xref: /openbsd/sys/dev/pci/drm/radeon/radeon_cs.c (revision 73471bf0)
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/list_sort.h>
29 #include <linux/pci.h>
30 #include <linux/uaccess.h>
31 
32 #include <drm/drm_device.h>
33 #include <drm/drm_file.h>
34 #include <drm/radeon_drm.h>
35 
36 #include "radeon.h"
37 #include "radeon_reg.h"
38 #include "radeon_trace.h"
39 
40 #define RADEON_CS_MAX_PRIORITY		32u
41 #define RADEON_CS_NUM_BUCKETS		(RADEON_CS_MAX_PRIORITY + 1)
42 
43 /* This is based on the bucket sort with O(n) time complexity.
44  * An item with priority "i" is added to bucket[i]. The lists are then
45  * concatenated in descending order.
46  */
47 struct radeon_cs_buckets {
48 	struct list_head bucket[RADEON_CS_NUM_BUCKETS];
49 };
50 
51 static void radeon_cs_buckets_init(struct radeon_cs_buckets *b)
52 {
53 	unsigned i;
54 
55 	for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++)
56 		INIT_LIST_HEAD(&b->bucket[i]);
57 }
58 
59 static void radeon_cs_buckets_add(struct radeon_cs_buckets *b,
60 				  struct list_head *item, unsigned priority)
61 {
62 	/* Since buffers which appear sooner in the relocation list are
63 	 * likely to be used more often than buffers which appear later
64 	 * in the list, the sort mustn't change the ordering of buffers
65 	 * with the same priority, i.e. it must be stable.
66 	 */
67 	list_add_tail(item, &b->bucket[min(priority, RADEON_CS_MAX_PRIORITY)]);
68 }
69 
70 static void radeon_cs_buckets_get_list(struct radeon_cs_buckets *b,
71 				       struct list_head *out_list)
72 {
73 	unsigned i;
74 
75 	/* Connect the sorted buckets in the output list. */
76 	for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++) {
77 		list_splice(&b->bucket[i], out_list);
78 	}
79 }
80 
81 static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
82 {
83 	struct radeon_cs_chunk *chunk;
84 	struct radeon_cs_buckets buckets;
85 	unsigned i;
86 	bool need_mmap_lock = false;
87 	int r;
88 
89 	if (p->chunk_relocs == NULL) {
90 		return 0;
91 	}
92 	chunk = p->chunk_relocs;
93 	p->dma_reloc_idx = 0;
94 	/* FIXME: we assume that each relocs use 4 dwords */
95 	p->nrelocs = chunk->length_dw / 4;
96 	p->relocs = kvmalloc_array(p->nrelocs, sizeof(struct radeon_bo_list),
97 			GFP_KERNEL | __GFP_ZERO);
98 	if (p->relocs == NULL) {
99 		return -ENOMEM;
100 	}
101 
102 	radeon_cs_buckets_init(&buckets);
103 
104 	for (i = 0; i < p->nrelocs; i++) {
105 		struct drm_radeon_cs_reloc *r;
106 		struct drm_gem_object *gobj;
107 		unsigned priority;
108 
109 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
110 		gobj = drm_gem_object_lookup(p->filp, r->handle);
111 		if (gobj == NULL) {
112 			DRM_ERROR("gem object lookup failed 0x%x\n",
113 				  r->handle);
114 			return -ENOENT;
115 		}
116 		p->relocs[i].robj = gem_to_radeon_bo(gobj);
117 
118 		/* The userspace buffer priorities are from 0 to 15. A higher
119 		 * number means the buffer is more important.
120 		 * Also, the buffers used for write have a higher priority than
121 		 * the buffers used for read only, which doubles the range
122 		 * to 0 to 31. 32 is reserved for the kernel driver.
