xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision a36954f5)
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 = drm_calloc_large(p->nrelocs, sizeof(struct radeon_bo_list));
91 	if (p->relocs == NULL) {
92 		return -ENOMEM;
93 	}
94 
95 	radeon_cs_buckets_init(&buckets);
96 
97 	for (i = 0; i < p->nrelocs; i++) {
98 		struct drm_radeon_cs_reloc *r;
99 		struct drm_gem_object *gobj;
100 		unsigned priority;
101 
102 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
103 		gobj = drm_gem_object_lookup(p->filp, r->handle);
104 		if (gobj == NULL) {
105 			DRM_ERROR("gem object lookup failed 0x%x\n",
106 				  r->handle);
107 			return -ENOENT;
108 		}
109 		p->relocs[i].robj = gem_to_radeon_bo(gobj);
110 
111 		/* The userspace buffer priorities are from 0 to 15. A higher
112 		 * number means the buffer is more important.
113 		 * Also, the buffers used for write have a higher priority than
114 		 * the buffers used for read only, which doubles the range
115 		 * to 0 to 31. 32 is reserved for the kernel driver.
116 		 */
117 		priority = (r->flags & RADEON_RELOC_PRIO_MASK) * 2
118 			   + !!r->write_domain;
119 
120 		/* The first reloc of an UVD job is the msg and that must be in
121 		 * VRAM, the second reloc is the DPB and for WMV that must be in
122 		 * VRAM as well. Also put everything into VRAM on AGP cards and older
123 		 * IGP chips to avoid image corruptions
124 		 */
125 		if (p->ring == R600_RING_TYPE_UVD_INDEX &&
126 		    (i <= 0 || pci_find_capability(p->rdev->ddev->pdev,
127 						   PCI_CAP_ID_AGP) ||
128 		     p->rdev->family == CHIP_RS780 ||
129 		     p->rdev->family == CHIP_RS880)) {
130 
131 			/* TODO: is this still needed for NI+ ? */
132 			p->relocs[i].prefered_domains =
133 				RADEON_GEM_DOMAIN_VRAM;
134 
135 			p->relocs[i].allowed_domains =
136 				RADEON_GEM_DOMAIN_VRAM;
137 
138 			/* prioritize this over any other relocation */
139 			priority = RADEON_CS_MAX_PRIORITY;
140 		} else {
141 			uint32_t domain = r->write_domain ?
142 				r->write_domain : r->read_domains;
143 
144 			if (domain & RADEON_GEM_DOMAIN_CPU) {
145 				DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
146 					  "for command submission\n");
147 				return -EINVAL;
148 			}
149 
150 			p->relocs[i].prefered_domains = domain;
151 			if (domain == RADEON_GEM_DOMAIN_VRAM)
152 				domain |= RADEON_GEM_DOMAIN_GTT;
153 			p->relocs[i].allowed_domains = domain;
154 		}
155 
156 		if (radeon_ttm_tt_has_userptr(p->relocs[i].robj->tbo.ttm)) {
157 			uint32_t domain = p->relocs[i].prefered_domains;
158 			if (!(domain & RADEON_GEM_DOMAIN_GTT)) {
159 				DRM_ERROR("Only RADEON_GEM_DOMAIN_GTT is "
160 					  "allowed for userptr BOs\n");
161 				return -EINVAL;
162 			}
163 			need_mmap_lock = true;
164 			domain = RADEON_GEM_DOMAIN_GTT;
165 			p->relocs[i].prefered_domains = domain;
166 			p->relocs[i].allowed_domains = domain;
167 		}
168 
169 		/* Objects shared as dma-bufs cannot be moved to VRAM */
170 		if (p->relocs[i].robj->prime_shared_count) {
171 			p->relocs[i].allowed_domains &= ~RADEON_GEM_DOMAIN_VRAM;
172 			if (!p->relocs[i].allowed_domains) {
173 				DRM_ERROR("BO associated with dma-buf cannot "
174 					  "be moved to VRAM\n");
175 				return -EINVAL;
176 			}
177 		}
178 
179 		p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
180 		p->relocs[i].tv.shared = !r->write_domain;
181 
182 		radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head,
183 				      priority);
184 	}
185 
186 	radeon_cs_buckets_get_list(&buckets, &p->validated);
187 
