xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision af958a38)
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 & RADEON_RELOC_PRIO_MASK) * 2
136 			   + !!r->write_domain;
137 
138 		/* the first reloc of an UVD job is the msg and that must be in
139 		   VRAM, also but everything into VRAM on AGP cards to avoid
140 		   image corruptions */
141 		if (p->ring == R600_RING_TYPE_UVD_INDEX &&
142 		    (i == 0 || drm_pci_device_is_agp(p->rdev->ddev))) {
143 			/* TODO: is this still needed for NI+ ? */
144 			p->relocs[i].prefered_domains =
145 				RADEON_GEM_DOMAIN_VRAM;
146 
147 			p->relocs[i].allowed_domains =
148 				RADEON_GEM_DOMAIN_VRAM;
149 
150 			/* prioritize this over any other relocation */
151 			priority = RADEON_CS_MAX_PRIORITY;
152 		} else {
153 			uint32_t domain = r->write_domain ?
154 				r->write_domain : r->read_domains;
155 
156 			if (domain & RADEON_GEM_DOMAIN_CPU) {
157 				DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
158 					  "for command submission\n");
159 				return -EINVAL;
160 			}
161 
162 			p->relocs[i].prefered_domains = domain;
163 			if (domain == RADEON_GEM_DOMAIN_VRAM)
164 				domain |= RADEON_GEM_DOMAIN_GTT;
165 			p->relocs[i].allowed_domains = domain;
166 		}
167 
168 		p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
169 		p->relocs[i].handle = r->handle;
170 
171 		radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head,
172 				      priority);
173 	}
174 
175 	radeon_cs_buckets_get_list(&buckets, &p->validated);
176 
177 	if (p->cs_flags & RADEON_CS_USE_VM)
178 		p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm,
179 					      &p->validated);
180 
181 	return radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
182 }
183 
184 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
185 {
186 	p->priority = priority;
187 
188 	switch (ring) {
189 	default:
190 		DRM_ERROR("unknown ring id: %d\n", ring);
191 		return -EINVAL;
192 	case RADEON_CS_RING_GFX:
193 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
194 		break;
195 	case RADEON_CS_RING_COMPUTE:
196 		if (p->rdev->family >= CHIP_TAHITI) {
197 			if (p->priority > 0)
198 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
199 			else
200 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
201 		} else
202 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
203 		break;
204 	case RADEON_CS_RING_DMA:
205 		if (p->rdev->family >= CHIP_CAYMAN) {
206 			if (p->priority > 0)
207 				p->ring = R600_RING_TYPE_DMA_INDEX;
208 			else
209 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
210 		} else if (p->rdev->family >= CHIP_RV770) {
211 			p->ring = R600_RING_TYPE_DMA_INDEX;
212 		} else {
213 			return -EINVAL;
214 		}
215 		break;
216 	case RADEON_CS_RING_UVD:
217 		p->ring = R600_RING_TYPE_UVD_INDEX;
218 		break;
219 	case RADEON_CS_RING_VCE:
220 		/* TODO: only use the low priority ring for now */
221 		p->ring = TN_RING_TYPE_VCE1_INDEX;
222 		break;
223 	}
224 	return 0;
225 }
226 
227 static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
228 {
229 	int i;
230 
231 	for (i = 0; i < p->nrelocs; i++) {
232 		if (!p->relocs[i].robj)
233 			continue;
234 
235 		radeon_semaphore_sync_to(p->ib.semaphore,
236 					 p->relocs[i].robj->tbo.sync_obj);
237 	}
238 }
239 
240 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
241 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
242 {
243 	struct drm_radeon_cs *cs = data;
244 	uint64_t *chunk_array_ptr;
245 	unsigned size, i;
246 	u32 ring = RADEON_CS_RING_GFX;
247 	s32 priority = 0;
248 
249 	if (!cs->num_chunks) {
250 		return 0;
251 	}
252 	/* get chunks */
253 	INIT_LIST_HEAD(&p->validated);
254 	p->idx = 0;
255 	p->ib.sa_bo = NULL;
256 	p->ib.semaphore = NULL;
257 	p->const_ib.sa_bo = NULL;
258 	p->const_ib.semaphore = NULL;
259 	p->chunk_ib_idx = -1;
260 	p->chunk_relocs_idx = -1;
261 	p->chunk_flags_idx = -1;
262 	p->chunk_const_ib_idx = -1;
263 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
264 	if (p->chunks_array == NULL) {
265 		return -ENOMEM;
266 	}
267 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
