xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision 0bf49ffb)
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 || pci_find_capability(p->rdev->pdev, PCI_CAP_ID_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 	if (need_mmap_lock)
198 		mmap_read_lock(current->mm);
199 
200 	r = radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
201 
202 	if (need_mmap_lock)
203 		mmap_read_unlock(current->mm);
204 
205 	return r;
206 }
207 
208 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
209 {
210 	p->priority = priority;
211 
212 	switch (ring) {
213 	default:
214 		DRM_ERROR("unknown ring id: %d\n", ring);
215 		return -EINVAL;
216 	case RADEON_CS_RING_GFX:
217 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
218 		break;
219 	case RADEON_CS_RING_COMPUTE:
220 		if (p->rdev->family >= CHIP_TAHITI) {
221 			if (p->priority > 0)
222 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
223 			else
224 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
225 		} else
226 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
227 		break;
228 	case RADEON_CS_RING_DMA:
229 		if (p->rdev->family >= CHIP_CAYMAN) {
230 			if (p->priority > 0)
231 				p->ring = R600_RING_TYPE_DMA_INDEX;
232 			else
233 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
234 		} else if (p->rdev->family >= CHIP_RV770) {
235 			p->ring = R600_RING_TYPE_DMA_INDEX;
236 		} else {
237 			return -EINVAL;
238 		}
239 		break;
240 	case RADEON_CS_RING_UVD:
241 		p->ring = R600_RING_TYPE_UVD_INDEX;
242 		break;
243 	case RADEON_CS_RING_VCE:
244 		/* TODO: only use the low priority ring for now */
245 		p->ring = TN_RING_TYPE_VCE1_INDEX;
246 		break;
247 	}
248 	return 0;
249 }
250 
251 static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
252 {
253 	struct radeon_bo_list *reloc;
254 	int r;
255 
256 	list_for_each_entry(reloc, &p->validated, tv.head) {
257 		struct dma_resv *resv;
258 
259 		resv = reloc->robj->tbo.base.resv;
260 		r = radeon_sync_resv(p->rdev, &p->ib.sync, resv,
261 				     reloc->tv.num_shared);
262 		if (r)
263 			return r;
264 	}
265 	return 0;
266 }
267 
268 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
269 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
270 {
271 	struct drm_radeon_cs *cs = data;
272 	uint64_t *chunk_array_ptr;
273 	unsigned size, i;
274 	u32 ring = RADEON_CS_RING_GFX;
275 	s32 priority = 0;
276 
277 	INIT_LIST_HEAD(&p->validated);
278 
279 	if (!cs->num_chunks) {
280 		return 0;
281 	}
282 
283 	/* get chunks */
284 	p->idx = 0;
285 	p->ib.sa_bo = NULL;
286 	p->const_ib.sa_bo = NULL;
287 	p->chunk_ib = NULL;
288 	p->chunk_relocs = NULL;
289 	p->chunk_flags = NULL;
290 	p->chunk_const_ib = NULL;
291 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
292 	if (p->chunks_array == NULL) {
293 		return -ENOMEM;
294 	}
295 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
296 	if (copy_from_user(p->chunks_array, chunk_array_ptr,
297 			       sizeof(uint64_t)*cs->num_chunks)) {
298 		return -EFAULT;
299 	}
300 	p->cs_flags = 0;
301 	p->nchunks = cs->num_chunks;
302 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
303 	if (p->chunks == NULL) {
304 		return -ENOMEM;
305 	}
306 	for (i = 0; i < p->nchunks; i++) {
307 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
308 		struct drm_radeon_cs_chunk user_chunk;
309 		uint32_t __user *cdata;
310 
311 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
312 		if (copy_from_user(&user_chunk, chunk_ptr,
313 				       sizeof(struct drm_radeon_cs_chunk))) {
314 			return -EFAULT;
315 		}
316 		p->chunks[i].length_dw = user_chunk.length_dw;
317 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_RELOCS) {
