpu_cs_bo_validate(struct amdgpu_cs_parser * p,struct amdgpu_bo * bo)404 static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
405 				 struct amdgpu_bo *bo)
406 {
407 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
408 	struct ttm_operation_ctx ctx = {
409 		.interruptible = true,
410 		.no_wait_gpu = false,
411 		.resv = bo->tbo.base.resv,
412 		.flags = 0
413 	};
414 	uint32_t domain;
415 	int r;
416 
417 	if (bo->pin_count)
418 		return 0;
419 
420 	/* Don't move this buffer if we have depleted our allowance
421 	 * to move it. Don't move anything if the threshold is zero.
422 	 */
423 	if (p->bytes_moved < p->bytes_moved_threshold) {
424 		if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
425 		    (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
426 			/* And don't move a CPU_ACCESS_REQUIRED BO to limited
427 			 * visible VRAM if we've depleted our allowance to do
428 			 * that.
429 			 */
430 			if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
431 				domain = bo->preferred_domains;
432 			else
433 				domain = bo->allowed_domains;
434 		} else {
435 			domain = bo->preferred_domains;
436 		}
437 	} else {
438 		domain = bo->allowed_domains;
439 	}
440 
441 retry:
442 	amdgpu_bo_placement_from_domain(bo, domain);
443 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
444 
445 	p->bytes_moved += ctx.bytes_moved;
446 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
447 	    amdgpu_bo_in_cpu_visible_vram(bo))
448 		p->bytes_moved_vis += ctx.bytes_moved;
449 
450 	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
451 		domain = bo->allowed_domains;
452 		goto retry;
453 	}
454 
455 	return r;
456 }
457 
amdgpu_cs_validate(void * param,struct amdgpu_bo * bo)458 static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
459 {
460 	struct amdgpu_cs_parser *p = param;
461 	int r;
462 
463 	r = amdgpu_cs_bo_validate(p, bo);
464 	if (r)
465 		return r;
466 
467 	if (bo->shadow)
468 		r = amdgpu_cs_bo_validate(p, bo->shadow);
469 
470 	return r;
471 }
472 
amdgpu_cs_list_validate(struct amdgpu_cs_parser * p,struct list_head * validated)473 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
474 			    struct list_head *validated)
475 {
476 	struct ttm_operation_ctx ctx = { true, false };
477 	struct amdgpu_bo_list_entry *lobj;
478 	int r;
479 
480 	list_for_each_entry(lobj, validated, tv.head) {
481 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
482 #ifdef __NetBSD__
483 		struct vmspace *usermm;
484 #else
485 		struct mm_struct *usermm;
486 #endif
487 
488 		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
489 #ifdef __NetBSD__
490 		if (usermm && usermm != curproc->p_vmspace)
491 #else
492 		if (usermm && usermm != current->mm)
493 #endif
494 			return -EPERM;
495 
496 		if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
497 		    lobj->user_invalidated && lobj->user_pages) {
498 			amdgpu_bo_placement_from_domain(bo,
499 							AMDGPU_GEM_DOMAIN_CPU);
500 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
501 			if (r)
502 				return r;
503 
504 			amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
505 						     lobj->user_pages);
506 		}
507 
508 		r = amdgpu_cs_validate(p, bo);
509 		if (r)
510 			return r;
511 
512 		kvfree(lobj->user_pages);
513 		lobj->user_pages = NULL;
514 	}
515 	return 0;
516 }
517 
amdgpu_cs_parser_bos(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)518 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
519 				union drm_amdgpu_cs *cs)
520 {
521 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
522 	struct amdgpu_vm *vm = &fpriv->vm;
523 	struct amdgpu_bo_list_entry *e;
524 	struct list_head duplicates;
525 	struct amdgpu_bo *gds;
526 	struct amdgpu_bo *gws;
527 	struct amdgpu_bo *oa;
528 	int r;
529 
530 	INIT_LIST_HEAD(&p->validated);
531 
532 	/* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
533 	if (cs->in.bo_list_handle) {
534 		if (p->bo_list)
535 			return -EINVAL;
536 
537 		r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
538 				       &p->bo_list);
539 		if (r)
540 			return r;
541 	} else if (!p->bo_list) {
542 		/* Create a empty bo_list when no handle is provided */
543 		r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
544 					  &p->bo_list);
545 		if (r)
546 			return r;
547 	}
548 
549 	/* One for TTM and one for the CS job */
550 	amdgpu_bo_list_for_each_entry(e, p->bo_list)
551 		e->tv.num_shared = 2;
552 
553 	amdgpu_bo_list_get_list(p->bo_list, &p->validated);
554 
555 	INIT_LIST_HEAD(&duplicates);
556 	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
557 
558 	if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
559 		list_add(&p->uf_entry.tv.head, &p->validated);
560 
561 	/* Get userptr backing pages. If pages are updated after registered
562 	 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
563 	 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
564 	 */
565 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
566 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
