• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2008 Jerome Glisse.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22  * DEALINGS IN THE SOFTWARE.
23  *
24  * Authors:
25  *    Jerome Glisse <glisse@freedesktop.org>
26  */
27 
28 #include <linux/list_sort.h>
29 #include <linux/pci.h>
30 #include <linux/uaccess.h>
31 
32 #include <drm/drm_device.h>
33 #include <drm/drm_file.h>
34 #include <drm/radeon_drm.h>
35 
36 #include "radeon.h"
37 #include "radeon_reg.h"
38 #include "radeon_trace.h"
39 
40 #define RADEON_CS_MAX_PRIORITY		32u
41 #define RADEON_CS_NUM_BUCKETS		(RADEON_CS_MAX_PRIORITY + 1)
42 
43 /* This is based on the bucket sort with O(n) time complexity.
44  * An item with priority "i" is added to bucket[i]. The lists are then
45  * concatenated in descending order.
46  */
47 struct radeon_cs_buckets {
48 	struct list_head bucket[RADEON_CS_NUM_BUCKETS];
49 };
50 
radeon_cs_buckets_init(struct radeon_cs_buckets * b)51 static void radeon_cs_buckets_init(struct radeon_cs_buckets *b)
52 {
53 	unsigned i;
54 
55 	for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++)
56 		INIT_LIST_HEAD(&b->bucket[i]);
57 }
58 
radeon_cs_buckets_add(struct radeon_cs_buckets * b,struct list_head * item,unsigned priority)59 static void radeon_cs_buckets_add(struct radeon_cs_buckets *b,
60 				  struct list_head *item, unsigned priority)
61 {
62 	/* Since buffers which appear sooner in the relocation list are
63 	 * likely to be used more often than buffers which appear later
64 	 * in the list, the sort mustn't change the ordering of buffers
65 	 * with the same priority, i.e. it must be stable.
66 	 */
67 	list_add_tail(item, &b->bucket[min(priority, RADEON_CS_MAX_PRIORITY)]);
68 }
69 
radeon_cs_buckets_get_list(struct radeon_cs_buckets * b,struct list_head * out_list)70 static void radeon_cs_buckets_get_list(struct radeon_cs_buckets *b,
71 				       struct list_head *out_list)
72 {
73 	unsigned i;
74 
75 	/* Connect the sorted buckets in the output list. */
76 	for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++) {
77 		list_splice(&b->bucket[i], out_list);
78 	}
79 }
80 
radeon_cs_parser_relocs(struct radeon_cs_parser * p)81 static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
82 {
83 	struct radeon_cs_chunk *chunk;
84 	struct radeon_cs_buckets buckets;
85 	unsigned i;
86 	bool need_mmap_lock = false;
87 	int r;
88 
89 	if (p->chunk_relocs == NULL) {
90 		return 0;
91 	}
92 	chunk = p->chunk_relocs;
93 	p->dma_reloc_idx = 0;
94 	/* FIXME: we assume that each relocs use 4 dwords */
95 	p->nrelocs = chunk->length_dw / 4;
96 	p->relocs = kvcalloc(p->nrelocs, sizeof(struct radeon_bo_list),
97 			GFP_KERNEL);
98 	if (p->relocs == NULL) {
99 		return -ENOMEM;
100 	}
101 
102 	radeon_cs_buckets_init(&buckets);
103 
104 	for (i = 0; i < p->nrelocs; i++) {
105 		struct drm_radeon_cs_reloc *r;
106 		struct drm_gem_object *gobj;
107 		unsigned priority;
108 
109 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
110 		gobj = drm_gem_object_lookup(p->filp, r->handle);
111 		if (gobj == NULL) {
112 			DRM_ERROR("gem object lookup failed 0x%x\n",
113 				  r->handle);
114 			return -ENOENT;
115 		}
116 		p->relocs[i].robj = gem_to_radeon_bo(gobj);
117 
118 		/* The userspace buffer priorities are from 0 to 15. A higher
119 		 * number means the buffer is more important.
120 		 * Also, the buffers used for write have a higher priority than
121 		 * the buffers used for read only, which doubles the range
122 		 * to 0 to 31. 32 is reserved for the kernel driver.
