xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision 62e7ca52)
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 drm_device *ddev = p->rdev->ddev;
78 	struct radeon_cs_chunk *chunk;
79 	struct radeon_cs_buckets buckets;
80 	unsigned i, j;
81 	bool duplicate;
82 
83 	if (p->chunk_relocs_idx == -1) {
84 		return 0;
85 	}
86 	chunk = &p->chunks[p->chunk_relocs_idx];
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_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
91 	if (p->relocs_ptr == NULL) {
92 		return -ENOMEM;
93 	}
94 	p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
95 	if (p->relocs == NULL) {
96 		return -ENOMEM;
97 	}
98 
99 	radeon_cs_buckets_init(&buckets);
100 
101 	for (i = 0; i < p->nrelocs; i++) {
102 		struct drm_radeon_cs_reloc *r;
103 		unsigned priority;
104 
105 		duplicate = false;
106 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
107 		for (j = 0; j < i; j++) {
108 			if (r->handle == p->relocs[j].handle) {
109 				p->relocs_ptr[i] = &p->relocs[j];
110 				duplicate = true;
111 				break;
112 			}
113 		}
114 		if (duplicate) {
115 			p->relocs[i].handle = 0;
116 			continue;
117 		}
118 
119 		p->relocs[i].gobj = drm_gem_object_lookup(ddev, p->filp,
120 							  r->handle);
121 		if (p->relocs[i].gobj == NULL) {
122 			DRM_ERROR("gem object lookup failed 0x%x\n",
123 				  r->handle);
124 			return -ENOENT;
125 		}
126 		p->relocs_ptr[i] = &p->relocs[i];
127 		p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
128 
129 		/* The userspace buffer priorities are from 0 to 15. A higher
130 		 * number means the buffer is more important.
131 		 * Also, the buffers used for write have a higher priority than
132 		 * the buffers used for read only, which doubles the range
133 		 * to 0 to 31. 32 is reserved for the kernel driver.
134 		 */
135 		priority = (r->flags & 0xf) * 2 + !!r->write_domain;
136 
137 		/* the first reloc of an UVD job is the msg and that must be in
138 		   VRAM, also but everything into VRAM on AGP cards to avoid
139 		   image corruptions */
140 		if (p->ring == R600_RING_TYPE_UVD_INDEX &&
141 		    (i == 0 || drm_pci_device_is_agp(p->rdev->ddev))) {
142 			/* TODO: is this still needed for NI+ ? */
143 			p->relocs[i].prefered_domains =
144 				RADEON_GEM_DOMAIN_VRAM;
145 
146 			p->relocs[i].allowed_domains =
147 				RADEON_GEM_DOMAIN_VRAM;
148 
149 			/* prioritize this over any other relocation */
150 			priority = RADEON_CS_MAX_PRIORITY;
151 		} else {
152 			uint32_t domain = r->write_domain ?
153 				r->write_domain : r->read_domains;
154 
155 			if (domain & RADEON_GEM_DOMAIN_CPU) {
156 				DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
157 					  "for command submission\n");
158 				return -EINVAL;
159 			}
160 
161 			p->relocs[i].prefered_domains = domain;
162 			if (domain == RADEON_GEM_DOMAIN_VRAM)
163 				domain |= RADEON_GEM_DOMAIN_GTT;
164 			p->relocs[i].allowed_domains = domain;
165 		}
166 
167 		p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
168 		p->relocs[i].handle = r->handle;
169 
170 		radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head,
171 				      priority);
172 	}
173 
174 	radeon_cs_buckets_get_list(&buckets, &p->validated);
175 
176 	if (p->cs_flags & RADEON_CS_USE_VM)
177 		p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm,
178 					      &p->validated);
179 
180 	return radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
181 }
182 
183 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
184 {
185 	p->priority = priority;
186 
187 	switch (ring) {
188 	default:
189 		DRM_ERROR("unknown ring id: %d\n", ring);
190 		return -EINVAL;
191 	case RADEON_CS_RING_GFX:
192 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
193 		break;
194 	case RADEON_CS_RING_COMPUTE:
195 		if (p->rdev->family >= CHIP_TAHITI) {
196 			if (p->priority > 0)
197 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
198 			else
199 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
200 		} else
201 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
202 		break;
203 	case RADEON_CS_RING_DMA:
204 		if (p->rdev->family >= CHIP_CAYMAN) {
205 			if (p->priority > 0)
206 				p->ring = R600_RING_TYPE_DMA_INDEX;
207 			else
208 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
209 		} else if (p->rdev->family >= CHIP_RV770) {
210 			p->ring = R600_RING_TYPE_DMA_INDEX;
211 		} else {
212 			return -EINVAL;
213 		}
214 		break;
215 	case RADEON_CS_RING_UVD:
216 		p->ring = R600_RING_TYPE_UVD_INDEX;
217 		break;
218 	case RADEON_CS_RING_VCE:
219 		/* TODO: only use the low priority ring for now */
220 		p->ring = TN_RING_TYPE_VCE1_INDEX;
221 		break;
222 	}
223 	return 0;
224 }
225 
226 static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
227 {