567 		bool userpage_invalidated = false;
568 		int i;
569 
570 		e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
571 					sizeof(struct page *),
572 					GFP_KERNEL | __GFP_ZERO);
573 		if (!e->user_pages) {
574 			DRM_ERROR("calloc failure\n");
575 			return -ENOMEM;
576 		}
577 
578 		r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
579 		if (r) {
580 			kvfree(e->user_pages);
581 			e->user_pages = NULL;
582 			return r;
583 		}
584 
585 		for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
586 			if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
587 				userpage_invalidated = true;
588 				break;
589 			}
590 		}
591 		e->user_invalidated = userpage_invalidated;
592 	}
593 
594 	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
595 				   &duplicates);
596 	if (unlikely(r != 0)) {
597 		if (r != -ERESTARTSYS)
598 			DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
599 		goto out;
600 	}
601 
602 	amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
603 					  &p->bytes_moved_vis_threshold);
604 	p->bytes_moved = 0;
605 	p->bytes_moved_vis = 0;
606 
607 	r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
608 				      amdgpu_cs_validate, p);
609 	if (r) {
610 		DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
611 		goto error_validate;
612 	}
613 
614 	r = amdgpu_cs_list_validate(p, &duplicates);
615 	if (r)
616 		goto error_validate;
617 
618 	r = amdgpu_cs_list_validate(p, &p->validated);
619 	if (r)
620 		goto error_validate;
621 
622 	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
623 				     p->bytes_moved_vis);
624 
625 	gds = p->bo_list->gds_obj;
626 	gws = p->bo_list->gws_obj;
627 	oa = p->bo_list->oa_obj;
628 
629 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
630 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
631 
632 		/* Make sure we use the exclusive slot for shared BOs */
633 		if (bo->prime_shared_count)
634 			e->tv.num_shared = 0;
635 		e->bo_va = amdgpu_vm_bo_find(vm, bo);
636 	}
637 
638 	if (gds) {
639 		p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
640 		p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
641 	}
642 	if (gws) {
643 		p->job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
644 		p->job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
645 	}
646 	if (oa) {
647 		p->job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
648 		p->job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
649 	}
650 
651 	if (!r && p->uf_entry.tv.bo) {
652 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
653 
654 		r = amdgpu_ttm_alloc_gart(&uf->tbo);
655 		p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
656 	}
657 
658 error_validate:
659 	if (r)
660 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
661 out:
662 	return r;
663 }
664 
amdgpu_cs_sync_rings(struct amdgpu_cs_parser * p)665 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
666 {
667 	struct amdgpu_bo_list_entry *e;
668 	int r;
669 
670 	list_for_each_entry(e, &p->validated, tv.head) {
671 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
672 		struct dma_resv *resv = bo->tbo.base.resv;
673 
674 		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp,
675 				     amdgpu_bo_explicit_sync(bo));
676 
677 		if (r)
678 			return r;
679 	}
680 	return 0;
681 }
682 
683 /**
684  * cs_parser_fini() - clean parser states
685  * @parser:	parser structure holding parsing context.
686  * @error:	error number
687  *
688  * If error is set than 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 
700 	for (i = 0; i < parser->num_post_deps; i++) {
701 		drm_syncobj_put(parser->post_deps[i].syncobj);
702 		kfree(parser->post_deps[i].chain);
703 	}
704 	kfree(parser->post_deps);
705 
706 	dma_fence_put(parser->fence);
707 
708 	if (parser->ctx) {
709 		mutex_unlock(&parser->ctx->lock);
710 		amdgpu_ctx_put(parser->ctx);
711 	}
712 	if (parser->bo_list)
713 		amdgpu_bo_list_put(parser->bo_list);
714 
715 	for (i = 0; i < parser->nchunks; i++)
716 		kvfree(parser->chunks[i].kdata);
717 	kfree(parser->chunks);
718 	if (parser->job)
719 		amdgpu_job_free(parser->job);
720 	if (parser->uf_entry.tv.bo) {
721 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
722 
723 		amdgpu_bo_unref(&uf);
724 	}
725 }
726 
amdgpu_cs_vm_handling(struct amdgpu_cs_parser * p)727 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
728 {
729 	struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
730 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
731 	struct amdgpu_device *adev = p->adev;
732 	struct amdgpu_vm *vm = &fpriv->vm;
733 	struct amdgpu_bo_list_entry *e;
734 	struct amdgpu_bo_va *bo_va;
735 	struct amdgpu_bo *bo;
736 	int r;
737 
738 	/* Only for UVD/VCE VM emulation */
739 	if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
740 		unsigned i, j;
741 
742 		for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
743 			struct drm_amdgpu_cs_chunk_ib *chunk_ib;
744 			struct amdgpu_bo_va_mapping *m;
