Linux Audio

Check our new training course

Loading...
v4.6
  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 */
 41struct radeon_cs_buckets {
 42	struct list_head bucket[RADEON_CS_NUM_BUCKETS];
 43};
 44
 45static 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
 53static 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
 64static 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
 75static 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;
 81	bool need_mmap_lock = false;
 82	int r;
 83
 84	if (p->chunk_relocs == NULL) {
 85		return 0;
 86	}
 87	chunk = p->chunk_relocs;
 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 = drm_calloc_large(p->nrelocs, sizeof(struct radeon_bo_list));
 
 92	if (p->relocs == NULL) {
 93		return -ENOMEM;
 94	}
 95
 96	radeon_cs_buckets_init(&buckets);
 97
 98	for (i = 0; i < p->nrelocs; i++) {
 99		struct drm_radeon_cs_reloc *r;
100		struct drm_gem_object *gobj;
101		unsigned priority;
102
103		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
104		gobj = drm_gem_object_lookup(ddev, p->filp, r->handle);
105		if (gobj == NULL) {
106			DRM_ERROR("gem object lookup failed 0x%x\n",
107				  r->handle);
108			return -ENOENT;
109		}
110		p->relocs[i].robj = gem_to_radeon_bo(gobj);
111
112		/* The userspace buffer priorities are from 0 to 15. A higher
113		 * number means the buffer is more important.
114		 * Also, the buffers used for write have a higher priority than
115		 * the buffers used for read only, which doubles the range
116		 * to 0 to 31. 32 is reserved for the kernel driver.
117		 */
118		priority = (r->flags & RADEON_RELOC_PRIO_MASK) * 2
119			   + !!r->write_domain;
120
121		/* the first reloc of an UVD job is the msg and that must be in
122		   VRAM, also but everything into VRAM on AGP cards and older
123		   IGP chips to avoid image corruptions */
 
 
124		if (p->ring == R600_RING_TYPE_UVD_INDEX &&
125		    (i == 0 || drm_pci_device_is_agp(p->rdev->ddev) ||
 
126		     p->rdev->family == CHIP_RS780 ||
127		     p->rdev->family == CHIP_RS880)) {
128
129			/* TODO: is this still needed for NI+ ? */
130			p->relocs[i].prefered_domains =
131				RADEON_GEM_DOMAIN_VRAM;
132
133			p->relocs[i].allowed_domains =
134				RADEON_GEM_DOMAIN_VRAM;
135
136			/* prioritize this over any other relocation */
137			priority = RADEON_CS_MAX_PRIORITY;
138		} else {
139			uint32_t domain = r->write_domain ?
140				r->write_domain : r->read_domains;
141
142			if (domain & RADEON_GEM_DOMAIN_CPU) {
143				DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
144					  "for command submission\n");
145				return -EINVAL;
146			}
147
148			p->relocs[i].prefered_domains = domain;
149			if (domain == RADEON_GEM_DOMAIN_VRAM)
150				domain |= RADEON_GEM_DOMAIN_GTT;
151			p->relocs[i].allowed_domains = domain;
152		}
153
154		if (radeon_ttm_tt_has_userptr(p->relocs[i].robj->tbo.ttm)) {
155			uint32_t domain = p->relocs[i].prefered_domains;
156			if (!(domain & RADEON_GEM_DOMAIN_GTT)) {
157				DRM_ERROR("Only RADEON_GEM_DOMAIN_GTT is "
158					  "allowed for userptr BOs\n");
159				return -EINVAL;
160			}
161			need_mmap_lock = true;
162			domain = RADEON_GEM_DOMAIN_GTT;
163			p->relocs[i].prefered_domains = domain;
164			p->relocs[i].allowed_domains = domain;
165		}
166
 
 
 
 
 
 
 
 
 
 
167		p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
168		p->relocs[i].tv.shared = !r->write_domain;
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	if (need_mmap_lock)
180		down_read(&current->mm->mmap_sem);
181
182	r = radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
183
184	if (need_mmap_lock)
185		up_read(&current->mm->mmap_sem);
186
187	return r;
188}
189
190static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
191{
192	p->priority = priority;
193
194	switch (ring) {
195	default:
196		DRM_ERROR("unknown ring id: %d\n", ring);
197		return -EINVAL;
198	case RADEON_CS_RING_GFX:
199		p->ring = RADEON_RING_TYPE_GFX_INDEX;
200		break;
201	case RADEON_CS_RING_COMPUTE:
202		if (p->rdev->family >= CHIP_TAHITI) {
203			if (p->priority > 0)
204				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
205			else
206				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
207		} else
208			p->ring = RADEON_RING_TYPE_GFX_INDEX;
209		break;
210	case RADEON_CS_RING_DMA:
211		if (p->rdev->family >= CHIP_CAYMAN) {
212			if (p->priority > 0)
213				p->ring = R600_RING_TYPE_DMA_INDEX;
214			else
215				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
216		} else if (p->rdev->family >= CHIP_RV770) {
217			p->ring = R600_RING_TYPE_DMA_INDEX;
218		} else {
219			return -EINVAL;
220		}
221		break;
222	case RADEON_CS_RING_UVD:
223		p->ring = R600_RING_TYPE_UVD_INDEX;
224		break;
225	case RADEON_CS_RING_VCE:
226		/* TODO: only use the low priority ring for now */
227		p->ring = TN_RING_TYPE_VCE1_INDEX;
228		break;
229	}
230	return 0;
231}
232
