xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision 65417d9f)
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 #include <linux/list_sort.h>
28 #include <drm/drmP.h>
29 #include <drm/radeon_drm.h>
30 #include "radeon_reg.h"
31 #include "radeon.h"
32 #include "radeon_trace.h"
33 
34 #define RADEON_CS_MAX_PRIORITY		32u
35 #define RADEON_CS_NUM_BUCKETS		(RADEON_CS_MAX_PRIORITY + 1)
36 
37 /* This is based on the bucket sort with O(n) time complexity.
38  * An item with priority "i" is added to bucket[i]. The lists are then
39  * concatenated in descending order.
40  */
41 struct radeon_cs_buckets {
42 	struct list_head bucket[RADEON_CS_NUM_BUCKETS];
43 };
44 
45 static void radeon_cs_buckets_init(struct radeon_cs_buckets *b)
46 {
47 	unsigned i;
48 
49 	for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++)
50 		INIT_LIST_HEAD(&b->bucket[i]);
51 }
52 
53 static void radeon_cs_buckets_add(struct radeon_cs_buckets *b,
54 				  struct list_head *item, unsigned priority)
55 {
56 	/* Since buffers which appear sooner in the relocation list are
57 	 * likely to be used more often than buffers which appear later
58 	 * in the list, the sort mustn't change the ordering of buffers
59 	 * with the same priority, i.e. it must be stable.
60 	 */
61 	list_add_tail(item, &b->bucket[min(priority, RADEON_CS_MAX_PRIORITY)]);
62 }
63 
64 static void radeon_cs_buckets_get_list(struct radeon_cs_buckets *b,
65 				       struct list_head *out_list)
66 {
67 	unsigned i;
68 
69 	/* Connect the sorted buckets in the output list. */
70 	for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++) {
71 		list_splice(&b->bucket[i], out_list);
72 	}
73 }
74 
75 static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
76 {
77 	struct radeon_cs_chunk *chunk;
78 	struct radeon_cs_buckets buckets;
79 	unsigned i;
80 	bool need_mmap_lock = false;
81 	int r;
82 
83 	if (p->chunk_relocs == NULL) {
84 		return 0;
85 	}
86 	chunk = p->chunk_relocs;
87 	p->dma_reloc_idx = 0;
88 	/* FIXME: we assume that each relocs use 4 dwords */
89 	p->nrelocs = chunk->length_dw / 4;
90 	p->relocs = kvmalloc_array(p->nrelocs, sizeof(struct radeon_bo_list),
91 			GFP_KERNEL | __GFP_ZERO);
92 	if (p->relocs == NULL) {
93 		return -ENOMEM;
94 	}
95 
96 	radeon_cs_buckets_init(&buckets);
97 
98 	for (i = 0; i < p->nrelocs; i++) {
99 		struct drm_radeon_cs_reloc *r;
100 		struct drm_gem_object *gobj;
101 		unsigned priority;
102 
103 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
104 		gobj = drm_gem_object_lookup(p->filp, r->handle);
105 		if (gobj == NULL) {
106 			DRM_ERROR("gem object lookup failed 0x%x\n",
107 				  r->handle);
108 			return -ENOENT;
109 		}
110 		p->relocs[i].robj = gem_to_radeon_bo(gobj);
111 
112 		/* The userspace buffer priorities are from 0 to 15. A higher
113 		 * number means the buffer is more important.
114 		 * Also, the buffers used for write have a higher priority than
115 		 * the buffers used for read only, which doubles the range
116 		 * to 0 to 31. 32 is reserved for the kernel driver.