188 	if (p->cs_flags & RADEON_CS_USE_VM)
189 		p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm,
190 					      &p->validated);
191 	if (need_mmap_lock)
192 		down_read(&current->mm->mmap_sem);
193 
194 	r = radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
195 
196 	if (need_mmap_lock)
197 		up_read(&current->mm->mmap_sem);
198 
199 	return r;
200 }
201 
202 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
203 {
204 	p->priority = priority;
205 
206 	switch (ring) {
207 	default:
208 		DRM_ERROR("unknown ring id: %d\n", ring);
209 		return -EINVAL;
210 	case RADEON_CS_RING_GFX:
211 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
212 		break;
213 	case RADEON_CS_RING_COMPUTE:
214 		if (p->rdev->family >= CHIP_TAHITI) {
215 			if (p->priority > 0)
216 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
217 			else
218 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
219 		} else
220 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
221 		break;
222 	case RADEON_CS_RING_DMA:
223 		if (p->rdev->family >= CHIP_CAYMAN) {
224 			if (p->priority > 0)
225 				p->ring = R600_RING_TYPE_DMA_INDEX;
226 			else
227 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
228 		} else if (p->rdev->family >= CHIP_RV770) {
229 			p->ring = R600_RING_TYPE_DMA_INDEX;
230 		} else {
231 			return -EINVAL;
232 		}
233 		break;
234 	case RADEON_CS_RING_UVD:
235 		p->ring = R600_RING_TYPE_UVD_INDEX;
236 		break;
237 	case RADEON_CS_RING_VCE:
238 		/* TODO: only use the low priority ring for now */
239 		p->ring = TN_RING_TYPE_VCE1_INDEX;
240 		break;
241 	}
242 	return 0;
243 }
244 
245 static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
246 {
247 	struct radeon_bo_list *reloc;
248 	int r;
249 
250 	list_for_each_entry(reloc, &p->validated, tv.head) {
251 		struct reservation_object *resv;
252 
253 		resv = reloc->robj->tbo.resv;
254 		r = radeon_sync_resv(p->rdev, &p->ib.sync, resv,
255 				     reloc->tv.shared);
256 		if (r)
257 			return r;
258 	}
259 	return 0;
260 }
261 
262 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
263 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
264 {
265 	struct drm_radeon_cs *cs = data;
266 	uint64_t *chunk_array_ptr;
267 	unsigned size, i;
268 	u32 ring = RADEON_CS_RING_GFX;
269 	s32 priority = 0;
270 
271 	INIT_LIST_HEAD(&p->validated);
272 
273 	if (!cs->num_chunks) {
274 		return 0;
275 	}
276 
277 	/* get chunks */
278 	p->idx = 0;
279 	p->ib.sa_bo = NULL;
280 	p->const_ib.sa_bo = NULL;
281 	p->chunk_ib = NULL;
282 	p->chunk_relocs = NULL;
283 	p->chunk_flags = NULL;
284 	p->chunk_const_ib = NULL;
285 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
286 	if (p->chunks_array == NULL) {
287 		return -ENOMEM;
288 	}
289 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
290 	if (copy_from_user(p->chunks_array, chunk_array_ptr,
291 			       sizeof(uint64_t)*cs->num_chunks)) {
292 		return -EFAULT;
293 	}
294 	p->cs_flags = 0;
295 	p->nchunks = cs->num_chunks;
296 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
297 	if (p->chunks == NULL) {
298 		return -ENOMEM;
299 	}
300 	for (i = 0; i < p->nchunks; i++) {
301 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
302 		struct drm_radeon_cs_chunk user_chunk;
303 		uint32_t __user *cdata;
304 
305 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
306 		if (copy_from_user(&user_chunk, chunk_ptr,
307 				       sizeof(struct drm_radeon_cs_chunk))) {
308 			return -EFAULT;
309 		}
310 		p->chunks[i].length_dw = user_chunk.length_dw;
311 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_RELOCS) {