318 			p->chunk_relocs = &p->chunks[i];
319 		}
320 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
321 			p->chunk_ib = &p->chunks[i];
322 			/* zero length IB isn't useful */
323 			if (p->chunks[i].length_dw == 0)
324 				return -EINVAL;
325 		}
326 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB) {
327 			p->chunk_const_ib = &p->chunks[i];
328 			/* zero length CONST IB isn't useful */
329 			if (p->chunks[i].length_dw == 0)
330 				return -EINVAL;
331 		}
332 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
333 			p->chunk_flags = &p->chunks[i];
334 			/* zero length flags aren't useful */
335 			if (p->chunks[i].length_dw == 0)
336 				return -EINVAL;
337 		}
338 
339 		size = p->chunks[i].length_dw;
340 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
341 		p->chunks[i].user_ptr = cdata;
342 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB)
343 			continue;
344 
345 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
346 			if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
347 				continue;
348 		}
349 
350 		p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
351 		size *= sizeof(uint32_t);
352 		if (p->chunks[i].kdata == NULL) {
353 			return -ENOMEM;
354 		}
355 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
356 			return -EFAULT;
357 		}
358 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
359 			p->cs_flags = p->chunks[i].kdata[0];
360 			if (p->chunks[i].length_dw > 1)
361 				ring = p->chunks[i].kdata[1];
362 			if (p->chunks[i].length_dw > 2)
363 				priority = (s32)p->chunks[i].kdata[2];
364 		}
365 	}
366 
367 	/* these are KMS only */
368 	if (p->rdev) {
369 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
370 		    !p->rdev->vm_manager.enabled) {
371 			DRM_ERROR("VM not active on asic!\n");
372 			return -EINVAL;
373 		}
374 
375 		if (radeon_cs_get_ring(p, ring, priority))
376 			return -EINVAL;
377 
378 		/* we only support VM on some SI+ rings */
379 		if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
380 			if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
381 				DRM_ERROR("Ring %d requires VM!\n", p->ring);
382 				return -EINVAL;
383 			}
384 		} else {
385 			if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
386 				DRM_ERROR("VM not supported on ring %d!\n",
387 					  p->ring);
388 				return -EINVAL;
389 			}
390 		}
391 	}
392 
393 	return 0;
394 }
395 
396 static int cmp_size_smaller_first(void *priv, struct list_head *a,
397 				  struct list_head *b)
398 {
399 	struct radeon_bo_list *la = list_entry(a, struct radeon_bo_list, tv.head);
400 	struct radeon_bo_list *lb = list_entry(b, struct radeon_bo_list, tv.head);
401 
402 	/* Sort A before B if A is smaller. */
403 	return (int)la->robj->tbo.mem.num_pages -
404 		(int)lb->robj->tbo.mem.num_pages;
405 }
406 
407 /**
408  * cs_parser_fini() - clean parser states
409  * @parser:	parser structure holding parsing context.
410  * @error:	error number
411  * @backoff:	indicator to backoff the reservation
412  *
413  * If error is set than unvalidate buffer, otherwise just free memory
414  * used by parsing context.
415  **/
416 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
417 {
418 	unsigned i;
419 
420 	if (!error) {
421 		/* Sort the buffer list from the smallest to largest buffer,
422 		 * which affects the order of buffers in the LRU list.
423 		 * This assures that the smallest buffers are added first
424 		 * to the LRU list, so they are likely to be later evicted
425 		 * first, instead of large buffers whose eviction is more
426 		 * expensive.
427 		 *
428 		 * This slightly lowers the number of bytes moved by TTM
429 		 * per frame under memory pressure.