228 	int i;
229 
230 	for (i = 0; i < p->nrelocs; i++) {
231 		if (!p->relocs[i].robj)
232 			continue;
233 
234 		radeon_semaphore_sync_to(p->ib.semaphore,
235 					 p->relocs[i].robj->tbo.sync_obj);
236 	}
237 }
238 
239 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
240 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
241 {
242 	struct drm_radeon_cs *cs = data;
243 	uint64_t *chunk_array_ptr;
244 	unsigned size, i;
245 	u32 ring = RADEON_CS_RING_GFX;
246 	s32 priority = 0;
247 
248 	if (!cs->num_chunks) {
249 		return 0;
250 	}
251 	/* get chunks */
252 	INIT_LIST_HEAD(&p->validated);
253 	p->idx = 0;
254 	p->ib.sa_bo = NULL;
255 	p->ib.semaphore = NULL;
256 	p->const_ib.sa_bo = NULL;
257 	p->const_ib.semaphore = NULL;
258 	p->chunk_ib_idx = -1;
259 	p->chunk_relocs_idx = -1;
260 	p->chunk_flags_idx = -1;
261 	p->chunk_const_ib_idx = -1;
262 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
263 	if (p->chunks_array == NULL) {
264 		return -ENOMEM;
265 	}
266 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
267 	if (copy_from_user(p->chunks_array, chunk_array_ptr,
268 			       sizeof(uint64_t)*cs->num_chunks)) {
269 		return -EFAULT;
270 	}
271 	p->cs_flags = 0;
272 	p->nchunks = cs->num_chunks;
273 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
274 	if (p->chunks == NULL) {
275 		return -ENOMEM;
276 	}
277 	for (i = 0; i < p->nchunks; i++) {
278 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
279 		struct drm_radeon_cs_chunk user_chunk;
280 		uint32_t __user *cdata;
281 
282 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
283 		if (copy_from_user(&user_chunk, chunk_ptr,
284 				       sizeof(struct drm_radeon_cs_chunk))) {
285 			return -EFAULT;
286 		}
287 		p->chunks[i].length_dw = user_chunk.length_dw;
288 		p->chunks[i].chunk_id = user_chunk.chunk_id;
289 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
290 			p->chunk_relocs_idx = i;
291 		}
292 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
293 			p->chunk_ib_idx = i;
294 			/* zero length IB isn't useful */
295 			if (p->chunks[i].length_dw == 0)
296 				return -EINVAL;
297 		}
298 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
299 			p->chunk_const_ib_idx = i;
300 			/* zero length CONST IB isn't useful */
301 			if (p->chunks[i].length_dw == 0)
302 				return -EINVAL;
303 		}
304 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
305 			p->chunk_flags_idx = i;
306 			/* zero length flags aren't useful */
307 			if (p->chunks[i].length_dw == 0)
308 				return -EINVAL;
309 		}
310 
311 		size = p->chunks[i].length_dw;
312 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
313 		p->chunks[i].user_ptr = cdata;
314 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB)
315 			continue;
316 
317 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
318 			if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
319 				continue;
320 		}
321 
322 		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
323 		size *= sizeof(uint32_t);
324 		if (p->chunks[i].kdata == NULL) {
325 			return -ENOMEM;
326 		}
327 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
328 			return -EFAULT;
329 		}
330 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
331 			p->cs_flags = p->chunks[i].kdata[0];
332 			if (p->chunks[i].length_dw > 1)
333 				ring = p->chunks[i].kdata[1];
334 			if (p->chunks[i].length_dw > 2)
335 				priority = (s32)p->chunks[i].kdata[2];
336 		}
337 	}
338 
339 	/* these are KMS only */
340 	if (p->rdev) {
341 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
342 		    !p->rdev->vm_manager.enabled) {
343 			DRM_ERROR("VM not active on asic!\n");
344 			return -EINVAL;
345 		}
346 
347 		if (radeon_cs_get_ring(p, ring, priority))
348 			return -EINVAL;
349 
350 		/* we only support VM on some SI+ rings */
351 		if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
352 			if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
353 				DRM_ERROR("Ring %d requires VM!\n", p->ring);
354 				return -EINVAL;
355 			}
356 		} else {
357 			if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
358 				DRM_ERROR("VM not supported on ring %d!\n",
359 					  p->ring);
360 				return -EINVAL;
361 			}
362 		}
363 	}
364 
365 	return 0;
366 }
367 
368 static int cmp_size_smaller_first(void *priv, struct list_head *a,
369 				  struct list_head *b)
370 {
371 	struct radeon_cs_reloc *la = list_entry(a, struct radeon_cs_reloc, tv.head);
372 	struct radeon_cs_reloc *lb = list_entry(b, struct radeon_cs_reloc, tv.head);
373 
374 	/* Sort A before B if A is smaller. */
375 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
376 }
377 
378 /**
379  * cs_parser_fini() - clean parser states
380  * @parser:	parser structure holding parsing context.