233static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
234{
235	struct radeon_bo_list *reloc;
236	int r;
237
238	list_for_each_entry(reloc, &p->validated, tv.head) {
239		struct reservation_object *resv;
240
241		resv = reloc->robj->tbo.resv;
242		r = radeon_sync_resv(p->rdev, &p->ib.sync, resv,
243				     reloc->tv.shared);
244		if (r)
245			return r;
246	}
247	return 0;
248}
249
250/* XXX: note that this is called from the legacy UMS CS ioctl as well */
251int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
252{
253	struct drm_radeon_cs *cs = data;
254	uint64_t *chunk_array_ptr;
255	unsigned size, i;
256	u32 ring = RADEON_CS_RING_GFX;
257	s32 priority = 0;
258
259	INIT_LIST_HEAD(&p->validated);
260
261	if (!cs->num_chunks) {
262		return 0;
263	}
264
265	/* get chunks */
266	p->idx = 0;
267	p->ib.sa_bo = NULL;
268	p->const_ib.sa_bo = NULL;
269	p->chunk_ib = NULL;
270	p->chunk_relocs = NULL;
271	p->chunk_flags = NULL;
272	p->chunk_const_ib = NULL;
273	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
274	if (p->chunks_array == NULL) {
275		return -ENOMEM;
276	}
277	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
278	if (copy_from_user(p->chunks_array, chunk_array_ptr,
279			       sizeof(uint64_t)*cs->num_chunks)) {
280		return -EFAULT;
281	}
282	p->cs_flags = 0;
283	p->nchunks = cs->num_chunks;
284	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
285	if (p->chunks == NULL) {
286		return -ENOMEM;
287	}
288	for (i = 0; i < p->nchunks; i++) {
289		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
290		struct drm_radeon_cs_chunk user_chunk;
291		uint32_t __user *cdata;
292
293		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
294		if (copy_from_user(&user_chunk, chunk_ptr,
295				       sizeof(struct drm_radeon_cs_chunk))) {
296			return -EFAULT;
297		}
298		p->chunks[i].length_dw = user_chunk.length_dw;
299		if (user_chunk.chunk_id == RADEON_CHUNK_ID_RELOCS) {
300			p->chunk_relocs = &p->chunks[i];
301		}
302		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
303			p->chunk_ib = &p->chunks[i];
304			/* zero length IB isn't useful */
305			if (p->chunks[i].length_dw == 0)
306				return -EINVAL;
307		}
308		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB) {
309			p->chunk_const_ib = &p->chunks[i];
310			/* zero length CONST IB isn't useful */
311			if (p->chunks[i].length_dw == 0)
312				return -EINVAL;
313		}
314		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
315			p->chunk_flags = &p->chunks[i];
316			/* zero length flags aren't useful */
317			if (p->chunks[i].length_dw == 0)
318				return -EINVAL;
319		}
320
321		size = p->chunks[i].length_dw;
322		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
323		p->chunks[i].user_ptr = cdata;
324		if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB)
325			continue;
326
327		if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
328			if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
329				continue;
330		}
331
332		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
333		size *= sizeof(uint32_t);
334		if (p->chunks[i].kdata == NULL) {
335			return -ENOMEM;
336		}
337		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
338			return -EFAULT;
339		}
340		if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
341			p->cs_flags = p->chunks[i].kdata[0];
342			if (p->chunks[i].length_dw > 1)
343				ring = p->chunks[i].kdata[1];
344			if (p->chunks[i].length_dw > 2)
345				priority = (s32)p->chunks[i].kdata[2];
346		}
347	}
348
349	/* these are KMS only */
350	if (p->rdev) {
351		if ((p->cs_flags & RADEON_CS_USE_VM) &&
352		    !p->rdev->vm_manager.enabled) {
353			DRM_ERROR("VM not active on asic!\n");
354			return -EINVAL;
355		}
356
357		if (radeon_cs_get_ring(p, ring, priority))
358			return -EINVAL;
359
360		/* we only support VM on some SI+ rings */
361		if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
362			if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
363				DRM_ERROR("Ring %d requires VM!\n", p->ring);
364				return -EINVAL;
365			}
366		} else {
367			if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
368				DRM_ERROR("VM not supported on ring %d!\n",
369					  p->ring);
370				return -EINVAL;
371			}
372		}
373	}
374
375	return 0;
376}
377
378static int cmp_size_smaller_first(void *priv, struct list_head *a,
379				  struct list_head *b)
380{
381	struct radeon_bo_list *la = list_entry(a, struct radeon_bo_list, tv.head);
382	struct radeon_bo_list *lb = list_entry(b, struct radeon_bo_list, tv.head);
383
384	/* Sort A before B if A is smaller. */
385	return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
386}
387
388/**
389 * cs_parser_fini() - clean parser states
390 * @parser:	parser structure holding parsing context.
391 * @error:	error number
392 *
393 * If error is set than unvalidate buffer, otherwise just free memory
394 * used by parsing context.