123 		 */
124 		priority = (r->flags & RADEON_RELOC_PRIO_MASK) * 2
125 			   + !!r->write_domain;
126 
127 		/* The first reloc of an UVD job is the msg and that must be in
128 		 * VRAM, the second reloc is the DPB and for WMV that must be in
129 		 * VRAM as well. Also put everything into VRAM on AGP cards and older
130 		 * IGP chips to avoid image corruptions
131 		 */
132 		if (p->ring == R600_RING_TYPE_UVD_INDEX &&
133 		    (i <= 0 || pci_find_capability(p->rdev->pdev, PCI_CAP_ID_AGP) ||
134 		     p->rdev->family == CHIP_RS780 ||
135 		     p->rdev->family == CHIP_RS880)) {
136 
137 			/* TODO: is this still needed for NI+ ? */
138 			p->relocs[i].preferred_domains =
139 				RADEON_GEM_DOMAIN_VRAM;
140 
141 			p->relocs[i].allowed_domains =
142 				RADEON_GEM_DOMAIN_VRAM;
143 
144 			/* prioritize this over any other relocation */
145 			priority = RADEON_CS_MAX_PRIORITY;
146 		} else {
147 			uint32_t domain = r->write_domain ?
148 				r->write_domain : r->read_domains;
149 
150 			if (domain & RADEON_GEM_DOMAIN_CPU) {
151 				DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
152 					  "for command submission\n");
153 				return -EINVAL;
154 			}
155 
156 			p->relocs[i].preferred_domains = domain;
157 			if (domain == RADEON_GEM_DOMAIN_VRAM)
158 				domain |= RADEON_GEM_DOMAIN_GTT;
159 			p->relocs[i].allowed_domains = domain;
160 		}
161 
162 		if (radeon_ttm_tt_has_userptr(p->rdev, p->relocs[i].robj->tbo.ttm)) {
163 			uint32_t domain = p->relocs[i].preferred_domains;
164 			if (!(domain & RADEON_GEM_DOMAIN_GTT)) {
165 				DRM_ERROR("Only RADEON_GEM_DOMAIN_GTT is "
166 					  "allowed for userptr BOs\n");
167 				return -EINVAL;
168 			}
169 			need_mmap_lock = true;
170 			domain = RADEON_GEM_DOMAIN_GTT;
171 			p->relocs[i].preferred_domains = domain;
172 			p->relocs[i].allowed_domains = domain;
173 		}
174 
175 		/* Objects shared as dma-bufs cannot be moved to VRAM */
176 		if (p->relocs[i].robj->prime_shared_count) {
177 			p->relocs[i].allowed_domains &= ~RADEON_GEM_DOMAIN_VRAM;
178 			if (!p->relocs[i].allowed_domains) {
179 				DRM_ERROR("BO associated with dma-buf cannot "
180 					  "be moved to VRAM\n");
181 				return -EINVAL;
182 			}
183 		}
184 
185 		p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
186 		p->relocs[i].tv.num_shared = !r->write_domain;
187 
188 		radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head,
189 				      priority);
190 	}
191 
192 	radeon_cs_buckets_get_list(&buckets, &p->validated);
193 
194 	if (p->cs_flags & RADEON_CS_USE_VM)
195 		p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm,
196 					      &p->validated);
197 	if (need_mmap_lock)
198 		mmap_read_lock(current->mm);
199 
200 	r = radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
201 
202 	if (need_mmap_lock)
203 		mmap_read_unlock(current->mm);
204 
205 	return r;
206 }
207 
radeon_cs_get_ring(struct radeon_cs_parser * p,u32 ring,s32 priority)208 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
209 {
210 	p->priority = priority;
211 
212 	switch (ring) {
213 	default:
214 		DRM_ERROR("unknown ring id: %d\n", ring);
215 		return -EINVAL;
216 	case RADEON_CS_RING_GFX:
217 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
218 		break;
219 	case RADEON_CS_RING_COMPUTE:
220 		if (p->rdev->family >= CHIP_TAHITI) {
221 			if (p->priority > 0)
222 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
223 			else
224 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
225 		} else
226 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
227 		break;
228 	case RADEON_CS_RING_DMA:
229 		if (p->rdev->family >= CHIP_CAYMAN) {
