xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision e1e0a9e6)
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 void r100_cs_dump_packet(struct radeon_cs_parser *p,
33 			 struct radeon_cs_packet *pkt);
34 
35 static 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 	p->dma_reloc_idx = 0;
47 	/* FIXME: we assume that each relocs use 4 dwords */
48 	p->nrelocs = chunk->length_dw / 4;
49 	p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
50 	if (p->relocs_ptr == NULL) {
51 		return -ENOMEM;
52 	}
53 	p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
54 	if (p->relocs == NULL) {
55 		return -ENOMEM;
56 	}
57 	for (i = 0; i < p->nrelocs; i++) {
58 		struct drm_radeon_cs_reloc *r;
59 
60 		duplicate = false;
61 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
62 		for (j = 0; j < i; j++) {
63 			if (r->handle == p->relocs[j].handle) {
64 				p->relocs_ptr[i] = &p->relocs[j];
65 				duplicate = true;
66 				break;
67 			}
68 		}
69 		if (!duplicate) {
70 			p->relocs[i].gobj = drm_gem_object_lookup(ddev,
71 								  p->filp,
72 								  r->handle);
73 			if (p->relocs[i].gobj == NULL) {
74 				DRM_ERROR("gem object lookup failed 0x%x\n",
75 					  r->handle);
76 				return -ENOENT;
77 			}
78 			p->relocs_ptr[i] = &p->relocs[i];
79 			p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
80 			p->relocs[i].lobj.bo = p->relocs[i].robj;
81 			p->relocs[i].lobj.wdomain = r->write_domain;
82 			p->relocs[i].lobj.rdomain = r->read_domains;
83 			p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
84 			p->relocs[i].handle = r->handle;
85 			p->relocs[i].flags = r->flags;
86 			radeon_bo_list_add_object(&p->relocs[i].lobj,
87 						  &p->validated);
88 
89 		} else
90 			p->relocs[i].handle = 0;
91 	}
92 	return radeon_bo_list_validate(&p->validated);
93 }
94 
95 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
96 {
97 	p->priority = priority;
98 
99 	switch (ring) {
100 	default:
101 		DRM_ERROR("unknown ring id: %d\n", ring);
102 		return -EINVAL;
103 	case RADEON_CS_RING_GFX:
104 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
105 		break;
106 	case RADEON_CS_RING_COMPUTE:
107 		if (p->rdev->family >= CHIP_TAHITI) {
108 			if (p->priority > 0)
109 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
110 			else
111 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
112 		} else
113 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
114 		break;
115 	case RADEON_CS_RING_DMA:
116 		if (p->rdev->family >= CHIP_CAYMAN) {
117 			if (p->priority > 0)
118 				p->ring = R600_RING_TYPE_DMA_INDEX;
119 			else
120 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
121 		} else if (p->rdev->family >= CHIP_R600) {
122 			p->ring = R600_RING_TYPE_DMA_INDEX;
123 		} else {
124 			return -EINVAL;
125 		}
126 		break;
127 	}
128 	return 0;
129 }
130 
131 static void radeon_cs_sync_to(struct radeon_cs_parser *p,
132 			      struct radeon_fence *fence)
133 {
134 	struct radeon_fence *other;
135 
136 	if (!fence)
137 		return;
138 
139 	other = p->ib.sync_to[fence->ring];
140 	p->ib.sync_to[fence->ring] = radeon_fence_later(fence, other);
141 }
142 
143 static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
144 {
145 	int i;
146 
147 	for (i = 0; i < p->nrelocs; i++) {
148 		if (!p->relocs[i].robj)
149 			continue;
150 
151 		radeon_cs_sync_to(p, p->relocs[i].robj->tbo.sync_obj);
152 	}
153 }
