xref: /linux/drivers/gpu/drm/radeon/radeon_cs.c (revision e6b5be2be4e30037eb551e0ed09dd97bd00d85d3)
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, need_mmap_lock = false;
82 	int r;
83 
84 	if (p->chunk_relocs_idx == -1) {
85 		return 0;
86 	}
87 	chunk = &p->chunks[p->chunk_relocs_idx];
88 	p->dma_reloc_idx = 0;
89 	/* FIXME: we assume that each relocs use 4 dwords */
90 	p->nrelocs = chunk->length_dw / 4;
91 	p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
92 	if (p->relocs_ptr == NULL) {
93 		return -ENOMEM;
94 	}
95 	p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
96 	if (p->relocs == NULL) {
97 		return -ENOMEM;
98 	}
99 
100 	radeon_cs_buckets_init(&buckets);
101 
102 	for (i = 0; i < p->nrelocs; i++) {
103 		struct drm_radeon_cs_reloc *r;
104 		unsigned priority;
105 
106 		duplicate = false;
107 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
108 		for (j = 0; j < i; j++) {
109 			if (r->handle == p->relocs[j].handle) {
110 				p->relocs_ptr[i] = &p->relocs[j];
111 				duplicate = true;
112 				break;
113 			}
114 		}
115 		if (duplicate) {
116 			p->relocs[i].handle = 0;
117 			continue;
118 		}
119 
120 		p->relocs[i].gobj = drm_gem_object_lookup(ddev, p->filp,
121 							  r->handle);
122 		if (p->relocs[i].gobj == NULL) {
123 			DRM_ERROR("gem object lookup failed 0x%x\n",
124 				  r->handle);
125 			return -ENOENT;
126 		}
127 		p->relocs_ptr[i] = &p->relocs[i];
128 		p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
129 
130 		/* The userspace buffer priorities are from 0 to 15. A higher
131 		 * number means the buffer is more important.
132 		 * Also, the buffers used for write have a higher priority than
133 		 * the buffers used for read only, which doubles the range
134 		 * to 0 to 31. 32 is reserved for the kernel driver.
135 		 */
136 		priority = (r->flags & RADEON_RELOC_PRIO_MASK) * 2
137 			   + !!r->write_domain;
138 
139 		/* the first reloc of an UVD job is the msg and that must be in
140 		   VRAM, also but everything into VRAM on AGP cards and older
141 		   IGP chips to avoid image corruptions */
142 		if (p->ring == R600_RING_TYPE_UVD_INDEX &&
143 		    (i == 0 || drm_pci_device_is_agp(p->rdev->ddev) ||
144 		     p->rdev->family == CHIP_RS780 ||
145 		     p->rdev->family == CHIP_RS880)) {
146 
147 			/* TODO: is this still needed for NI+ ? */
148 			p->relocs[i].prefered_domains =
149 				RADEON_GEM_DOMAIN_VRAM;
150 
151 			p->relocs[i].allowed_domains =
152 				RADEON_GEM_DOMAIN_VRAM;
153 
154 			/* prioritize this over any other relocation */
155 			priority = RADEON_CS_MAX_PRIORITY;
156 		} else {
157 			uint32_t domain = r->write_domain ?