745 			struct amdgpu_bo *aobj = NULL;
746 			struct amdgpu_cs_chunk *chunk;
747 			uint64_t offset, va_start;
748 			struct amdgpu_ib *ib;
749 			uint8_t *kptr;
750 
751 			chunk = &p->chunks[i];
752 			ib = &p->job->ibs[j];
753 			chunk_ib = chunk->kdata;
754 
755 			if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
756 				continue;
757 
758 			va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
759 			r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
760 			if (r) {
761 				DRM_ERROR("IB va_start is invalid\n");
762 				return r;
763 			}
764 
765 			if ((va_start + chunk_ib->ib_bytes) >
766 			    (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
767 				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
768 				return -EINVAL;
769 			}
770 
771 			/* the IB should be reserved at this point */
772 			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
773 			if (r) {
774 				return r;
775 			}
776 
777 			offset = m->start * AMDGPU_GPU_PAGE_SIZE;
778 			kptr += va_start - offset;
779 
780 			if (ring->funcs->parse_cs) {
781 				memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
782 				amdgpu_bo_kunmap(aobj);
783 
784 				r = amdgpu_ring_parse_cs(ring, p, j);
785 				if (r)
786 					return r;
787 			} else {
788 				ib->ptr = (uint32_t *)kptr;
789 				r = amdgpu_ring_patch_cs_in_place(ring, p, j);
790 				amdgpu_bo_kunmap(aobj);
791 				if (r)
792 					return r;
793 			}
794 
795 			j++;
796 		}
797 	}
798 
799 	if (!p->job->vm)
800 		return amdgpu_cs_sync_rings(p);
801 
802 
803 	r = amdgpu_vm_clear_freed(adev, vm, NULL);
804 	if (r)
805 		return r;
806 
807 	r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
808 	if (r)
809 		return r;
810 
811 	r = amdgpu_sync_vm_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
812 	if (r)
813 		return r;
814 
815 	if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
816 		bo_va = fpriv->csa_va;
817 		BUG_ON(!bo_va);
818 		r = amdgpu_vm_bo_update(adev, bo_va, false);
819 		if (r)
820 			return r;
821 
822 		r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
823 		if (r)
824 			return r;
825 	}
826 
827 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
828 		/* ignore duplicates */
829 		bo = ttm_to_amdgpu_bo(e->tv.bo);
830 		if (!bo)
831 			continue;
832 
833 		bo_va = e->bo_va;
834 		if (bo_va == NULL)
835 			continue;
836 
837 		r = amdgpu_vm_bo_update(adev, bo_va, false);
838 		if (r)
839 			return r;
840 
841 		r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
842 		if (r)
843 			return r;
844 	}
845 
846 	r = amdgpu_vm_handle_moved(adev, vm);
847 	if (r)
848 		return r;
849 
850 	r = amdgpu_vm_update_pdes(adev, vm, false);
851 	if (r)
852 		return r;
853 
854 	r = amdgpu_sync_vm_fence(&p->job->sync, vm->last_update);
855 	if (r)
856 		return r;
857 
858 	p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.base.bo);
859 
860 	if (amdgpu_vm_debug) {
861 		/* Invalidate all BOs to test for userspace bugs */
862 		amdgpu_bo_list_for_each_entry(e, p->bo_list) {
863 			struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
864 
865 			/* ignore duplicates */
866 			if (!bo)
867 				continue;
868 
869 			amdgpu_vm_bo_invalidate(adev, bo, false);
870 		}
871 	}
872 
873 	return amdgpu_cs_sync_rings(p);
874 }
875 
amdgpu_cs_ib_fill(struct amdgpu_device * adev,struct amdgpu_cs_parser * parser)876 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
877 			     struct amdgpu_cs_parser *parser)
878 {
879 	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
880 	struct amdgpu_vm *vm = &fpriv->vm;
881 	int r, ce_preempt = 0, de_preempt = 0;
882 	struct amdgpu_ring *ring;
883 	int i, j;
884 
885 	for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
886 		struct amdgpu_cs_chunk *chunk;
887 		struct amdgpu_ib *ib;
888 		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
889 		struct drm_sched_entity *entity;
890 
891 		chunk = &parser->chunks[i];
892 		ib = &parser->job->ibs[j];
893 		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
894 
895 		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
896 			continue;
897 
898 		if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
899 		    (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
900 			if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
901 				if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
902 					ce_preempt++;
903 				else
904 					de_preempt++;
905 			}
906 
907 			/* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
908 			if (ce_preempt > 1 || de_preempt > 1)
909 				return -EINVAL;
910 		}
911 
912 		r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
913 					  chunk_ib->ip_instance, chunk_ib->ring,
914 					  &entity);
915 		if (r)
916 			return r;
917 
918 		if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
919 			parser->job->preamble_status |=
920 				AMDGPU_PREAMBLE_IB_PRESENT;
921 
922 		if (parser->entity && parser->entity != entity)
923 			return -EINVAL;
924 
925 		/* Return if there is no run queue associated with this entity.