230 			if (p->priority > 0)
231 				p->ring = R600_RING_TYPE_DMA_INDEX;
232 			else
233 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
234 		} else if (p->rdev->family >= CHIP_RV770) {
235 			p->ring = R600_RING_TYPE_DMA_INDEX;
236 		} else {
237 			return -EINVAL;
238 		}
239 		break;
240 	case RADEON_CS_RING_UVD:
241 		p->ring = R600_RING_TYPE_UVD_INDEX;
242 		break;
243 	case RADEON_CS_RING_VCE:
244 		/* TODO: only use the low priority ring for now */
245 		p->ring = TN_RING_TYPE_VCE1_INDEX;
246 		break;
247 	}
248 	return 0;
249 }
250 
radeon_cs_sync_rings(struct radeon_cs_parser * p)251 static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
252 {
253 	struct radeon_bo_list *reloc;
254 	int r;
255 
256 	list_for_each_entry(reloc, &p->validated, tv.head) {
257 		struct dma_resv *resv;
258 
259 		resv = reloc->robj->tbo.base.resv;
260 		r = radeon_sync_resv(p->rdev, &p->ib.sync, resv,
261 				     reloc->tv.num_shared);
262 		if (r)
263 			return r;
264 	}
265 	return 0;
266 }
267 
268 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
radeon_cs_parser_init(struct radeon_cs_parser * p,void * data)269 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
270 {
271 	struct drm_radeon_cs *cs = data;
272 	uint64_t *chunk_array_ptr;
273 	u64 size;
274 	unsigned i;
275 	u32 ring = RADEON_CS_RING_GFX;
276 	s32 priority = 0;
277 
278 	INIT_LIST_HEAD(&p->validated);
279 
280 	if (!cs->num_chunks) {
281 		return 0;
282 	}
283 
284 	/* get chunks */
285 	p->idx = 0;
286 	p->ib.sa_bo = NULL;
287 	p->const_ib.sa_bo = NULL;
288 	p->chunk_ib = NULL;
289 	p->chunk_relocs = NULL;
290 	p->chunk_flags = NULL;
291 	p->chunk_const_ib = NULL;
292 	p->chunks_array = kvmalloc_array(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 = kvcalloc(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 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_RELOCS) {
319 			p->chunk_relocs = &p->chunks[i];
320 		}
321 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
322 			p->chunk_ib = &p->chunks[i];
323 			/* zero length IB isn't useful */
324 			if (p->chunks[i].length_dw == 0)
325 				return -EINVAL;
326 		}
327 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB) {
328 			p->chunk_const_ib = &p->chunks[i];
329 			/* zero length CONST IB isn't useful */
330 			if (p->chunks[i].length_dw == 0)
331 				return -EINVAL;
332 		}
333 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
334 			p->chunk_flags = &p->chunks[i];
335 			/* zero length flags aren't useful */
336 			if (p->chunks[i].length_dw == 0)
337 				return -EINVAL;
338 		}
339 
340 		size = p->chunks[i].length_dw;
341 		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
342 		p->chunks[i].user_ptr = cdata;
343 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB)
344 			continue;
345 
346 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
347 			if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
348 				continue;
349 		}
350 
351 		p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
352 		size *= sizeof(uint32_t);
353 		if (p->chunks[i].kdata == NULL) {
354 			return -ENOMEM;
355 		}
356 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
357 			return -EFAULT;
358 		}
359 		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
360 			p->cs_flags = p->chunks[i].kdata[0];
361 			if (p->chunks[i].length_dw > 1)
362 				ring = p->chunks[i].kdata[1];
363 			if (p->chunks[i].length_dw > 2)
364 				priority = (s32)p->chunks[i].kdata[2];
365 		}
366 	}
367 
368 	/* these are KMS only */
369 	if (p->rdev) {
370 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
371 		    !p->rdev->vm_manager.enabled) {
