xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision 089a49b6)
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 <drm/drmP.h>
28 #include <drm/radeon_drm.h>
29 #include "radeon_reg.h"
30 #include "radeon.h"
31 #include "radeon_trace.h"
32 
33 static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
34 {
35 	struct drm_device *ddev = p->rdev->ddev;
36 	struct radeon_cs_chunk *chunk;
37 	unsigned i, j;
38 	bool duplicate;
39 
40 	if (p->chunk_relocs_idx == -1) {
41 		return 0;
42 	}
43 	chunk = &p->chunks[p->chunk_relocs_idx];
44 	p->dma_reloc_idx = 0;
45 	/* FIXME: we assume that each relocs use 4 dwords */
46 	p->nrelocs = chunk->length_dw / 4;
47 	p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
48 	if (p->relocs_ptr == NULL) {
49 		return -ENOMEM;
50 	}
51 	p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
52 	if (p->relocs == NULL) {
53 		return -ENOMEM;
54 	}
55 	for (i = 0; i < p->nrelocs; i++) {
56 		struct drm_radeon_cs_reloc *r;
57 
58 		duplicate = false;
59 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
60 		for (j = 0; j < i; j++) {
61 			if (r->handle == p->relocs[j].handle) {
62 				p->relocs_ptr[i] = &p->relocs[j];
63 				duplicate = true;
64 				break;
65 			}
66 		}
67 		if (duplicate) {
68 			p->relocs[i].handle = 0;
69 			continue;
70 		}
71 
72 		p->relocs[i].gobj = drm_gem_object_lookup(ddev, p->filp,
73 							  r->handle);
74 		if (p->relocs[i].gobj == NULL) {
75 			DRM_ERROR("gem object lookup failed 0x%x\n",
76 				  r->handle);
77 			return -ENOENT;
78 		}
79 		p->relocs_ptr[i] = &p->relocs[i];
80 		p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
81 		p->relocs[i].lobj.bo = p->relocs[i].robj;
82 		p->relocs[i].lobj.written = !!r->write_domain;
83 
84 		/* the first reloc of an UVD job is the msg and that must be in
85 		   VRAM, also but everything into VRAM on AGP cards to avoid
86 		   image corruptions */
87 		if (p->ring == R600_RING_TYPE_UVD_INDEX &&
88 		    p->rdev->family < CHIP_PALM &&
89 		    (i == 0 || drm_pci_device_is_agp(p->rdev->ddev))) {
90 
91 			p->relocs[i].lobj.domain =
92 				RADEON_GEM_DOMAIN_VRAM;
93 
94 			p->relocs[i].lobj.alt_domain =
95 				RADEON_GEM_DOMAIN_VRAM;
96 
97 		} else {
98 			uint32_t domain = r->write_domain ?
99 				r->write_domain : r->read_domains;
100 
101 			p->relocs[i].lobj.domain = domain;
102 			if (domain == RADEON_GEM_DOMAIN_VRAM)
103 				domain |= RADEON_GEM_DOMAIN_GTT;
104 			p->relocs[i].lobj.alt_domain = domain;
105 		}
106 
107 		p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
108 		p->relocs[i].handle = r->handle;
109 
110 		radeon_bo_list_add_object(&p->relocs[i].lobj,
111 					  &p->validated);
112 	}
113 	return radeon_bo_list_validate(&p->ticket, &p->validated, p->ring);
114 }
115 
116 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
117 {
118 	p->priority = priority;
119 
120 	switch (ring) {
121 	default:
122 		DRM_ERROR("unknown ring id: %d\n", ring);
123 		return -EINVAL;
124 	case RADEON_CS_RING_GFX:
125 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
126 		break;
127 	case RADEON_CS_RING_COMPUTE:
128 		if (p->rdev->family >= CHIP_TAHITI) {
129 			if (p->priority > 0)
130 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
131 			else
132 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
133 		} else
134 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
135 		break;
136 	case RADEON_CS_RING_DMA:
137 		if (p->rdev->family >= CHIP_CAYMAN) {
138 			if (p->priority > 0)
139 				p->ring = R600_RING_TYPE_DMA_INDEX;
140 			else
141 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
142 		} else if (p->rdev->family >= CHIP_R600) {
143 			p->ring = R600_RING_TYPE_DMA_INDEX;