268 	if (copy_from_user(p->chunks_array, chunk_array_ptr,
269 			       sizeof(uint64_t)*cs->num_chunks)) {
270 		return -EFAULT;
271 	}
272 	p->cs_flags = 0;
273 	p->nchunks = cs->num_chunks;
274 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
275 	if (p->chunks == NULL) {
276 		return -ENOMEM;
277 	}
278 	for (i = 0; i < p->nchunks; i++) {
279 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
280 		struct drm_radeon_cs_chunk user_chunk;
281 		uint32_t __user *cdata;
282 
283 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
284 		if (copy_from_user(&user_chunk, chunk_ptr,
285 				       sizeof(struct drm_radeon_cs_chunk))) {
286 			return -EFAULT;
287 		}
288 		p->chunks[i].length_dw = user_chunk.length_dw;
289 		p->chunks[i].chunk_id = user_chunk.chunk_id;
290 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
291 			p->chunk_relocs_idx = i;
292 		}
293 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
294 			p->chunk_ib_idx = i;
295 			/* zero length IB isn't useful */
296 			if (p->chunks[i].length_dw == 0)
297 				return -EINVAL;
298 		}
299 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
300 			p->chunk_const_ib_idx = i;
301 			/* zero length CONST IB isn't useful */
302 			if (p->chunks[i].length_dw == 0)
303 				return -EINVAL;
304 		}
305 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
306 			p->chunk_flags_idx = i;
307 			/* zero length flags aren't useful */
308 			if (p->chunks[i].length_dw == 0)
309 				return -EINVAL;
310 		}
311 
312 		size = p->chunks[i].length_dw;
313 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
314 		p->chunks[i].user_ptr = cdata;
315 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB)
316 			continue;
317 
318 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
319 			if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
320 				continue;
321 		}
322 
323 		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
324 		size *= sizeof(uint32_t);
325 		if (p->chunks[i].kdata == NULL) {
326 			return -ENOMEM;
327 		}
328 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
329 			return -EFAULT;
330 		}
331 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
332 			p->cs_flags = p->chunks[i].kdata[0];
333 			if (p->chunks[i].length_dw > 1)
334 				ring = p->chunks[i].kdata[1];
335 			if (p->chunks[i].length_dw > 2)
336 				priority = (s32)p->chunks[i].kdata[2];
337 		}
338 	}
339 
340 	/* these are KMS only */
341 	if (p->rdev) {
342 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
343 		    !p->rdev->vm_manager.enabled) {
344 			DRM_ERROR("VM not active on asic!\n");
345 			return -EINVAL;
346 		}
347 
348 		if (radeon_cs_get_ring(p, ring, priority))
349 			return -EINVAL;
350 
351 		/* we only support VM on some SI+ rings */
352 		if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
353 			if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
354 				DRM_ERROR("Ring %d requires VM!\n", p->ring);
355 				return -EINVAL;
356 			}
357 		} else {
358 			if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
359 				DRM_ERROR("VM not supported on ring %d!\n",
360 					  p->ring);
361 				return -EINVAL;
362 			}
363 		}
364 	}
365 
366 	return 0;
367 }
368 
369 static int cmp_size_smaller_first(void *priv, struct list_head *a,
370 				  struct list_head *b)
371 {
372 	struct radeon_cs_reloc *la = list_entry(a, struct radeon_cs_reloc, tv.head);
373 	struct radeon_cs_reloc *lb = list_entry(b, struct radeon_cs_reloc, tv.head);
374 
375 	/* Sort A before B if A is smaller. */
376 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
377 }
378 
379 /**
380  * cs_parser_fini() - clean parser states
381  * @parser:	parser structure holding parsing context.
382  * @error:	error number
383  *
384  * If error is set than unvalidate buffer, otherwise just free memory
385  * used by parsing context.
386  **/
387 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
388 {
389 	unsigned i;
390 
391 	if (!error) {
392 		/* Sort the buffer list from the smallest to largest buffer,
393 		 * which affects the order of buffers in the LRU list.