926 		 * Possibly because of disabled HW IP*/
927 		if (entity->rq == NULL)
928 			return -EINVAL;
929 
930 		parser->entity = entity;
931 
932 		ring = to_amdgpu_ring(entity->rq->sched);
933 		r =  amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
934 				   chunk_ib->ib_bytes : 0, ib);
935 		if (r) {
936 			DRM_ERROR("Failed to get ib !\n");
937 			return r;
938 		}
939 
940 		ib->gpu_addr = chunk_ib->va_start;
941 		ib->length_dw = chunk_ib->ib_bytes / 4;
942 		ib->flags = chunk_ib->flags;
943 
944 		j++;
945 	}
946 
947 	/* MM engine doesn't support user fences */
948 	ring = to_amdgpu_ring(parser->entity->rq->sched);
949 	if (parser->job->uf_addr && ring->funcs->no_user_fence)
950 		return -EINVAL;
951 
952 	return amdgpu_ctx_wait_prev_fence(parser->ctx, parser->entity);
953 }
954 
amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)955 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
956 				       struct amdgpu_cs_chunk *chunk)
957 {
958 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
959 	unsigned num_deps;
960 	int i, r;
961 	struct drm_amdgpu_cs_chunk_dep *deps;
962 
963 	deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
964 	num_deps = chunk->length_dw * 4 /
965 		sizeof(struct drm_amdgpu_cs_chunk_dep);
966 
967 	for (i = 0; i < num_deps; ++i) {
968 		struct amdgpu_ctx *ctx;
969 		struct drm_sched_entity *entity;
970 		struct dma_fence *fence;
971 
972 		ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
973 		if (ctx == NULL)
974 			return -EINVAL;
975 
976 		r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
977 					  deps[i].ip_instance,
978 					  deps[i].ring, &entity);
979 		if (r) {
980 			amdgpu_ctx_put(ctx);
981 			return r;
982 		}
983 
984 		fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
985 		amdgpu_ctx_put(ctx);
986 
987 		if (IS_ERR(fence))
988 			return PTR_ERR(fence);
989 		else if (!fence)
990 			continue;
991 
992 		if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
993 			struct drm_sched_fence *s_fence;
994 			struct dma_fence *old = fence;
995 
996 			s_fence = to_drm_sched_fence(fence);
997 			fence = dma_fence_get(&s_fence->scheduled);
998 			dma_fence_put(old);
999 		}
1000 
1001 		r = amdgpu_sync_fence(&p->job->sync, fence, true);
1002 		dma_fence_put(fence);
1003 		if (r)
1004 			return r;
1005 	}
1006 	return 0;
1007 }
1008 
amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser * p,uint32_t handle,u64 point,u64 flags)1009 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1010 						 uint32_t handle, u64 point,
1011 						 u64 flags)
1012 {
1013 	struct dma_fence *fence;
1014 	int r;
1015 
1016 	r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
1017 	if (r) {
1018 		DRM_ERROR("syncobj %u failed to find fence @ %"PRIu64" (%d)!\n",
1019 			  handle, point, r);
1020 		return r;
1021 	}
1022 
1023 	r = amdgpu_sync_fence(&p->job->sync, fence, true);
1024 	dma_fence_put(fence);
1025 
1026 	return r;
1027 }
1028 
amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)1029 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1030 					    struct amdgpu_cs_chunk *chunk)
1031 {
1032 	struct drm_amdgpu_cs_chunk_sem *deps;
1033 	unsigned num_deps;
1034 	int i, r;
1035 
1036 	deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1037 	num_deps = chunk->length_dw * 4 /
1038 		sizeof(struct drm_amdgpu_cs_chunk_sem);
1039 	for (i = 0; i < num_deps; ++i) {
1040 		r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
1041 							  0, 0);
1042 		if (r)
1043 			return r;
1044 	}
1045 
1046 	return 0;
1047 }
1048 
1049 
amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)1050 static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
1051 						     struct amdgpu_cs_chunk *chunk)
1052 {
1053 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1054 	unsigned num_deps;
1055 	int i, r;
1056 
1057 	syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1058 	num_deps = chunk->length_dw * 4 /
1059 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1060 	for (i = 0; i < num_deps; ++i) {
1061 		r = amdgpu_syncobj_lookup_and_add_to_sync(p,
1062 							  syncobj_deps[i].handle,
1063 							  syncobj_deps[i].point,
1064 							  syncobj_deps[i].flags);
1065 		if (r)
1066 			return r;
1067 	}
1068 
1069 	return 0;
1070 }
1071 
amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)1072 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1073 					     struct amdgpu_cs_chunk *chunk)
1074 {
1075 	struct drm_amdgpu_cs_chunk_sem *deps;
1076 	unsigned num_deps;
1077 	int i;
1078 