117 		 */
118 		priority = (r->flags & RADEON_RELOC_PRIO_MASK) * 2
119 			   + !!r->write_domain;
120 
121 		/* The first reloc of an UVD job is the msg and that must be in
122 		 * VRAM, the second reloc is the DPB and for WMV that must be in
123 		 * VRAM as well. Also put everything into VRAM on AGP cards and older
124 		 * IGP chips to avoid image corruptions
125 		 */
126 		if (p->ring == R600_RING_TYPE_UVD_INDEX &&
127 		    (i <= 0 || pci_find_capability(p->rdev->ddev->pdev,
128 						   PCI_CAP_ID_AGP) ||
129 		     p->rdev->family == CHIP_RS780 ||
130 		     p->rdev->family == CHIP_RS880)) {
131 
132 			/* TODO: is this still needed for NI+ ? */
133 			p->relocs[i].preferred_domains =
134 				RADEON_GEM_DOMAIN_VRAM;
135 
136 			p->relocs[i].allowed_domains =
137 				RADEON_GEM_DOMAIN_VRAM;
138 
139 			/* prioritize this over any other relocation */
140 			priority = RADEON_CS_MAX_PRIORITY;
141 		} else {
142 			uint32_t domain = r->write_domain ?
143 				r->write_domain : r->read_domains;
144 
145 			if (domain & RADEON_GEM_DOMAIN_CPU) {
146 				DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
147 					  "for command submission\n");
148 				return -EINVAL;
149 			}
150 
151 			p->relocs[i].preferred_domains = domain;
152 			if (domain == RADEON_GEM_DOMAIN_VRAM)
153 				domain |= RADEON_GEM_DOMAIN_GTT;
154 			p->relocs[i].allowed_domains = domain;
155 		}
156 
157 		if (radeon_ttm_tt_has_userptr(p->relocs[i].robj->tbo.ttm)) {
158 			uint32_t domain = p->relocs[i].preferred_domains;
159 			if (!(domain & RADEON_GEM_DOMAIN_GTT)) {
160 				DRM_ERROR("Only RADEON_GEM_DOMAIN_GTT is "
161 					  "allowed for userptr BOs\n");
162 				return -EINVAL;
163 			}
164 			need_mmap_lock = true;
165 			domain = RADEON_GEM_DOMAIN_GTT;
166 			p->relocs[i].preferred_domains = domain;
167 			p->relocs[i].allowed_domains = domain;
168 		}
169 
170 		/* Objects shared as dma-bufs cannot be moved to VRAM */
171 		if (p->relocs[i].robj->prime_shared_count) {
172 			p->relocs[i].allowed_domains &= ~RADEON_GEM_DOMAIN_VRAM;
173 			if (!p->relocs[i].allowed_domains) {
174 				DRM_ERROR("BO associated with dma-buf cannot "
175 					  "be moved to VRAM\n");
176 				return -EINVAL;
177 			}
178 		}
179 
180 		p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
181 		p->relocs[i].tv.shared = !r->write_domain;
182 
183 		radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head,
184 				      priority);
185 	}
186 
187 	radeon_cs_buckets_get_list(&buckets, &p->validated);
188 
189 	if (p->cs_flags & RADEON_CS_USE_VM)
190 		p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm,
191 					      &p->validated);
192 	if (need_mmap_lock)
193 		down_read(&current->mm->mmap_sem);
194 
195 	r = radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
196 
197 	if (need_mmap_lock)
198 		up_read(&current->mm->mmap_sem);
199 
200 	return r;
201 }
202 
203 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
204 {
205 	p->priority = priority;
206 
207 	switch (ring) {
208 	default:
209 		DRM_ERROR("unknown ring id: %d\n", ring);
210 		return -EINVAL;
211 	case RADEON_CS_RING_GFX:
212 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
213 		break;
214 	case RADEON_CS_RING_COMPUTE:
215 		if (p->rdev->family >= CHIP_TAHITI) {
216 			if (p->priority > 0)
217 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
218 			else
219 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
220 		} else
221 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
222 		break;
223 	case RADEON_CS_RING_DMA:
224 		if (p->rdev->family >= CHIP_CAYMAN) {