395 **/
396static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
397{
398	unsigned i;
399
400	if (!error) {
401		/* Sort the buffer list from the smallest to largest buffer,
402		 * which affects the order of buffers in the LRU list.
403		 * This assures that the smallest buffers are added first
404		 * to the LRU list, so they are likely to be later evicted
405		 * first, instead of large buffers whose eviction is more
406		 * expensive.
407		 *
408		 * This slightly lowers the number of bytes moved by TTM
409		 * per frame under memory pressure.
410		 */
411		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
412
413		ttm_eu_fence_buffer_objects(&parser->ticket,
414					    &parser->validated,
415					    &parser->ib.fence->base);
416	} else if (backoff) {
417		ttm_eu_backoff_reservation(&parser->ticket,
418					   &parser->validated);
419	}
420
421	if (parser->relocs != NULL) {
422		for (i = 0; i < parser->nrelocs; i++) {
423			struct radeon_bo *bo = parser->relocs[i].robj;
424			if (bo == NULL)
425				continue;
426
427			drm_gem_object_unreference_unlocked(&bo->gem_base);
428		}
429	}
430	kfree(parser->track);
431	drm_free_large(parser->relocs);
432	drm_free_large(parser->vm_bos);
433	for (i = 0; i < parser->nchunks; i++)
434		drm_free_large(parser->chunks[i].kdata);
435	kfree(parser->chunks);
436	kfree(parser->chunks_array);
437	radeon_ib_free(parser->rdev, &parser->ib);
438	radeon_ib_free(parser->rdev, &parser->const_ib);
439}
440
441static int radeon_cs_ib_chunk(struct radeon_device *rdev,
442			      struct radeon_cs_parser *parser)
443{
444	int r;
445
446	if (parser->chunk_ib == NULL)
447		return 0;
448
449	if (parser->cs_flags & RADEON_CS_USE_VM)
450		return 0;
451
452	r = radeon_cs_parse(rdev, parser->ring, parser);
453	if (r || parser->parser_error) {
454		DRM_ERROR("Invalid command stream !\n");
455		return r;
456	}
457
458	r = radeon_cs_sync_rings(parser);
459	if (r) {
460		if (r != -ERESTARTSYS)
461			DRM_ERROR("Failed to sync rings: %i\n", r);
462		return r;
463	}
464
465	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
466		radeon_uvd_note_usage(rdev);
467	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
468		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
469		radeon_vce_note_usage(rdev);
470
471	r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
472	if (r) {
473		DRM_ERROR("Failed to schedule IB !\n");
474	}
475	return r;
476}
477
478static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
479				   struct radeon_vm *vm)
480{
481	struct radeon_device *rdev = p->rdev;
482	struct radeon_bo_va *bo_va;
483	int i, r;
484
485	r = radeon_vm_update_page_directory(rdev, vm);
486	if (r)
487		return r;
488
489	r = radeon_vm_clear_freed(rdev, vm);
490	if (r)
491		return r;
492
493	if (vm->ib_bo_va == NULL) {
494		DRM_ERROR("Tmp BO not in VM!\n");
495		return -EINVAL;
496	}
497
498	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
499				&rdev->ring_tmp_bo.bo->tbo.mem);
500	if (r)
501		return r;
502
503	for (i = 0; i < p->nrelocs; i++) {
504		struct radeon_bo *bo;
505
506		bo = p->relocs[i].robj;
507		bo_va = radeon_vm_bo_find(vm, bo);
508		if (bo_va == NULL) {
509			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
510			return -EINVAL;
511		}
512
513		r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
514		if (r)
515			return r;
516
517		radeon_sync_fence(&p->ib.sync, bo_va->last_pt_update);
518	}
519
520	return radeon_vm_clear_invalids(rdev, vm);
521}
522
523static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
524				 struct radeon_cs_parser *parser)
525{
526	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
527	struct radeon_vm *vm = &fpriv->vm;
528	int r;
529
530	if (parser->chunk_ib == NULL)
531		return 0;
532	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
533		return 0;
534
535	if (parser->const_ib.length_dw) {
536		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
537		if (r) {
538			return r;
539		}
540	}
541
542	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
543	if (r) {
544		return r;
545	}
546
547	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
548		radeon_uvd_note_usage(rdev);
549
550	mutex_lock(&vm->mutex);
551	r = radeon_bo_vm_update_pte(parser, vm);
552	if (r) {
553		goto out;
554	}
555
556	r = radeon_cs_sync_rings(parser);
557	if (r) {
558		if (r != -ERESTARTSYS)
559			DRM_ERROR("Failed to sync rings: %i\n", r);
560		goto out;
561	}
562
563	if ((rdev->family >= CHIP_TAHITI) &&
564	    (parser->chunk_const_ib != NULL)) {
565		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
566	} else {
567		r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
568	}
569
570out:
571	mutex_unlock(&vm->mutex);
572	return r;
573}
574
575static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
576{
577	if (r == -EDEADLK) {
578		r = radeon_gpu_reset(rdev);
579		if (!r)
580			r = -EAGAIN;
581	}
582	return r;
583}
584
585static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
586{
587	struct radeon_cs_chunk *ib_chunk;
588	struct radeon_vm *vm = NULL;
589	int r;
590
591	if (parser->chunk_ib == NULL)
592		return 0;
593
594	if (parser->cs_flags & RADEON_CS_USE_VM) {