430 		 */
431 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
432 
433 		ttm_eu_fence_buffer_objects(&parser->ticket,
434 					    &parser->validated,
435 					    &parser->ib.fence->base);
436 	} else if (backoff) {
437 		ttm_eu_backoff_reservation(&parser->ticket,
438 					   &parser->validated);
439 	}
440 
441 	if (parser->relocs != NULL) {
442 		for (i = 0; i < parser->nrelocs; i++) {
443 			struct radeon_bo *bo = parser->relocs[i].robj;
444 			if (bo == NULL)
445 				continue;
446 
447 			drm_gem_object_put(&bo->tbo.base);
448 		}
449 	}
450 	kfree(parser->track);
451 	kvfree(parser->relocs);
452 	kvfree(parser->vm_bos);
453 	for (i = 0; i < parser->nchunks; i++)
454 		kvfree(parser->chunks[i].kdata);
455 	kfree(parser->chunks);
456 	kfree(parser->chunks_array);
457 	radeon_ib_free(parser->rdev, &parser->ib);
458 	radeon_ib_free(parser->rdev, &parser->const_ib);
459 }
460 
461 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
462 			      struct radeon_cs_parser *parser)
463 {
464 	int r;
465 
466 	if (parser->chunk_ib == NULL)
467 		return 0;
468 
469 	if (parser->cs_flags & RADEON_CS_USE_VM)
470 		return 0;
471 
472 	r = radeon_cs_parse(rdev, parser->ring, parser);
473 	if (r || parser->parser_error) {
474 		DRM_ERROR("Invalid command stream !\n");
475 		return r;
476 	}
477 
478 	r = radeon_cs_sync_rings(parser);
479 	if (r) {
480 		if (r != -ERESTARTSYS)
481 			DRM_ERROR("Failed to sync rings: %i\n", r);
482 		return r;
483 	}
484 
485 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
486 		radeon_uvd_note_usage(rdev);
487 	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
488 		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
489 		radeon_vce_note_usage(rdev);
490 
491 	r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
492 	if (r) {
493 		DRM_ERROR("Failed to schedule IB !\n");
494 	}
495 	return r;
496 }
497 
498 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
499 				   struct radeon_vm *vm)
500 {
501 	struct radeon_device *rdev = p->rdev;
502 	struct radeon_bo_va *bo_va;
503 	int i, r;
504 
505 	r = radeon_vm_update_page_directory(rdev, vm);
506 	if (r)
507 		return r;
508 
509 	r = radeon_vm_clear_freed(rdev, vm);
510 	if (r)
511 		return r;
512 
513 	if (vm->ib_bo_va == NULL) {
514 		DRM_ERROR("Tmp BO not in VM!\n");
515 		return -EINVAL;
516 	}
517 
518 	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
519 				&rdev->ring_tmp_bo.bo->tbo.mem);
520 	if (r)
521 		return r;
522 
523 	for (i = 0; i < p->nrelocs; i++) {
524 		struct radeon_bo *bo;
525 
526 		bo = p->relocs[i].robj;
527 		bo_va = radeon_vm_bo_find(vm, bo);
528 		if (bo_va == NULL) {
529 			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
530 			return -EINVAL;
531 		}
532 
533 		r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
534 		if (r)
535 			return r;
536 
537 		radeon_sync_fence(&p->ib.sync, bo_va->last_pt_update);
538 	}
539 
540 	return radeon_vm_clear_invalids(rdev, vm);
541 }
542 
543 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
544 				 struct radeon_cs_parser *parser)
545 {
546 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
547 	struct radeon_vm *vm = &fpriv->vm;
548 	int r;
549 
550 	if (parser->chunk_ib == NULL)
551 		return 0;
552 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
553 		return 0;
554 
555 	if (parser->const_ib.length_dw) {
556 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
557 		if (r) {
558 			return r;
559 		}
560 	}
561 
562 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
563 	if (r) {
564 		return r;
565 	}
566 
567 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
568 		radeon_uvd_note_usage(rdev);
569 
570 	mutex_lock(&vm->mutex);
571 	r = radeon_bo_vm_update_pte(parser, vm);
572 	if (r) {
573 		goto out;
574 	}
575 
576 	r = radeon_cs_sync_rings(parser);
577 	if (r) {
578 		if (r != -ERESTARTSYS)
579 			DRM_ERROR("Failed to sync rings: %i\n", r);
580 		goto out;
581 	}
582 
583 	if ((rdev->family >= CHIP_TAHITI) &&
584 	    (parser->chunk_const_ib != NULL)) {
585 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
586 	} else {