225 			if (p->priority > 0)
226 				p->ring = R600_RING_TYPE_DMA_INDEX;
227 			else
228 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
229 		} else if (p->rdev->family >= CHIP_RV770) {
230 			p->ring = R600_RING_TYPE_DMA_INDEX;
231 		} else {
232 			return -EINVAL;
233 		}
234 		break;
235 	case RADEON_CS_RING_UVD:
236 		p->ring = R600_RING_TYPE_UVD_INDEX;
237 		break;
238 	case RADEON_CS_RING_VCE:
239 		/* TODO: only use the low priority ring for now */
240 		p->ring = TN_RING_TYPE_VCE1_INDEX;
241 		break;
242 	}
243 	return 0;
244 }
245 
246 static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
247 {
248 	struct radeon_bo_list *reloc;
249 	int r;
250 
251 	list_for_each_entry(reloc, &p->validated, tv.head) {
252 		struct reservation_object *resv;
253 
254 		resv = reloc->robj->tbo.resv;
255 		r = radeon_sync_resv(p->rdev, &p->ib.sync, resv,
256 				     reloc->tv.shared);
257 		if (r)
258 			return r;
259 	}
260 	return 0;
261 }
262 
263 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
264 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
265 {
266 	struct drm_radeon_cs *cs = data;
267 	uint64_t *chunk_array_ptr;
268 	unsigned size, i;
269 	u32 ring = RADEON_CS_RING_GFX;
270 	s32 priority = 0;
271 
272 	INIT_LIST_HEAD(&p->validated);
273 
274 	if (!cs->num_chunks) {
275 		return 0;
276 	}
277 
278 	/* get chunks */
279 	p->idx = 0;
280 	p->ib.sa_bo = NULL;
281 	p->const_ib.sa_bo = NULL;
282 	p->chunk_ib = NULL;
283 	p->chunk_relocs = NULL;
284 	p->chunk_flags = NULL;
285 	p->chunk_const_ib = NULL;
286 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
287 	if (p->chunks_array == NULL) {
288 		return -ENOMEM;
289 	}
290 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
291 	if (copy_from_user(p->chunks_array, chunk_array_ptr,
292 			       sizeof(uint64_t)*cs->num_chunks)) {
293 		return -EFAULT;
294 	}
295 	p->cs_flags = 0;
296 	p->nchunks = cs->num_chunks;
297 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
298 	if (p->chunks == NULL) {
299 		return -ENOMEM;
300 	}
301 	for (i = 0; i < p->nchunks; i++) {
302 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
303 		struct drm_radeon_cs_chunk user_chunk;
304 		uint32_t __user *cdata;
305 
306 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
307 		if (copy_from_user(&user_chunk, chunk_ptr,
308 				       sizeof(struct drm_radeon_cs_chunk))) {
309 			return -EFAULT;
310 		}
311 		p->chunks[i].length_dw = user_chunk.length_dw;
312 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_RELOCS) {
313 			p->chunk_relocs = &p->chunks[i];
314 		}
315 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
316 			p->chunk_ib = &p->chunks[i];
317 			/* zero length IB isn't useful */
318 			if (p->chunks[i].length_dw == 0)
319 				return -EINVAL;
320 		}
321 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB) {
322 			p->chunk_const_ib = &p->chunks[i];
323 			/* zero length CONST IB isn't useful */
324 			if (p->chunks[i].length_dw == 0)
325 				return -EINVAL;
326 		}
327 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
328 			p->chunk_flags = &p->chunks[i];
329 			/* zero length flags aren't useful */
330 			if (p->chunks[i].length_dw == 0)
331 				return -EINVAL;
332 		}
333 
334 		size = p->chunks[i].length_dw;
335 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
336 		p->chunks[i].user_ptr = cdata;
337 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB)
338 			continue;
339 
340 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
341 			if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
342 				continue;
343 		}
344 