144 		} else {
145 			return -EINVAL;
146 		}
147 		break;
148 	case RADEON_CS_RING_UVD:
149 		p->ring = R600_RING_TYPE_UVD_INDEX;
150 		break;
151 	}
152 	return 0;
153 }
154 
155 static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
156 {
157 	int i;
158 
159 	for (i = 0; i < p->nrelocs; i++) {
160 		if (!p->relocs[i].robj)
161 			continue;
162 
163 		radeon_ib_sync_to(&p->ib, p->relocs[i].robj->tbo.sync_obj);
164 	}
165 }
166 
167 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
168 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
169 {
170 	struct drm_radeon_cs *cs = data;
171 	uint64_t *chunk_array_ptr;
172 	unsigned size, i;
173 	u32 ring = RADEON_CS_RING_GFX;
174 	s32 priority = 0;
175 
176 	if (!cs->num_chunks) {
177 		return 0;
178 	}
179 	/* get chunks */
180 	INIT_LIST_HEAD(&p->validated);
181 	p->idx = 0;
182 	p->ib.sa_bo = NULL;
183 	p->ib.semaphore = NULL;
184 	p->const_ib.sa_bo = NULL;
185 	p->const_ib.semaphore = NULL;
186 	p->chunk_ib_idx = -1;
187 	p->chunk_relocs_idx = -1;
188 	p->chunk_flags_idx = -1;
189 	p->chunk_const_ib_idx = -1;
190 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
191 	if (p->chunks_array == NULL) {
192 		return -ENOMEM;
193 	}
194 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
195 	if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
196 			       sizeof(uint64_t)*cs->num_chunks)) {
197 		return -EFAULT;
198 	}
199 	p->cs_flags = 0;
200 	p->nchunks = cs->num_chunks;
201 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
202 	if (p->chunks == NULL) {
203 		return -ENOMEM;
204 	}
205 	for (i = 0; i < p->nchunks; i++) {
206 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
207 		struct drm_radeon_cs_chunk user_chunk;
208 		uint32_t __user *cdata;
209 
210 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
211 		if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
212 				       sizeof(struct drm_radeon_cs_chunk))) {
213 			return -EFAULT;
214 		}
215 		p->chunks[i].length_dw = user_chunk.length_dw;
216 		p->chunks[i].kdata = NULL;
217 		p->chunks[i].chunk_id = user_chunk.chunk_id;
218 		p->chunks[i].user_ptr = (void __user *)(unsigned long)user_chunk.chunk_data;
219 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
220 			p->chunk_relocs_idx = i;
221 		}
222 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
223 			p->chunk_ib_idx = i;
224 			/* zero length IB isn't useful */
225 			if (p->chunks[i].length_dw == 0)
226 				return -EINVAL;
227 		}
228 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
229 			p->chunk_const_ib_idx = i;
230 			/* zero length CONST IB isn't useful */
231 			if (p->chunks[i].length_dw == 0)
232 				return -EINVAL;
233 		}
234 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
235 			p->chunk_flags_idx = i;
236 			/* zero length flags aren't useful */
237 			if (p->chunks[i].length_dw == 0)
238 				return -EINVAL;
239 		}
240 
241 		cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
242 		if ((p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) ||
243 		    (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS)) {
244 			size = p->chunks[i].length_dw * sizeof(uint32_t);
245 			p->chunks[i].kdata = kmalloc(size, GFP_KERNEL);
246 			if (p->chunks[i].kdata == NULL) {
247 				return -ENOMEM;
248 			}
249 			if (DRM_COPY_FROM_USER(p->chunks[i].kdata,
250 					       p->chunks[i].user_ptr, size)) {
251 				return -EFAULT;
252 			}
253 			if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
254 				p->cs_flags = p->chunks[i].kdata[0];
255 				if (p->chunks[i].length_dw > 1)
256 					ring = p->chunks[i].kdata[1];
257 				if (p->chunks[i].length_dw > 2)
258 					priority = (s32)p->chunks[i].kdata[2];