381  * @error:	error number
382  *
383  * If error is set than unvalidate buffer, otherwise just free memory
384  * used by parsing context.
385  **/
386 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
387 {
388 	unsigned i;
389 
390 	if (!error) {
391 		/* Sort the buffer list from the smallest to largest buffer,
392 		 * which affects the order of buffers in the LRU list.
393 		 * This assures that the smallest buffers are added first
394 		 * to the LRU list, so they are likely to be later evicted
395 		 * first, instead of large buffers whose eviction is more
396 		 * expensive.
397 		 *
398 		 * This slightly lowers the number of bytes moved by TTM
399 		 * per frame under memory pressure.
400 		 */
401 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
402 
403 		ttm_eu_fence_buffer_objects(&parser->ticket,
404 					    &parser->validated,
405 					    parser->ib.fence);
406 	} else if (backoff) {
407 		ttm_eu_backoff_reservation(&parser->ticket,
408 					   &parser->validated);
409 	}
410 
411 	if (parser->relocs != NULL) {
412 		for (i = 0; i < parser->nrelocs; i++) {
413 			if (parser->relocs[i].gobj)
414 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
415 		}
416 	}
417 	kfree(parser->track);
418 	kfree(parser->relocs);
419 	kfree(parser->relocs_ptr);
420 	kfree(parser->vm_bos);
421 	for (i = 0; i < parser->nchunks; i++)
422 		drm_free_large(parser->chunks[i].kdata);
423 	kfree(parser->chunks);
424 	kfree(parser->chunks_array);
425 	radeon_ib_free(parser->rdev, &parser->ib);
426 	radeon_ib_free(parser->rdev, &parser->const_ib);
427 }
428 
429 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
430 			      struct radeon_cs_parser *parser)
431 {
432 	int r;
433 
434 	if (parser->chunk_ib_idx == -1)
435 		return 0;
436 
437 	if (parser->cs_flags & RADEON_CS_USE_VM)
438 		return 0;
439 
440 	r = radeon_cs_parse(rdev, parser->ring, parser);
441 	if (r || parser->parser_error) {
442 		DRM_ERROR("Invalid command stream !\n");
443 		return r;
444 	}
445 
446 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
447 		radeon_uvd_note_usage(rdev);
448 	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
449 		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
450 		radeon_vce_note_usage(rdev);
451 
452 	radeon_cs_sync_rings(parser);
453 	r = radeon_ib_schedule(rdev, &parser->ib, NULL);
454 	if (r) {
455 		DRM_ERROR("Failed to schedule IB !\n");
456 	}
457 	return r;
458 }
459 
460 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
461 				   struct radeon_vm *vm)
462 {
463 	struct radeon_device *rdev = p->rdev;
464 	struct radeon_bo_va *bo_va;
465 	int i, r;
466 
467 	r = radeon_vm_update_page_directory(rdev, vm);
468 	if (r)
469 		return r;
470 
471 	r = radeon_vm_clear_freed(rdev, vm);
472 	if (r)
473 		return r;
474 
475 	if (vm->ib_bo_va == NULL) {
476 		DRM_ERROR("Tmp BO not in VM!\n");
477 		return -EINVAL;
478 	}
479 
480 	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
481 				&rdev->ring_tmp_bo.bo->tbo.mem);
482 	if (r)
483 		return r;
484 
485 	for (i = 0; i < p->nrelocs; i++) {
486 		struct radeon_bo *bo;
487 
488 		/* ignore duplicates */
489 		if (p->relocs_ptr[i] != &p->relocs[i])
490 			continue;
491 
492 		bo = p->relocs[i].robj;
493 		bo_va = radeon_vm_bo_find(vm, bo);
494 		if (bo_va == NULL) {
495 			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
496 			return -EINVAL;
497 		}
498 
499 		r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
500 		if (r)
501 			return r;
502 	}
503 	return 0;
504 }
505 
506 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
507 				 struct radeon_cs_parser *parser)
508 {
509 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
510 	struct radeon_vm *vm = &fpriv->vm;
511 	int r;
512 
513 	if (parser->chunk_ib_idx == -1)
514 		return 0;
515 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
516 		return 0;
517 
518 	if (parser->const_ib.length_dw) {
519 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
520 		if (r) {
521 			return r;
522 		}
523 	}
524 
525 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
526 	if (r) {
527 		return r;
528 	}
529 
530 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