345 		p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
346 		size *= sizeof(uint32_t);
347 		if (p->chunks[i].kdata == NULL) {
348 			return -ENOMEM;
349 		}
350 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
351 			return -EFAULT;
352 		}
353 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
354 			p->cs_flags = p->chunks[i].kdata[0];
355 			if (p->chunks[i].length_dw > 1)
356 				ring = p->chunks[i].kdata[1];
357 			if (p->chunks[i].length_dw > 2)
358 				priority = (s32)p->chunks[i].kdata[2];
359 		}
360 	}
361 
362 	/* these are KMS only */
363 	if (p->rdev) {
364 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
365 		    !p->rdev->vm_manager.enabled) {
366 			DRM_ERROR("VM not active on asic!\n");
367 			return -EINVAL;
368 		}
369 
370 		if (radeon_cs_get_ring(p, ring, priority))
371 			return -EINVAL;
372 
373 		/* we only support VM on some SI+ rings */
374 		if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
375 			if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
376 				DRM_ERROR("Ring %d requires VM!\n", p->ring);
377 				return -EINVAL;
378 			}
379 		} else {
380 			if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
381 				DRM_ERROR("VM not supported on ring %d!\n",
382 					  p->ring);
383 				return -EINVAL;
384 			}
385 		}
386 	}
387 
388 	return 0;
389 }
390 
391 static int cmp_size_smaller_first(void *priv, struct list_head *a,
392 				  struct list_head *b)
393 {
394 	struct radeon_bo_list *la = list_entry(a, struct radeon_bo_list, tv.head);
395 	struct radeon_bo_list *lb = list_entry(b, struct radeon_bo_list, tv.head);
396 
397 	/* Sort A before B if A is smaller. */
398 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
399 }
400 
401 /**
402  * cs_parser_fini() - clean parser states
403  * @parser:	parser structure holding parsing context.
404  * @error:	error number
405  *
406  * If error is set than unvalidate buffer, otherwise just free memory
407  * used by parsing context.
408  **/
409 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
410 {
411 	unsigned i;
412 
413 	if (!error) {
414 		/* Sort the buffer list from the smallest to largest buffer,
415 		 * which affects the order of buffers in the LRU list.
416 		 * This assures that the smallest buffers are added first
417 		 * to the LRU list, so they are likely to be later evicted
418 		 * first, instead of large buffers whose eviction is more
419 		 * expensive.
420 		 *
421 		 * This slightly lowers the number of bytes moved by TTM
422 		 * per frame under memory pressure.
423 		 */
424 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
425 
426 		ttm_eu_fence_buffer_objects(&parser->ticket,
427 					    &parser->validated,
428 					    &parser->ib.fence->base);
429 	} else if (backoff) {
430 		ttm_eu_backoff_reservation(&parser->ticket,
431 					   &parser->validated);
432 	}
433 
434 	if (parser->relocs != NULL) {
435 		for (i = 0; i < parser->nrelocs; i++) {
436 			struct radeon_bo *bo = parser->relocs[i].robj;
437 			if (bo == NULL)
438 				continue;
439 
440 			drm_gem_object_put_unlocked(&bo->gem_base);
441 		}
442 	}
443 	kfree(parser->track);
444 	kvfree(parser->relocs);
445 	kvfree(parser->vm_bos);
446 	for (i = 0; i < parser->nchunks; i++)
447 		kvfree(parser->chunks[i].kdata);
448 	kfree(parser->chunks);
449 	kfree(parser->chunks_array);
450 	radeon_ib_free(parser->rdev, &parser->ib);
451 	radeon_ib_free(parser->rdev, &parser->const_ib);
452 }
453 
454 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
455 			      struct radeon_cs_parser *parser)
456 {
457 	int r;
458 