312 			p->chunk_relocs = &p->chunks[i];
313 		}
314 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
315 			p->chunk_ib = &p->chunks[i];
316 			/* zero length IB isn't useful */
317 			if (p->chunks[i].length_dw == 0)
318 				return -EINVAL;
319 		}
320 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB) {
321 			p->chunk_const_ib = &p->chunks[i];
322 			/* zero length CONST 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_FLAGS) {
327 			p->chunk_flags = &p->chunks[i];
328 			/* zero length flags aren't useful */
329 			if (p->chunks[i].length_dw == 0)
330 				return -EINVAL;
331 		}
332 
333 		size = p->chunks[i].length_dw;
334 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
335 		p->chunks[i].user_ptr = cdata;
336 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB)
337 			continue;
338 
339 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
340 			if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
341 				continue;
342 		}
343 
344 		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
345 		size *= sizeof(uint32_t);
346 		if (p->chunks[i].kdata == NULL) {
347 			return -ENOMEM;
348 		}
349 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
350 			return -EFAULT;
351 		}
352 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
353 			p->cs_flags = p->chunks[i].kdata[0];
354 			if (p->chunks[i].length_dw > 1)
355 				ring = p->chunks[i].kdata[1];
356 			if (p->chunks[i].length_dw > 2)
357 				priority = (s32)p->chunks[i].kdata[2];
358 		}
359 	}
360 
361 	/* these are KMS only */
362 	if (p->rdev) {
363 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
364 		    !p->rdev->vm_manager.enabled) {
365 			DRM_ERROR("VM not active on asic!\n");
366 			return -EINVAL;
367 		}
368 
369 		if (radeon_cs_get_ring(p, ring, priority))
370 			return -EINVAL;
371 
372 		/* we only support VM on some SI+ rings */
373 		if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
374 			if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
375 				DRM_ERROR("Ring %d requires VM!\n", p->ring);
376 				return -EINVAL;
377 			}
378 		} else {
379 			if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
380 				DRM_ERROR("VM not supported on ring %d!\n",
381 					  p->ring);
382 				return -EINVAL;
383 			}
384 		}
385 	}
386 
387 	return 0;
388 }
389 
390 static int cmp_size_smaller_first(void *priv, struct list_head *a,
391 				  struct list_head *b)
392 {
393 	struct radeon_bo_list *la = list_entry(a, struct radeon_bo_list, tv.head);
394 	struct radeon_bo_list *lb = list_entry(b, struct radeon_bo_list, tv.head);
395 
396 	/* Sort A before B if A is smaller. */
397 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
398 }
399 
400 /**
401  * cs_parser_fini() - clean parser states
402  * @parser:	parser structure holding parsing context.
403  * @error:	error number
404  *
405  * If error is set than unvalidate buffer, otherwise just free memory
406  * used by parsing context.
407  **/
408 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
409 {
410 	unsigned i;
411 
412 	if (!error) {
413 		/* Sort the buffer list from the smallest to largest buffer,
414 		 * which affects the order of buffers in the LRU list.
415 		 * This assures that the smallest buffers are added first
416 		 * to the LRU list, so they are likely to be later evicted
417 		 * first, instead of large buffers whose eviction is more
418 		 * expensive.
419 		 *
420 		 * This slightly lowers the number of bytes moved by TTM
421 		 * per frame under memory pressure.