1079 	deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1080 	num_deps = chunk->length_dw * 4 /
1081 		sizeof(struct drm_amdgpu_cs_chunk_sem);
1082 
1083 	if (p->post_deps)
1084 		return -EINVAL;
1085 
1086 	p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1087 				     GFP_KERNEL);
1088 	p->num_post_deps = 0;
1089 
1090 	if (!p->post_deps)
1091 		return -ENOMEM;
1092 
1093 
1094 	for (i = 0; i < num_deps; ++i) {
1095 		p->post_deps[i].syncobj =
1096 			drm_syncobj_find(p->filp, deps[i].handle);
1097 		if (!p->post_deps[i].syncobj)
1098 			return -EINVAL;
1099 		p->post_deps[i].chain = NULL;
1100 		p->post_deps[i].point = 0;
1101 		p->num_post_deps++;
1102 	}
1103 
1104 	return 0;
1105 }
1106 
1107 
amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser * p,struct amdgpu_cs_chunk * chunk)1108 static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
1109 						      struct amdgpu_cs_chunk *chunk)
1110 {
1111 	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1112 	unsigned num_deps;
1113 	int i;
1114 
1115 	syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1116 	num_deps = chunk->length_dw * 4 /
1117 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1118 
1119 	if (p->post_deps)
1120 		return -EINVAL;
1121 
1122 	p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1123 				     GFP_KERNEL);
1124 	p->num_post_deps = 0;
1125 
1126 	if (!p->post_deps)
1127 		return -ENOMEM;
1128 
1129 	for (i = 0; i < num_deps; ++i) {
1130 		struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
1131 
1132 		dep->chain = NULL;
1133 		if (syncobj_deps[i].point) {
1134 			dep->chain = kmalloc(sizeof(*dep->chain), GFP_KERNEL);
1135 			if (!dep->chain)
1136 				return -ENOMEM;
1137 		}
1138 
1139 		dep->syncobj = drm_syncobj_find(p->filp,
1140 						syncobj_deps[i].handle);
1141 		if (!dep->syncobj) {
1142 			kfree(dep->chain);
1143 			return -EINVAL;
1144 		}
1145 		dep->point = syncobj_deps[i].point;
1146 		p->num_post_deps++;
1147 	}
1148 
1149 	return 0;
1150 }
1151 
amdgpu_cs_dependencies(struct amdgpu_device * adev,struct amdgpu_cs_parser * p)1152 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1153 				  struct amdgpu_cs_parser *p)
1154 {
1155 	int i, r;
1156 
1157 	for (i = 0; i < p->nchunks; ++i) {
1158 		struct amdgpu_cs_chunk *chunk;
1159 
1160 		chunk = &p->chunks[i];
1161 
1162 		switch (chunk->chunk_id) {
1163 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
1164 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
1165 			r = amdgpu_cs_process_fence_dep(p, chunk);
1166 			if (r)
1167 				return r;
1168 			break;
1169 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
1170 			r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1171 			if (r)
1172 				return r;
1173 			break;
1174 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
1175 			r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1176 			if (r)
1177 				return r;
1178 			break;
1179 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
1180 			r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
1181 			if (r)
1182 				return r;
1183 			break;
1184 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
1185 			r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
1186 			if (r)
1187 				return r;
1188 			break;
1189 		}
1190 	}
1191 
1192 	return 0;
1193 }
1194 
amdgpu_cs_post_dependencies(struct amdgpu_cs_parser * p)1195 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1196 {
1197 	int i;
1198 
1199 	for (i = 0; i < p->num_post_deps; ++i) {
1200 		if (p->post_deps[i].chain && p->post_deps[i].point) {
1201 			drm_syncobj_add_point(p->post_deps[i].syncobj,
1202 					      p->post_deps[i].chain,
1203 					      p->fence, p->post_deps[i].point);
1204 			p->post_deps[i].chain = NULL;
1205 		} else {
1206 			drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1207 						  p->fence);
1208 		}
1209 	}
1210 }
1211 
amdgpu_cs_submit(struct amdgpu_cs_parser * p,union drm_amdgpu_cs * cs)1212 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1213 			    union drm_amdgpu_cs *cs)
1214 {
1215 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1216 	struct drm_sched_entity *entity = p->entity;
1217 	enum drm_sched_priority priority;
1218 	struct amdgpu_ring *ring;
1219 	struct amdgpu_bo_list_entry *e;
1220 	struct amdgpu_job *job;
1221 	uint64_t seq;
1222 	int r;
1223 
1224 	job = p->job;
1225 	p->job = NULL;
1226 
1227 	r = drm_sched_job_init(&job->base, entity, p->filp);
1228 	if (r)
1229 		goto error_unlock;
1230 
1231 	/* No memory allocation is allowed while holding the notifier lock.
