xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision 95e9fd10)
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 "drmP.h"
28 #include "radeon_drm.h"
29 #include "radeon_reg.h"
30 #include "radeon.h"
31 
32 void r100_cs_dump_packet(struct radeon_cs_parser *p,
33 			 struct radeon_cs_packet *pkt);
34 
35 int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
36 {
37 	struct drm_device *ddev = p->rdev->ddev;
38 	struct radeon_cs_chunk *chunk;
39 	unsigned i, j;
40 	bool duplicate;
41 
42 	if (p->chunk_relocs_idx == -1) {
43 		return 0;
44 	}
45 	chunk = &p->chunks[p->chunk_relocs_idx];
46 	/* FIXME: we assume that each relocs use 4 dwords */
47 	p->nrelocs = chunk->length_dw / 4;
48 	p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
49 	if (p->relocs_ptr == NULL) {
50 		return -ENOMEM;
51 	}
52 	p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
53 	if (p->relocs == NULL) {
54 		return -ENOMEM;
55 	}
56 	for (i = 0; i < p->nrelocs; i++) {
57 		struct drm_radeon_cs_reloc *r;
58 
59 		duplicate = false;
60 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
61 		for (j = 0; j < i; j++) {
62 			if (r->handle == p->relocs[j].handle) {
63 				p->relocs_ptr[i] = &p->relocs[j];
64 				duplicate = true;
65 				break;
66 			}
67 		}
68 		if (!duplicate) {
69 			p->relocs[i].gobj = drm_gem_object_lookup(ddev,
70 								  p->filp,
71 								  r->handle);
72 			if (p->relocs[i].gobj == NULL) {
73 				DRM_ERROR("gem object lookup failed 0x%x\n",
74 					  r->handle);
75 				return -ENOENT;
76 			}
77 			p->relocs_ptr[i] = &p->relocs[i];
78 			p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
79 			p->relocs[i].lobj.bo = p->relocs[i].robj;
80 			p->relocs[i].lobj.wdomain = r->write_domain;
81 			p->relocs[i].lobj.rdomain = r->read_domains;
82 			p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
83 			p->relocs[i].handle = r->handle;
84 			p->relocs[i].flags = r->flags;
85 			radeon_bo_list_add_object(&p->relocs[i].lobj,
86 						  &p->validated);
87 
88 		} else
89 			p->relocs[i].handle = 0;
90 	}
91 	return radeon_bo_list_validate(&p->validated);
92 }
93 
94 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
95 {
96 	p->priority = priority;
97 
98 	switch (ring) {
99 	default:
100 		DRM_ERROR("unknown ring id: %d\n", ring);
101 		return -EINVAL;
102 	case RADEON_CS_RING_GFX:
103 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
104 		break;
105 	case RADEON_CS_RING_COMPUTE:
106 		if (p->rdev->family >= CHIP_TAHITI) {
107 			if (p->priority > 0)
108 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
109 			else
110 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
111 		} else
112 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
113 		break;
114 	}
115 	return 0;
116 }
117 
118 static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
119 {
120 	int i;
121 
122 	for (i = 0; i < p->nrelocs; i++) {
123 		struct radeon_fence *a, *b;
124 
125 		if (!p->relocs[i].robj || !p->relocs[i].robj->tbo.sync_obj)
126 			continue;
127 
128 		a = p->relocs[i].robj->tbo.sync_obj;
129 		b = p->ib.sync_to[a->ring];
130 		p->ib.sync_to[a->ring] = radeon_fence_later(a, b);
131 	}
132 }
133 
134 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
135 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
136 {
137 	struct drm_radeon_cs *cs = data;
138 	uint64_t *chunk_array_ptr;
139 	unsigned size, i;
140 	u32 ring = RADEON_CS_RING_GFX;
141 	s32 priority = 0;
142 
143 	if (!cs->num_chunks) {
144 		return 0;
145 	}
146 	/* get chunks */
147 	INIT_LIST_HEAD(&p->validated);