422 		 */
423 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
424 
425 		ttm_eu_fence_buffer_objects(&parser->ticket,
426 					    &parser->validated,
427 					    &parser->ib.fence->base);
428 	} else if (backoff) {
429 		ttm_eu_backoff_reservation(&parser->ticket,
430 					   &parser->validated);
431 	}
432 
433 	if (parser->relocs != NULL) {
434 		for (i = 0; i < parser->nrelocs; i++) {
435 			struct radeon_bo *bo = parser->relocs[i].robj;
436 			if (bo == NULL)
437 				continue;
438 
439 			drm_gem_object_unreference_unlocked(&bo->gem_base);
440 		}
441 	}
442 	kfree(parser->track);
443 	drm_free_large(parser->relocs);
444 	drm_free_large(parser->vm_bos);
445 	for (i = 0; i < parser->nchunks; i++)
446 		drm_free_large(parser->chunks[i].kdata);
447 	kfree(parser->chunks);
448 	kfree(parser->chunks_array);
449 	radeon_ib_free(parser->rdev, &parser->ib);
450 	radeon_ib_free(parser->rdev, &parser->const_ib);
451 }
452 
453 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
454 			      struct radeon_cs_parser *parser)
455 {
456 	int r;
457 
458 	if (parser->chunk_ib == NULL)
459 		return 0;
460 
461 	if (parser->cs_flags & RADEON_CS_USE_VM)
462 		return 0;
463 
464 	r = radeon_cs_parse(rdev, parser->ring, parser);
465 	if (r || parser->parser_error) {
466 		DRM_ERROR("Invalid command stream !\n");
467 		return r;
468 	}
469 
470 	r = radeon_cs_sync_rings(parser);
471 	if (r) {
472 		if (r != -ERESTARTSYS)
473 			DRM_ERROR("Failed to sync rings: %i\n", r);
474 		return r;
475 	}
476 
477 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
478 		radeon_uvd_note_usage(rdev);
479 	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
480 		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
481 		radeon_vce_note_usage(rdev);
482 
483 	r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
484 	if (r) {
485 		DRM_ERROR("Failed to schedule IB !\n");
486 	}
487 	return r;
488 }
489 
490 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
491 				   struct radeon_vm *vm)
492 {
493 	struct radeon_device *rdev = p->rdev;
494 	struct radeon_bo_va *bo_va;
495 	int i, r;
496 
497 	r = radeon_vm_update_page_directory(rdev, vm);
498 	if (r)
499 		return r;
500 
501 	r = radeon_vm_clear_freed(rdev, vm);
502 	if (r)
503 		return r;
504 
505 	if (vm->ib_bo_va == NULL) {
506 		DRM_ERROR("Tmp BO not in VM!\n");
507 		return -EINVAL;
508 	}
509 
510 	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
511 				&rdev->ring_tmp_bo.bo->tbo.mem);
512 	if (r)
513 		return r;
514 
515 	for (i = 0; i < p->nrelocs; i++) {
516 		struct radeon_bo *bo;
517 
518 		bo = p->relocs[i].robj;
519 		bo_va = radeon_vm_bo_find(vm, bo);
520 		if (bo_va == NULL) {
521 			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
522 			return -EINVAL;
523 		}
524 
525 		r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
526 		if (r)
527 			return r;
528 
529 		radeon_sync_fence(&p->ib.sync, bo_va->last_pt_update);
530 	}
531 
532 	return radeon_vm_clear_invalids(rdev, vm);
533 }
534 
535 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
536 				 struct radeon_cs_parser *parser)
537 {
538 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
539 	struct radeon_vm *vm = &fpriv->vm;
540 	int r;
541 
542 	if (parser->chunk_ib == NULL)
543 		return 0;
544 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
545 		return 0;
546 
547 	if (parser->const_ib.length_dw) {
548 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
549 		if (r) {
550 			return r;
551 		}
552 	}
553 
554 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
555 	if (r) {
556 		return r;
557 	}
558 
559 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
560 		radeon_uvd_note_usage(rdev);
561 
562 	mutex_lock(&vm->mutex);
563 	r = radeon_bo_vm_update_pte(parser, vm);
564 	if (r) {
565 		goto out;
566 	}
567 
568 	r = radeon_cs_sync_rings(parser);
569 	if (r) {
570 		if (r != -ERESTARTSYS)
571 			DRM_ERROR("Failed to sync rings: %i\n", r);
572 		goto out;
573 	}
574 
575 	if ((rdev->family >= CHIP_TAHITI) &&
576 	    (parser->chunk_const_ib != NULL)) {
577 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