1232 	 * The lock is held until amdgpu_cs_submit is finished and fence is
1233 	 * added to BOs.
1234 	 */
1235 	mutex_lock(&p->adev->notifier_lock);
1236 
1237 	/* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1238 	 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1239 	 */
1240 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1241 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1242 
1243 		r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1244 	}
1245 	if (r) {
1246 		r = -EAGAIN;
1247 		goto error_abort;
1248 	}
1249 
1250 	p->fence = dma_fence_get(&job->base.s_fence->finished);
1251 
1252 	amdgpu_ctx_add_fence(p->ctx, entity, p->fence, &seq);
1253 	amdgpu_cs_post_dependencies(p);
1254 
1255 	if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1256 	    !p->ctx->preamble_presented) {
1257 		job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1258 		p->ctx->preamble_presented = true;
1259 	}
1260 
1261 	cs->out.handle = seq;
1262 	job->uf_sequence = seq;
1263 
1264 	amdgpu_job_free_resources(job);
1265 
1266 	trace_amdgpu_cs_ioctl(job);
1267 	amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1268 	priority = job->base.s_priority;
1269 	drm_sched_entity_push_job(&job->base, entity);
1270 
1271 	ring = to_amdgpu_ring(entity->rq->sched);
1272 	amdgpu_ring_priority_get(ring, priority);
1273 
1274 	amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1275 
1276 	ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1277 	mutex_unlock(&p->adev->notifier_lock);
1278 
1279 	return 0;
1280 
1281 error_abort:
1282 	drm_sched_job_cleanup(&job->base);
1283 	mutex_unlock(&p->adev->notifier_lock);
1284 
1285 error_unlock:
1286 	amdgpu_job_free(job);
1287 	return r;
1288 }
1289 
amdgpu_cs_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1290 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1291 {
1292 	struct amdgpu_device *adev = dev->dev_private;
1293 	union drm_amdgpu_cs *cs = data;
1294 	struct amdgpu_cs_parser parser = {};
1295 	bool reserved_buffers = false;
1296 	int i, r;
1297 
1298 	if (amdgpu_ras_intr_triggered())
1299 		return -EHWPOISON;
1300 
1301 	if (!adev->accel_working)
1302 		return -EBUSY;
1303 
1304 	parser.adev = adev;
1305 	parser.filp = filp;
1306 
1307 	r = amdgpu_cs_parser_init(&parser, data);
1308 	if (r) {
1309 		DRM_ERROR("Failed to initialize parser %d!\n", r);
1310 		goto out;
1311 	}
1312 
1313 	r = amdgpu_cs_ib_fill(adev, &parser);
1314 	if (r)
1315 		goto out;
1316 
1317 	r = amdgpu_cs_dependencies(adev, &parser);
1318 	if (r) {
1319 		DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1320 		goto out;
1321 	}
1322 
1323 	r = amdgpu_cs_parser_bos(&parser, data);
1324 	if (r) {
1325 		if (r == -ENOMEM)
1326 			DRM_ERROR("Not enough memory for command submission!\n");
1327 		else if (r != -ERESTARTSYS && r != -EAGAIN)
1328 			DRM_ERROR("Failed to process the buffer list %d!\n", r);
1329 		goto out;
1330 	}
1331 
1332 	reserved_buffers = true;
1333 
1334 	for (i = 0; i < parser.job->num_ibs; i++)
1335 		trace_amdgpu_cs(&parser, i);
1336 
1337 	r = amdgpu_cs_vm_handling(&parser);
1338 	if (r)
1339 		goto out;
1340 
1341 	r = amdgpu_cs_submit(&parser, cs);
1342 
1343 out:
1344 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1345 
1346 	return r;
1347 }
1348 
1349 /**
1350  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1351  *
1352  * @dev: drm device
1353  * @data: data from userspace
1354  * @filp: file private
1355  *
1356  * Wait for the command submission identified by handle to finish.