595		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
596		vm = &fpriv->vm;
597
598		if ((rdev->family >= CHIP_TAHITI) &&
599		    (parser->chunk_const_ib != NULL)) {
600			ib_chunk = parser->chunk_const_ib;
601			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
602				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
603				return -EINVAL;
604			}
605			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
606					   vm, ib_chunk->length_dw * 4);
607			if (r) {
608				DRM_ERROR("Failed to get const ib !\n");
609				return r;
610			}
611			parser->const_ib.is_const_ib = true;
612			parser->const_ib.length_dw = ib_chunk->length_dw;
613			if (copy_from_user(parser->const_ib.ptr,
614					       ib_chunk->user_ptr,
615					       ib_chunk->length_dw * 4))
616				return -EFAULT;
617		}
618
619		ib_chunk = parser->chunk_ib;
620		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
621			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
622			return -EINVAL;
623		}
624	}
625	ib_chunk = parser->chunk_ib;
626
627	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
628			   vm, ib_chunk->length_dw * 4);
629	if (r) {
630		DRM_ERROR("Failed to get ib !\n");
631		return r;
632	}
633	parser->ib.length_dw = ib_chunk->length_dw;
634	if (ib_chunk->kdata)
635		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
636	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
637		return -EFAULT;
638	return 0;
639}
640
641int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
642{
643	struct radeon_device *rdev = dev->dev_private;
644	struct radeon_cs_parser parser;
645	int r;
646
647	down_read(&rdev->exclusive_lock);
648	if (!rdev->accel_working) {
649		up_read(&rdev->exclusive_lock);
650		return -EBUSY;
651	}
652	if (rdev->in_reset) {
653		up_read(&rdev->exclusive_lock);
654		r = radeon_gpu_reset(rdev);
655		if (!r)
656			r = -EAGAIN;
657		return r;
658	}
659	/* initialize parser */
660	memset(&parser, 0, sizeof(struct radeon_cs_parser));
661	parser.filp = filp;
662	parser.rdev = rdev;
663	parser.dev = rdev->dev;
664	parser.family = rdev->family;
665	r = radeon_cs_parser_init(&parser, data);
666	if (r) {
667		DRM_ERROR("Failed to initialize parser !\n");
668		radeon_cs_parser_fini(&parser, r, false);
669		up_read(&rdev->exclusive_lock);
670		r = radeon_cs_handle_lockup(rdev, r);
671		return r;
672	}
673
674	r = radeon_cs_ib_fill(rdev, &parser);
675	if (!r) {
676		r = radeon_cs_parser_relocs(&parser);
677		if (r && r != -ERESTARTSYS)
678			DRM_ERROR("Failed to parse relocation %d!\n", r);
679	}
680
681	if (r) {
682		radeon_cs_parser_fini(&parser, r, false);
683		up_read(&rdev->exclusive_lock);
684		r = radeon_cs_handle_lockup(rdev, r);
685		return r;
686	}
687
688	trace_radeon_cs(&parser);
689
690	r = radeon_cs_ib_chunk(rdev, &parser);
691	if (r) {
692		goto out;
693	}
694	r = radeon_cs_ib_vm_chunk(rdev, &parser);
695	if (r) {
696		goto out;
697	}
698out:
699	radeon_cs_parser_fini(&parser, r, true);
700	up_read(&rdev->exclusive_lock);
701	r = radeon_cs_handle_lockup(rdev, r);
702	return r;
703}
704
705/**
706 * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
707 * @parser:	parser structure holding parsing context.
708 * @pkt:	where to store packet information
709 *
710 * Assume that chunk_ib_index is properly set. Will return -EINVAL
711 * if packet is bigger than remaining ib size. or if packets is unknown.
712 **/
713int radeon_cs_packet_parse(struct radeon_cs_parser *p,
714			   struct radeon_cs_packet *pkt,
715			   unsigned idx)
716{
717	struct radeon_cs_chunk *ib_chunk = p->chunk_ib;
718	struct radeon_device *rdev = p->rdev;
719	uint32_t header;
720	int ret = 0, i;
721
722	if (idx >= ib_chunk->length_dw) {
723		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
724			  idx, ib_chunk->length_dw);
725		return -EINVAL;
726	}
727	header = radeon_get_ib_value(p, idx);
728	pkt->idx = idx;
729	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
730	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
731	pkt->one_reg_wr = 0;
732	switch (pkt->type) {
733	case RADEON_PACKET_TYPE0:
734		if (rdev->family < CHIP_R600) {
735			pkt->reg = R100_CP_PACKET0_GET_REG(header);
736			pkt->one_reg_wr =
737				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
738		} else
739			pkt->reg = R600_CP_PACKET0_GET_REG(header);
740		break;
741	case RADEON_PACKET_TYPE3:
742		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
743		break;
744	case RADEON_PACKET_TYPE2:
745		pkt->count = -1;
746		break;
747	default:
748		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
749		ret = -EINVAL;
750		goto dump_ib;
751	}
752	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
753		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
754			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
755		ret = -EINVAL;
756		goto dump_ib;
757	}
758	return 0;
759
760dump_ib:
761	for (i = 0; i < ib_chunk->length_dw; i++) {
762		if (i == idx)
763			printk("\t0x%08x <---\n", radeon_get_ib_value(p, i));
764		else
765			printk("\t0x%08x\n", radeon_get_ib_value(p, i));
766	}
767	return ret;
768}
769
770/**
771 * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
772 * @p:		structure holding the parser context.
