xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision 97da55fc)
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 
32 static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
33 {
34 	struct drm_device *ddev = p->rdev->ddev;
35 	struct radeon_cs_chunk *chunk;
36 	unsigned i, j;
37 	bool duplicate;
38 
39 	if (p->chunk_relocs_idx == -1) {
40 		return 0;
41 	}
42 	chunk = &p->chunks[p->chunk_relocs_idx];
43 	p->dma_reloc_idx = 0;
44 	/* FIXME: we assume that each relocs use 4 dwords */
45 	p->nrelocs = chunk->length_dw / 4;
46 	p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
47 	if (p->relocs_ptr == NULL) {
48 		return -ENOMEM;
49 	}
50 	p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
51 	if (p->relocs == NULL) {
52 		return -ENOMEM;
53 	}
54 	for (i = 0; i < p->nrelocs; i++) {
55 		struct drm_radeon_cs_reloc *r;
56 
57 		duplicate = false;
58 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
59 		for (j = 0; j < i; j++) {
60 			if (r->handle == p->relocs[j].handle) {
61 				p->relocs_ptr[i] = &p->relocs[j];
62 				duplicate = true;
63 				break;
64 			}
65 		}
66 		if (!duplicate) {
67 			p->relocs[i].gobj = drm_gem_object_lookup(ddev,
68 								  p->filp,
69 								  r->handle);
70 			if (p->relocs[i].gobj == NULL) {
71 				DRM_ERROR("gem object lookup failed 0x%x\n",
72 					  r->handle);
73 				return -ENOENT;
74 			}
75 			p->relocs_ptr[i] = &p->relocs[i];
76 			p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
77 			p->relocs[i].lobj.bo = p->relocs[i].robj;
78 			p->relocs[i].lobj.wdomain = r->write_domain;
79 			p->relocs[i].lobj.rdomain = r->read_domains;
80 			p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
81 			p->relocs[i].handle = r->handle;
82 			p->relocs[i].flags = r->flags;
83 			radeon_bo_list_add_object(&p->relocs[i].lobj,
84 						  &p->validated);
85 
86 		} else
87 			p->relocs[i].handle = 0;
88 	}
89 	return radeon_bo_list_validate(&p->validated);
90 }
91 
92 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
93 {
94 	p->priority = priority;
95 
96 	switch (ring) {
97 	default:
98 		DRM_ERROR("unknown ring id: %d\n", ring);
99 		return -EINVAL;
100 	case RADEON_CS_RING_GFX:
101 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
102 		break;
103 	case RADEON_CS_RING_COMPUTE:
104 		if (p->rdev->family >= CHIP_TAHITI) {
105 			if (p->priority > 0)
106 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
107 			else
108 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
109 		} else
110 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
111 		break;
112 	case RADEON_CS_RING_DMA:
113 		if (p->rdev->family >= CHIP_CAYMAN) {
114 			if (p->priority > 0)
115 				p->ring = R600_RING_TYPE_DMA_INDEX;
116 			else
117 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
118 		} else if (p->rdev->family >= CHIP_R600) {
119 			p->ring = R600_RING_TYPE_DMA_INDEX;
120 		} else {
121 			return -EINVAL;
122 		}
123 		break;
124 	}
125 	return 0;
126 }
127 
128 static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
129 {
130 	int i;
131 
132 	for (i = 0; i < p->nrelocs; i++) {
133 		if (!p->relocs[i].robj)
134 			continue;
135 
136 		radeon_ib_sync_to(&p->ib, p->relocs[i].robj->tbo.sync_obj);
137 	}
138 }
139 
140 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
141 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