259 			}
260 		}
261 	}
262 
263 	/* these are KMS only */
264 	if (p->rdev) {
265 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
266 		    !p->rdev->vm_manager.enabled) {
267 			DRM_ERROR("VM not active on asic!\n");
268 			return -EINVAL;
269 		}
270 
271 		if (radeon_cs_get_ring(p, ring, priority))
272 			return -EINVAL;
273 
274 		/* we only support VM on some SI+ rings */
275 		if ((p->rdev->asic->ring[p->ring]->cs_parse == NULL) &&
276 		   ((p->cs_flags & RADEON_CS_USE_VM) == 0)) {
277 			DRM_ERROR("Ring %d requires VM!\n", p->ring);
278 			return -EINVAL;
279 		}
280 	}
281 
282 	/* deal with non-vm */
283 	if ((p->chunk_ib_idx != -1) &&
284 	    ((p->cs_flags & RADEON_CS_USE_VM) == 0) &&
285 	    (p->chunks[p->chunk_ib_idx].chunk_id == RADEON_CHUNK_ID_IB)) {
286 		if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
287 			DRM_ERROR("cs IB too big: %d\n",
288 				  p->chunks[p->chunk_ib_idx].length_dw);
289 			return -EINVAL;
290 		}
291 		if (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) {
292 			p->chunks[p->chunk_ib_idx].kpage[0] = kmalloc(PAGE_SIZE, GFP_KERNEL);
293 			p->chunks[p->chunk_ib_idx].kpage[1] = kmalloc(PAGE_SIZE, GFP_KERNEL);
294 			if (p->chunks[p->chunk_ib_idx].kpage[0] == NULL ||
295 			    p->chunks[p->chunk_ib_idx].kpage[1] == NULL) {
296 				kfree(p->chunks[p->chunk_ib_idx].kpage[0]);
297 				kfree(p->chunks[p->chunk_ib_idx].kpage[1]);
298 				p->chunks[p->chunk_ib_idx].kpage[0] = NULL;
299 				p->chunks[p->chunk_ib_idx].kpage[1] = NULL;
300 				return -ENOMEM;
301 			}
302 		}
303 		p->chunks[p->chunk_ib_idx].kpage_idx[0] = -1;
304 		p->chunks[p->chunk_ib_idx].kpage_idx[1] = -1;
305 		p->chunks[p->chunk_ib_idx].last_copied_page = -1;
306 		p->chunks[p->chunk_ib_idx].last_page_index =
307 			((p->chunks[p->chunk_ib_idx].length_dw * 4) - 1) / PAGE_SIZE;
308 	}
309 
310 	return 0;
311 }
312 
313 /**
314  * cs_parser_fini() - clean parser states
315  * @parser:	parser structure holding parsing context.
316  * @error:	error number
317  *
318  * If error is set than unvalidate buffer, otherwise just free memory
319  * used by parsing context.
320  **/
321 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
322 {
323 	unsigned i;
324 
325 	if (!error) {
326 		ttm_eu_fence_buffer_objects(&parser->ticket,
327 					    &parser->validated,
328 					    parser->ib.fence);
329 	} else if (backoff) {
330 		ttm_eu_backoff_reservation(&parser->ticket,
331 					   &parser->validated);
332 	}
333 
334 	if (parser->relocs != NULL) {
335 		for (i = 0; i < parser->nrelocs; i++) {
336 			if (parser->relocs[i].gobj)
337 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
338 		}
339 	}
340 	kfree(parser->track);
341 	kfree(parser->relocs);
342 	kfree(parser->relocs_ptr);
343 	for (i = 0; i < parser->nchunks; i++) {
344 		kfree(parser->chunks[i].kdata);
345 		if ((parser->rdev->flags & RADEON_IS_AGP)) {
346 			kfree(parser->chunks[i].kpage[0]);
347 			kfree(parser->chunks[i].kpage[1]);
348 		}
349 	}
350 	kfree(parser->chunks);
351 	kfree(parser->chunks_array);
352 	radeon_ib_free(parser->rdev, &parser->ib);
353 	radeon_ib_free(parser->rdev, &parser->const_ib);
354 }
355 
356 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
357 			      struct radeon_cs_parser *parser)
358 {
359 	struct radeon_cs_chunk *ib_chunk;
360 	int r;
361 
362 	if (parser->chunk_ib_idx == -1)
363 		return 0;
364 
365 	if (parser->cs_flags & RADEON_CS_USE_VM)
366 		return 0;
367 
368 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
369 	/* Copy the packet into the IB, the parser will read from the
370 	 * input memory (cached) and write to the IB (which can be
371 	 * uncached).