531 		radeon_uvd_note_usage(rdev);
532 
533 	mutex_lock(&vm->mutex);
534 	r = radeon_bo_vm_update_pte(parser, vm);
535 	if (r) {
536 		goto out;
537 	}
538 	radeon_cs_sync_rings(parser);
539 	radeon_semaphore_sync_to(parser->ib.semaphore, vm->fence);
540 
541 	if ((rdev->family >= CHIP_TAHITI) &&
542 	    (parser->chunk_const_ib_idx != -1)) {
543 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib);
544 	} else {
545 		r = radeon_ib_schedule(rdev, &parser->ib, NULL);
546 	}
547 
548 out:
549 	mutex_unlock(&vm->mutex);
550 	return r;
551 }
552 
553 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
554 {
555 	if (r == -EDEADLK) {
556 		r = radeon_gpu_reset(rdev);
557 		if (!r)
558 			r = -EAGAIN;
559 	}
560 	return r;
561 }
562 
563 static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
564 {
565 	struct radeon_cs_chunk *ib_chunk;
566 	struct radeon_vm *vm = NULL;
567 	int r;
568 
569 	if (parser->chunk_ib_idx == -1)
570 		return 0;
571 
572 	if (parser->cs_flags & RADEON_CS_USE_VM) {
573 		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
574 		vm = &fpriv->vm;
575 
576 		if ((rdev->family >= CHIP_TAHITI) &&
577 		    (parser->chunk_const_ib_idx != -1)) {
578 			ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
579 			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
580 				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
581 				return -EINVAL;
582 			}
583 			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
584 					   vm, ib_chunk->length_dw * 4);
585 			if (r) {
586 				DRM_ERROR("Failed to get const ib !\n");
587 				return r;
588 			}
589 			parser->const_ib.is_const_ib = true;
590 			parser->const_ib.length_dw = ib_chunk->length_dw;
591 			if (copy_from_user(parser->const_ib.ptr,
592 					       ib_chunk->user_ptr,
593 					       ib_chunk->length_dw * 4))
594 				return -EFAULT;
595 		}
596 
597 		ib_chunk = &parser->chunks[parser->chunk_ib_idx];
598 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
599 			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
600 			return -EINVAL;
601 		}
602 	}
603 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
604 
605 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
606 			   vm, ib_chunk->length_dw * 4);
607 	if (r) {
608 		DRM_ERROR("Failed to get ib !\n");
609 		return r;
610 	}
611 	parser->ib.length_dw = ib_chunk->length_dw;
612 	if (ib_chunk->kdata)
613 		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
614 	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
615 		return -EFAULT;
616 	return 0;
617 }
618 
619 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
620 {
621 	struct radeon_device *rdev = dev->dev_private;
622 	struct radeon_cs_parser parser;
623 	int r;
624 
625 	down_read(&rdev->exclusive_lock);
626 	if (!rdev->accel_working) {
627 		up_read(&rdev->exclusive_lock);
628 		return -EBUSY;
629 	}
630 	/* initialize parser */
631 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
632 	parser.filp = filp;
633 	parser.rdev = rdev;
634 	parser.dev = rdev->dev;
635 	parser.family = rdev->family;
636 	r = radeon_cs_parser_init(&parser, data);
637 	if (r) {
638 		DRM_ERROR("Failed to initialize parser !\n");
639 		radeon_cs_parser_fini(&parser, r, false);
640 		up_read(&rdev->exclusive_lock);
641 		r = radeon_cs_handle_lockup(rdev, r);
642 		return r;
643 	}
644 
645 	r = radeon_cs_ib_fill(rdev, &parser);
646 	if (!r) {
647 		r = radeon_cs_parser_relocs(&parser);
648 		if (r && r != -ERESTARTSYS)
649 			DRM_ERROR("Failed to parse relocation %d!\n", r);
650 	}
651 
652 	if (r) {
653 		radeon_cs_parser_fini(&parser, r, false);
654 		up_read(&rdev->exclusive_lock);
655 		r = radeon_cs_handle_lockup(rdev, r);
656 		return r;
657 	}
658 
659 	trace_radeon_cs(&parser);
660 
661 	r = radeon_cs_ib_chunk(rdev, &parser);
662 	if (r) {
663 		goto out;
664 	}
665 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
666 	if (r) {
667 		goto out;
668 	}
669 out:
670 	radeon_cs_parser_fini(&parser, r, true);
671 	up_read(&rdev->exclusive_lock);
672 	r = radeon_cs_handle_lockup(rdev, r);
673 	return r;
674 }
675 
676 /**
677  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
678  * @parser:	parser structure holding parsing context.