394 		 * This assures that the smallest buffers are added first
395 		 * to the LRU list, so they are likely to be later evicted
396 		 * first, instead of large buffers whose eviction is more
397 		 * expensive.
398 		 *
399 		 * This slightly lowers the number of bytes moved by TTM
400 		 * per frame under memory pressure.
401 		 */
402 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
403 
404 		ttm_eu_fence_buffer_objects(&parser->ticket,
405 					    &parser->validated,
406 					    parser->ib.fence);
407 	} else if (backoff) {
408 		ttm_eu_backoff_reservation(&parser->ticket,
409 					   &parser->validated);
410 	}
411 
412 	if (parser->relocs != NULL) {
413 		for (i = 0; i < parser->nrelocs; i++) {
414 			if (parser->relocs[i].gobj)
415 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
416 		}
417 	}
418 	kfree(parser->track);
419 	kfree(parser->relocs);
420 	kfree(parser->relocs_ptr);
421 	kfree(parser->vm_bos);
422 	for (i = 0; i < parser->nchunks; i++)
423 		drm_free_large(parser->chunks[i].kdata);
424 	kfree(parser->chunks);
425 	kfree(parser->chunks_array);
426 	radeon_ib_free(parser->rdev, &parser->ib);
427 	radeon_ib_free(parser->rdev, &parser->const_ib);
428 }
429 
430 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
431 			      struct radeon_cs_parser *parser)
432 {
433 	int r;
434 
435 	if (parser->chunk_ib_idx == -1)
436 		return 0;
437 
438 	if (parser->cs_flags & RADEON_CS_USE_VM)
439 		return 0;
440 
441 	r = radeon_cs_parse(rdev, parser->ring, parser);
442 	if (r || parser->parser_error) {
443 		DRM_ERROR("Invalid command stream !\n");
444 		return r;
445 	}
446 
447 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
448 		radeon_uvd_note_usage(rdev);
449 	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
450 		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
451 		radeon_vce_note_usage(rdev);
452 
453 	radeon_cs_sync_rings(parser);
454 	r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
455 	if (r) {
456 		DRM_ERROR("Failed to schedule IB !\n");
457 	}
458 	return r;
459 }
460 
461 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
462 				   struct radeon_vm *vm)
463 {
464 	struct radeon_device *rdev = p->rdev;
465 	struct radeon_bo_va *bo_va;
466 	int i, r;
467 
468 	r = radeon_vm_update_page_directory(rdev, vm);
469 	if (r)
470 		return r;
471 
472 	r = radeon_vm_clear_freed(rdev, vm);
473 	if (r)
474 		return r;
475 
476 	if (vm->ib_bo_va == NULL) {
477 		DRM_ERROR("Tmp BO not in VM!\n");
478 		return -EINVAL;
479 	}
480 
481 	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
482 				&rdev->ring_tmp_bo.bo->tbo.mem);
483 	if (r)
484 		return r;
485 
486 	for (i = 0; i < p->nrelocs; i++) {
487 		struct radeon_bo *bo;
488 
489 		/* ignore duplicates */
490 		if (p->relocs_ptr[i] != &p->relocs[i])
491 			continue;
492 
493 		bo = p->relocs[i].robj;
494 		bo_va = radeon_vm_bo_find(vm, bo);
495 		if (bo_va == NULL) {
496 			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
497 			return -EINVAL;
498 		}
499 
500 		r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
501 		if (r)
502 			return r;
503 	}
504 
505 	return radeon_vm_clear_invalids(rdev, vm);
506 }
507 
508 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
509 				 struct radeon_cs_parser *parser)
510 {
511 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
512 	struct radeon_vm *vm = &fpriv->vm;
513 	int r;
514 
515 	if (parser->chunk_ib_idx == -1)
516 		return 0;
517 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
518 		return 0;
519 
520 	if (parser->const_ib.length_dw) {
521 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
522 		if (r) {
523 			return r;
524 		}
525 	}
526 
527 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
528 	if (r) {
529 		return r;
530 	}
531 
532 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
533 		radeon_uvd_note_usage(rdev);
534 
535 	mutex_lock(&vm->mutex);
536 	r = radeon_bo_vm_update_pte(parser, vm);
537 	if (r) {
538 		goto out;
539 	}
540 	radeon_cs_sync_rings(parser);