142 {
143 	struct drm_radeon_cs *cs = data;
144 	uint64_t *chunk_array_ptr;
145 	unsigned size, i;
146 	u32 ring = RADEON_CS_RING_GFX;
147 	s32 priority = 0;
148 
149 	if (!cs->num_chunks) {
150 		return 0;
151 	}
152 	/* get chunks */
153 	INIT_LIST_HEAD(&p->validated);
154 	p->idx = 0;
155 	p->ib.sa_bo = NULL;
156 	p->ib.semaphore = NULL;
157 	p->const_ib.sa_bo = NULL;
158 	p->const_ib.semaphore = NULL;
159 	p->chunk_ib_idx = -1;
160 	p->chunk_relocs_idx = -1;
161 	p->chunk_flags_idx = -1;
162 	p->chunk_const_ib_idx = -1;
163 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
164 	if (p->chunks_array == NULL) {
165 		return -ENOMEM;
166 	}
167 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
168 	if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
169 			       sizeof(uint64_t)*cs->num_chunks)) {
170 		return -EFAULT;
171 	}
172 	p->cs_flags = 0;
173 	p->nchunks = cs->num_chunks;
174 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
175 	if (p->chunks == NULL) {
176 		return -ENOMEM;
177 	}
178 	for (i = 0; i < p->nchunks; i++) {
179 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
180 		struct drm_radeon_cs_chunk user_chunk;
181 		uint32_t __user *cdata;
182 
183 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
184 		if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
185 				       sizeof(struct drm_radeon_cs_chunk))) {
186 			return -EFAULT;
187 		}
188 		p->chunks[i].length_dw = user_chunk.length_dw;
189 		p->chunks[i].kdata = NULL;
190 		p->chunks[i].chunk_id = user_chunk.chunk_id;
191 		p->chunks[i].user_ptr = (void __user *)(unsigned long)user_chunk.chunk_data;
192 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
193 			p->chunk_relocs_idx = i;
194 		}
195 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
196 			p->chunk_ib_idx = i;
197 			/* zero length IB isn't useful */
198 			if (p->chunks[i].length_dw == 0)
199 				return -EINVAL;
200 		}
201 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
202 			p->chunk_const_ib_idx = i;
203 			/* zero length CONST IB isn't useful */
204 			if (p->chunks[i].length_dw == 0)
205 				return -EINVAL;
206 		}
207 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
208 			p->chunk_flags_idx = i;
209 			/* zero length flags aren't useful */
210 			if (p->chunks[i].length_dw == 0)
211 				return -EINVAL;
212 		}
213 
214 		cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
215 		if ((p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) ||
216 		    (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS)) {
217 			size = p->chunks[i].length_dw * sizeof(uint32_t);
218 			p->chunks[i].kdata = kmalloc(size, GFP_KERNEL);
219 			if (p->chunks[i].kdata == NULL) {
220 				return -ENOMEM;
221 			}
222 			if (DRM_COPY_FROM_USER(p->chunks[i].kdata,
223 					       p->chunks[i].user_ptr, size)) {
224 				return -EFAULT;
225 			}
226 			if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
227 				p->cs_flags = p->chunks[i].kdata[0];
228 				if (p->chunks[i].length_dw > 1)
229 					ring = p->chunks[i].kdata[1];
230 				if (p->chunks[i].length_dw > 2)
231 					priority = (s32)p->chunks[i].kdata[2];
232 			}
233 		}
234 	}
235 
236 	/* these are KMS only */
237 	if (p->rdev) {
238 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
239 		    !p->rdev->vm_manager.enabled) {
240 			DRM_ERROR("VM not active on asic!\n");