154 
155 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
156 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
157 {
158 	struct drm_radeon_cs *cs = data;
159 	uint64_t *chunk_array_ptr;
160 	unsigned size, i;
161 	u32 ring = RADEON_CS_RING_GFX;
162 	s32 priority = 0;
163 
164 	if (!cs->num_chunks) {
165 		return 0;
166 	}
167 	/* get chunks */
168 	INIT_LIST_HEAD(&p->validated);
169 	p->idx = 0;
170 	p->ib.sa_bo = NULL;
171 	p->ib.semaphore = NULL;
172 	p->const_ib.sa_bo = NULL;
173 	p->const_ib.semaphore = NULL;
174 	p->chunk_ib_idx = -1;
175 	p->chunk_relocs_idx = -1;
176 	p->chunk_flags_idx = -1;
177 	p->chunk_const_ib_idx = -1;
178 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
179 	if (p->chunks_array == NULL) {
180 		return -ENOMEM;
181 	}
182 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
183 	if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
184 			       sizeof(uint64_t)*cs->num_chunks)) {
185 		return -EFAULT;
186 	}
187 	p->cs_flags = 0;
188 	p->nchunks = cs->num_chunks;
189 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
190 	if (p->chunks == NULL) {
191 		return -ENOMEM;
192 	}
193 	for (i = 0; i < p->nchunks; i++) {
194 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
195 		struct drm_radeon_cs_chunk user_chunk;
196 		uint32_t __user *cdata;
197 
198 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
199 		if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
200 				       sizeof(struct drm_radeon_cs_chunk))) {
201 			return -EFAULT;
202 		}
203 		p->chunks[i].length_dw = user_chunk.length_dw;
204 		p->chunks[i].kdata = NULL;
205 		p->chunks[i].chunk_id = user_chunk.chunk_id;
206 
207 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
208 			p->chunk_relocs_idx = i;
209 		}
210 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
211 			p->chunk_ib_idx = i;
212 			/* zero length IB isn't useful */
213 			if (p->chunks[i].length_dw == 0)
214 				return -EINVAL;
215 		}
216 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
217 			p->chunk_const_ib_idx = i;
218 			/* zero length CONST IB isn't useful */
219 			if (p->chunks[i].length_dw == 0)
220 				return -EINVAL;
221 		}
222 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
223 			p->chunk_flags_idx = i;
224 			/* zero length flags aren't useful */
225 			if (p->chunks[i].length_dw == 0)
226 				return -EINVAL;
227 		}
228 
229 		p->chunks[i].length_dw = user_chunk.length_dw;
230 		p->chunks[i].user_ptr = (void __user *)(unsigned long)user_chunk.chunk_data;
231 
232 		cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
233 		if ((p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) ||
234 		    (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS)) {
235 			size = p->chunks[i].length_dw * sizeof(uint32_t);
236 			p->chunks[i].kdata = kmalloc(size, GFP_KERNEL);
237 			if (p->chunks[i].kdata == NULL) {
238 				return -ENOMEM;
239 			}
240 			if (DRM_COPY_FROM_USER(p->chunks[i].kdata,
241 					       p->chunks[i].user_ptr, size)) {
242 				return -EFAULT;
243 			}
244 			if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
245 				p->cs_flags = p->chunks[i].kdata[0];
246 				if (p->chunks[i].length_dw > 1)
247 					ring = p->chunks[i].kdata[1];
248 				if (p->chunks[i].length_dw > 2)
249 					priority = (s32)p->chunks[i].kdata[2];
250 			}
251 		}
252 	}
253 
254 	/* these are KMS only */