773 *
774 * Check if the next packet is NOP relocation packet3.
775 **/
776bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
777{
778	struct radeon_cs_packet p3reloc;
779	int r;
780
781	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
782	if (r)
783		return false;
784	if (p3reloc.type != RADEON_PACKET_TYPE3)
785		return false;
786	if (p3reloc.opcode != RADEON_PACKET3_NOP)
787		return false;
788	return true;
789}
790
791/**
792 * radeon_cs_dump_packet() - dump raw packet context
793 * @p:		structure holding the parser context.
794 * @pkt:	structure holding the packet.
795 *
796 * Used mostly for debugging and error reporting.
797 **/
798void radeon_cs_dump_packet(struct radeon_cs_parser *p,
799			   struct radeon_cs_packet *pkt)
800{
801	volatile uint32_t *ib;
802	unsigned i;
803	unsigned idx;
804
805	ib = p->ib.ptr;
806	idx = pkt->idx;
807	for (i = 0; i <= (pkt->count + 1); i++, idx++)
808		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
809}
810
811/**
812 * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
813 * @parser:		parser structure holding parsing context.
814 * @data:		pointer to relocation data
815 * @offset_start:	starting offset
816 * @offset_mask:	offset mask (to align start offset on)
817 * @reloc:		reloc informations
818 *
819 * Check if next packet is relocation packet3, do bo validation and compute
820 * GPU offset using the provided start.
821 **/
822int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
823				struct radeon_bo_list **cs_reloc,
824				int nomm)
825{
826	struct radeon_cs_chunk *relocs_chunk;
827	struct radeon_cs_packet p3reloc;
828	unsigned idx;
829	int r;
830
831	if (p->chunk_relocs == NULL) {
832		DRM_ERROR("No relocation chunk !\n");
833		return -EINVAL;
834	}
835	*cs_reloc = NULL;
836	relocs_chunk = p->chunk_relocs;
837	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
838	if (r)
839		return r;
840	p->idx += p3reloc.count + 2;
841	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
842	    p3reloc.opcode != RADEON_PACKET3_NOP) {
843		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
844			  p3reloc.idx);
845		radeon_cs_dump_packet(p, &p3reloc);
846		return -EINVAL;
847	}
848	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
849	if (idx >= relocs_chunk->length_dw) {
850		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
851			  idx, relocs_chunk->length_dw);
852		radeon_cs_dump_packet(p, &p3reloc);
853		return -EINVAL;
854	}
855	/* FIXME: we assume reloc size is 4 dwords */
856	if (nomm) {
857		*cs_reloc = p->relocs;
858		(*cs_reloc)->gpu_offset =
859			(u64)relocs_chunk->kdata[idx + 3] << 32;
860		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
861	} else
862		*cs_reloc = &p->relocs[(idx / 4)];
863	return 0;
864}
v5.4
  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/uaccess.h>
 30
 31#include <drm/drm_device.h>
 32#include <drm/drm_file.h>
 33#include <drm/drm_pci.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 */
 47struct radeon_cs_buckets {
 48	struct list_head bucket[RADEON_CS_NUM_BUCKETS];
 49};
 50
 51static 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
 59static 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
 70static 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
 81static 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 = kvmalloc_array(p->nrelocs, sizeof(struct radeon_bo_list),
 97			GFP_KERNEL | __GFP_ZERO);
 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->ddev->pdev,
134						   PCI_CAP_ID_AGP) ||
135		     p->rdev->family == CHIP_RS780 ||
136		     p->rdev->family == CHIP_RS880)) {
137
138			/* TODO: is this still needed for NI+ ? */
139			p->relocs[i].preferred_domains =
140				RADEON_GEM_DOMAIN_VRAM;
141
142			p->relocs[i].allowed_domains =
143				RADEON_GEM_DOMAIN_VRAM;
144
145			/* prioritize this over any other relocation */
146			priority = RADEON_CS_MAX_PRIORITY;
147		} else {
148			uint32_t domain = r->write_domain ?