158 				r->write_domain : r->read_domains;
159 
160 			if (domain & RADEON_GEM_DOMAIN_CPU) {
161 				DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
162 					  "for command submission\n");
163 				return -EINVAL;
164 			}
165 
166 			p->relocs[i].prefered_domains = domain;
167 			if (domain == RADEON_GEM_DOMAIN_VRAM)
168 				domain |= RADEON_GEM_DOMAIN_GTT;
169 			p->relocs[i].allowed_domains = domain;
170 		}
171 
172 		if (radeon_ttm_tt_has_userptr(p->relocs[i].robj->tbo.ttm)) {
173 			uint32_t domain = p->relocs[i].prefered_domains;
174 			if (!(domain & RADEON_GEM_DOMAIN_GTT)) {
175 				DRM_ERROR("Only RADEON_GEM_DOMAIN_GTT is "
176 					  "allowed for userptr BOs\n");
177 				return -EINVAL;
178 			}
179 			need_mmap_lock = true;
180 			domain = RADEON_GEM_DOMAIN_GTT;
181 			p->relocs[i].prefered_domains = domain;
182 			p->relocs[i].allowed_domains = domain;
183 		}
184 
185 		p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
186 		p->relocs[i].tv.shared = !r->write_domain;
187 		p->relocs[i].handle = r->handle;
188 
189 		radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head,
190 				      priority);
191 	}
192 
193 	radeon_cs_buckets_get_list(&buckets, &p->validated);
194 
195 	if (p->cs_flags & RADEON_CS_USE_VM)
196 		p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm,
197 					      &p->validated);
198 	if (need_mmap_lock)
199 		down_read(&current->mm->mmap_sem);
200 
201 	r = radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
202 
203 	if (need_mmap_lock)
204 		up_read(&current->mm->mmap_sem);
205 
206 	return r;
207 }
208 
209 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
210 {
211 	p->priority = priority;
212 
213 	switch (ring) {
214 	default:
215 		DRM_ERROR("unknown ring id: %d\n", ring);
216 		return -EINVAL;
217 	case RADEON_CS_RING_GFX:
218 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
219 		break;
220 	case RADEON_CS_RING_COMPUTE:
221 		if (p->rdev->family >= CHIP_TAHITI) {
222 			if (p->priority > 0)
223 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
224 			else
225 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
226 		} else
227 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
228 		break;
229 	case RADEON_CS_RING_DMA:
230 		if (p->rdev->family >= CHIP_CAYMAN) {
231 			if (p->priority > 0)
232 				p->ring = R600_RING_TYPE_DMA_INDEX;
233 			else
234 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
235 		} else if (p->rdev->family >= CHIP_RV770) {
236 			p->ring = R600_RING_TYPE_DMA_INDEX;
237 		} else {
238 			return -EINVAL;
239 		}
240 		break;
241 	case RADEON_CS_RING_UVD:
242 		p->ring = R600_RING_TYPE_UVD_INDEX;
243 		break;
244 	case RADEON_CS_RING_VCE:
245 		/* TODO: only use the low priority ring for now */
246 		p->ring = TN_RING_TYPE_VCE1_INDEX;
247 		break;
248 	}
249 	return 0;
250 }
251 
252 static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
253 {
254 	struct radeon_cs_reloc *reloc;
255 	int r;
256 
257 	list_for_each_entry(reloc, &p->validated, tv.head) {
258 		struct reservation_object *resv;
259 
260 		resv = reloc->robj->tbo.resv;
261 		r = radeon_semaphore_sync_resv(p->rdev, p->ib.semaphore, resv,
262 					       reloc->tv.shared);
263 		if (r)
264 			return r;
265 	}
266 	return 0;
267 }
268 
269 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
270 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
271 {
272 	struct drm_radeon_cs *cs = data;
273 	uint64_t *chunk_array_ptr;
274 	unsigned size, i;
275 	u32 ring = RADEON_CS_RING_GFX;
276 	s32 priority = 0;
277 
278 	if (!cs->num_chunks) {
279 		return 0;
280 	}
281 	/* get chunks */
282 	INIT_LIST_HEAD(&p->validated);
283 	p->idx = 0;
284 	p->ib.sa_bo = NULL;
285 	p->ib.semaphore = NULL;
286 	p->const_ib.sa_bo = NULL;
287 	p->const_ib.semaphore = NULL;
288 	p->chunk_ib_idx = -1;
289 	p->chunk_relocs_idx = -1;
290 	p->chunk_flags_idx = -1;
291 	p->chunk_const_ib_idx = -1;
292 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
293 	if (p->chunks_array == NULL) {
294 		return -ENOMEM;
295 	}
296 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
297 	if (copy_from_user(p->chunks_array, chunk_array_ptr,
298 			       sizeof(uint64_t)*cs->num_chunks)) {
299 		return -EFAULT;
300 	}
301 	p->cs_flags = 0;
302 	p->nchunks = cs->num_chunks;
303 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
304 	if (p->chunks == NULL) {
305 		return -ENOMEM;
306 	}
307 	for (i = 0; i < p->nchunks; i++) {
308 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
309 		struct drm_radeon_cs_chunk user_chunk;
310 		uint32_t __user *cdata;
311 
312 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
313 		if (copy_from_user(&user_chunk, chunk_ptr,
314 				       sizeof(struct drm_radeon_cs_chunk))) {
315 			return -EFAULT;
316 		}
317 		p->chunks[i].length_dw = user_chunk.length_dw;
318 		p->chunks[i].chunk_id = user_chunk.chunk_id;