148 	p->idx = 0;
149 	p->ib.sa_bo = NULL;
150 	p->ib.semaphore = NULL;
151 	p->const_ib.sa_bo = NULL;
152 	p->const_ib.semaphore = NULL;
153 	p->chunk_ib_idx = -1;
154 	p->chunk_relocs_idx = -1;
155 	p->chunk_flags_idx = -1;
156 	p->chunk_const_ib_idx = -1;
157 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
158 	if (p->chunks_array == NULL) {
159 		return -ENOMEM;
160 	}
161 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
162 	if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
163 			       sizeof(uint64_t)*cs->num_chunks)) {
164 		return -EFAULT;
165 	}
166 	p->cs_flags = 0;
167 	p->nchunks = cs->num_chunks;
168 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
169 	if (p->chunks == NULL) {
170 		return -ENOMEM;
171 	}
172 	for (i = 0; i < p->nchunks; i++) {
173 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
174 		struct drm_radeon_cs_chunk user_chunk;
175 		uint32_t __user *cdata;
176 
177 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
178 		if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
179 				       sizeof(struct drm_radeon_cs_chunk))) {
180 			return -EFAULT;
181 		}
182 		p->chunks[i].length_dw = user_chunk.length_dw;
183 		p->chunks[i].kdata = NULL;
184 		p->chunks[i].chunk_id = user_chunk.chunk_id;
185 
186 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
187 			p->chunk_relocs_idx = i;
188 		}
189 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
190 			p->chunk_ib_idx = i;
191 			/* zero length IB isn't useful */
192 			if (p->chunks[i].length_dw == 0)
193 				return -EINVAL;
194 		}
195 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
196 			p->chunk_const_ib_idx = i;
197 			/* zero length CONST 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_FLAGS) {
202 			p->chunk_flags_idx = i;
203 			/* zero length flags aren't useful */
204 			if (p->chunks[i].length_dw == 0)
205 				return -EINVAL;
206 		}
207 
208 		p->chunks[i].length_dw = user_chunk.length_dw;
209 		p->chunks[i].user_ptr = (void __user *)(unsigned long)user_chunk.chunk_data;
210 
211 		cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
212 		if ((p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) ||
213 		    (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS)) {
214 			size = p->chunks[i].length_dw * sizeof(uint32_t);
215 			p->chunks[i].kdata = kmalloc(size, GFP_KERNEL);
216 			if (p->chunks[i].kdata == NULL) {
217 				return -ENOMEM;
218 			}
219 			if (DRM_COPY_FROM_USER(p->chunks[i].kdata,
220 					       p->chunks[i].user_ptr, size)) {
221 				return -EFAULT;
222 			}
223 			if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
224 				p->cs_flags = p->chunks[i].kdata[0];
225 				if (p->chunks[i].length_dw > 1)
226 					ring = p->chunks[i].kdata[1];
227 				if (p->chunks[i].length_dw > 2)
228 					priority = (s32)p->chunks[i].kdata[2];
229 			}
230 		}
231 	}
232 
233 	/* these are KMS only */
234 	if (p->rdev) {
235 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
236 		    !p->rdev->vm_manager.enabled) {
237 			DRM_ERROR("VM not active on asic!\n");
238 			return -EINVAL;
239 		}
240 
241 		/* we only support VM on SI+ */
242 		if ((p->rdev->family >= CHIP_TAHITI) &&
243 		    ((p->cs_flags & RADEON_CS_USE_VM) == 0)) {
244 			DRM_ERROR("VM required on SI+!\n");
245 			return -EINVAL;
246 		}
247 
248 		if (radeon_cs_get_ring(p, ring, priority))
249 			return -EINVAL;
250 	}
251 
252 	/* deal with non-vm */
253 	if ((p->chunk_ib_idx != -1) &&
254 	    ((p->cs_flags & RADEON_CS_USE_VM) == 0) &&