149				r->write_domain : r->read_domains;
150
151			if (domain & RADEON_GEM_DOMAIN_CPU) {
152				DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
153					  "for command submission\n");
154				return -EINVAL;
155			}
156
157			p->relocs[i].preferred_domains = domain;
158			if (domain == RADEON_GEM_DOMAIN_VRAM)
159				domain |= RADEON_GEM_DOMAIN_GTT;
160			p->relocs[i].allowed_domains = domain;
161		}
162
163		if (radeon_ttm_tt_has_userptr(p->relocs[i].robj->tbo.ttm)) {
164			uint32_t domain = p->relocs[i].preferred_domains;
165			if (!(domain & RADEON_GEM_DOMAIN_GTT)) {
166				DRM_ERROR("Only RADEON_GEM_DOMAIN_GTT is "
167					  "allowed for userptr BOs\n");
168				return -EINVAL;
169			}
170			need_mmap_lock = true;
171			domain = RADEON_GEM_DOMAIN_GTT;
172			p->relocs[i].preferred_domains = domain;
173			p->relocs[i].allowed_domains = domain;
174		}
175
176		/* Objects shared as dma-bufs cannot be moved to VRAM */
177		if (p->relocs[i].robj->prime_shared_count) {
178			p->relocs[i].allowed_domains &= ~RADEON_GEM_DOMAIN_VRAM;
179			if (!p->relocs[i].allowed_domains) {
180				DRM_ERROR("BO associated with dma-buf cannot "
181					  "be moved to VRAM\n");
182				return -EINVAL;
183			}
184		}
185
186		p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
187		p->relocs[i].tv.num_shared = !r->write_domain;
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
209static 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
252static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
253{
254	struct radeon_bo_list *reloc;
255	int r;
256
257	list_for_each_entry(reloc, &p->validated, tv.head) {
258		struct dma_resv *resv;
259
260		resv = reloc->robj->tbo.base.resv;
261		r = radeon_sync_resv(p->rdev, &p->ib.sync, resv,
262				     reloc->tv.num_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 */
270int 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	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 = 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		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
397static int cmp_size_smaller_first(void *priv, struct list_head *a,
398				  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.num_pages - (int)lb->robj->tbo.num_pages;
405}
406
407/**
408 * cs_parser_fini() - clean parser states
409 * @parser:	parser structure holding parsing context.
410 * @error:	error number
411 *
412 * If error is set than unvalidate buffer, otherwise just free memory
413 * used by parsing context.
414 **/
415static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
416{
417	unsigned i;
418
419	if (!error) {
420		/* Sort the buffer list from the smallest to largest buffer,
421		 * which affects the order of buffers in the LRU list.
422		 * This assures that the smallest buffers are added first
423		 * to the LRU list, so they are likely to be later evicted
424		 * first, instead of large buffers whose eviction is more
425		 * expensive.
426		 *
427		 * This slightly lowers the number of bytes moved by TTM
428		 * per frame under memory pressure.
429		 */
430		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
431
432		ttm_eu_fence_buffer_objects(&parser->ticket,
433					    &parser->validated,
434					    &parser->ib.fence->base);
435	} else if (backoff) {
436		ttm_eu_backoff_reservation(&parser->ticket,
437					   &parser->validated);
438	}
439
440	if (parser->relocs != NULL) {
441		for (i = 0; i < parser->nrelocs; i++) {
442			struct radeon_bo *bo = parser->relocs[i].robj;
443			if (bo == NULL)
444				continue;
445
446			drm_gem_object_put_unlocked(&bo->tbo.base);
447		}
448	}
449	kfree(parser->track);
450	kvfree(parser->relocs);
451	kvfree(parser->vm_bos);
452	for (i = 0; i < parser->nchunks; i++)
453		kvfree(parser->chunks[i].kdata);
454	kfree(parser->chunks);
455	kfree(parser->chunks_array);
456	radeon_ib_free(parser->rdev, &parser->ib);
457	radeon_ib_free(parser->rdev, &parser->const_ib);
458}
459
460static int radeon_cs_ib_chunk(struct radeon_device *rdev,
461			      struct radeon_cs_parser *parser)
462{
463	int r;
464
465	if (parser->chunk_ib == NULL)
466		return 0;
467
468	if (parser->cs_flags & RADEON_CS_USE_VM)
469		return 0;
470
471	r = radeon_cs_parse(rdev, parser->ring, parser);
472	if (r || parser->parser_error) {
473		DRM_ERROR("Invalid command stream !\n");
474		return r;
475	}
476
477	r = radeon_cs_sync_rings(parser);
478	if (r) {
479		if (r != -ERESTARTSYS)
480			DRM_ERROR("Failed to sync rings: %i\n", r);
481		return r;
482	}
483
484	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
485		radeon_uvd_note_usage(rdev);
486	else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
487		 (parser->ring == TN_RING_TYPE_VCE2_INDEX))
488		radeon_vce_note_usage(rdev);
489
490	r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
491	if (r) {
492		DRM_ERROR("Failed to schedule IB !\n");
493	}
494	return r;
495}
496
497static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
498				   struct radeon_vm *vm)
499{
500	struct radeon_device *rdev = p->rdev;
501	struct radeon_bo_va *bo_va;
502	int i, r;
503
504	r = radeon_vm_update_page_directory(rdev, vm);
505	if (r)
506		return r;
507
508	r = radeon_vm_clear_freed(rdev, vm);
509	if (r)
510		return r;
511
512	if (vm->ib_bo_va == NULL) {
513		DRM_ERROR("Tmp BO not in VM!\n");
514		return -EINVAL;
515	}
516
517	r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