459 	if (parser->chunk_ib == NULL)
460 		return 0;
461 
462 	if (parser->cs_flags & RADEON_CS_USE_VM)
463 		return 0;
464 
465 	r = radeon_cs_parse(rdev, parser->ring, parser);
466 	if (r || parser->parser_error) {
467 		DRM_ERROR("Invalid command stream !\n");
468 		return r;
469 	}
470 
471 	r = radeon_cs_sync_rings(parser);
472 	if (r) {
473 		if (r != -ERESTARTSYS)
474 			DRM_ERROR("Failed to sync rings: %i\n", r);
475 		return r;
476 	}
477 
478 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
479 		radeon_uvd_note_usage(rdev);
480 	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
481 		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
482 		radeon_vce_note_usage(rdev);
483 
484 	r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
485 	if (r) {
486 		DRM_ERROR("Failed to schedule IB !\n");
487 	}
488 	return r;
489 }
490 
491 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
492 				   struct radeon_vm *vm)
493 {
494 	struct radeon_device *rdev = p->rdev;
495 	struct radeon_bo_va *bo_va;
496 	int i, r;
497 
498 	r = radeon_vm_update_page_directory(rdev, vm);
499 	if (r)
500 		return r;
501 
502 	r = radeon_vm_clear_freed(rdev, vm);
503 	if (r)
504 		return r;
505 
506 	if (vm->ib_bo_va == NULL) {
507 		DRM_ERROR("Tmp BO not in VM!\n");
508 		return -EINVAL;
509 	}
510 
511 	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
512 				&rdev->ring_tmp_bo.bo->tbo.mem);
513 	if (r)
514 		return r;
515 
516 	for (i = 0; i < p->nrelocs; i++) {
517 		struct radeon_bo *bo;
518 
519 		bo = p->relocs[i].robj;
520 		bo_va = radeon_vm_bo_find(vm, bo);
521 		if (bo_va == NULL) {
522 			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
523 			return -EINVAL;
524 		}
525 
526 		r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
527 		if (r)
528 			return r;
529 
530 		radeon_sync_fence(&p->ib.sync, bo_va->last_pt_update);
531 	}
532 
533 	return radeon_vm_clear_invalids(rdev, vm);
534 }
535 
536 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
537 				 struct radeon_cs_parser *parser)
538 {
539 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
540 	struct radeon_vm *vm = &fpriv->vm;
541 	int r;
542 
543 	if (parser->chunk_ib == NULL)
544 		return 0;
545 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
546 		return 0;
547 
548 	if (parser->const_ib.length_dw) {
549 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
550 		if (r) {
551 			return r;
552 		}
553 	}
554 
555 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
556 	if (r) {
557 		return r;
558 	}
559 
560 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
561 		radeon_uvd_note_usage(rdev);
562 
563 	mutex_lock(&vm->mutex);
564 	r = radeon_bo_vm_update_pte(parser, vm);
565 	if (r) {
566 		goto out;
567 	}
568 
569 	r = radeon_cs_sync_rings(parser);
570 	if (r) {
571 		if (r != -ERESTARTSYS)
572 			DRM_ERROR("Failed to sync rings: %i\n", r);
573 		goto out;
574 	}
575 
576 	if ((rdev->family >= CHIP_TAHITI) &&
577 	    (parser->chunk_const_ib != NULL)) {
578 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
579 	} else {
580 		r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
581 	}
582 
583 out:
584 	mutex_unlock(&vm->mutex);
585 	return r;
586 }
587 
588 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
589 {
590 	if (r == -EDEADLK) {
591 		r = radeon_gpu_reset(rdev);
592 		if (!r)
593 			r = -EAGAIN;
594 	}
595 	return r;
596 }
597 
598 static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
599 {
600 	struct radeon_cs_chunk *ib_chunk;
601 	struct radeon_vm *vm = NULL;
602 	int r;
603 