255 	    (p->chunks[p->chunk_ib_idx].chunk_id == RADEON_CHUNK_ID_IB)) {
256 		if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
257 			DRM_ERROR("cs IB too big: %d\n",
258 				  p->chunks[p->chunk_ib_idx].length_dw);
259 			return -EINVAL;
260 		}
261 		if ((p->rdev->flags & RADEON_IS_AGP)) {
262 			p->chunks[p->chunk_ib_idx].kpage[0] = kmalloc(PAGE_SIZE, GFP_KERNEL);
263 			p->chunks[p->chunk_ib_idx].kpage[1] = kmalloc(PAGE_SIZE, GFP_KERNEL);
264 			if (p->chunks[p->chunk_ib_idx].kpage[0] == NULL ||
265 			    p->chunks[p->chunk_ib_idx].kpage[1] == NULL) {
266 				kfree(p->chunks[i].kpage[0]);
267 				kfree(p->chunks[i].kpage[1]);
268 				return -ENOMEM;
269 			}
270 		}
271 		p->chunks[p->chunk_ib_idx].kpage_idx[0] = -1;
272 		p->chunks[p->chunk_ib_idx].kpage_idx[1] = -1;
273 		p->chunks[p->chunk_ib_idx].last_copied_page = -1;
274 		p->chunks[p->chunk_ib_idx].last_page_index =
275 			((p->chunks[p->chunk_ib_idx].length_dw * 4) - 1) / PAGE_SIZE;
276 	}
277 
278 	return 0;
279 }
280 
281 static void radeon_bo_vm_fence_va(struct radeon_cs_parser *parser,
282 				  struct radeon_fence *fence)
283 {
284 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
285 	struct radeon_vm *vm = &fpriv->vm;
286 	struct radeon_bo_list *lobj;
287 
288 	if (parser->chunk_ib_idx == -1) {
289 		return;
290 	}
291 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0) {
292 		return;
293 	}
294 
295 	list_for_each_entry(lobj, &parser->validated, tv.head) {
296 		struct radeon_bo_va *bo_va;
297 		struct radeon_bo *rbo = lobj->bo;
298 
299 		bo_va = radeon_bo_va(rbo, vm);
300 		radeon_fence_unref(&bo_va->fence);
301 		bo_va->fence = radeon_fence_ref(fence);
302 	}
303 }
304 
305 /**
306  * cs_parser_fini() - clean parser states
307  * @parser:	parser structure holding parsing context.
308  * @error:	error number
309  *
310  * If error is set than unvalidate buffer, otherwise just free memory
311  * used by parsing context.
312  **/
313 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
314 {
315 	unsigned i;
316 
317 	if (!error) {
318 		/* fence all bo va before ttm_eu_fence_buffer_objects so bo are still reserved */
319 		radeon_bo_vm_fence_va(parser, parser->ib.fence);
320 		ttm_eu_fence_buffer_objects(&parser->validated,
321 					    parser->ib.fence);
322 	} else {
323 		ttm_eu_backoff_reservation(&parser->validated);
324 	}
325 
326 	if (parser->relocs != NULL) {
327 		for (i = 0; i < parser->nrelocs; i++) {
328 			if (parser->relocs[i].gobj)
329 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
330 		}
331 	}
332 	kfree(parser->track);
333 	kfree(parser->relocs);
334 	kfree(parser->relocs_ptr);
335 	for (i = 0; i < parser->nchunks; i++) {
336 		kfree(parser->chunks[i].kdata);
337 		if ((parser->rdev->flags & RADEON_IS_AGP)) {
338 			kfree(parser->chunks[i].kpage[0]);
339 			kfree(parser->chunks[i].kpage[1]);
340 		}
341 	}
342 	kfree(parser->chunks);
343 	kfree(parser->chunks_array);
344 	radeon_ib_free(parser->rdev, &parser->ib);
345 	radeon_ib_free(parser->rdev, &parser->const_ib);
346 }
347 
348 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
349 			      struct radeon_cs_parser *parser)
350 {
351 	struct radeon_cs_chunk *ib_chunk;
352 	int r;
353 
354 	if (parser->chunk_ib_idx == -1)
355 		return 0;
356 
357 	if (parser->cs_flags & RADEON_CS_USE_VM)
358 		return 0;
359 
360 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
361 	/* Copy the packet into the IB, the parser will read from the
362 	 * input memory (cached) and write to the IB (which can be
363 	 * uncached).