123 		 */
124 		priority = (r->flags & RADEON_RELOC_PRIO_MASK) * 2
125 			   + !!r->write_domain;
126 
127 		/* The first reloc of an UVD job is the msg and that must be in
128 		 * VRAM, the second reloc is the DPB and for WMV that must be in
129 		 * VRAM as well. Also put everything into VRAM on AGP cards and older
130 		 * IGP chips to avoid image corruptions
131 		 */
132 		if (p->ring == R600_RING_TYPE_UVD_INDEX &&
133 		    (i <= 0 || (p->rdev->flags & RADEON_IS_AGP) ||
134 		     p->rdev->family == CHIP_RS780 ||
135 		     p->rdev->family == CHIP_RS880)) {
136 
137 			/* TODO: is this still needed for NI+ ? */
138 			p->relocs[i].preferred_domains =
139 				RADEON_GEM_DOMAIN_VRAM;
140 
141 			p->relocs[i].allowed_domains =
142 				RADEON_GEM_DOMAIN_VRAM;
143 
144 			/* prioritize this over any other relocation */
145 			priority = RADEON_CS_MAX_PRIORITY;
146 		} else {
147 			uint32_t domain = r->write_domain ?
148 				r->write_domain : r->read_domains;
149 
150 			if (domain & RADEON_GEM_DOMAIN_CPU) {
151 				DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
152 					  "for command submission\n");
153 				return -EINVAL;
154 			}
155 
156 			p->relocs[i].preferred_domains = domain;
157 			if (domain == RADEON_GEM_DOMAIN_VRAM)
158 				domain |= RADEON_GEM_DOMAIN_GTT;
159 			p->relocs[i].allowed_domains = domain;
160 		}
161 
162 		if (radeon_ttm_tt_has_userptr(p->rdev, p->relocs[i].robj->tbo.ttm)) {
163 			uint32_t domain = p->relocs[i].preferred_domains;
164 			if (!(domain & RADEON_GEM_DOMAIN_GTT)) {
165 				DRM_ERROR("Only RADEON_GEM_DOMAIN_GTT is "
166 					  "allowed for userptr BOs\n");
167 				return -EINVAL;
168 			}
169 			need_mmap_lock = true;
170 			domain = RADEON_GEM_DOMAIN_GTT;
171 			p->relocs[i].preferred_domains = domain;
172 			p->relocs[i].allowed_domains = domain;
173 		}
174 
175 		/* Objects shared as dma-bufs cannot be moved to VRAM */
176 		if (p->relocs[i].robj->prime_shared_count) {
177 			p->relocs[i].allowed_domains &= ~RADEON_GEM_DOMAIN_VRAM;
178 			if (!p->relocs[i].allowed_domains) {
179 				DRM_ERROR("BO associated with dma-buf cannot "
180 					  "be moved to VRAM\n");
181 				return -EINVAL;
182 			}
183 		}
184 
185 		p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
186 		p->relocs[i].tv.num_shared = !r->write_domain;
187 
188 		radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head,
189 				      priority);
190 	}
191 
192 	radeon_cs_buckets_get_list(&buckets, &p->validated);
193 
194 	if (p->cs_flags & RADEON_CS_USE_VM)
195 		p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm,
196 					      &p->validated);
197 #ifdef notyet
198 	if (need_mmap_lock)
199 		mmap_read_lock(current->mm);
200 #endif
201 
202 	r = radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
203 
204 #ifdef notyet
205 	if (need_mmap_lock)
206 		mmap_read_unlock(current->mm);
207 #endif
208 
209 	return r;
210 }
211 
212 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
213 {
214 	p->priority = priority;
215 
216 	switch (ring) {
217 	default:
218 		DRM_ERROR("unknown ring id: %d\n", ring);
219 		return -EINVAL;
220 	case RADEON_CS_RING_GFX:
221 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
222 		break;
223 	case RADEON_CS_RING_COMPUTE:
224 		if (p->rdev->family >= CHIP_TAHITI) {
225 			if (p->priority > 0)
226 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
227 			else
228 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
229 		} else
230 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
231 		break;
232 	case RADEON_CS_RING_DMA:
233 		if (p->rdev->family >= CHIP_CAYMAN) {