578 	} else {
579 		r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
580 	}
581 
582 out:
583 	mutex_unlock(&vm->mutex);
584 	return r;
585 }
586 
587 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
588 {
589 	if (r == -EDEADLK) {
590 		r = radeon_gpu_reset(rdev);
591 		if (!r)
592 			r = -EAGAIN;
593 	}
594 	return r;
595 }
596 
597 static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
598 {
599 	struct radeon_cs_chunk *ib_chunk;
600 	struct radeon_vm *vm = NULL;
601 	int r;
602 
603 	if (parser->chunk_ib == NULL)
604 		return 0;
605 
606 	if (parser->cs_flags & RADEON_CS_USE_VM) {
607 		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
608 		vm = &fpriv->vm;
609 
610 		if ((rdev->family >= CHIP_TAHITI) &&
611 		    (parser->chunk_const_ib != NULL)) {
612 			ib_chunk = parser->chunk_const_ib;
613 			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
614 				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
615 				return -EINVAL;
616 			}
617 			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
618 					   vm, ib_chunk->length_dw * 4);
619 			if (r) {
620 				DRM_ERROR("Failed to get const ib !\n");
621 				return r;
622 			}
623 			parser->const_ib.is_const_ib = true;
624 			parser->const_ib.length_dw = ib_chunk->length_dw;
625 			if (copy_from_user(parser->const_ib.ptr,
626 					       ib_chunk->user_ptr,
627 					       ib_chunk->length_dw * 4))
628 				return -EFAULT;
629 		}
630 
631 		ib_chunk = parser->chunk_ib;
632 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
633 			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
634 			return -EINVAL;
635 		}
636 	}
637 	ib_chunk = parser->chunk_ib;
638 
639 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
640 			   vm, ib_chunk->length_dw * 4);
641 	if (r) {
642 		DRM_ERROR("Failed to get ib !\n");
643 		return r;
644 	}
645 	parser->ib.length_dw = ib_chunk->length_dw;
646 	if (ib_chunk->kdata)
647 		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
648 	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
649 		return -EFAULT;
650 	return 0;
651 }
652 
653 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
654 {
655 	struct radeon_device *rdev = dev->dev_private;
656 	struct radeon_cs_parser parser;
657 	int r;
658 
659 	down_read(&rdev->exclusive_lock);
660 	if (!rdev->accel_working) {
661 		up_read(&rdev->exclusive_lock);
662 		return -EBUSY;
663 	}
664 	if (rdev->in_reset) {
665 		up_read(&rdev->exclusive_lock);
666 		r = radeon_gpu_reset(rdev);
667 		if (!r)
668 			r = -EAGAIN;
669 		return r;
670 	}
671 	/* initialize parser */
672 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
673 	parser.filp = filp;
674 	parser.rdev = rdev;
675 	parser.dev = rdev->dev;
676 	parser.family = rdev->family;
677 	r = radeon_cs_parser_init(&parser, data);
678 	if (r) {
679 		DRM_ERROR("Failed to initialize parser !\n");
680 		radeon_cs_parser_fini(&parser, r, false);
681 		up_read(&rdev->exclusive_lock);
682 		r = radeon_cs_handle_lockup(rdev, r);
683 		return r;
684 	}
685 
686 	r = radeon_cs_ib_fill(rdev, &parser);
687 	if (!r) {
688 		r = radeon_cs_parser_relocs(&parser);
689 		if (r && r != -ERESTARTSYS)
690 			DRM_ERROR("Failed to parse relocation %d!\n", r);
691 	}
692 
693 	if (r) {
694 		radeon_cs_parser_fini(&parser, r, false);
695 		up_read(&rdev->exclusive_lock);
696 		r = radeon_cs_handle_lockup(rdev, r);
697 		return r;
698 	}
699 
700 	trace_radeon_cs(&parser);
701 
702 	r = radeon_cs_ib_chunk(rdev, &parser);
703 	if (r) {
704 		goto out;
705 	}
706 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
707 	if (r) {
708 		goto out;
709 	}
710 out:
711 	radeon_cs_parser_fini(&parser, r, true);
712 	up_read(&rdev->exclusive_lock);
713 	r = radeon_cs_handle_lockup(rdev, r);
714 	return r;
715 }
716 
717 /**
718  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
719  * @parser:	parser structure holding parsing context.
720  * @pkt:	where to store packet information
721  *
722  * Assume that chunk_ib_index is properly set. Will return -EINVAL
723  * if packet is bigger than remaining ib size. or if packets is unknown.