241 			return -EINVAL;
242 		}
243 
244 		/* we only support VM on SI+ */
245 		if ((p->rdev->family >= CHIP_TAHITI) &&
246 		    ((p->cs_flags & RADEON_CS_USE_VM) == 0)) {
247 			DRM_ERROR("VM required on SI+!\n");
248 			return -EINVAL;
249 		}
250 
251 		if (radeon_cs_get_ring(p, ring, priority))
252 			return -EINVAL;
253 	}
254 
255 	/* deal with non-vm */
256 	if ((p->chunk_ib_idx != -1) &&
257 	    ((p->cs_flags & RADEON_CS_USE_VM) == 0) &&
258 	    (p->chunks[p->chunk_ib_idx].chunk_id == RADEON_CHUNK_ID_IB)) {
259 		if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
260 			DRM_ERROR("cs IB too big: %d\n",
261 				  p->chunks[p->chunk_ib_idx].length_dw);
262 			return -EINVAL;
263 		}
264 		if (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) {
265 			p->chunks[p->chunk_ib_idx].kpage[0] = kmalloc(PAGE_SIZE, GFP_KERNEL);
266 			p->chunks[p->chunk_ib_idx].kpage[1] = kmalloc(PAGE_SIZE, GFP_KERNEL);
267 			if (p->chunks[p->chunk_ib_idx].kpage[0] == NULL ||
268 			    p->chunks[p->chunk_ib_idx].kpage[1] == NULL) {
269 				kfree(p->chunks[p->chunk_ib_idx].kpage[0]);
270 				kfree(p->chunks[p->chunk_ib_idx].kpage[1]);
271 				p->chunks[p->chunk_ib_idx].kpage[0] = NULL;
272 				p->chunks[p->chunk_ib_idx].kpage[1] = NULL;
273 				return -ENOMEM;
274 			}
275 		}
276 		p->chunks[p->chunk_ib_idx].kpage_idx[0] = -1;
277 		p->chunks[p->chunk_ib_idx].kpage_idx[1] = -1;
278 		p->chunks[p->chunk_ib_idx].last_copied_page = -1;
279 		p->chunks[p->chunk_ib_idx].last_page_index =
280 			((p->chunks[p->chunk_ib_idx].length_dw * 4) - 1) / PAGE_SIZE;
281 	}
282 
283 	return 0;
284 }
285 
286 /**
287  * cs_parser_fini() - clean parser states
288  * @parser:	parser structure holding parsing context.
289  * @error:	error number
290  *
291  * If error is set than unvalidate buffer, otherwise just free memory
292  * used by parsing context.
293  **/
294 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
295 {
296 	unsigned i;
297 
298 	if (!error) {
299 		ttm_eu_fence_buffer_objects(&parser->validated,
300 					    parser->ib.fence);
301 	} else {
302 		ttm_eu_backoff_reservation(&parser->validated);
303 	}
304 
305 	if (parser->relocs != NULL) {
306 		for (i = 0; i < parser->nrelocs; i++) {
307 			if (parser->relocs[i].gobj)
308 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
309 		}
310 	}
311 	kfree(parser->track);
312 	kfree(parser->relocs);
313 	kfree(parser->relocs_ptr);
314 	for (i = 0; i < parser->nchunks; i++) {
315 		kfree(parser->chunks[i].kdata);
316 		if ((parser->rdev->flags & RADEON_IS_AGP)) {
317 			kfree(parser->chunks[i].kpage[0]);
318 			kfree(parser->chunks[i].kpage[1]);
319 		}
320 	}
321 	kfree(parser->chunks);
322 	kfree(parser->chunks_array);
323 	radeon_ib_free(parser->rdev, &parser->ib);
324 	radeon_ib_free(parser->rdev, &parser->const_ib);
325 }
326 
327 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
328 			      struct radeon_cs_parser *parser)
329 {
330 	struct radeon_cs_chunk *ib_chunk;
331 	int r;
332 
333 	if (parser->chunk_ib_idx == -1)
334 		return 0;
335 
336 	if (parser->cs_flags & RADEON_CS_USE_VM)
337 		return 0;
338 
339 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
340 	/* Copy the packet into the IB, the parser will read from the
341 	 * input memory (cached) and write to the IB (which can be
342 	 * uncached).