518				&rdev->ring_tmp_bo.bo->tbo.mem);
519	if (r)
520		return r;
521
522	for (i = 0; i < p->nrelocs; i++) {
523		struct radeon_bo *bo;
524
525		bo = p->relocs[i].robj;
526		bo_va = radeon_vm_bo_find(vm, bo);
527		if (bo_va == NULL) {
528			dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
529			return -EINVAL;
530		}
531
532		r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
533		if (r)
534			return r;
535
536		radeon_sync_fence(&p->ib.sync, bo_va->last_pt_update);
537	}
538
539	return radeon_vm_clear_invalids(rdev, vm);
540}
541
542static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
543				 struct radeon_cs_parser *parser)
544{
545	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
546	struct radeon_vm *vm = &fpriv->vm;
547	int r;
548
549	if (parser->chunk_ib == NULL)
550		return 0;
551	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
552		return 0;
553
554	if (parser->const_ib.length_dw) {
555		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
556		if (r) {
557			return r;
558		}
559	}
560
561	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
562	if (r) {
563		return r;
564	}
565
566	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
567		radeon_uvd_note_usage(rdev);
568
569	mutex_lock(&vm->mutex);
570	r = radeon_bo_vm_update_pte(parser, vm);
571	if (r) {
572		goto out;
573	}
574
575	r = radeon_cs_sync_rings(parser);
576	if (r) {
577		if (r != -ERESTARTSYS)
578			DRM_ERROR("Failed to sync rings: %i\n", r);
579		goto out;
580	}
581
582	if ((rdev->family >= CHIP_TAHITI) &&
583	    (parser->chunk_const_ib != NULL)) {
584		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
585	} else {
586		r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
587	}
588
589out:
590	mutex_unlock(&vm->mutex);
591	return r;
592}
593
594static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
595{
596	if (r == -EDEADLK) {
597		r = radeon_gpu_reset(rdev);
598		if (!r)
599			r = -EAGAIN;
600	}
601	return r;
602}
603
604static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
605{
606	struct radeon_cs_chunk *ib_chunk;
607	struct radeon_vm *vm = NULL;
608	int r;
609
610	if (parser->chunk_ib == NULL)
611		return 0;
612
613	if (parser->cs_flags & RADEON_CS_USE_VM) {
614		struct radeon_fpriv *fpriv = parser->filp->driver_priv;
615		vm = &fpriv->vm;
616
617		if ((rdev->family >= CHIP_TAHITI) &&
618		    (parser->chunk_const_ib != NULL)) {
619			ib_chunk = parser->chunk_const_ib;
620			if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
621				DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
622				return -EINVAL;
623			}
624			r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
625					   vm, ib_chunk->length_dw * 4);
626			if (r) {
627				DRM_ERROR("Failed to get const ib !\n");
628				return r;
629			}
630			parser->const_ib.is_const_ib = true;
631			parser->const_ib.length_dw = ib_chunk->length_dw;
632			if (copy_from_user(parser->const_ib.ptr,
633					       ib_chunk->user_ptr,
634					       ib_chunk->length_dw * 4))
635				return -EFAULT;
636		}
637
638		ib_chunk = parser->chunk_ib;
639		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
640			DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
641			return -EINVAL;
642		}
643	}
644	ib_chunk = parser->chunk_ib;
645
646	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
647			   vm, ib_chunk->length_dw * 4);
648	if (r) {
649		DRM_ERROR("Failed to get ib !\n");
650		return r;
651	}
652	parser->ib.length_dw = ib_chunk->length_dw;
653	if (ib_chunk->kdata)
654		memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
655	else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
656		return -EFAULT;
657	return 0;
658}
659
660int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
661{
662	struct radeon_device *rdev = dev->dev_private;
663	struct radeon_cs_parser parser;
664	int r;
665
666	down_read(&rdev->exclusive_lock);
667	if (!rdev->accel_working) {
668		up_read(&rdev->exclusive_lock);
669		return -EBUSY;
670	}
671	if (rdev->in_reset) {
672		up_read(&rdev->exclusive_lock);
673		r = radeon_gpu_reset(rdev);
674		if (!r)
675			r = -EAGAIN;
676		return r;
677	}
678	/* initialize parser */
679	memset(&parser, 0, sizeof(struct radeon_cs_parser));
680	parser.filp = filp;
681	parser.rdev = rdev;
682	parser.dev = rdev->dev;
683	parser.family = rdev->family;
684	r = radeon_cs_parser_init(&parser, data);
685	if (r) {
686		DRM_ERROR("Failed to initialize parser !\n");
687		radeon_cs_parser_fini(&parser, r, false);
688		up_read(&rdev->exclusive_lock);
689		r = radeon_cs_handle_lockup(rdev, r);
690		return r;
691	}
692
693	r = radeon_cs_ib_fill(rdev, &parser);
694	if (!r) {
695		r = radeon_cs_parser_relocs(&parser);
696		if (r && r != -ERESTARTSYS)
697			DRM_ERROR("Failed to parse relocation %d!\n", r);
698	}
699
700	if (r) {
701		radeon_cs_parser_fini(&parser, r, false);
702		up_read(&rdev->exclusive_lock);
703		r = radeon_cs_handle_lockup(rdev, r);
704		return r;
705	}
706
707	trace_radeon_cs(&parser);
708
709	r = radeon_cs_ib_chunk(rdev, &parser);
710	if (r) {
711		goto out;
712	}
713	r = radeon_cs_ib_vm_chunk(rdev, &parser);
714	if (r) {
715		goto out;
716	}
717out:
718	radeon_cs_parser_fini(&parser, r, true);
719	up_read(&rdev->exclusive_lock);
720	r = radeon_cs_handle_lockup(rdev, r);
721	return r;
722}
723
724/**
725 * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
726 * @parser:	parser structure holding parsing context.