604 	if (parser->chunk_ib == NULL)
605 		return 0;
606 
607 	if (parser->cs_flags & RADEON_CS_USE_VM) {
608 		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
609 		vm = &fpriv->vm;
610 
611 		if ((rdev->family >= CHIP_TAHITI) &&
612 		    (parser->chunk_const_ib != NULL)) {
613 			ib_chunk = parser->chunk_const_ib;
614 			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
615 				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
616 				return -EINVAL;
617 			}
618 			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
619 					   vm, ib_chunk->length_dw * 4);
620 			if (r) {
621 				DRM_ERROR("Failed to get const ib !\n");
622 				return r;
623 			}
624 			parser->const_ib.is_const_ib = true;
625 			parser->const_ib.length_dw = ib_chunk->length_dw;
626 			if (copy_from_user(parser->const_ib.ptr,
627 					       ib_chunk->user_ptr,
628 					       ib_chunk->length_dw * 4))
629 				return -EFAULT;
630 		}
631 
632 		ib_chunk = parser->chunk_ib;
633 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
634 			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
635 			return -EINVAL;
636 		}
637 	}
638 	ib_chunk = parser->chunk_ib;
639 
640 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
641 			   vm, ib_chunk->length_dw * 4);
642 	if (r) {
643 		DRM_ERROR("Failed to get ib !\n");
644 		return r;
645 	}
646 	parser->ib.length_dw = ib_chunk->length_dw;
647 	if (ib_chunk->kdata)
648 		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
649 	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
650 		return -EFAULT;
651 	return 0;
652 }
653 
654 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
655 {
656 	struct radeon_device *rdev = dev->dev_private;
657 	struct radeon_cs_parser parser;
658 	int r;
659 
660 	down_read(&rdev->exclusive_lock);
661 	if (!rdev->accel_working) {
662 		up_read(&rdev->exclusive_lock);
663 		return -EBUSY;
664 	}
665 	if (rdev->in_reset) {
666 		up_read(&rdev->exclusive_lock);
667 		r = radeon_gpu_reset(rdev);
668 		if (!r)
669 			r = -EAGAIN;
670 		return r;
671 	}
672 	/* initialize parser */
673 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
674 	parser.filp = filp;
675 	parser.rdev = rdev;
676 	parser.dev = rdev->dev;
677 	parser.family = rdev->family;
678 	r = radeon_cs_parser_init(&parser, data);
679 	if (r) {
680 		DRM_ERROR("Failed to initialize parser !\n");
681 		radeon_cs_parser_fini(&parser, r, false);
682 		up_read(&rdev->exclusive_lock);
683 		r = radeon_cs_handle_lockup(rdev, r);
684 		return r;
685 	}
686 
687 	r = radeon_cs_ib_fill(rdev, &parser);
688 	if (!r) {
689 		r = radeon_cs_parser_relocs(&parser);
690 		if (r && r != -ERESTARTSYS)
691 			DRM_ERROR("Failed to parse relocation %d!\n", r);
692 	}
693 
694 	if (r) {
695 		radeon_cs_parser_fini(&parser, r, false);
696 		up_read(&rdev->exclusive_lock);
697 		r = radeon_cs_handle_lockup(rdev, r);
698 		return r;
699 	}
700 
701 	trace_radeon_cs(&parser);
702 
703 	r = radeon_cs_ib_chunk(rdev, &parser);
704 	if (r) {
705 		goto out;
706 	}
707 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
708 	if (r) {
709 		goto out;
710 	}
711 out:
712 	radeon_cs_parser_fini(&parser, r, true);
713 	up_read(&rdev->exclusive_lock);
714 	r = radeon_cs_handle_lockup(rdev, r);
715 	return r;
716 }
717 
718 /**
719  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
720  * @parser:	parser structure holding parsing context.
721  * @pkt:	where to store packet information
722  *
723  * Assume that chunk_ib_index is properly set. Will return -EINVAL
724  * if packet is bigger than remaining ib size. or if packets is unknown.