234 			if (p->priority > 0)
235 				p->ring = R600_RING_TYPE_DMA_INDEX;
236 			else
237 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
238 		} else if (p->rdev->family >= CHIP_RV770) {
239 			p->ring = R600_RING_TYPE_DMA_INDEX;
240 		} else {
241 			return -EINVAL;
242 		}
243 		break;
244 	case RADEON_CS_RING_UVD:
245 		p->ring = R600_RING_TYPE_UVD_INDEX;
246 		break;
247 	case RADEON_CS_RING_VCE:
248 		/* TODO: only use the low priority ring for now */
249 		p->ring = TN_RING_TYPE_VCE1_INDEX;
250 		break;
251 	}
252 	return 0;
253 }
254 
255 static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
256 {
257 	struct radeon_bo_list *reloc;
258 	int r;
259 
260 	list_for_each_entry(reloc, &p->validated, tv.head) {
261 		struct dma_resv *resv;
262 
263 		resv = reloc->robj->tbo.base.resv;
264 		r = radeon_sync_resv(p->rdev, &p->ib.sync, resv,
265 				     reloc->tv.num_shared);
266 		if (r)
267 			return r;
268 	}
269 	return 0;
270 }
271 
272 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
273 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
274 {
275 	struct drm_radeon_cs *cs = data;
276 	uint64_t *chunk_array_ptr;
277 	unsigned size, i;
278 	u32 ring = RADEON_CS_RING_GFX;
279 	s32 priority = 0;
280 
281 	INIT_LIST_HEAD(&p->validated);
282 
283 	if (!cs->num_chunks) {
284 		return 0;
285 	}
286 
287 	/* get chunks */
288 	p->idx = 0;
289 	p->ib.sa_bo = NULL;
290 	p->const_ib.sa_bo = NULL;
291 	p->chunk_ib = NULL;
292 	p->chunk_relocs = NULL;
293 	p->chunk_flags = NULL;
294 	p->chunk_const_ib = NULL;
295 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
296 	if (p->chunks_array == NULL) {
297 		return -ENOMEM;
298 	}
299 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
300 	if (copy_from_user(p->chunks_array, chunk_array_ptr,
301 			       sizeof(uint64_t)*cs->num_chunks)) {
302 		return -EFAULT;
303 	}
304 	p->cs_flags = 0;
305 	p->nchunks = cs->num_chunks;
306 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
307 	if (p->chunks == NULL) {
308 		return -ENOMEM;
309 	}
310 	for (i = 0; i < p->nchunks; i++) {
311 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
312 		struct drm_radeon_cs_chunk user_chunk;
313 		uint32_t __user *cdata;
314 
315 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
316 		if (copy_from_user(&user_chunk, chunk_ptr,
317 				       sizeof(struct drm_radeon_cs_chunk))) {
318 			return -EFAULT;
319 		}
320 		p->chunks[i].length_dw = user_chunk.length_dw;
321 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_RELOCS) {
322 			p->chunk_relocs = &p->chunks[i];
323 		}
324 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
325 			p->chunk_ib = &p->chunks[i];
326 			/* zero length IB isn't useful */
327 			if (p->chunks[i].length_dw == 0)
328 				return -EINVAL;
329 		}
330 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB) {
331 			p->chunk_const_ib = &p->chunks[i];
332 			/* zero length CONST IB isn't useful */
333 			if (p->chunks[i].length_dw == 0)
334 				return -EINVAL;
335 		}
336 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
337 			p->chunk_flags = &p->chunks[i];
338 			/* zero length flags aren't useful */
339 			if (p->chunks[i].length_dw == 0)
340 				return -EINVAL;
341 		}
342 
343 		size = p->chunks[i].length_dw;
344 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
345 		p->chunks[i].user_ptr = cdata;
346 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB)
347 			continue;
348 
349 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
350 			if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
351 				continue;