364 	 */
365 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
366 			   ib_chunk->length_dw * 4);
367 	if (r) {
368 		DRM_ERROR("Failed to get ib !\n");
369 		return r;
370 	}
371 	parser->ib.length_dw = ib_chunk->length_dw;
372 	r = radeon_cs_parse(rdev, parser->ring, parser);
373 	if (r || parser->parser_error) {
374 		DRM_ERROR("Invalid command stream !\n");
375 		return r;
376 	}
377 	r = radeon_cs_finish_pages(parser);
378 	if (r) {
379 		DRM_ERROR("Invalid command stream !\n");
380 		return r;
381 	}
382 	radeon_cs_sync_rings(parser);
383 	parser->ib.vm_id = 0;
384 	r = radeon_ib_schedule(rdev, &parser->ib, NULL);
385 	if (r) {
386 		DRM_ERROR("Failed to schedule IB !\n");
387 	}
388 	return r;
389 }
390 
391 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
392 				   struct radeon_vm *vm)
393 {
394 	struct radeon_bo_list *lobj;
395 	struct radeon_bo *bo;
396 	int r;
397 
398 	list_for_each_entry(lobj, &parser->validated, tv.head) {
399 		bo = lobj->bo;
400 		r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
401 		if (r) {
402 			return r;
403 		}
404 	}
405 	return 0;
406 }
407 
408 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
409 				 struct radeon_cs_parser *parser)
410 {
411 	struct radeon_cs_chunk *ib_chunk;
412 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
413 	struct radeon_vm *vm = &fpriv->vm;
414 	int r;
415 
416 	if (parser->chunk_ib_idx == -1)
417 		return 0;
418 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
419 		return 0;
420 
421 	if ((rdev->family >= CHIP_TAHITI) &&
422 	    (parser->chunk_const_ib_idx != -1)) {
423 		ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
424 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
425 			DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
426 			return -EINVAL;
427 		}
428 		r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
429 				   ib_chunk->length_dw * 4);
430 		if (r) {
431 			DRM_ERROR("Failed to get const ib !\n");
432 			return r;
433 		}
434 		parser->const_ib.is_const_ib = true;
435 		parser->const_ib.length_dw = ib_chunk->length_dw;
436 		/* Copy the packet into the IB */
437 		if (DRM_COPY_FROM_USER(parser->const_ib.ptr, ib_chunk->user_ptr,
438 				       ib_chunk->length_dw * 4)) {
439 			return -EFAULT;
440 		}
441 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
442 		if (r) {
443 			return r;
444 		}
445 	}
446 
447 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
448 	if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
449 		DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
450 		return -EINVAL;
451 	}
452 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
453 			   ib_chunk->length_dw * 4);
454 	if (r) {
455 		DRM_ERROR("Failed to get ib !\n");
456 		return r;
457 	}
458 	parser->ib.length_dw = ib_chunk->length_dw;
459 	/* Copy the packet into the IB */
460 	if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
461 			       ib_chunk->length_dw * 4)) {
462 		return -EFAULT;
463 	}
464 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
465 	if (r) {
466 		return r;
467 	}
468 
469 	mutex_lock(&rdev->vm_manager.lock);
470 	mutex_lock(&vm->mutex);
471 	r = radeon_vm_bind(rdev, vm);
472 	if (r) {
473 		goto out;
474 	}
475 	r = radeon_bo_vm_update_pte(parser, vm);
476 	if (r) {
477 		goto out;
478 	}
479 	radeon_cs_sync_rings(parser);
480 
481 	parser->ib.vm_id = vm->id;
482 	/* ib pool is bind at 0 in virtual address space,
483 	 * so gpu_addr is the offset inside the pool bo
484 	 */
485 	parser->ib.gpu_addr = parser->ib.sa_bo->soffset;
486 
487 	if ((rdev->family >= CHIP_TAHITI) &&
488 	    (parser->chunk_const_ib_idx != -1)) {
489 		parser->const_ib.vm_id = vm->id;
490 		/* ib pool is bind at 0 in virtual address space,
491 		 * so gpu_addr is the offset inside the pool bo
492 		 */
493 		parser->const_ib.gpu_addr = parser->const_ib.sa_bo->soffset;
494 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib);
495 	} else {
496 		r = radeon_ib_schedule(rdev, &parser->ib, NULL);
497 	}
498 
499 out:
500 	if (!r) {
501 		if (vm->fence) {
502 			radeon_fence_unref(&vm->fence);
503 		}
504 		vm->fence = radeon_fence_ref(parser->ib.fence);