372 			DRM_ERROR("VM not active on asic!\n");
373 			return -EINVAL;
374 		}
375 
376 		if (radeon_cs_get_ring(p, ring, priority))
377 			return -EINVAL;
378 
379 		/* we only support VM on some SI+ rings */
380 		if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
381 			if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
382 				DRM_ERROR("Ring %d requires VM!\n", p->ring);
383 				return -EINVAL;
384 			}
385 		} else {
386 			if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
387 				DRM_ERROR("VM not supported on ring %d!\n",
388 					  p->ring);
389 				return -EINVAL;
390 			}
391 		}
392 	}
393 
394 	return 0;
395 }
396 
cmp_size_smaller_first(void * priv,const struct list_head * a,const struct list_head * b)397 static int cmp_size_smaller_first(void *priv, const struct list_head *a,
398 				  const struct list_head *b)
399 {
400 	struct radeon_bo_list *la = list_entry(a, struct radeon_bo_list, tv.head);
401 	struct radeon_bo_list *lb = list_entry(b, struct radeon_bo_list, tv.head);
402 
403 	/* Sort A before B if A is smaller. */
404 	return (int)la->robj->tbo.resource->num_pages -
405 		(int)lb->robj->tbo.resource->num_pages;
406 }
407 
408 /**
409  * radeon_cs_parser_fini() - clean parser states
410  * @parser:	parser structure holding parsing context.
411  * @error:	error number
412  * @backoff:	indicator to backoff the reservation
413  *
414  * If error is set than unvalidate buffer, otherwise just free memory
415  * used by parsing context.
416  **/
radeon_cs_parser_fini(struct radeon_cs_parser * parser,int error,bool backoff)417 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
418 {
419 	unsigned i;
420 
421 	if (!error) {
422 		/* Sort the buffer list from the smallest to largest buffer,
423 		 * which affects the order of buffers in the LRU list.
424 		 * This assures that the smallest buffers are added first
425 		 * to the LRU list, so they are likely to be later evicted
426 		 * first, instead of large buffers whose eviction is more
427 		 * expensive.
428 		 *
429 		 * This slightly lowers the number of bytes moved by TTM
430 		 * per frame under memory pressure.
431 		 */
432 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
433 
434 		ttm_eu_fence_buffer_objects(&parser->ticket,
435 					    &parser->validated,
436 					    &parser->ib.fence->base);
437 	} else if (backoff) {
438 		ttm_eu_backoff_reservation(&parser->ticket,
439 					   &parser->validated);
440 	}
441 
442 	if (parser->relocs != NULL) {
443 		for (i = 0; i < parser->nrelocs; i++) {
444 			struct radeon_bo *bo = parser->relocs[i].robj;
445 			if (bo == NULL)
446 				continue;
447 
448 			drm_gem_object_put(&bo->tbo.base);
449 		}
450 	}
451 	kfree(parser->track);
452 	kvfree(parser->relocs);
453 	kvfree(parser->vm_bos);
454 	for (i = 0; i < parser->nchunks; i++)
455 		kvfree(parser->chunks[i].kdata);
456 	kvfree(parser->chunks);
457 	kvfree(parser->chunks_array);
458 	radeon_ib_free(parser->rdev, &parser->ib);
459 	radeon_ib_free(parser->rdev, &parser->const_ib);
460 }
461 
radeon_cs_ib_chunk(struct radeon_device * rdev,struct radeon_cs_parser * parser)462 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
463 			      struct radeon_cs_parser *parser)
464 {
465 	int r;
466 
467 	if (parser->chunk_ib == NULL)
468 		return 0;
469 
470 	if (parser->cs_flags & RADEON_CS_USE_VM)
471 		return 0;
472 
473 	r = radeon_cs_parse(rdev, parser->ring, parser);
474 	if (r || parser->parser_error) {
475 		DRM_ERROR("Invalid command stream !\n");
476 		return r;
477 	}
478 
479 	r = radeon_cs_sync_rings(parser);
480 	if (r) {
481 		if (r != -ERESTARTSYS)