352 		}
353 
354 		p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
355 		size *= sizeof(uint32_t);
356 		if (p->chunks[i].kdata == NULL) {
357 			return -ENOMEM;
358 		}
359 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
360 			return -EFAULT;
361 		}
362 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
363 			p->cs_flags = p->chunks[i].kdata[0];
364 			if (p->chunks[i].length_dw > 1)
365 				ring = p->chunks[i].kdata[1];
366 			if (p->chunks[i].length_dw > 2)
367 				priority = (s32)p->chunks[i].kdata[2];
368 		}
369 	}
370 
371 	/* these are KMS only */
372 	if (p->rdev) {
373 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
374 		    !p->rdev->vm_manager.enabled) {
375 			DRM_ERROR("VM not active on asic!\n");
376 			return -EINVAL;
377 		}
378 
379 		if (radeon_cs_get_ring(p, ring, priority))
380 			return -EINVAL;
381 
382 		/* we only support VM on some SI+ rings */
383 		if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
384 			if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
385 				DRM_ERROR("Ring %d requires VM!\n", p->ring);
386 				return -EINVAL;
387 			}
388 		} else {
389 			if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
390 				DRM_ERROR("VM not supported on ring %d!\n",
391 					  p->ring);
392 				return -EINVAL;
393 			}
394 		}
395 	}
396 
397 	return 0;
398 }
399 
400 static int cmp_size_smaller_first(void *priv, const struct list_head *a,
401 				  const struct list_head *b)
402 {
403 	struct radeon_bo_list *la = list_entry(a, struct radeon_bo_list, tv.head);
404 	struct radeon_bo_list *lb = list_entry(b, struct radeon_bo_list, tv.head);
405 
406 	/* Sort A before B if A is smaller. */
407 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
408 }
409 
410 /**
411  * cs_parser_fini() - clean parser states
412  * @parser:	parser structure holding parsing context.
413  * @error:	error number
414  *
415  * If error is set than unvalidate buffer, otherwise just free memory
416  * used by parsing context.
417  **/
418 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
419 {
420 	unsigned i;
421 
422 	if (!error) {
423 		/* Sort the buffer list from the smallest to largest buffer,
424 		 * which affects the order of buffers in the LRU list.
425 		 * This assures that the smallest buffers are added first
426 		 * to the LRU list, so they are likely to be later evicted
427 		 * first, instead of large buffers whose eviction is more
428 		 * expensive.
429 		 *
430 		 * This slightly lowers the number of bytes moved by TTM
431 		 * per frame under memory pressure.
432 		 */
433 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
434 
435 		ttm_eu_fence_buffer_objects(&parser->ticket,
436 					    &parser->validated,
437 					    &parser->ib.fence->base);
438 	} else if (backoff) {
439 		ttm_eu_backoff_reservation(&parser->ticket,
440 					   &parser->validated);
441 	}
442 
443 	if (parser->relocs != NULL) {
444 		for (i = 0; i < parser->nrelocs; i++) {
445 			struct radeon_bo *bo = parser->relocs[i].robj;
446 			if (bo == NULL)
447 				continue;
448 
449 			drm_gem_object_put(&bo->tbo.base);
450 		}
451 	}
452 	kfree(parser->track);
453 	kvfree(parser->relocs);
454 	kvfree(parser->vm_bos);
455 	for (i = 0; i < parser->nchunks; i++)
456 		kvfree(parser->chunks[i].kdata);
457 	kfree(parser->chunks);
458 	kfree(parser->chunks_array);
459 	radeon_ib_free(parser->rdev, &parser->ib);
460 	radeon_ib_free(parser->rdev, &parser->const_ib);
461 }
462 
463 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
464 			      struct radeon_cs_parser *parser)
465 {
466 	int r;