319 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
320 			p->chunk_relocs_idx = i;
321 		}
322 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
323 			p->chunk_ib_idx = i;
324 			/* zero length IB isn't useful */
325 			if (p->chunks[i].length_dw == 0)
326 				return -EINVAL;
327 		}
328 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
329 			p->chunk_const_ib_idx = i;
330 			/* zero length CONST IB isn't useful */
331 			if (p->chunks[i].length_dw == 0)
332 				return -EINVAL;
333 		}
334 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
335 			p->chunk_flags_idx = i;
336 			/* zero length flags aren't useful */
337 			if (p->chunks[i].length_dw == 0)
338 				return -EINVAL;
339 		}
340 
341 		size = p->chunks[i].length_dw;
342 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
343 		p->chunks[i].user_ptr = cdata;
344 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB)
345 			continue;
346 
347 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
348 			if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
349 				continue;
350 		}
351 
352 		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
353 		size *= sizeof(uint32_t);
354 		if (p->chunks[i].kdata == NULL) {
355 			return -ENOMEM;
356 		}
357 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
358 			return -EFAULT;
359 		}
360 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
361 			p->cs_flags = p->chunks[i].kdata[0];
362 			if (p->chunks[i].length_dw > 1)
363 				ring = p->chunks[i].kdata[1];
364 			if (p->chunks[i].length_dw > 2)
365 				priority = (s32)p->chunks[i].kdata[2];
366 		}
367 	}
368 
369 	/* these are KMS only */
370 	if (p->rdev) {
371 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
372 		    !p->rdev->vm_manager.enabled) {
373 			DRM_ERROR("VM not active on asic!\n");
374 			return -EINVAL;
375 		}
376 
377 		if (radeon_cs_get_ring(p, ring, priority))
378 			return -EINVAL;
379 
380 		/* we only support VM on some SI+ rings */
381 		if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
382 			if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
383 				DRM_ERROR("Ring %d requires VM!\n", p->ring);
384 				return -EINVAL;
385 			}
386 		} else {
387 			if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
388 				DRM_ERROR("VM not supported on ring %d!\n",
389 					  p->ring);
390 				return -EINVAL;
391 			}
392 		}
393 	}
394 
395 	return 0;
396 }
397 
398 static int cmp_size_smaller_first(void *priv, struct list_head *a,
399 				  struct list_head *b)
400 {
401 	struct radeon_cs_reloc *la = list_entry(a, struct radeon_cs_reloc, tv.head);
402 	struct radeon_cs_reloc *lb = list_entry(b, struct radeon_cs_reloc, tv.head);
403 
404 	/* Sort A before B if A is smaller. */
405 	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
406 }
407 
408 /**
409  * cs_parser_fini() - clean parser states
410  * @parser:	parser structure holding parsing context.
411  * @error:	error number
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 			if (parser->relocs[i].gobj)
444 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
445 		}
446 	}
447 	kfree(parser->track);
448 	kfree(parser->relocs);
449 	kfree(parser->relocs_ptr);
450 	drm_free_large(parser->vm_bos);
451 	for (i = 0; i < parser->nchunks; i++)
452 		drm_free_large(parser->chunks[i].kdata);
453 	kfree(parser->chunks);
454 	kfree(parser->chunks_array);
455 	radeon_ib_free(parser->rdev, &parser->ib);
456 	radeon_ib_free(parser->rdev, &parser->const_ib);
457 }
458 
459 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
460 			      struct radeon_cs_parser *parser)
461 {
462 	int r;
463 
464 	if (parser->chunk_ib_idx == -1)
465 		return 0;
466 
467 	if (parser->cs_flags & RADEON_CS_USE_VM)
468 		return 0;
469 
470 	r = radeon_cs_parse(rdev, parser->ring, parser);
471 	if (r || parser->parser_error) {
472 		DRM_ERROR("Invalid command stream !\n");
473 		return r;
474 	}
475 
476 	r = radeon_cs_sync_rings(parser);
477 	if (r) {
478 		if (r != -ERESTARTSYS)
479 			DRM_ERROR("Failed to sync rings: %i\n", r);
480 		return r;
481 	}
482 
483 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
484 		radeon_uvd_note_usage(rdev);
485 	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
486 		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
487 		radeon_vce_note_usage(rdev);
488 
489 	r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
490 	if (r) {
491 		DRM_ERROR("Failed to schedule IB !\n");
492 	}
493 	return r;
494 }
495 
496 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
497 				   struct radeon_vm *vm)
498 {
499 	struct radeon_device *rdev = p->rdev;
500 	struct radeon_bo_va *bo_va;
501 	int i, r;
502 
503 	r = radeon_vm_update_page_directory(rdev, vm);
504 	if (r)
505 		return r;
506 
507 	r = radeon_vm_clear_freed(rdev, vm);
508 	if (r)
509 		return r;
510 
511 	if (vm->ib_bo_va == NULL) {
512 		DRM_ERROR("Tmp BO not in VM!\n");
513 		return -EINVAL;