482 			DRM_ERROR("Failed to sync rings: %i\n", r);
483 		return r;
484 	}
485 
486 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
487 		radeon_uvd_note_usage(rdev);
488 	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
489 		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
490 		radeon_vce_note_usage(rdev);
491 
492 	r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
493 	if (r) {
494 		DRM_ERROR("Failed to schedule IB !\n");
495 	}
496 	return r;
497 }
498 
radeon_bo_vm_update_pte(struct radeon_cs_parser * p,struct radeon_vm * vm)499 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
500 				   struct radeon_vm *vm)
501 {
502 	struct radeon_device *rdev = p->rdev;
503 	struct radeon_bo_va *bo_va;
504 	int i, r;
505 
506 	r = radeon_vm_update_page_directory(rdev, vm);
507 	if (r)
508 		return r;
509 
510 	r = radeon_vm_clear_freed(rdev, vm);
511 	if (r)
512 		return r;
513 
514 	if (vm->ib_bo_va == NULL) {
515 		DRM_ERROR("Tmp BO not in VM!\n");
516 		return -EINVAL;
517 	}
518 
519 	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
520 				rdev->ring_tmp_bo.bo->tbo.resource);
521 	if (r)
522 		return r;
523 
524 	for (i = 0; i < p->nrelocs; i++) {
525 		struct radeon_bo *bo;
526 
527 		bo = p->relocs[i].robj;
528 		bo_va = radeon_vm_bo_find(vm, bo);
529 		if (bo_va == NULL) {
530 			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
531 			return -EINVAL;
532 		}
533 
534 		r = radeon_vm_bo_update(rdev, bo_va, bo->tbo.resource);
535 		if (r)
536 			return r;
537 
538 		radeon_sync_fence(&p->ib.sync, bo_va->last_pt_update);
539 
540 		r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
541 		if (r)
542 			return r;
543 	}
544 
545 	return radeon_vm_clear_invalids(rdev, vm);
546 }
547 
radeon_cs_ib_vm_chunk(struct radeon_device * rdev,struct radeon_cs_parser * parser)548 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
549 				 struct radeon_cs_parser *parser)
550 {
551 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
552 	struct radeon_vm *vm = &fpriv->vm;
553 	int r;
554 
555 	if (parser->chunk_ib == NULL)
556 		return 0;
557 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
558 		return 0;
559 
560 	if (parser->const_ib.length_dw) {
561 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
562 		if (r) {
563 			return r;
564 		}
565 	}
566 
567 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
568 	if (r) {
569 		return r;
570 	}
571 
572 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
573 		radeon_uvd_note_usage(rdev);
574 
575 	mutex_lock(&vm->mutex);
576 	r = radeon_bo_vm_update_pte(parser, vm);
577 	if (r) {
578 		goto out;
579 	}
580 
581 	r = radeon_cs_sync_rings(parser);
582 	if (r) {
583 		if (r != -ERESTARTSYS)
584 			DRM_ERROR("Failed to sync rings: %i\n", r);
585 		goto out;
586 	}
587 
588 	if ((rdev->family >= CHIP_TAHITI) &&
589 	    (parser->chunk_const_ib != NULL)) {
590 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
591 	} else {
592 		r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
593 	}
594 
595 out:
596 	mutex_unlock(&vm->mutex);
597 	return r;
598 }
599 
radeon_cs_handle_lockup(struct radeon_device * rdev,int r)600 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
601 {
602 	if (r == -EDEADLK) {
603 		r = radeon_gpu_reset(rdev);
604 		if (!r)
605 			r = -EAGAIN;
606 	}
607 	return r;
608 }
609 
radeon_cs_ib_fill(struct radeon_device * rdev,struct radeon_cs_parser * parser)610 static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
611 {
612 	struct radeon_cs_chunk *ib_chunk;
613 	struct radeon_vm *vm = NULL;
614 	int r;
615 
616 	if (parser->chunk_ib == NULL)
617 		return 0;