541 	radeon_semaphore_sync_to(parser->ib.semaphore, vm->fence);
542 
543 	if ((rdev->family >= CHIP_TAHITI) &&
544 	    (parser->chunk_const_ib_idx != -1)) {
545 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
546 	} else {
547 		r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
548 	}
549 
550 out:
551 	mutex_unlock(&vm->mutex);
552 	return r;
553 }
554 
555 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
556 {
557 	if (r == -EDEADLK) {
558 		r = radeon_gpu_reset(rdev);
559 		if (!r)
560 			r = -EAGAIN;
561 	}
562 	return r;
563 }
564 
565 static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
566 {
567 	struct radeon_cs_chunk *ib_chunk;
568 	struct radeon_vm *vm = NULL;
569 	int r;
570 
571 	if (parser->chunk_ib_idx == -1)
572 		return 0;
573 
574 	if (parser->cs_flags & RADEON_CS_USE_VM) {
575 		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
576 		vm = &fpriv->vm;
577 
578 		if ((rdev->family >= CHIP_TAHITI) &&
579 		    (parser->chunk_const_ib_idx != -1)) {
580 			ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
581 			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
582 				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
583 				return -EINVAL;
584 			}
585 			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
586 					   vm, ib_chunk->length_dw * 4);
587 			if (r) {
588 				DRM_ERROR("Failed to get const ib !\n");
589 				return r;
590 			}
591 			parser->const_ib.is_const_ib = true;
592 			parser->const_ib.length_dw = ib_chunk->length_dw;
593 			if (copy_from_user(parser->const_ib.ptr,
594 					       ib_chunk->user_ptr,
595 					       ib_chunk->length_dw * 4))
596 				return -EFAULT;
597 		}
598 
599 		ib_chunk = &parser->chunks[parser->chunk_ib_idx];
600 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
601 			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
602 			return -EINVAL;
603 		}
604 	}
605 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
606 
607 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
608 			   vm, ib_chunk->length_dw * 4);
609 	if (r) {
610 		DRM_ERROR("Failed to get ib !\n");
611 		return r;
612 	}
613 	parser->ib.length_dw = ib_chunk->length_dw;
614 	if (ib_chunk->kdata)
615 		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
616 	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
617 		return -EFAULT;
618 	return 0;
619 }
620 
621 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
622 {
623 	struct radeon_device *rdev = dev->dev_private;
624 	struct radeon_cs_parser parser;
625 	int r;
626 
627 	down_read(&rdev->exclusive_lock);
628 	if (!rdev->accel_working) {
629 		up_read(&rdev->exclusive_lock);
630 		return -EBUSY;
631 	}
632 	/* initialize parser */
633 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
634 	parser.filp = filp;
635 	parser.rdev = rdev;
636 	parser.dev = rdev->dev;
637 	parser.family = rdev->family;
638 	r = radeon_cs_parser_init(&parser, data);
639 	if (r) {
640 		DRM_ERROR("Failed to initialize parser !\n");
641 		radeon_cs_parser_fini(&parser, r, false);
642 		up_read(&rdev->exclusive_lock);
643 		r = radeon_cs_handle_lockup(rdev, r);
644 		return r;
645 	}
646 
647 	r = radeon_cs_ib_fill(rdev, &parser);
648 	if (!r) {
649 		r = radeon_cs_parser_relocs(&parser);
650 		if (r && r != -ERESTARTSYS)
651 			DRM_ERROR("Failed to parse relocation %d!\n", r);
652 	}
653 
654 	if (r) {
655 		radeon_cs_parser_fini(&parser, r, false);
656 		up_read(&rdev->exclusive_lock);
657 		r = radeon_cs_handle_lockup(rdev, r);
658 		return r;
659 	}
660 
661 	trace_radeon_cs(&parser);
662 
663 	r = radeon_cs_ib_chunk(rdev, &parser);
664 	if (r) {
665 		goto out;
666 	}
667 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
668 	if (r) {
669 		goto out;
670 	}
671 out:
672 	radeon_cs_parser_fini(&parser, r, true);
673 	up_read(&rdev->exclusive_lock);
674 	r = radeon_cs_handle_lockup(rdev, r);
675 	return r;
676 }
677 
678 /**
679  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
680  * @parser:	parser structure holding parsing context.