467 
468 	if (parser->chunk_ib == NULL)
469 		return 0;
470 
471 	if (parser->cs_flags & RADEON_CS_USE_VM)
472 		return 0;
473 
474 	r = radeon_cs_parse(rdev, parser->ring, parser);
475 	if (r || parser->parser_error) {
476 		DRM_ERROR("Invalid command stream !\n");
477 		return r;
478 	}
479 
480 	r = radeon_cs_sync_rings(parser);
481 	if (r) {
482 		if (r != -ERESTARTSYS)
483 			DRM_ERROR("Failed to sync rings: %i\n", r);
484 		return r;
485 	}
486 
487 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
488 		radeon_uvd_note_usage(rdev);
489 	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
490 		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
491 		radeon_vce_note_usage(rdev);
492 
493 	r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
494 	if (r) {
495 		DRM_ERROR("Failed to schedule IB !\n");
496 	}
497 	return r;
498 }
499 
500 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
501 				   struct radeon_vm *vm)
502 {
503 	struct radeon_device *rdev = p->rdev;
504 	struct radeon_bo_va *bo_va;
505 	int i, r;
506 
507 	r = radeon_vm_update_page_directory(rdev, vm);
508 	if (r)
509 		return r;
510 
511 	r = radeon_vm_clear_freed(rdev, vm);
512 	if (r)
513 		return r;
514 
515 	if (vm->ib_bo_va == NULL) {
516 		DRM_ERROR("Tmp BO not in VM!\n");
517 		return -EINVAL;
518 	}
519 
520 	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
521 				&rdev->ring_tmp_bo.bo->tbo.mem);
522 	if (r)
523 		return r;
524 
525 	for (i = 0; i < p->nrelocs; i++) {
526 		struct radeon_bo *bo;
527 
528 		bo = p->relocs[i].robj;
529 		bo_va = radeon_vm_bo_find(vm, bo);
530 		if (bo_va == NULL) {
531 			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
532 			return -EINVAL;
533 		}
534 
535 		r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
536 		if (r)
537 			return r;
538 
539 		radeon_sync_fence(&p->ib.sync, bo_va->last_pt_update);
540 	}
541 
542 	return radeon_vm_clear_invalids(rdev, vm);
543 }
544 
545 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
546 				 struct radeon_cs_parser *parser)
547 {
548 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
549 	struct radeon_vm *vm = &fpriv->vm;
550 	int r;
551 
552 	if (parser->chunk_ib == NULL)
553 		return 0;
554 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
555 		return 0;
556 
557 	if (parser->const_ib.length_dw) {
558 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
559 		if (r) {
560 			return r;
561 		}
562 	}
563 
564 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
565 	if (r) {
566 		return r;
567 	}
568 
569 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
570 		radeon_uvd_note_usage(rdev);
571 
572 	mutex_lock(&vm->mutex);
573 	r = radeon_bo_vm_update_pte(parser, vm);
574 	if (r) {
575 		goto out;
576 	}
577 
578 	r = radeon_cs_sync_rings(parser);
579 	if (r) {
580 		if (r != -ERESTARTSYS)
581 			DRM_ERROR("Failed to sync rings: %i\n", r);
582 		goto out;
583 	}
584 
585 	if ((rdev->family >= CHIP_TAHITI) &&
586 	    (parser->chunk_const_ib != NULL)) {
587 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
588 	} else {
589 		r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
590 	}
591 
592 out:
593 	mutex_unlock(&vm->mutex);
594 	return r;
595 }
596 
597 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
598 {
599 	if (r == -EDEADLK) {
600 		r = radeon_gpu_reset(rdev);
601 		if (!r)
602 			r = -EAGAIN;
603 	}
604 	return r;
605 }
606 
607 static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
608 {
609 	struct radeon_cs_chunk *ib_chunk;
610 	struct radeon_vm *vm = NULL;
611 	int r;
612 