679  * @pkt:	where to store packet information
680  *
681  * Assume that chunk_ib_index is properly set. Will return -EINVAL
682  * if packet is bigger than remaining ib size. or if packets is unknown.
683  **/
684 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
685 			   struct radeon_cs_packet *pkt,
686 			   unsigned idx)
687 {
688 	struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
689 	struct radeon_device *rdev = p->rdev;
690 	uint32_t header;
691 
692 	if (idx >= ib_chunk->length_dw) {
693 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
694 			  idx, ib_chunk->length_dw);
695 		return -EINVAL;
696 	}
697 	header = radeon_get_ib_value(p, idx);
698 	pkt->idx = idx;
699 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
700 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
701 	pkt->one_reg_wr = 0;
702 	switch (pkt->type) {
703 	case RADEON_PACKET_TYPE0:
704 		if (rdev->family < CHIP_R600) {
705 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
706 			pkt->one_reg_wr =
707 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
708 		} else
709 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
710 		break;
711 	case RADEON_PACKET_TYPE3:
712 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
713 		break;
714 	case RADEON_PACKET_TYPE2:
715 		pkt->count = -1;
716 		break;
717 	default:
718 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
719 		return -EINVAL;
720 	}
721 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
722 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
723 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
724 		return -EINVAL;
725 	}
726 	return 0;
727 }
728 
729 /**
730  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
731  * @p:		structure holding the parser context.
732  *
733  * Check if the next packet is NOP relocation packet3.
734  **/
735 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
736 {
737 	struct radeon_cs_packet p3reloc;
738 	int r;
739 
740 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
741 	if (r)
742 		return false;
743 	if (p3reloc.type != RADEON_PACKET_TYPE3)
744 		return false;
745 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
746 		return false;
747 	return true;
748 }
749 
750 /**
751  * radeon_cs_dump_packet() - dump raw packet context
752  * @p:		structure holding the parser context.
753  * @pkt:	structure holding the packet.
754  *
755  * Used mostly for debugging and error reporting.
756  **/
757 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
758 			   struct radeon_cs_packet *pkt)
759 {
760 	volatile uint32_t *ib;
761 	unsigned i;
762 	unsigned idx;
763 
764 	ib = p->ib.ptr;
765 	idx = pkt->idx;
766 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
767 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
768 }
769 
770 /**
771  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
772  * @parser:		parser structure holding parsing context.
773  * @data:		pointer to relocation data
774  * @offset_start:	starting offset
775  * @offset_mask:	offset mask (to align start offset on)
776  * @reloc:		reloc informations
777  *
778  * Check if next packet is relocation packet3, do bo validation and compute
779  * GPU offset using the provided start.
780  **/
781 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
782 				struct radeon_cs_reloc **cs_reloc,
783 				int nomm)
784 {
785 	struct radeon_cs_chunk *relocs_chunk;
786 	struct radeon_cs_packet p3reloc;
787 	unsigned idx;
788 	int r;
789 
790 	if (p->chunk_relocs_idx == -1) {
791 		DRM_ERROR("No relocation chunk !\n");
792 		return -EINVAL;
793 	}
794 	*cs_reloc = NULL;
795 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
796 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
797 	if (r)
798 		return r;
799 	p->idx += p3reloc.count + 2;
800 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
801 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
802 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
803 			  p3reloc.idx);
804 		radeon_cs_dump_packet(p, &p3reloc);
805 		return -EINVAL;
806 	}
807 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
808 	if (idx >= relocs_chunk->length_dw) {
809 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
810 			  idx, relocs_chunk->length_dw);
811 		radeon_cs_dump_packet(p, &p3reloc);
812 		return -EINVAL;
813 	}
814 	/* FIXME: we assume reloc size is 4 dwords */
815 	if (nomm) {
816 		*cs_reloc = p->relocs;
817 		(*cs_reloc)->gpu_offset =
818 			(u64)relocs_chunk->kdata[idx + 3] << 32;
819 		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
820 	} else
821 		*cs_reloc = p->relocs_ptr[(idx / 4)];
822 	return 0;
823 }
824