343 	 */
344 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
345 			   NULL, ib_chunk->length_dw * 4);
346 	if (r) {
347 		DRM_ERROR("Failed to get ib !\n");
348 		return r;
349 	}
350 	parser->ib.length_dw = ib_chunk->length_dw;
351 	r = radeon_cs_parse(rdev, parser->ring, parser);
352 	if (r || parser->parser_error) {
353 		DRM_ERROR("Invalid command stream !\n");
354 		return r;
355 	}
356 	r = radeon_cs_finish_pages(parser);
357 	if (r) {
358 		DRM_ERROR("Invalid command stream !\n");
359 		return r;
360 	}
361 	radeon_cs_sync_rings(parser);
362 	r = radeon_ib_schedule(rdev, &parser->ib, NULL);
363 	if (r) {
364 		DRM_ERROR("Failed to schedule IB !\n");
365 	}
366 	return r;
367 }
368 
369 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
370 				   struct radeon_vm *vm)
371 {
372 	struct radeon_device *rdev = parser->rdev;
373 	struct radeon_bo_list *lobj;
374 	struct radeon_bo *bo;
375 	int r;
376 
377 	r = radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo, &rdev->ring_tmp_bo.bo->tbo.mem);
378 	if (r) {
379 		return r;
380 	}
381 	list_for_each_entry(lobj, &parser->validated, tv.head) {
382 		bo = lobj->bo;
383 		r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
384 		if (r) {
385 			return r;
386 		}
387 	}
388 	return 0;
389 }
390 
391 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
392 				 struct radeon_cs_parser *parser)
393 {
394 	struct radeon_cs_chunk *ib_chunk;
395 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
396 	struct radeon_vm *vm = &fpriv->vm;
397 	int r;
398 
399 	if (parser->chunk_ib_idx == -1)
400 		return 0;
401 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
402 		return 0;
403 
404 	if ((rdev->family >= CHIP_TAHITI) &&
405 	    (parser->chunk_const_ib_idx != -1)) {
406 		ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
407 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
408 			DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
409 			return -EINVAL;
410 		}
411 		r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
412 				   vm, ib_chunk->length_dw * 4);
413 		if (r) {
414 			DRM_ERROR("Failed to get const ib !\n");
415 			return r;
416 		}
417 		parser->const_ib.is_const_ib = true;
418 		parser->const_ib.length_dw = ib_chunk->length_dw;
419 		/* Copy the packet into the IB */
420 		if (DRM_COPY_FROM_USER(parser->const_ib.ptr, ib_chunk->user_ptr,
421 				       ib_chunk->length_dw * 4)) {
422 			return -EFAULT;
423 		}
424 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
425 		if (r) {
426 			return r;
427 		}
428 	}
429 
430 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
431 	if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
432 		DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
433 		return -EINVAL;
434 	}
435 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
436 			   vm, ib_chunk->length_dw * 4);
437 	if (r) {
438 		DRM_ERROR("Failed to get ib !\n");
439 		return r;
440 	}
441 	parser->ib.length_dw = ib_chunk->length_dw;
442 	/* Copy the packet into the IB */
443 	if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
444 			       ib_chunk->length_dw * 4)) {
445 		return -EFAULT;
446 	}
447 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
448 	if (r) {
449 		return r;
450 	}
451 
452 	mutex_lock(&rdev->vm_manager.lock);
453 	mutex_lock(&vm->mutex);
454 	r = radeon_vm_alloc_pt(rdev, vm);
455 	if (r) {
456 		goto out;
457 	}
458 	r = radeon_bo_vm_update_pte(parser, vm);
459 	if (r) {
460 		goto out;
461 	}
462 	radeon_cs_sync_rings(parser);
463 	radeon_ib_sync_to(&parser->ib, vm->fence);
464 	radeon_ib_sync_to(&parser->ib, radeon_vm_grab_id(
465 		rdev, vm, parser->ring));
466 
467 	if ((rdev->family >= CHIP_TAHITI) &&
468 	    (parser->chunk_const_ib_idx != -1)) {
469 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib);
470 	} else {
471 		r = radeon_ib_schedule(rdev, &parser->ib, NULL);
472 	}
473 
474 	if (!r) {
475 		radeon_vm_fence(rdev, vm, parser->ib.fence);
476 	}
477 
478 out:
479 	radeon_vm_add_to_lru(rdev, vm);
480 	mutex_unlock(&vm->mutex);
481 	mutex_unlock(&rdev->vm_manager.lock);
482 	return r;
483 }
484 
485 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