587 		r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
588 	}
589 
590 out:
591 	mutex_unlock(&vm->mutex);
592 	return r;
593 }
594 
595 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
596 {
597 	if (r == -EDEADLK) {
598 		r = radeon_gpu_reset(rdev);
599 		if (!r)
600 			r = -EAGAIN;
601 	}
602 	return r;
603 }
604 
605 static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
606 {
607 	struct radeon_cs_chunk *ib_chunk;
608 	struct radeon_vm *vm = NULL;
609 	int r;
610 
611 	if (parser->chunk_ib == NULL)
612 		return 0;
613 
614 	if (parser->cs_flags & RADEON_CS_USE_VM) {
615 		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
616 		vm = &fpriv->vm;
617 
618 		if ((rdev->family >= CHIP_TAHITI) &&
619 		    (parser->chunk_const_ib != NULL)) {
620 			ib_chunk = parser->chunk_const_ib;
621 			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
622 				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
623 				return -EINVAL;
624 			}
625 			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
626 					   vm, ib_chunk->length_dw * 4);
627 			if (r) {
628 				DRM_ERROR("Failed to get const ib !\n");
629 				return r;
630 			}
631 			parser->const_ib.is_const_ib = true;
632 			parser->const_ib.length_dw = ib_chunk->length_dw;
633 			if (copy_from_user(parser->const_ib.ptr,
634 					       ib_chunk->user_ptr,
635 					       ib_chunk->length_dw * 4))
636 				return -EFAULT;
637 		}
638 
639 		ib_chunk = parser->chunk_ib;
640 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
641 			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
642 			return -EINVAL;
643 		}
644 	}
645 	ib_chunk = parser->chunk_ib;
646 
647 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
648 			   vm, ib_chunk->length_dw * 4);
649 	if (r) {
650 		DRM_ERROR("Failed to get ib !\n");
651 		return r;
652 	}
653 	parser->ib.length_dw = ib_chunk->length_dw;
654 	if (ib_chunk->kdata)
655 		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
656 	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
657 		return -EFAULT;
658 	return 0;
659 }
660 
661 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
662 {
663 	struct radeon_device *rdev = dev->dev_private;
664 	struct radeon_cs_parser parser;
665 	int r;
666 
667 	down_read(&rdev->exclusive_lock);
668 	if (!rdev->accel_working) {
669 		up_read(&rdev->exclusive_lock);
670 		return -EBUSY;
671 	}
672 	if (rdev->in_reset) {
673 		up_read(&rdev->exclusive_lock);
674 		r = radeon_gpu_reset(rdev);
675 		if (!r)
676 			r = -EAGAIN;
677 		return r;
678 	}
679 	/* initialize parser */
680 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
681 	parser.filp = filp;
682 	parser.rdev = rdev;
683 	parser.dev = rdev->dev;
684 	parser.family = rdev->family;
685 	r = radeon_cs_parser_init(&parser, data);
686 	if (r) {
687 		DRM_ERROR("Failed to initialize parser !\n");
688 		radeon_cs_parser_fini(&parser, r, false);
689 		up_read(&rdev->exclusive_lock);
690 		r = radeon_cs_handle_lockup(rdev, r);
691 		return r;
692 	}
693 
694 	r = radeon_cs_ib_fill(rdev, &parser);
695 	if (!r) {
696 		r = radeon_cs_parser_relocs(&parser);
697 		if (r && r != -ERESTARTSYS)
698 			DRM_ERROR("Failed to parse relocation %d!\n", r);
699 	}
700 
701 	if (r) {
702 		radeon_cs_parser_fini(&parser, r, false);
703 		up_read(&rdev->exclusive_lock);
704 		r = radeon_cs_handle_lockup(rdev, r);
705 		return r;
706 	}
707 
708 	trace_radeon_cs(&parser);
709 
710 	r = radeon_cs_ib_chunk(rdev, &parser);
711 	if (r) {
712 		goto out;
713 	}
714 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
715 	if (r) {
716 		goto out;
717 	}
718 out:
719 	radeon_cs_parser_fini(&parser, r, true);
720 	up_read(&rdev->exclusive_lock);
721 	r = radeon_cs_handle_lockup(rdev, r);
722 	return r;
723 }
724 
725 /**
726  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
727  * @p:		parser structure holding parsing context.
728  * @pkt:	where to store packet information
729  * @idx:	packet index
730  *
731  * Assume that chunk_ib_index is properly set. Will return -EINVAL
732  * if packet is bigger than remaining ib size. or if packets is unknown.