372 	 */
373 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
374 			   NULL, ib_chunk->length_dw * 4);
375 	if (r) {
376 		DRM_ERROR("Failed to get ib !\n");
377 		return r;
378 	}
379 	parser->ib.length_dw = ib_chunk->length_dw;
380 	r = radeon_cs_parse(rdev, parser->ring, parser);
381 	if (r || parser->parser_error) {
382 		DRM_ERROR("Invalid command stream !\n");
383 		return r;
384 	}
385 	r = radeon_cs_finish_pages(parser);
386 	if (r) {
387 		DRM_ERROR("Invalid command stream !\n");
388 		return r;
389 	}
390 
391 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
392 		radeon_uvd_note_usage(rdev);
393 
394 	radeon_cs_sync_rings(parser);
395 	r = radeon_ib_schedule(rdev, &parser->ib, NULL);
396 	if (r) {
397 		DRM_ERROR("Failed to schedule IB !\n");
398 	}
399 	return r;
400 }
401 
402 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
403 				   struct radeon_vm *vm)
404 {
405 	struct radeon_device *rdev = parser->rdev;
406 	struct radeon_bo_list *lobj;
407 	struct radeon_bo *bo;
408 	int r;
409 
410 	r = radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo, &rdev->ring_tmp_bo.bo->tbo.mem);
411 	if (r) {
412 		return r;
413 	}
414 	list_for_each_entry(lobj, &parser->validated, tv.head) {
415 		bo = lobj->bo;
416 		r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
417 		if (r) {
418 			return r;
419 		}
420 	}
421 	return 0;
422 }
423 
424 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
425 				 struct radeon_cs_parser *parser)
426 {
427 	struct radeon_cs_chunk *ib_chunk;
428 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
429 	struct radeon_vm *vm = &fpriv->vm;
430 	int r;
431 
432 	if (parser->chunk_ib_idx == -1)
433 		return 0;
434 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
435 		return 0;
436 
437 	if ((rdev->family >= CHIP_TAHITI) &&
438 	    (parser->chunk_const_ib_idx != -1)) {
439 		ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
440 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
441 			DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
442 			return -EINVAL;
443 		}
444 		r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
445 				   vm, ib_chunk->length_dw * 4);
446 		if (r) {
447 			DRM_ERROR("Failed to get const ib !\n");
448 			return r;
449 		}
450 		parser->const_ib.is_const_ib = true;
451 		parser->const_ib.length_dw = ib_chunk->length_dw;
452 		/* Copy the packet into the IB */
453 		if (DRM_COPY_FROM_USER(parser->const_ib.ptr, ib_chunk->user_ptr,
454 				       ib_chunk->length_dw * 4)) {
455 			return -EFAULT;
456 		}
457 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
458 		if (r) {
459 			return r;
460 		}
461 	}
462 
463 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
464 	if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
465 		DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
466 		return -EINVAL;
467 	}
468 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
469 			   vm, ib_chunk->length_dw * 4);
470 	if (r) {
471 		DRM_ERROR("Failed to get ib !\n");
472 		return r;
473 	}
474 	parser->ib.length_dw = ib_chunk->length_dw;
475 	/* Copy the packet into the IB */
476 	if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
477 			       ib_chunk->length_dw * 4)) {
478 		return -EFAULT;
479 	}
480 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
481 	if (r) {
482 		return r;
483 	}
484 
485 	if (parser->ring == R600_RING_TYPE_UVD_INDEX)
486 		radeon_uvd_note_usage(rdev);
487 
488 	mutex_lock(&rdev->vm_manager.lock);
489 	mutex_lock(&vm->mutex);
490 	r = radeon_vm_alloc_pt(rdev, vm);
491 	if (r) {
492 		goto out;
493 	}
494 	r = radeon_bo_vm_update_pte(parser, vm);
495 	if (r) {
496 		goto out;
497 	}
498 	radeon_cs_sync_rings(parser);
499 	radeon_ib_sync_to(&parser->ib, vm->fence);
500 	radeon_ib_sync_to(&parser->ib, radeon_vm_grab_id(
501 		rdev, vm, parser->ring));
502 
503 	if ((rdev->family >= CHIP_TAHITI) &&
504 	    (parser->chunk_const_ib_idx != -1)) {
505 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib);
506 	} else {
507 		r = radeon_ib_schedule(rdev, &parser->ib, NULL);
508 	}
509 
510 	if (!r) {
511 		radeon_vm_fence(rdev, vm, parser->ib.fence);
512 	}
513 
514 out:
515 	radeon_vm_add_to_lru(rdev, vm);
516 	mutex_unlock(&vm->mutex);
517 	mutex_unlock(&rdev->vm_manager.lock);