486 {
487 	if (r == -EDEADLK) {
488 		r = radeon_gpu_reset(rdev);
489 		if (!r)
490 			r = -EAGAIN;
491 	}
492 	return r;
493 }
494 
495 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
496 {
497 	struct radeon_device *rdev = dev->dev_private;
498 	struct radeon_cs_parser parser;
499 	int r;
500 
501 	down_read(&rdev->exclusive_lock);
502 	if (!rdev->accel_working) {
503 		up_read(&rdev->exclusive_lock);
504 		return -EBUSY;
505 	}
506 	/* initialize parser */
507 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
508 	parser.filp = filp;
509 	parser.rdev = rdev;
510 	parser.dev = rdev->dev;
511 	parser.family = rdev->family;
512 	r = radeon_cs_parser_init(&parser, data);
513 	if (r) {
514 		DRM_ERROR("Failed to initialize parser !\n");
515 		radeon_cs_parser_fini(&parser, r);
516 		up_read(&rdev->exclusive_lock);
517 		r = radeon_cs_handle_lockup(rdev, r);
518 		return r;
519 	}
520 	r = radeon_cs_parser_relocs(&parser);
521 	if (r) {
522 		if (r != -ERESTARTSYS)
523 			DRM_ERROR("Failed to parse relocation %d!\n", r);
524 		radeon_cs_parser_fini(&parser, r);
525 		up_read(&rdev->exclusive_lock);
526 		r = radeon_cs_handle_lockup(rdev, r);
527 		return r;
528 	}
529 	r = radeon_cs_ib_chunk(rdev, &parser);
530 	if (r) {
531 		goto out;
532 	}
533 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
534 	if (r) {
535 		goto out;
536 	}
537 out:
538 	radeon_cs_parser_fini(&parser, r);
539 	up_read(&rdev->exclusive_lock);
540 	r = radeon_cs_handle_lockup(rdev, r);
541 	return r;
542 }
543 
544 int radeon_cs_finish_pages(struct radeon_cs_parser *p)
545 {
546 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
547 	int i;
548 	int size = PAGE_SIZE;
549 
550 	for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
551 		if (i == ibc->last_page_index) {
552 			size = (ibc->length_dw * 4) % PAGE_SIZE;
553 			if (size == 0)
554 				size = PAGE_SIZE;
555 		}
556 
557 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
558 				       ibc->user_ptr + (i * PAGE_SIZE),
559 				       size))
560 			return -EFAULT;
561 	}
562 	return 0;
563 }
564 
565 static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
566 {
567 	int new_page;
568 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
569 	int i;
570 	int size = PAGE_SIZE;
571 	bool copy1 = (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) ?
572 		false : true;
573 
574 	for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
575 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
576 				       ibc->user_ptr + (i * PAGE_SIZE),
577 				       PAGE_SIZE)) {
578 			p->parser_error = -EFAULT;
579 			return 0;
580 		}
581 	}
582 
583 	if (pg_idx == ibc->last_page_index) {
584 		size = (ibc->length_dw * 4) % PAGE_SIZE;
585 		if (size == 0)
586 			size = PAGE_SIZE;
587 	}
588 
589 	new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
590 	if (copy1)
591 		ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
592 
593 	if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
594 			       ibc->user_ptr + (pg_idx * PAGE_SIZE),
595 			       size)) {
596 		p->parser_error = -EFAULT;
597 		return 0;
598 	}
599 
600 	/* copy to IB for non single case */
601 	if (!copy1)
602 		memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
603 
604 	ibc->last_copied_page = pg_idx;
605 	ibc->kpage_idx[new_page] = pg_idx;
606 
607 	return new_page;
608 }
609 
610 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
611 {
612 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
613 	u32 pg_idx, pg_offset;
614 	u32 idx_value = 0;
615 	int new_page;
616 
617 	pg_idx = (idx * 4) / PAGE_SIZE;
618 	pg_offset = (idx * 4) % PAGE_SIZE;
619 
620 	if (ibc->kpage_idx[0] == pg_idx)
621 		return ibc->kpage[0][pg_offset/4];
622 	if (ibc->kpage_idx[1] == pg_idx)
623 		return ibc->kpage[1][pg_offset/4];
624 
625 	new_page = radeon_cs_update_pages(p, pg_idx);
626 	if (new_page < 0) {
627 		p->parser_error = new_page;
628 		return 0;
629 	}
630 
631 	idx_value = ibc->kpage[new_page][pg_offset/4];
632 	return idx_value;
633 }
634 
635 /**
636  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
637  * @parser:	parser structure holding parsing context.