733  **/
734 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
735 			   struct radeon_cs_packet *pkt,
736 			   unsigned idx)
737 {
738 	struct radeon_cs_chunk *ib_chunk = p->chunk_ib;
739 	struct radeon_device *rdev = p->rdev;
740 	uint32_t header;
741 	int ret = 0, i;
742 
743 	if (idx >= ib_chunk->length_dw) {
744 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
745 			  idx, ib_chunk->length_dw);
746 		return -EINVAL;
747 	}
748 	header = radeon_get_ib_value(p, idx);
749 	pkt->idx = idx;
750 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
751 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
752 	pkt->one_reg_wr = 0;
753 	switch (pkt->type) {
754 	case RADEON_PACKET_TYPE0:
755 		if (rdev->family < CHIP_R600) {
756 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
757 			pkt->one_reg_wr =
758 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
759 		} else
760 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
761 		break;
762 	case RADEON_PACKET_TYPE3:
763 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
764 		break;
765 	case RADEON_PACKET_TYPE2:
766 		pkt->count = -1;
767 		break;
768 	default:
769 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
770 		ret = -EINVAL;
771 		goto dump_ib;
772 	}
773 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
774 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
775 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
776 		ret = -EINVAL;
777 		goto dump_ib;
778 	}
779 	return 0;
780 
781 dump_ib:
782 	for (i = 0; i < ib_chunk->length_dw; i++) {
783 		if (i == idx)
784 			printk("\t0x%08x <---\n", radeon_get_ib_value(p, i));
785 		else
786 			printk("\t0x%08x\n", radeon_get_ib_value(p, i));
787 	}
788 	return ret;
789 }
790 
791 /**
792  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
793  * @p:		structure holding the parser context.
794  *
795  * Check if the next packet is NOP relocation packet3.
796  **/
797 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
798 {
799 	struct radeon_cs_packet p3reloc;
800 	int r;
801 
802 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
803 	if (r)
804 		return false;
805 	if (p3reloc.type != RADEON_PACKET_TYPE3)
806 		return false;
807 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
808 		return false;
809 	return true;
810 }
811 
812 /**
813  * radeon_cs_dump_packet() - dump raw packet context
814  * @p:		structure holding the parser context.
815  * @pkt:	structure holding the packet.
816  *
817  * Used mostly for debugging and error reporting.
818  **/
819 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
820 			   struct radeon_cs_packet *pkt)
821 {
822 	volatile uint32_t *ib;
823 	unsigned i;
824 	unsigned idx;
825 
826 	ib = p->ib.ptr;
827 	idx = pkt->idx;
828 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
829 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
830 }
831 
832 /**
833  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
834  * @p:			parser structure holding parsing context.
835  * @cs_reloc:		reloc informations
836  * @nomm:		no memory management for debugging
837  *
838  * Check if next packet is relocation packet3, do bo validation and compute
839  * GPU offset using the provided start.
840  **/
841 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
842 				struct radeon_bo_list **cs_reloc,
843 				int nomm)
844 {
845 	struct radeon_cs_chunk *relocs_chunk;
846 	struct radeon_cs_packet p3reloc;
847 	unsigned idx;
848 	int r;
849 
850 	if (p->chunk_relocs == NULL) {
851 		DRM_ERROR("No relocation chunk !\n");
852 		return -EINVAL;
853 	}
854 	*cs_reloc = NULL;
855 	relocs_chunk = p->chunk_relocs;
856 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
857 	if (r)
858 		return r;
859 	p->idx += p3reloc.count + 2;
860 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
861 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
862 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
863 			  p3reloc.idx);
864 		radeon_cs_dump_packet(p, &p3reloc);
865 		return -EINVAL;
866 	}
867 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
868 	if (idx >= relocs_chunk->length_dw) {
869 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
870 			  idx, relocs_chunk->length_dw);
871 		radeon_cs_dump_packet(p, &p3reloc);
872 		return -EINVAL;
873 	}
874 	/* FIXME: we assume reloc size is 4 dwords */
875 	if (nomm) {
876 		*cs_reloc = p->relocs;
877 		(*cs_reloc)->gpu_offset =
878 			(u64)relocs_chunk->kdata[idx + 3] << 32;
879 		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
880 	} else
881 		*cs_reloc = &p->relocs[(idx / 4)];
882 	return 0;
883 }
884