518 	return r;
519 }
520 
521 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
522 {
523 	if (r == -EDEADLK) {
524 		r = radeon_gpu_reset(rdev);
525 		if (!r)
526 			r = -EAGAIN;
527 	}
528 	return r;
529 }
530 
531 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
532 {
533 	struct radeon_device *rdev = dev->dev_private;
534 	struct radeon_cs_parser parser;
535 	int r;
536 
537 	down_read(&rdev->exclusive_lock);
538 	if (!rdev->accel_working) {
539 		up_read(&rdev->exclusive_lock);
540 		return -EBUSY;
541 	}
542 	/* initialize parser */
543 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
544 	parser.filp = filp;
545 	parser.rdev = rdev;
546 	parser.dev = rdev->dev;
547 	parser.family = rdev->family;
548 	r = radeon_cs_parser_init(&parser, data);
549 	if (r) {
550 		DRM_ERROR("Failed to initialize parser !\n");
551 		radeon_cs_parser_fini(&parser, r, false);
552 		up_read(&rdev->exclusive_lock);
553 		r = radeon_cs_handle_lockup(rdev, r);
554 		return r;
555 	}
556 	r = radeon_cs_parser_relocs(&parser);
557 	if (r) {
558 		if (r != -ERESTARTSYS)
559 			DRM_ERROR("Failed to parse relocation %d!\n", r);
560 		radeon_cs_parser_fini(&parser, r, false);
561 		up_read(&rdev->exclusive_lock);
562 		r = radeon_cs_handle_lockup(rdev, r);
563 		return r;
564 	}
565 
566 	trace_radeon_cs(&parser);
567 
568 	r = radeon_cs_ib_chunk(rdev, &parser);
569 	if (r) {
570 		goto out;
571 	}
572 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
573 	if (r) {
574 		goto out;
575 	}
576 out:
577 	radeon_cs_parser_fini(&parser, r, true);
578 	up_read(&rdev->exclusive_lock);
579 	r = radeon_cs_handle_lockup(rdev, r);
580 	return r;
581 }
582 
583 int radeon_cs_finish_pages(struct radeon_cs_parser *p)
584 {
585 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
586 	int i;
587 	int size = PAGE_SIZE;
588 
589 	for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
590 		if (i == ibc->last_page_index) {
591 			size = (ibc->length_dw * 4) % PAGE_SIZE;
592 			if (size == 0)
593 				size = PAGE_SIZE;
594 		}
595 
596 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
597 				       ibc->user_ptr + (i * PAGE_SIZE),
598 				       size))
599 			return -EFAULT;
600 	}
601 	return 0;
602 }
603 
604 static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
605 {
606 	int new_page;
607 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
608 	int i;
609 	int size = PAGE_SIZE;
610 	bool copy1 = (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) ?
611 		false : true;
612 
613 	for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
614 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
615 				       ibc->user_ptr + (i * PAGE_SIZE),
616 				       PAGE_SIZE)) {
617 			p->parser_error = -EFAULT;
618 			return 0;
619 		}
620 	}
621 
622 	if (pg_idx == ibc->last_page_index) {
623 		size = (ibc->length_dw * 4) % PAGE_SIZE;
624 		if (size == 0)
625 			size = PAGE_SIZE;
626 	}
627 
628 	new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
629 	if (copy1)
630 		ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
631 
632 	if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
633 			       ibc->user_ptr + (pg_idx * PAGE_SIZE),
634 			       size)) {
635 		p->parser_error = -EFAULT;
636 		return 0;
637 	}
638 
639 	/* copy to IB for non single case */
640 	if (!copy1)
641 		memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
642 
643 	ibc->last_copied_page = pg_idx;
644 	ibc->kpage_idx[new_page] = pg_idx;
645 
646 	return new_page;
647 }
648 
649 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
650 {
651 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
652 	u32 pg_idx, pg_offset;
653 	u32 idx_value = 0;
654 	int new_page;
655 
656 	pg_idx = (idx * 4) / PAGE_SIZE;
657 	pg_offset = (idx * 4) % PAGE_SIZE;
658 
659 	if (ibc->kpage_idx[0] == pg_idx)
660 		return ibc->kpage[0][pg_offset/4];
661 	if (ibc->kpage_idx[1] == pg_idx)
662 		return ibc->kpage[1][pg_offset/4];
663 
664 	new_page = radeon_cs_update_pages(p, pg_idx);
665 	if (new_page < 0) {
666 		p->parser_error = new_page;
667 		return 0;
668 	}
669 
670 	idx_value = ibc->kpage[new_page][pg_offset/4];
671 	return idx_value;
672 }
673 
674 /**
675  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
676  * @parser:	parser structure holding parsing context.