255 	if (p->rdev) {
256 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
257 		    !p->rdev->vm_manager.enabled) {
258 			DRM_ERROR("VM not active on asic!\n");
259 			return -EINVAL;
260 		}
261 
262 		/* we only support VM on SI+ */
263 		if ((p->rdev->family >= CHIP_TAHITI) &&
264 		    ((p->cs_flags & RADEON_CS_USE_VM) == 0)) {
265 			DRM_ERROR("VM required on SI+!\n");
266 			return -EINVAL;
267 		}
268 
269 		if (radeon_cs_get_ring(p, ring, priority))
270 			return -EINVAL;
271 	}
272 
273 	/* deal with non-vm */
274 	if ((p->chunk_ib_idx != -1) &&
275 	    ((p->cs_flags & RADEON_CS_USE_VM) == 0) &&
276 	    (p->chunks[p->chunk_ib_idx].chunk_id == RADEON_CHUNK_ID_IB)) {
277 		if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
278 			DRM_ERROR("cs IB too big: %d\n",
279 				  p->chunks[p->chunk_ib_idx].length_dw);
280 			return -EINVAL;
281 		}
282 		if (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) {
283 			p->chunks[p->chunk_ib_idx].kpage[0] = kmalloc(PAGE_SIZE, GFP_KERNEL);
284 			p->chunks[p->chunk_ib_idx].kpage[1] = kmalloc(PAGE_SIZE, GFP_KERNEL);
285 			if (p->chunks[p->chunk_ib_idx].kpage[0] == NULL ||
286 			    p->chunks[p->chunk_ib_idx].kpage[1] == NULL) {
287 				kfree(p->chunks[p->chunk_ib_idx].kpage[0]);
288 				kfree(p->chunks[p->chunk_ib_idx].kpage[1]);
289 				return -ENOMEM;
290 			}
291 		}
292 		p->chunks[p->chunk_ib_idx].kpage_idx[0] = -1;
293 		p->chunks[p->chunk_ib_idx].kpage_idx[1] = -1;
294 		p->chunks[p->chunk_ib_idx].last_copied_page = -1;
295 		p->chunks[p->chunk_ib_idx].last_page_index =
296 			((p->chunks[p->chunk_ib_idx].length_dw * 4) - 1) / PAGE_SIZE;
297 	}
298 
299 	return 0;
300 }
301 
302 /**
303  * cs_parser_fini() - clean parser states
304  * @parser:	parser structure holding parsing context.
305  * @error:	error number
306  *
307  * If error is set than unvalidate buffer, otherwise just free memory
308  * used by parsing context.
309  **/
310 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
311 {
312 	unsigned i;
313 
314 	if (!error) {
315 		ttm_eu_fence_buffer_objects(&parser->validated,
316 					    parser->ib.fence);
317 	} else {
318 		ttm_eu_backoff_reservation(&parser->validated);
319 	}
320 
321 	if (parser->relocs != NULL) {
322 		for (i = 0; i < parser->nrelocs; i++) {
323 			if (parser->relocs[i].gobj)
324 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
325 		}
326 	}
327 	kfree(parser->track);
328 	kfree(parser->relocs);
329 	kfree(parser->relocs_ptr);
330 	for (i = 0; i < parser->nchunks; i++) {
331 		kfree(parser->chunks[i].kdata);
332 		if ((parser->rdev->flags & RADEON_IS_AGP)) {
333 			kfree(parser->chunks[i].kpage[0]);
334 			kfree(parser->chunks[i].kpage[1]);
335 		}
336 	}
337 	kfree(parser->chunks);
338 	kfree(parser->chunks_array);
339 	radeon_ib_free(parser->rdev, &parser->ib);
340 	radeon_ib_free(parser->rdev, &parser->const_ib);
341 }
342 
343 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
344 			      struct radeon_cs_parser *parser)
345 {
346 	struct radeon_cs_chunk *ib_chunk;
347 	int r;
348 
349 	if (parser->chunk_ib_idx == -1)
350 		return 0;
351 
352 	if (parser->cs_flags & RADEON_CS_USE_VM)
353 		return 0;
354 
355 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
356 	/* Copy the packet into the IB, the parser will read from the
357 	 * input memory (cached) and write to the IB (which can be
358 	 * uncached).