613 	if (parser->chunk_ib == NULL)
614 		return 0;
615 
616 	if (parser->cs_flags & RADEON_CS_USE_VM) {
617 		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
618 		vm = &fpriv->vm;
619 
620 		if ((rdev->family >= CHIP_TAHITI) &&
621 		    (parser->chunk_const_ib != NULL)) {
622 			ib_chunk = parser->chunk_const_ib;
623 			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
624 				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
625 				return -EINVAL;
626 			}
627 			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
628 					   vm, ib_chunk->length_dw * 4);
629 			if (r) {
630 				DRM_ERROR("Failed to get const ib !\n");
631 				return r;
632 			}
633 			parser->const_ib.is_const_ib = true;
634 			parser->const_ib.length_dw = ib_chunk->length_dw;
635 			if (copy_from_user(parser->const_ib.ptr,
636 					       ib_chunk->user_ptr,
637 					       ib_chunk->length_dw * 4))
638 				return -EFAULT;
639 		}
640 
641 		ib_chunk = parser->chunk_ib;
642 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
643 			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
644 			return -EINVAL;
645 		}
646 	}
647 	ib_chunk = parser->chunk_ib;
648 
649 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
650 			   vm, ib_chunk->length_dw * 4);
651 	if (r) {
652 		DRM_ERROR("Failed to get ib !\n");
653 		return r;
654 	}
655 	parser->ib.length_dw = ib_chunk->length_dw;
656 	if (ib_chunk->kdata)
657 		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
658 	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
659 		return -EFAULT;
660 	return 0;
661 }
662 
663 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
664 {
665 	struct radeon_device *rdev = dev->dev_private;
666 	struct radeon_cs_parser parser;
667 	int r;
668 
669 	down_read(&rdev->exclusive_lock);
670 	if (!rdev->accel_working) {
671 		up_read(&rdev->exclusive_lock);
672 		return -EBUSY;
673 	}
674 	if (rdev->in_reset) {
675 		up_read(&rdev->exclusive_lock);
676 		r = radeon_gpu_reset(rdev);
677 		if (!r)
678 			r = -EAGAIN;
679 		return r;
680 	}
681 	/* initialize parser */
682 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
683 	parser.filp = filp;
684 	parser.rdev = rdev;
685 	parser.dev = rdev->dev;
686 	parser.family = rdev->family;
687 	r = radeon_cs_parser_init(&parser, data);
688 	if (r) {
689 		DRM_ERROR("Failed to initialize parser !\n");
690 		radeon_cs_parser_fini(&parser, r, false);
691 		up_read(&rdev->exclusive_lock);
692 		r = radeon_cs_handle_lockup(rdev, r);
693 		return r;
694 	}
695 
696 	r = radeon_cs_ib_fill(rdev, &parser);
697 	if (!r) {
698 		r = radeon_cs_parser_relocs(&parser);
699 		if (r && r != -ERESTARTSYS)
700 			DRM_ERROR("Failed to parse relocation %d!\n", r);
701 	}
702 
703 	if (r) {
704 		radeon_cs_parser_fini(&parser, r, false);
705 		up_read(&rdev->exclusive_lock);
706 		r = radeon_cs_handle_lockup(rdev, r);
707 		return r;
708 	}
709 
710 	trace_radeon_cs(&parser);
711 
712 	r = radeon_cs_ib_chunk(rdev, &parser);
713 	if (r) {
714 		goto out;
715 	}
716 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
717 	if (r) {
718 		goto out;
719 	}
720 out:
721 	radeon_cs_parser_fini(&parser, r, true);
722 	up_read(&rdev->exclusive_lock);
723 	r = radeon_cs_handle_lockup(rdev, r);
724 	return r;
725 }
726 
727 /**
728  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
729  * @parser:	parser structure holding parsing context.
730  * @pkt:	where to store packet information
731  *
732  * Assume that chunk_ib_index is properly set. Will return -EINVAL
733  * if packet is bigger than remaining ib size. or if packets is unknown.