638  * @pkt:	where to store packet information
639  *
640  * Assume that chunk_ib_index is properly set. Will return -EINVAL
641  * if packet is bigger than remaining ib size. or if packets is unknown.
642  **/
643 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
644 			   struct radeon_cs_packet *pkt,
645 			   unsigned idx)
646 {
647 	struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
648 	struct radeon_device *rdev = p->rdev;
649 	uint32_t header;
650 
651 	if (idx >= ib_chunk->length_dw) {
652 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
653 			  idx, ib_chunk->length_dw);
654 		return -EINVAL;
655 	}
656 	header = radeon_get_ib_value(p, idx);
657 	pkt->idx = idx;
658 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
659 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
660 	pkt->one_reg_wr = 0;
661 	switch (pkt->type) {
662 	case RADEON_PACKET_TYPE0:
663 		if (rdev->family < CHIP_R600) {
664 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
665 			pkt->one_reg_wr =
666 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
667 		} else
668 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
669 		break;
670 	case RADEON_PACKET_TYPE3:
671 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
672 		break;
673 	case RADEON_PACKET_TYPE2:
674 		pkt->count = -1;
675 		break;
676 	default:
677 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
678 		return -EINVAL;
679 	}
680 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
681 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
682 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
683 		return -EINVAL;
684 	}
685 	return 0;
686 }
687 
688 /**
689  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
690  * @p:		structure holding the parser context.
691  *
692  * Check if the next packet is NOP relocation packet3.
693  **/
694 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
695 {
696 	struct radeon_cs_packet p3reloc;
697 	int r;
698 
699 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
700 	if (r)
701 		return false;
702 	if (p3reloc.type != RADEON_PACKET_TYPE3)
703 		return false;
704 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
705 		return false;
706 	return true;
707 }
708 
709 /**
710  * radeon_cs_dump_packet() - dump raw packet context
711  * @p:		structure holding the parser context.
712  * @pkt:	structure holding the packet.
713  *
714  * Used mostly for debugging and error reporting.
715  **/
716 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
717 			   struct radeon_cs_packet *pkt)
718 {
719 	volatile uint32_t *ib;
720 	unsigned i;
721 	unsigned idx;
722 
723 	ib = p->ib.ptr;
724 	idx = pkt->idx;
725 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
726 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
727 }
728 
729 /**
730  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
731  * @parser:		parser structure holding parsing context.
732  * @data:		pointer to relocation data
733  * @offset_start:	starting offset
734  * @offset_mask:	offset mask (to align start offset on)
735  * @reloc:		reloc informations
736  *
737  * Check if next packet is relocation packet3, do bo validation and compute
738  * GPU offset using the provided start.
739  **/
740 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
741 				struct radeon_cs_reloc **cs_reloc,
742 				int nomm)
743 {
744 	struct radeon_cs_chunk *relocs_chunk;
745 	struct radeon_cs_packet p3reloc;
746 	unsigned idx;
747 	int r;
748 
749 	if (p->chunk_relocs_idx == -1) {
750 		DRM_ERROR("No relocation chunk !\n");
751 		return -EINVAL;
752 	}
753 	*cs_reloc = NULL;
754 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
755 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
756 	if (r)
757 		return r;
758 	p->idx += p3reloc.count + 2;
759 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
760 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
761 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
762 			  p3reloc.idx);
763 		radeon_cs_dump_packet(p, &p3reloc);
764 		return -EINVAL;
765 	}
766 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
767 	if (idx >= relocs_chunk->length_dw) {
768 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
769 			  idx, relocs_chunk->length_dw);
770 		radeon_cs_dump_packet(p, &p3reloc);
771 		return -EINVAL;
772 	}
773 	/* FIXME: we assume reloc size is 4 dwords */
774 	if (nomm) {
775 		*cs_reloc = p->relocs;
776 		(*cs_reloc)->lobj.gpu_offset =
777 			(u64)relocs_chunk->kdata[idx + 3] << 32;
778 		(*cs_reloc)->lobj.gpu_offset |= relocs_chunk->kdata[idx + 0];
779 	} else
780 		*cs_reloc = p->relocs_ptr[(idx / 4)];
781 	return 0;
782 }
783