618 
619 	if (parser->cs_flags & RADEON_CS_USE_VM) {
620 		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
621 		vm = &fpriv->vm;
622 
623 		if ((rdev->family >= CHIP_TAHITI) &&
624 		    (parser->chunk_const_ib != NULL)) {
625 			ib_chunk = parser->chunk_const_ib;
626 			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
627 				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
628 				return -EINVAL;
629 			}
630 			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
631 					   vm, ib_chunk->length_dw * 4);
632 			if (r) {
633 				DRM_ERROR("Failed to get const ib !\n");
634 				return r;
635 			}
636 			parser->const_ib.is_const_ib = true;
637 			parser->const_ib.length_dw = ib_chunk->length_dw;
638 			if (copy_from_user(parser->const_ib.ptr,
639 					       ib_chunk->user_ptr,
640 					       ib_chunk->length_dw * 4))
641 				return -EFAULT;
642 		}
643 
644 		ib_chunk = parser->chunk_ib;
645 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
646 			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
647 			return -EINVAL;
648 		}
649 	}
650 	ib_chunk = parser->chunk_ib;
651 
652 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
653 			   vm, ib_chunk->length_dw * 4);
654 	if (r) {
655 		DRM_ERROR("Failed to get ib !\n");
656 		return r;
657 	}
658 	parser->ib.length_dw = ib_chunk->length_dw;
659 	if (ib_chunk->kdata)
660 		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
661 	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
662 		return -EFAULT;
663 	return 0;
664 }
665 
radeon_cs_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)666 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
667 {
668 	struct radeon_device *rdev = dev->dev_private;
669 	struct radeon_cs_parser parser;
670 	int r;
671 
672 	down_read(&rdev->exclusive_lock);
673 	if (!rdev->accel_working) {
674 		up_read(&rdev->exclusive_lock);
675 		return -EBUSY;
676 	}
677 	if (rdev->in_reset) {
678 		up_read(&rdev->exclusive_lock);
679 		r = radeon_gpu_reset(rdev);
680 		if (!r)
681 			r = -EAGAIN;
682 		return r;
683 	}
684 	/* initialize parser */
685 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
686 	parser.filp = filp;
687 	parser.rdev = rdev;
688 	parser.dev = rdev->dev;
689 	parser.family = rdev->family;
690 	r = radeon_cs_parser_init(&parser, data);
691 	if (r) {
692 		DRM_ERROR("Failed to initialize parser !\n");
693 		radeon_cs_parser_fini(&parser, r, false);
694 		up_read(&rdev->exclusive_lock);
695 		r = radeon_cs_handle_lockup(rdev, r);
696 		return r;
697 	}
698 
699 	r = radeon_cs_ib_fill(rdev, &parser);
700 	if (!r) {
701 		r = radeon_cs_parser_relocs(&parser);
702 		if (r && r != -ERESTARTSYS)
703 			DRM_ERROR("Failed to parse relocation %d!\n", r);
704 	}
705 
706 	if (r) {
707 		radeon_cs_parser_fini(&parser, r, false);
708 		up_read(&rdev->exclusive_lock);
709 		r = radeon_cs_handle_lockup(rdev, r);
710 		return r;
711 	}
712 
713 	trace_radeon_cs(&parser);
714 
715 	r = radeon_cs_ib_chunk(rdev, &parser);
716 	if (r) {
717 		goto out;
718 	}
719 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
720 	if (r) {
721 		goto out;
722 	}
723 out:
724 	radeon_cs_parser_fini(&parser, r, true);
725 	up_read(&rdev->exclusive_lock);
726 	r = radeon_cs_handle_lockup(rdev, r);
727 	return r;
728 }
729 
730 /**
731  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
732  * @p:		parser structure holding parsing context.
733  * @pkt:	where to store packet information
734  * @idx:	packet index
735  *
736  * Assume that chunk_ib_index is properly set. Will return -EINVAL
737  * if packet is bigger than remaining ib size. or if packets is unknown.