514 	}
515 
516 	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
517 				&rdev->ring_tmp_bo.bo->tbo.mem);
518 	if (r)
519 		return r;
520 
521 	for (i = 0; i < p->nrelocs; i++) {
522 		struct radeon_bo *bo;
523 
524 		/* ignore duplicates */
525 		if (p->relocs_ptr[i] != &p->relocs[i])
526 			continue;
527 
528 		bo = p->relocs[i].robj;
529 		bo_va = radeon_vm_bo_find(vm, bo);
530 		if (bo_va == NULL) {
531 			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
532 			return -EINVAL;
533 		}
534 
535 		r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
536 		if (r)
537 			return r;
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_idx == -1)
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 	radeon_semaphore_sync_fence(parser->ib.semaphore, vm->fence);
583 
584 	if ((rdev->family >= CHIP_TAHITI) &&
585 	    (parser->chunk_const_ib_idx != -1)) {
586 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
587 	} else {
588 		r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
589 	}
590 
591 out:
592 	mutex_unlock(&vm->mutex);
593 	return r;
594 }
595 
596 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
597 {
598 	if (r == -EDEADLK) {
599 		r = radeon_gpu_reset(rdev);
600 		if (!r)
601 			r = -EAGAIN;
602 	}
603 	return r;
604 }
605 
606 static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
607 {
608 	struct radeon_cs_chunk *ib_chunk;
609 	struct radeon_vm *vm = NULL;
610 	int r;
611 
612 	if (parser->chunk_ib_idx == -1)
613 		return 0;
614 
615 	if (parser->cs_flags & RADEON_CS_USE_VM) {
616 		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
617 		vm = &fpriv->vm;
618 
619 		if ((rdev->family >= CHIP_TAHITI) &&
620 		    (parser->chunk_const_ib_idx != -1)) {
621 			ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
622 			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
623 				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
624 				return -EINVAL;
625 			}
626 			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
627 					   vm, ib_chunk->length_dw * 4);
628 			if (r) {
629 				DRM_ERROR("Failed to get const ib !\n");
630 				return r;
631 			}
632 			parser->const_ib.is_const_ib = true;
633 			parser->const_ib.length_dw = ib_chunk->length_dw;
634 			if (copy_from_user(parser->const_ib.ptr,
635 					       ib_chunk->user_ptr,
636 					       ib_chunk->length_dw * 4))
637 				return -EFAULT;
638 		}
639 
640 		ib_chunk = &parser->chunks[parser->chunk_ib_idx];
641 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
642 			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
643 			return -EINVAL;
644 		}
645 	}
646 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
647 
648 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
649 			   vm, ib_chunk->length_dw * 4);
650 	if (r) {
651 		DRM_ERROR("Failed to get ib !\n");
652 		return r;
653 	}
654 	parser->ib.length_dw = ib_chunk->length_dw;
655 	if (ib_chunk->kdata)
656 		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
657 	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
658 		return -EFAULT;
659 	return 0;
660 }
661 
662 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
663 {
664 	struct radeon_device *rdev = dev->dev_private;
665 	struct radeon_cs_parser parser;
666 	int r;
667 
668 	down_read(&rdev->exclusive_lock);
669 	if (!rdev->accel_working) {
670 		up_read(&rdev->exclusive_lock);
671 		return -EBUSY;
672 	}
673 	if (rdev->in_reset) {
674 		up_read(&rdev->exclusive_lock);
675 		r = radeon_gpu_reset(rdev);
676 		if (!r)
677 			r = -EAGAIN;
678 		return r;
679 	}
680 	/* initialize parser */
681 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
682 	parser.filp = filp;
683 	parser.rdev = rdev;
684 	parser.dev = rdev->dev;
685 	parser.family = rdev->family;
686 	r = radeon_cs_parser_init(&parser, data);
687 	if (r) {
688 		DRM_ERROR("Failed to initialize parser !\n");
689 		radeon_cs_parser_fini(&parser, r, false);
690 		up_read(&rdev->exclusive_lock);
691 		r = radeon_cs_handle_lockup(rdev, r);
692 		return r;
693 	}
694 
695 	r = radeon_cs_ib_fill(rdev, &parser);
696 	if (!r) {
697 		r = radeon_cs_parser_relocs(&parser);
698 		if (r && r != -ERESTARTSYS)
699 			DRM_ERROR("Failed to parse relocation %d!\n", r);
700 	}
701 
702 	if (r) {
703 		radeon_cs_parser_fini(&parser, r, false);
704 		up_read(&rdev->exclusive_lock);
705 		r = radeon_cs_handle_lockup(rdev, r);
706 		return r;
707 	}
708 
709 	trace_radeon_cs(&parser);
710 
711 	r = radeon_cs_ib_chunk(rdev, &parser);
712 	if (r) {
713 		goto out;
714 	}
715 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
716 	if (r) {
717 		goto out;
718 	}
719 out:
720 	radeon_cs_parser_fini(&parser, r, true);
721 	up_read(&rdev->exclusive_lock);
722 	r = radeon_cs_handle_lockup(rdev, r);
723 	return r;
724 }
725 
726 /**
727  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
728  * @parser:	parser structure holding parsing context.