734  **/
735 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
736 			   struct radeon_cs_packet *pkt,
737 			   unsigned idx)
738 {
739 	struct radeon_cs_chunk *ib_chunk = p->chunk_ib;
740 	struct radeon_device *rdev = p->rdev;
741 	uint32_t header;
742 	int ret = 0, i;
743 
744 	if (idx >= ib_chunk->length_dw) {
745 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
746 			  idx, ib_chunk->length_dw);
747 		return -EINVAL;
748 	}
749 	header = radeon_get_ib_value(p, idx);
750 	pkt->idx = idx;
751 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
752 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
753 	pkt->one_reg_wr = 0;
754 	switch (pkt->type) {
755 	case RADEON_PACKET_TYPE0:
756 		if (rdev->family < CHIP_R600) {
757 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
758 			pkt->one_reg_wr =
759 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
760 		} else
761 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
762 		break;
763 	case RADEON_PACKET_TYPE3:
764 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
765 		break;
766 	case RADEON_PACKET_TYPE2:
767 		pkt->count = -1;
768 		break;
769 	default:
770 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
771 		ret = -EINVAL;
772 		goto dump_ib;
773 	}
774 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
775 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
776 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
777 		ret = -EINVAL;
778 		goto dump_ib;
779 	}
780 	return 0;
781 
782 dump_ib:
783 	for (i = 0; i < ib_chunk->length_dw; i++) {
784 		if (i == idx)
785 			printk("\t0x%08x <---\n", radeon_get_ib_value(p, i));
786 		else
787 			printk("\t0x%08x\n", radeon_get_ib_value(p, i));
788 	}
789 	return ret;
790 }
791 
792 /**
793  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
794  * @p:		structure holding the parser context.
795  *
796  * Check if the next packet is NOP relocation packet3.
797  **/
798 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
799 {
800 	struct radeon_cs_packet p3reloc;
801 	int r;
802 
803 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
804 	if (r)
805 		return false;
806 	if (p3reloc.type != RADEON_PACKET_TYPE3)
807 		return false;
808 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
809 		return false;
810 	return true;
811 }
812 
813 /**
814  * radeon_cs_dump_packet() - dump raw packet context
815  * @p:		structure holding the parser context.
816  * @pkt:	structure holding the packet.
817  *
818  * Used mostly for debugging and error reporting.
819  **/
820 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
821 			   struct radeon_cs_packet *pkt)
822 {
823 	volatile uint32_t *ib;
824 	unsigned i;
825 	unsigned idx;
826 
827 	ib = p->ib.ptr;
828 	idx = pkt->idx;
829 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
830 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
831 }
832 
833 /**
834  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
835  * @parser:		parser structure holding parsing context.
836  * @data:		pointer to relocation data
837  * @offset_start:	starting offset
838  * @offset_mask:	offset mask (to align start offset on)
839  * @reloc:		reloc informations
840  *
841  * Check if next packet is relocation packet3, do bo validation and compute
842  * GPU offset using the provided start.
843  **/
844 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
845 				struct radeon_bo_list **cs_reloc,
846 				int nomm)
847 {
848 	struct radeon_cs_chunk *relocs_chunk;
849 	struct radeon_cs_packet p3reloc;
850 	unsigned idx;
851 	int r;
852 
853 	if (p->chunk_relocs == NULL) {
854 		DRM_ERROR("No relocation chunk !\n");
855 		return -EINVAL;
856 	}
857 	*cs_reloc = NULL;
858 	relocs_chunk = p->chunk_relocs;
859 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
860 	if (r)
861 		return r;
862 	p->idx += p3reloc.count + 2;
863 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
864 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
865 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
866 			  p3reloc.idx);
867 		radeon_cs_dump_packet(p, &p3reloc);
868 		return -EINVAL;
869 	}
870 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
871 	if (idx >= relocs_chunk->length_dw) {
872 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
873 			  idx, relocs_chunk->length_dw);
874 		radeon_cs_dump_packet(p, &p3reloc);
875 		return -EINVAL;
876 	}
877 	/* FIXME: we assume reloc size is 4 dwords */
878 	if (nomm) {
879 		*cs_reloc = p->relocs;
880 		(*cs_reloc)->gpu_offset =
881 			(u64)relocs_chunk->kdata[idx + 3] << 32;
882 		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
883 	} else
884 		*cs_reloc = &p->relocs[(idx / 4)];
885 	return 0;
886 }
887