677  * @pkt:	where to store packet information
678  *
679  * Assume that chunk_ib_index is properly set. Will return -EINVAL
680  * if packet is bigger than remaining ib size. or if packets is unknown.
681  **/
682 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
683 			   struct radeon_cs_packet *pkt,
684 			   unsigned idx)
685 {
686 	struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
687 	struct radeon_device *rdev = p->rdev;
688 	uint32_t header;
689 
690 	if (idx >= ib_chunk->length_dw) {
691 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
692 			  idx, ib_chunk->length_dw);
693 		return -EINVAL;
694 	}
695 	header = radeon_get_ib_value(p, idx);
696 	pkt->idx = idx;
697 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
698 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
699 	pkt->one_reg_wr = 0;
700 	switch (pkt->type) {
701 	case RADEON_PACKET_TYPE0:
702 		if (rdev->family < CHIP_R600) {
703 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
704 			pkt->one_reg_wr =
705 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
706 		} else
707 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
708 		break;
709 	case RADEON_PACKET_TYPE3:
710 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
711 		break;
712 	case RADEON_PACKET_TYPE2:
713 		pkt->count = -1;
714 		break;
715 	default:
716 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
717 		return -EINVAL;
718 	}
719 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
720 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
721 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
722 		return -EINVAL;
723 	}
724 	return 0;
725 }
726 
727 /**
728  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
729  * @p:		structure holding the parser context.
730  *
731  * Check if the next packet is NOP relocation packet3.
732  **/
733 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
734 {
735 	struct radeon_cs_packet p3reloc;
736 	int r;
737 
738 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
739 	if (r)
740 		return false;
741 	if (p3reloc.type != RADEON_PACKET_TYPE3)
742 		return false;
743 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
744 		return false;
745 	return true;
746 }
747 
748 /**
749  * radeon_cs_dump_packet() - dump raw packet context
750  * @p:		structure holding the parser context.
751  * @pkt:	structure holding the packet.
752  *
753  * Used mostly for debugging and error reporting.
754  **/
755 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
756 			   struct radeon_cs_packet *pkt)
757 {
758 	volatile uint32_t *ib;
759 	unsigned i;
760 	unsigned idx;
761 
762 	ib = p->ib.ptr;
763 	idx = pkt->idx;
764 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
765 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
766 }
767 
768 /**
769  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
770  * @parser:		parser structure holding parsing context.
771  * @data:		pointer to relocation data
772  * @offset_start:	starting offset
773  * @offset_mask:	offset mask (to align start offset on)
774  * @reloc:		reloc informations
775  *
776  * Check if next packet is relocation packet3, do bo validation and compute
777  * GPU offset using the provided start.
778  **/
779 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
780 				struct radeon_cs_reloc **cs_reloc,
781 				int nomm)
782 {
783 	struct radeon_cs_chunk *relocs_chunk;
784 	struct radeon_cs_packet p3reloc;
785 	unsigned idx;
786 	int r;
787 
788 	if (p->chunk_relocs_idx == -1) {
789 		DRM_ERROR("No relocation chunk !\n");
790 		return -EINVAL;
791 	}
792 	*cs_reloc = NULL;
793 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
794 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
795 	if (r)
796 		return r;
797 	p->idx += p3reloc.count + 2;
798 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
799 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
800 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
801 			  p3reloc.idx);
802 		radeon_cs_dump_packet(p, &p3reloc);
803 		return -EINVAL;
804 	}
805 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
806 	if (idx >= relocs_chunk->length_dw) {
807 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
808 			  idx, relocs_chunk->length_dw);
809 		radeon_cs_dump_packet(p, &p3reloc);
810 		return -EINVAL;
811 	}
812 	/* FIXME: we assume reloc size is 4 dwords */
813 	if (nomm) {
814 		*cs_reloc = p->relocs;
815 		(*cs_reloc)->lobj.gpu_offset =
816 			(u64)relocs_chunk->kdata[idx + 3] << 32;
817 		(*cs_reloc)->lobj.gpu_offset |= relocs_chunk->kdata[idx + 0];
818 	} else
819 		*cs_reloc = p->relocs_ptr[(idx / 4)];
820 	return 0;
821 }
822