727 * @pkt:	where to store packet information
728 *
729 * Assume that chunk_ib_index is properly set. Will return -EINVAL
730 * if packet is bigger than remaining ib size. or if packets is unknown.
731 **/
732int radeon_cs_packet_parse(struct radeon_cs_parser *p,
733			   struct radeon_cs_packet *pkt,
734			   unsigned idx)
735{
736	struct radeon_cs_chunk *ib_chunk = p->chunk_ib;
737	struct radeon_device *rdev = p->rdev;
738	uint32_t header;
739	int ret = 0, i;
740
741	if (idx >= ib_chunk->length_dw) {
742		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
743			  idx, ib_chunk->length_dw);
744		return -EINVAL;
745	}
746	header = radeon_get_ib_value(p, idx);
747	pkt->idx = idx;
748	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
749	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
750	pkt->one_reg_wr = 0;
751	switch (pkt->type) {
752	case RADEON_PACKET_TYPE0:
753		if (rdev->family < CHIP_R600) {
754			pkt->reg = R100_CP_PACKET0_GET_REG(header);
755			pkt->one_reg_wr =
756				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
757		} else
758			pkt->reg = R600_CP_PACKET0_GET_REG(header);
759		break;
760	case RADEON_PACKET_TYPE3:
761		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
762		break;
763	case RADEON_PACKET_TYPE2:
764		pkt->count = -1;
765		break;
766	default:
767		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
768		ret = -EINVAL;
769		goto dump_ib;
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		ret = -EINVAL;
775		goto dump_ib;
776	}
777	return 0;
778
779dump_ib:
780	for (i = 0; i < ib_chunk->length_dw; i++) {
781		if (i == idx)
782			printk("\t0x%08x <---\n", radeon_get_ib_value(p, i));
783		else
784			printk("\t0x%08x\n", radeon_get_ib_value(p, i));
785	}
786	return ret;
787}
788
789/**
790 * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
791 * @p:		structure holding the parser context.
792 *
793 * Check if the next packet is NOP relocation packet3.
794 **/
795bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
796{
797	struct radeon_cs_packet p3reloc;
798	int r;
799
800	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
801	if (r)
802		return false;
803	if (p3reloc.type != RADEON_PACKET_TYPE3)
804		return false;
805	if (p3reloc.opcode != RADEON_PACKET3_NOP)
806		return false;
807	return true;
808}
809
810/**
811 * radeon_cs_dump_packet() - dump raw packet context
812 * @p:		structure holding the parser context.
813 * @pkt:	structure holding the packet.
814 *
815 * Used mostly for debugging and error reporting.
816 **/
817void radeon_cs_dump_packet(struct radeon_cs_parser *p,
818			   struct radeon_cs_packet *pkt)
819{
820	volatile uint32_t *ib;
821	unsigned i;
822	unsigned idx;
823
824	ib = p->ib.ptr;
825	idx = pkt->idx;
826	for (i = 0; i <= (pkt->count + 1); i++, idx++)
827		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
828}
829
830/**
831 * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
832 * @parser:		parser structure holding parsing context.
833 * @data:		pointer to relocation data
834 * @offset_start:	starting offset
835 * @offset_mask:	offset mask (to align start offset on)
836 * @reloc:		reloc informations
837 *
838 * Check if next packet is relocation packet3, do bo validation and compute
839 * GPU offset using the provided start.
840 **/
841int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
842				struct radeon_bo_list **cs_reloc,
843				int nomm)
844{
845	struct radeon_cs_chunk *relocs_chunk;
846	struct radeon_cs_packet p3reloc;
847	unsigned idx;
848	int r;
849
850	if (p->chunk_relocs == NULL) {
851		DRM_ERROR("No relocation chunk !\n");
852		return -EINVAL;
853	}
854	*cs_reloc = NULL;
855	relocs_chunk = p->chunk_relocs;
856	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
857	if (r)
858		return r;
859	p->idx += p3reloc.count + 2;
860	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
861	    p3reloc.opcode != RADEON_PACKET3_NOP) {
862		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
863			  p3reloc.idx);
864		radeon_cs_dump_packet(p, &p3reloc);
865		return -EINVAL;
866	}
867	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
868	if (idx >= relocs_chunk->length_dw) {
869		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
870			  idx, relocs_chunk->length_dw);
871		radeon_cs_dump_packet(p, &p3reloc);
872		return -EINVAL;
873	}
874	/* FIXME: we assume reloc size is 4 dwords */
875	if (nomm) {
876		*cs_reloc = p->relocs;
877		(*cs_reloc)->gpu_offset =
878			(u64)relocs_chunk->kdata[idx + 3] << 32;
879		(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
880	} else
881		*cs_reloc = &p->relocs[(idx / 4)];
882	return 0;
883}