725  **/
726 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
727 			   struct radeon_cs_packet *pkt,
728 			   unsigned idx)
729 {
730 	struct radeon_cs_chunk *ib_chunk = p->chunk_ib;
731 	struct radeon_device *rdev = p->rdev;
732 	uint32_t header;
733 	int ret = 0, i;
734 
735 	if (idx >= ib_chunk->length_dw) {
736 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
737 			  idx, ib_chunk->length_dw);
738 		return -EINVAL;
739 	}
740 	header = radeon_get_ib_value(p, idx);
741 	pkt->idx = idx;
742 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
743 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
744 	pkt->one_reg_wr = 0;
745 	switch (pkt->type) {
746 	case RADEON_PACKET_TYPE0:
747 		if (rdev->family < CHIP_R600) {
748 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
749 			pkt->one_reg_wr =
750 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
751 		} else
752 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
753 		break;
754 	case RADEON_PACKET_TYPE3:
755 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
756 		break;
757 	case RADEON_PACKET_TYPE2:
758 		pkt->count = -1;
759 		break;
760 	default:
761 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
762 		ret = -EINVAL;
763 		goto dump_ib;
764 	}
765 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
766 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
767 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
768 		ret = -EINVAL;
769 		goto dump_ib;
770 	}
771 	return 0;
772 
773 dump_ib:
774 	for (i = 0; i < ib_chunk->length_dw; i++) {
775 		if (i == idx)
776 			printk("\t0x%08x <---\n", radeon_get_ib_value(p, i));
777 		else
778 			printk("\t0x%08x\n", radeon_get_ib_value(p, i));
779 	}
780 	return ret;
781 }
782 
783 /**
784  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
785  * @p:		structure holding the parser context.
786  *
787  * Check if the next packet is NOP relocation packet3.
788  **/
789 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
790 {
791 	struct radeon_cs_packet p3reloc;
792 	int r;
793 
794 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
795 	if (r)
796 		return false;
797 	if (p3reloc.type != RADEON_PACKET_TYPE3)
798 		return false;
799 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
800 		return false;
801 	return true;
802 }
803 
804 /**
805  * radeon_cs_dump_packet() - dump raw packet context
806  * @p:		structure holding the parser context.
807  * @pkt:	structure holding the packet.
808  *
809  * Used mostly for debugging and error reporting.
810  **/
811 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
812 			   struct radeon_cs_packet *pkt)
813 {
814 	volatile uint32_t *ib;
815 	unsigned i;
816 	unsigned idx;
817 
818 	ib = p->ib.ptr;
819 	idx = pkt->idx;
820 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
821 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
822 }
823 
824 /**
825  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
826  * @parser:		parser structure holding parsing context.
827  * @data:		pointer to relocation data
828  * @offset_start:	starting offset
829  * @offset_mask:	offset mask (to align start offset on)
830  * @reloc:		reloc informations
831  *
832  * Check if next packet is relocation packet3, do bo validation and compute
833  * GPU offset using the provided start.
834  **/
835 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
836 				struct radeon_bo_list **cs_reloc,
837 				int nomm)
838 {
839 	struct radeon_cs_chunk *relocs_chunk;
840 	struct radeon_cs_packet p3reloc;
841 	unsigned idx;
842 	int r;
843 
844 	if (p->chunk_relocs == NULL) {
845 		DRM_ERROR("No relocation chunk !\n");
846 		return -EINVAL;
847 	}
848 	*cs_reloc = NULL;
849 	relocs_chunk = p->chunk_relocs;
850 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
851 	if (r)
852 		return r;
853 	p->idx += p3reloc.count + 2;
854 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
855 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
856 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
857 			  p3reloc.idx);
858 		radeon_cs_dump_packet(p, &p3reloc);
859 		return -EINVAL;
860 	}
861 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
862 	if (idx >= relocs_chunk->length_dw) {
863 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
864 			  idx, relocs_chunk->length_dw);
865 		radeon_cs_dump_packet(p, &p3reloc);
866 		return -EINVAL;
867 	}
868 	/* FIXME: we assume reloc size is 4 dwords */
869 	if (nomm) {
870 		*cs_reloc = p->relocs;
871 		(*cs_reloc)->gpu_offset =
872 			(u64)relocs_chunk->kdata[idx + 3] << 32;
873 		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
874 	} else
875 		*cs_reloc = &p->relocs[(idx / 4)];
876 	return 0;
877 }
878