729  * @pkt:	where to store packet information
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->chunks[p->chunk_ib_idx];
739 	struct radeon_device *rdev = p->rdev;
740 	uint32_t header;
741 
742 	if (idx >= ib_chunk->length_dw) {
743 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
744 			  idx, ib_chunk->length_dw);
745 		return -EINVAL;
746 	}
747 	header = radeon_get_ib_value(p, idx);
748 	pkt->idx = idx;
749 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
750 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
751 	pkt->one_reg_wr = 0;
752 	switch (pkt->type) {
753 	case RADEON_PACKET_TYPE0:
754 		if (rdev->family < CHIP_R600) {
755 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
756 			pkt->one_reg_wr =
757 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
758 		} else
759 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
760 		break;
761 	case RADEON_PACKET_TYPE3:
762 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
763 		break;
764 	case RADEON_PACKET_TYPE2:
765 		pkt->count = -1;
766 		break;
767 	default:
768 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
769 		return -EINVAL;
770 	}
771 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
772 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
773 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
774 		return -EINVAL;
775 	}
776 	return 0;
777 }
778 
779 /**
780  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
781  * @p:		structure holding the parser context.
782  *
783  * Check if the next packet is NOP relocation packet3.
784  **/
785 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
786 {
787 	struct radeon_cs_packet p3reloc;
788 	int r;
789 
790 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
791 	if (r)
792 		return false;
793 	if (p3reloc.type != RADEON_PACKET_TYPE3)
794 		return false;
795 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
796 		return false;
797 	return true;
798 }
799 
800 /**
801  * radeon_cs_dump_packet() - dump raw packet context
802  * @p:		structure holding the parser context.
803  * @pkt:	structure holding the packet.
804  *
805  * Used mostly for debugging and error reporting.
806  **/
807 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
808 			   struct radeon_cs_packet *pkt)
809 {
810 	volatile uint32_t *ib;
811 	unsigned i;
812 	unsigned idx;
813 
814 	ib = p->ib.ptr;
815 	idx = pkt->idx;
816 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
817 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
818 }
819 
820 /**
821  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
822  * @parser:		parser structure holding parsing context.
823  * @data:		pointer to relocation data
824  * @offset_start:	starting offset
825  * @offset_mask:	offset mask (to align start offset on)
826  * @reloc:		reloc informations
827  *
828  * Check if next packet is relocation packet3, do bo validation and compute
829  * GPU offset using the provided start.
830  **/
831 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
832 				struct radeon_cs_reloc **cs_reloc,
833 				int nomm)
834 {
835 	struct radeon_cs_chunk *relocs_chunk;
836 	struct radeon_cs_packet p3reloc;
837 	unsigned idx;
838 	int r;
839 
840 	if (p->chunk_relocs_idx == -1) {
841 		DRM_ERROR("No relocation chunk !\n");
842 		return -EINVAL;
843 	}
844 	*cs_reloc = NULL;
845 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
846 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
847 	if (r)
848 		return r;
849 	p->idx += p3reloc.count + 2;
850 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
851 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
852 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
853 			  p3reloc.idx);
854 		radeon_cs_dump_packet(p, &p3reloc);
855 		return -EINVAL;
856 	}
857 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
858 	if (idx >= relocs_chunk->length_dw) {
859 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
860 			  idx, relocs_chunk->length_dw);
861 		radeon_cs_dump_packet(p, &p3reloc);
862 		return -EINVAL;
863 	}
864 	/* FIXME: we assume reloc size is 4 dwords */
865 	if (nomm) {
866 		*cs_reloc = p->relocs;
867 		(*cs_reloc)->gpu_offset =
868 			(u64)relocs_chunk->kdata[idx + 3] << 32;
869 		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
870 	} else
871 		*cs_reloc = p->relocs_ptr[(idx / 4)];
872 	return 0;
873 }
874