681  * @pkt:	where to store packet information
682  *
683  * Assume that chunk_ib_index is properly set. Will return -EINVAL
684  * if packet is bigger than remaining ib size. or if packets is unknown.
685  **/
686 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
687 			   struct radeon_cs_packet *pkt,
688 			   unsigned idx)
689 {
690 	struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
691 	struct radeon_device *rdev = p->rdev;
692 	uint32_t header;
693 
694 	if (idx >= ib_chunk->length_dw) {
695 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
696 			  idx, ib_chunk->length_dw);
697 		return -EINVAL;
698 	}
699 	header = radeon_get_ib_value(p, idx);
700 	pkt->idx = idx;
701 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
702 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
703 	pkt->one_reg_wr = 0;
704 	switch (pkt->type) {
705 	case RADEON_PACKET_TYPE0:
706 		if (rdev->family < CHIP_R600) {
707 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
708 			pkt->one_reg_wr =
709 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
710 		} else
711 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
712 		break;
713 	case RADEON_PACKET_TYPE3:
714 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
715 		break;
716 	case RADEON_PACKET_TYPE2:
717 		pkt->count = -1;
718 		break;
719 	default:
720 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
721 		return -EINVAL;
722 	}
723 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
724 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
725 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
726 		return -EINVAL;
727 	}
728 	return 0;
729 }
730 
731 /**
732  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
733  * @p:		structure holding the parser context.
734  *
735  * Check if the next packet is NOP relocation packet3.
736  **/
737 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
738 {
739 	struct radeon_cs_packet p3reloc;
740 	int r;
741 
742 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
743 	if (r)
744 		return false;
745 	if (p3reloc.type != RADEON_PACKET_TYPE3)
746 		return false;
747 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
748 		return false;
749 	return true;
750 }
751 
752 /**
753  * radeon_cs_dump_packet() - dump raw packet context
754  * @p:		structure holding the parser context.
755  * @pkt:	structure holding the packet.
756  *
757  * Used mostly for debugging and error reporting.
758  **/
759 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
760 			   struct radeon_cs_packet *pkt)
761 {
762 	volatile uint32_t *ib;
763 	unsigned i;
764 	unsigned idx;
765 
766 	ib = p->ib.ptr;
767 	idx = pkt->idx;
768 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
769 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
770 }
771 
772 /**
773  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
774  * @parser:		parser structure holding parsing context.
775  * @data:		pointer to relocation data
776  * @offset_start:	starting offset
777  * @offset_mask:	offset mask (to align start offset on)
778  * @reloc:		reloc informations
779  *
780  * Check if next packet is relocation packet3, do bo validation and compute
781  * GPU offset using the provided start.
782  **/
783 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
784 				struct radeon_cs_reloc **cs_reloc,
785 				int nomm)
786 {
787 	struct radeon_cs_chunk *relocs_chunk;
788 	struct radeon_cs_packet p3reloc;
789 	unsigned idx;
790 	int r;
791 
792 	if (p->chunk_relocs_idx == -1) {
793 		DRM_ERROR("No relocation chunk !\n");
794 		return -EINVAL;
795 	}
796 	*cs_reloc = NULL;
797 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
798 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
799 	if (r)
800 		return r;
801 	p->idx += p3reloc.count + 2;
802 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
803 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
804 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
805 			  p3reloc.idx);
806 		radeon_cs_dump_packet(p, &p3reloc);
807 		return -EINVAL;
808 	}
809 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
810 	if (idx >= relocs_chunk->length_dw) {
811 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
812 			  idx, relocs_chunk->length_dw);
813 		radeon_cs_dump_packet(p, &p3reloc);
814 		return -EINVAL;
815 	}
816 	/* FIXME: we assume reloc size is 4 dwords */
817 	if (nomm) {
818 		*cs_reloc = p->relocs;
819 		(*cs_reloc)->gpu_offset =
820 			(u64)relocs_chunk->kdata[idx + 3] << 32;
821 		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
822 	} else
823 		*cs_reloc = p->relocs_ptr[(idx / 4)];
824 	return 0;
825 }
826