505 	}
506 	mutex_unlock(&vm->mutex);
507 	mutex_unlock(&rdev->vm_manager.lock);
508 	return r;
509 }
510 
511 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
512 {
513 	if (r == -EDEADLK) {
514 		r = radeon_gpu_reset(rdev);
515 		if (!r)
516 			r = -EAGAIN;
517 	}
518 	return r;
519 }
520 
521 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
522 {
523 	struct radeon_device *rdev = dev->dev_private;
524 	struct radeon_cs_parser parser;
525 	int r;
526 
527 	down_read(&rdev->exclusive_lock);
528 	if (!rdev->accel_working) {
529 		up_read(&rdev->exclusive_lock);
530 		return -EBUSY;
531 	}
532 	/* initialize parser */
533 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
534 	parser.filp = filp;
535 	parser.rdev = rdev;
536 	parser.dev = rdev->dev;
537 	parser.family = rdev->family;
538 	r = radeon_cs_parser_init(&parser, data);
539 	if (r) {
540 		DRM_ERROR("Failed to initialize parser !\n");
541 		radeon_cs_parser_fini(&parser, r);
542 		up_read(&rdev->exclusive_lock);
543 		r = radeon_cs_handle_lockup(rdev, r);
544 		return r;
545 	}
546 	r = radeon_cs_parser_relocs(&parser);
547 	if (r) {
548 		if (r != -ERESTARTSYS)
549 			DRM_ERROR("Failed to parse relocation %d!\n", r);
550 		radeon_cs_parser_fini(&parser, r);
551 		up_read(&rdev->exclusive_lock);
552 		r = radeon_cs_handle_lockup(rdev, r);
553 		return r;
554 	}
555 	r = radeon_cs_ib_chunk(rdev, &parser);
556 	if (r) {
557 		goto out;
558 	}
559 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
560 	if (r) {
561 		goto out;
562 	}
563 out:
564 	radeon_cs_parser_fini(&parser, r);
565 	up_read(&rdev->exclusive_lock);
566 	r = radeon_cs_handle_lockup(rdev, r);
567 	return r;
568 }
569 
570 int radeon_cs_finish_pages(struct radeon_cs_parser *p)
571 {
572 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
573 	int i;
574 	int size = PAGE_SIZE;
575 
576 	for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
577 		if (i == ibc->last_page_index) {
578 			size = (ibc->length_dw * 4) % PAGE_SIZE;
579 			if (size == 0)
580 				size = PAGE_SIZE;
581 		}
582 
583 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
584 				       ibc->user_ptr + (i * PAGE_SIZE),
585 				       size))
586 			return -EFAULT;
587 	}
588 	return 0;
589 }
590 
591 static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
592 {
593 	int new_page;
594 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
595 	int i;
596 	int size = PAGE_SIZE;
597 	bool copy1 = (p->rdev->flags & RADEON_IS_AGP) ? false : true;
598 
599 	for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
600 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
601 				       ibc->user_ptr + (i * PAGE_SIZE),
602 				       PAGE_SIZE)) {
603 			p->parser_error = -EFAULT;
604 			return 0;
605 		}
606 	}
607 
608 	if (pg_idx == ibc->last_page_index) {
609 		size = (ibc->length_dw * 4) % PAGE_SIZE;
610 		if (size == 0)
611 			size = PAGE_SIZE;
612 	}
613 
614 	new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
615 	if (copy1)
616 		ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
617 
618 	if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
619 			       ibc->user_ptr + (pg_idx * PAGE_SIZE),
620 			       size)) {
621 		p->parser_error = -EFAULT;
622 		return 0;
623 	}
624 
625 	/* copy to IB for non single case */
626 	if (!copy1)
627 		memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
628 
629 	ibc->last_copied_page = pg_idx;
630 	ibc->kpage_idx[new_page] = pg_idx;
631 
632 	return new_page;
633 }
634 
635 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
636 {
637 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
638 	u32 pg_idx, pg_offset;
639 	u32 idx_value = 0;
640 	int new_page;
641 
642 	pg_idx = (idx * 4) / PAGE_SIZE;
643 	pg_offset = (idx * 4) % PAGE_SIZE;
644 
645 	if (ibc->kpage_idx[0] == pg_idx)
646 		return ibc->kpage[0][pg_offset/4];
647 	if (ibc->kpage_idx[1] == pg_idx)
648 		return ibc->kpage[1][pg_offset/4];
649 
650 	new_page = radeon_cs_update_pages(p, pg_idx);
651 	if (new_page < 0) {
652 		p->parser_error = new_page;
653 		return 0;
654 	}
655 
656 	idx_value = ibc->kpage[new_page][pg_offset/4];
657 	return idx_value;
658 }
659