724  **/
725 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
726 			   struct radeon_cs_packet *pkt,
727 			   unsigned idx)
728 {
729 	struct radeon_cs_chunk *ib_chunk = p->chunk_ib;
730 	struct radeon_device *rdev = p->rdev;
731 	uint32_t header;
732 	int ret = 0, i;
733 
734 	if (idx >= ib_chunk->length_dw) {
735 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
736 			  idx, ib_chunk->length_dw);
737 		return -EINVAL;
738 	}
739 	header = radeon_get_ib_value(p, idx);
740 	pkt->idx = idx;
741 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
742 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
743 	pkt->one_reg_wr = 0;
744 	switch (pkt->type) {
745 	case RADEON_PACKET_TYPE0:
746 		if (rdev->family < CHIP_R600) {
747 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
748 			pkt->one_reg_wr =
749 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
750 		} else
751 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
752 		break;
753 	case RADEON_PACKET_TYPE3:
754 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
755 		break;
756 	case RADEON_PACKET_TYPE2:
757 		pkt->count = -1;
758 		break;
759 	default:
760 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
761 		ret = -EINVAL;
762 		goto dump_ib;
763 	}
764 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
765 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
766 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
767 		ret = -EINVAL;
768 		goto dump_ib;
769 	}
770 	return 0;
771 
772 dump_ib:
773 	for (i = 0; i < ib_chunk->length_dw; i++) {
774 		if (i == idx)
775 			printk("\t0x%08x <---\n", radeon_get_ib_value(p, i));
776 		else
777 			printk("\t0x%08x\n", radeon_get_ib_value(p, i));
778 	}
779 	return ret;
780 }
781 
782 /**
783  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
784  * @p:		structure holding the parser context.
785  *
786  * Check if the next packet is NOP relocation packet3.
787  **/
788 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
789 {
790 	struct radeon_cs_packet p3reloc;
791 	int r;
792 
793 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
794 	if (r)
795 		return false;
796 	if (p3reloc.type != RADEON_PACKET_TYPE3)
797 		return false;
798 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
799 		return false;
800 	return true;
801 }
802 
803 /**
804  * radeon_cs_dump_packet() - dump raw packet context
805  * @p:		structure holding the parser context.
806  * @pkt:	structure holding the packet.
807  *
808  * Used mostly for debugging and error reporting.
809  **/
810 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
811 			   struct radeon_cs_packet *pkt)
812 {
813 	volatile uint32_t *ib;
814 	unsigned i;
815 	unsigned idx;
816 
817 	ib = p->ib.ptr;
818 	idx = pkt->idx;
819 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
820 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
821 }
822 
823 /**
824  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
825  * @parser:		parser structure holding parsing context.
826  * @data:		pointer to relocation data
827  * @offset_start:	starting offset
828  * @offset_mask:	offset mask (to align start offset on)
829  * @reloc:		reloc informations
830  *
831  * Check if next packet is relocation packet3, do bo validation and compute
832  * GPU offset using the provided start.
833  **/
834 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
835 				struct radeon_bo_list **cs_reloc,
836 				int nomm)
837 {
838 	struct radeon_cs_chunk *relocs_chunk;
839 	struct radeon_cs_packet p3reloc;
840 	unsigned idx;
841 	int r;
842 
843 	if (p->chunk_relocs == NULL) {
844 		DRM_ERROR("No relocation chunk !\n");
845 		return -EINVAL;
846 	}
847 	*cs_reloc = NULL;
848 	relocs_chunk = p->chunk_relocs;
849 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
850 	if (r)
851 		return r;
852 	p->idx += p3reloc.count + 2;
853 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
854 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
855 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
856 			  p3reloc.idx);
857 		radeon_cs_dump_packet(p, &p3reloc);
858 		return -EINVAL;
859 	}
860 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
861 	if (idx >= relocs_chunk->length_dw) {
862 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
863 			  idx, relocs_chunk->length_dw);
864 		radeon_cs_dump_packet(p, &p3reloc);
865 		return -EINVAL;
866 	}
867 	/* FIXME: we assume reloc size is 4 dwords */
868 	if (nomm) {
869 		*cs_reloc = p->relocs;
870 		(*cs_reloc)->gpu_offset =
871 			(u64)relocs_chunk->kdata[idx + 3] << 32;
872 		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
873 	} else
874 		*cs_reloc = &p->relocs[(idx / 4)];
875 	return 0;
876 }
877