1357  */
amdgpu_cs_wait_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1358 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1359 			 struct drm_file *filp)
1360 {
1361 	union drm_amdgpu_wait_cs *wait = data;
1362 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1363 	struct drm_sched_entity *entity;
1364 	struct amdgpu_ctx *ctx;
1365 	struct dma_fence *fence;
1366 	long r;
1367 
1368 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1369 	if (ctx == NULL)
1370 		return -EINVAL;
1371 
1372 	r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1373 				  wait->in.ring, &entity);
1374 	if (r) {
1375 		amdgpu_ctx_put(ctx);
1376 		return r;
1377 	}
1378 
1379 	fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1380 	if (IS_ERR(fence))
1381 		r = PTR_ERR(fence);
1382 	else if (fence) {
1383 		r = dma_fence_wait_timeout(fence, true, timeout);
1384 		if (r > 0 && fence->error)
1385 			r = fence->error;
1386 		dma_fence_put(fence);
1387 	} else
1388 		r = 1;
1389 
1390 	amdgpu_ctx_put(ctx);
1391 	if (r < 0)
1392 		return r;
1393 
1394 	memset(wait, 0, sizeof(*wait));
1395 	wait->out.status = (r == 0);
1396 
1397 	return 0;
1398 }
1399 
1400 /**
1401  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1402  *
1403  * @adev: amdgpu device
1404  * @filp: file private
1405  * @user: drm_amdgpu_fence copied from user space
1406  */
amdgpu_cs_get_fence(struct amdgpu_device * adev,struct drm_file * filp,struct drm_amdgpu_fence * user)1407 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1408 					     struct drm_file *filp,
1409 					     struct drm_amdgpu_fence *user)
1410 {
1411 	struct drm_sched_entity *entity;
1412 	struct amdgpu_ctx *ctx;
1413 	struct dma_fence *fence;
1414 	int r;
1415 
1416 	ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1417 	if (ctx == NULL)
1418 		return ERR_PTR(-EINVAL);
1419 
1420 	r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1421 				  user->ring, &entity);
1422 	if (r) {
1423 		amdgpu_ctx_put(ctx);
1424 		return ERR_PTR(r);
1425 	}
1426 
1427 	fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1428 	amdgpu_ctx_put(ctx);
1429 
1430 	return fence;
1431 }
1432 
amdgpu_cs_fence_to_handle_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1433 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1434 				    struct drm_file *filp)
1435 {
1436 	struct amdgpu_device *adev = dev->dev_private;
1437 	union drm_amdgpu_fence_to_handle *info = data;
1438 	struct dma_fence *fence;
1439 	struct drm_syncobj *syncobj;
1440 	struct sync_file *sync_file;
1441 	int fd, r;
1442 
1443 	fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1444 	if (IS_ERR(fence))
1445 		return PTR_ERR(fence);
1446 
1447 	if (!fence)
1448 		fence = dma_fence_get_stub();
1449 
1450 	switch (info->in.what) {
1451 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1452 		r = drm_syncobj_create(&syncobj, 0, fence);
1453 		dma_fence_put(fence);
1454 		if (r)
1455 			return r;
1456 		r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1457 		drm_syncobj_put(syncobj);
1458 		return r;
1459 
1460 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1461 		r = drm_syncobj_create(&syncobj, 0, fence);
1462 		dma_fence_put(fence);
1463 		if (r)
1464 			return r;
1465 		r = drm_syncobj_get_fd(syncobj, (int*)&info->out.handle);
1466 		drm_syncobj_put(syncobj);
1467 		return r;
1468 
1469 	case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1470 #ifdef __NetBSD__
1471 	   {
1472 		struct file *fp = NULL;
1473 
1474 		/* XXX errno NetBSD->Linux */
1475 		r = -fd_allocfile(&fp, &fd);
1476 		if (r)
1477 			goto out;
1478 		sync_file = sync_file_create(fence, fp);
1479 		if (sync_file == NULL)
1480 			goto out;
1481 		fd_affix(curproc, fp, fd);
1482 		fp = NULL;	/* consumed by sync_file */
1483 
1484 out:		if (fp) {
1485 			fd_abort(curproc, fp, fd);
1486 			fd = -1;
1487 		}
1488 		dma_fence_put(fence);
1489 	   }
1490 #else
1491 		fd = get_unused_fd_flags(O_CLOEXEC);
1492 		if (fd < 0) {
1493 			dma_fence_put(fence);
1494 			return fd;
1495 		}
1496 
1497 		sync_file = sync_file_create(fence);
1498 		dma_fence_put(fence);
1499 		if (!sync_file) {
1500 			put_unused_fd(fd);
1501 			return -ENOMEM;
1502 		}
1503 
1504 		fd_install(fd, sync_file->file);
1505 #endif
1506 		info->out.handle = fd;
1507 		return 0;
1508 
1509 	default:
1510 		return -EINVAL;
1511 	}
1512 }
1513 
1514 /**
1515  * amdgpu_cs_wait_all_fence - wait on all fences to signal
1516  *
1517  * @adev: amdgpu device
1518  * @filp: file private