738  **/
radeon_cs_packet_parse(struct radeon_cs_parser * p,struct radeon_cs_packet * pkt,unsigned idx)739 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
740 			   struct radeon_cs_packet *pkt,
741 			   unsigned idx)
742 {
743 	struct radeon_cs_chunk *ib_chunk = p->chunk_ib;
744 	struct radeon_device *rdev = p->rdev;
745 	uint32_t header;
746 	int ret = 0, i;
747 
748 	if (idx >= ib_chunk->length_dw) {
749 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
750 			  idx, ib_chunk->length_dw);
751 		return -EINVAL;
752 	}
753 	header = radeon_get_ib_value(p, idx);
754 	pkt->idx = idx;
755 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
756 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
757 	pkt->one_reg_wr = 0;
758 	switch (pkt->type) {
759 	case RADEON_PACKET_TYPE0:
760 		if (rdev->family < CHIP_R600) {
761 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
762 			pkt->one_reg_wr =
763 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
764 		} else
765 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
766 		break;
767 	case RADEON_PACKET_TYPE3:
768 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
769 		break;
770 	case RADEON_PACKET_TYPE2:
771 		pkt->count = -1;
772 		break;
773 	default:
774 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
775 		ret = -EINVAL;
776 		goto dump_ib;
777 	}
778 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
779 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
780 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
781 		ret = -EINVAL;
782 		goto dump_ib;
783 	}
784 	return 0;
785 
786 dump_ib:
787 	for (i = 0; i < ib_chunk->length_dw; i++) {
788 		if (i == idx)
789 			printk("\t0x%08x <---\n", radeon_get_ib_value(p, i));
790 		else
791 			printk("\t0x%08x\n", radeon_get_ib_value(p, i));
792 	}
793 	return ret;
794 }
795 
796 /**
797  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
798  * @p:		structure holding the parser context.
799  *
800  * Check if the next packet is NOP relocation packet3.
801  **/
radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser * p)802 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
803 {
804 	struct radeon_cs_packet p3reloc;
805 	int r;
806 
807 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
808 	if (r)
809 		return false;
810 	if (p3reloc.type != RADEON_PACKET_TYPE3)
811 		return false;
812 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
813 		return false;
814 	return true;
815 }
816 
817 /**
818  * radeon_cs_dump_packet() - dump raw packet context
819  * @p:		structure holding the parser context.
820  * @pkt:	structure holding the packet.
821  *
822  * Used mostly for debugging and error reporting.
823  **/
radeon_cs_dump_packet(struct radeon_cs_parser * p,struct radeon_cs_packet * pkt)824 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
825 			   struct radeon_cs_packet *pkt)
826 {
827 	volatile uint32_t *ib;
828 	unsigned i;
829 	unsigned idx;
830 
831 	ib = p->ib.ptr;
832 	idx = pkt->idx;
833 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
834 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
835 }
836 
837 /**
838  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
839  * @p:			parser structure holding parsing context.
840  * @cs_reloc:		reloc informations
841  * @nomm:		no memory management for debugging
842  *
843  * Check if next packet is relocation packet3, do bo validation and compute
844  * GPU offset using the provided start.
845  **/
radeon_cs_packet_next_reloc(struct radeon_cs_parser * p,struct radeon_bo_list ** cs_reloc,int nomm)846 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
847 				struct radeon_bo_list **cs_reloc,
848 				int nomm)
849 {
850 	struct radeon_cs_chunk *relocs_chunk;
851 	struct radeon_cs_packet p3reloc;
852 	unsigned idx;
853 	int r;
854 
855 	if (p->chunk_relocs == NULL) {
856 		DRM_ERROR("No relocation chunk !\n");
857 		return -EINVAL;
858 	}
859 	*cs_reloc = NULL;
860 	relocs_chunk = p->chunk_relocs;
861 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
862 	if (r)
863 		return r;
864 	p->idx += p3reloc.count + 2;
865 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
866 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
867 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
868 			  p3reloc.idx);
869 		radeon_cs_dump_packet(p, &p3reloc);
870 		return -EINVAL;
871 	}
872 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
873 	if (idx >= relocs_chunk->length_dw) {
874 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
875 			  idx, relocs_chunk->length_dw);
876 		radeon_cs_dump_packet(p, &p3reloc);
877 		return -EINVAL;
878 	}
879 	/* FIXME: we assume reloc size is 4 dwords */
880 	if (nomm) {
881 		*cs_reloc = p->relocs;
882 		(*cs_reloc)->gpu_offset =
883 			(u64)relocs_chunk->kdata[idx + 3] << 32;
884 		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
885 	} else
886 		*cs_reloc = &p->relocs[(idx / 4)];
887 	return 0;
888 }
889