1519  * @wait: wait parameters
1520  * @fences: array of drm_amdgpu_fence
1521  */
amdgpu_cs_wait_all_fences(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1522 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1523 				     struct drm_file *filp,
1524 				     union drm_amdgpu_wait_fences *wait,
1525 				     struct drm_amdgpu_fence *fences)
1526 {
1527 	uint32_t fence_count = wait->in.fence_count;
1528 	unsigned int i;
1529 	long r = 1;
1530 
1531 	for (i = 0; i < fence_count; i++) {
1532 		struct dma_fence *fence;
1533 		unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1534 
1535 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1536 		if (IS_ERR(fence))
1537 			return PTR_ERR(fence);
1538 		else if (!fence)
1539 			continue;
1540 
1541 		r = dma_fence_wait_timeout(fence, true, timeout);
1542 		dma_fence_put(fence);
1543 		if (r < 0)
1544 			return r;
1545 
1546 		if (r == 0)
1547 			break;
1548 
1549 		if (fence->error)
1550 			return fence->error;
1551 	}
1552 
1553 	memset(wait, 0, sizeof(*wait));
1554 	wait->out.status = (r > 0);
1555 
1556 	return 0;
1557 }
1558 
1559 /**
1560  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1561  *
1562  * @adev: amdgpu device
1563  * @filp: file private
1564  * @wait: wait parameters
1565  * @fences: array of drm_amdgpu_fence
1566  */
amdgpu_cs_wait_any_fence(struct amdgpu_device * adev,struct drm_file * filp,union drm_amdgpu_wait_fences * wait,struct drm_amdgpu_fence * fences)1567 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1568 				    struct drm_file *filp,
1569 				    union drm_amdgpu_wait_fences *wait,
1570 				    struct drm_amdgpu_fence *fences)
1571 {
1572 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1573 	uint32_t fence_count = wait->in.fence_count;
1574 	uint32_t first = ~0;
1575 	struct dma_fence **array;
1576 	unsigned int i;
1577 	long r;
1578 
1579 	/* Prepare the fence array */
1580 	array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1581 
1582 	if (array == NULL)
1583 		return -ENOMEM;
1584 
1585 	for (i = 0; i < fence_count; i++) {
1586 		struct dma_fence *fence;
1587 
1588 		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1589 		if (IS_ERR(fence)) {
1590 			r = PTR_ERR(fence);
1591 			goto err_free_fence_array;
1592 		} else if (fence) {
1593 			array[i] = fence;
1594 		} else { /* NULL, the fence has been already signaled */
1595 			r = 1;
1596 			first = i;
1597 			goto out;
1598 		}
1599 	}
1600 
1601 	r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1602 				       &first);
1603 	if (r < 0)
1604 		goto err_free_fence_array;
1605 
1606 out:
1607 	memset(wait, 0, sizeof(*wait));
1608 	wait->out.status = (r > 0);
1609 	wait->out.first_signaled = first;
1610 
1611 	if (first < fence_count && array[first])
1612 		r = array[first]->error;
1613 	else
1614 		r = 0;
1615 
1616 err_free_fence_array:
1617 	for (i = 0; i < fence_count; i++)
1618 		dma_fence_put(array[i]);
1619 	kfree(array);
1620 
1621 	return r;
1622 }
1623 
1624 /**
1625  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1626  *
1627  * @dev: drm device
1628  * @data: data from userspace
1629  * @filp: file private
1630  */
amdgpu_cs_wait_fences_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)1631 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1632 				struct drm_file *filp)
1633 {
1634 	struct amdgpu_device *adev = dev->dev_private;
1635 	union drm_amdgpu_wait_fences *wait = data;
1636 	uint32_t fence_count = wait->in.fence_count;
1637 	struct drm_amdgpu_fence *fences_user;
1638 	struct drm_amdgpu_fence *fences;
1639 	int r;
1640 
1641 	/* Get the fences from userspace */
1642 	fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1643 			GFP_KERNEL);
1644 	if (fences == NULL)
1645 		return -ENOMEM;
1646 
1647 	fences_user = u64_to_user_ptr(wait->in.fences);
1648 	if (copy_from_user(fences, fences_user,
1649 		sizeof(struct drm_amdgpu_fence) * fence_count)) {
1650 		r = -EFAULT;
1651 		goto err_free_fences;
1652 	}
1653 
1654 	if (wait->in.wait_all)
1655 		r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1656 	else
1657 		r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1658 
1659 err_free_fences:
1660 	kfree(fences);
1661 
1662 	return r;
1663 }
1664 
1665 /**
1666  * amdgpu_cs_find_bo_va - find bo_va for VM address
1667  *
1668  * @parser: command submission parser context
1669  * @addr: VM address
1670  * @bo: resulting BO of the mapping found
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