359 	 */
360 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
361 			   NULL, ib_chunk->length_dw * 4);
362 	if (r) {
363 		DRM_ERROR("Failed to get ib !\n");
364 		return r;
365 	}
366 	parser->ib.length_dw = ib_chunk->length_dw;
367 	r = radeon_cs_parse(rdev, parser->ring, parser);
368 	if (r || parser->parser_error) {
369 		DRM_ERROR("Invalid command stream !\n");
370 		return r;
371 	}
372 	r = radeon_cs_finish_pages(parser);
373 	if (r) {
374 		DRM_ERROR("Invalid command stream !\n");
375 		return r;
376 	}
377 	radeon_cs_sync_rings(parser);
378 	r = radeon_ib_schedule(rdev, &parser->ib, NULL);
379 	if (r) {
380 		DRM_ERROR("Failed to schedule IB !\n");
381 	}
382 	return r;
383 }
384 
385 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
386 				   struct radeon_vm *vm)
387 {
388 	struct radeon_device *rdev = parser->rdev;
389 	struct radeon_bo_list *lobj;
390 	struct radeon_bo *bo;
391 	int r;
392 
393 	r = radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo, &rdev->ring_tmp_bo.bo->tbo.mem);
394 	if (r) {
395 		return r;
396 	}
397 	list_for_each_entry(lobj, &parser->validated, tv.head) {
398 		bo = lobj->bo;
399 		r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
400 		if (r) {
401 			return r;
402 		}
403 	}
404 	return 0;
405 }
406 
407 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
408 				 struct radeon_cs_parser *parser)
409 {
410 	struct radeon_cs_chunk *ib_chunk;
411 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
412 	struct radeon_vm *vm = &fpriv->vm;
413 	int r;
414 
415 	if (parser->chunk_ib_idx == -1)
416 		return 0;
417 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
418 		return 0;
419 
420 	if ((rdev->family >= CHIP_TAHITI) &&
421 	    (parser->chunk_const_ib_idx != -1)) {
422 		ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
423 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
424 			DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
425 			return -EINVAL;
426 		}
427 		r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
428 				   vm, ib_chunk->length_dw * 4);
429 		if (r) {
430 			DRM_ERROR("Failed to get const ib !\n");
431 			return r;
432 		}
433 		parser->const_ib.is_const_ib = true;
434 		parser->const_ib.length_dw = ib_chunk->length_dw;
435 		/* Copy the packet into the IB */
436 		if (DRM_COPY_FROM_USER(parser->const_ib.ptr, ib_chunk->user_ptr,
437 				       ib_chunk->length_dw * 4)) {
438 			return -EFAULT;
439 		}
440 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
441 		if (r) {
442 			return r;
443 		}
444 	}
445 
446 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
447 	if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
448 		DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
449 		return -EINVAL;
450 	}
451 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
452 			   vm, ib_chunk->length_dw * 4);
453 	if (r) {
454 		DRM_ERROR("Failed to get ib !\n");
455 		return r;
456 	}
457 	parser->ib.length_dw = ib_chunk->length_dw;
458 	/* Copy the packet into the IB */
459 	if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
460 			       ib_chunk->length_dw * 4)) {
461 		return -EFAULT;
462 	}
463 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
464 	if (r) {
465 		return r;
466 	}
467 
468 	mutex_lock(&rdev->vm_manager.lock);
469 	mutex_lock(&vm->mutex);
470 	r = radeon_vm_alloc_pt(rdev, vm);
471 	if (r) {
472 		goto out;
473 	}
474 	r = radeon_bo_vm_update_pte(parser, vm);
475 	if (r) {
476 		goto out;
477 	}
478 	radeon_cs_sync_rings(parser);
479 	radeon_cs_sync_to(parser, vm->fence);
480 	radeon_cs_sync_to(parser, radeon_vm_grab_id(rdev, vm, parser->ring));
481 
482 	if ((rdev->family >= CHIP_TAHITI) &&
483 	    (parser->chunk_const_ib_idx != -1)) {
484 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib);
485 	} else {
486 		r = radeon_ib_schedule(rdev, &parser->ib, NULL);
487 	}
488 
489 	if (!r) {
490 		radeon_vm_fence(rdev, vm, parser->ib.fence);
491 	}
492 
493 out:
494 	radeon_vm_add_to_lru(rdev, vm);
495 	mutex_unlock(&vm->mutex);
496 	mutex_unlock(&rdev->vm_manager.lock);
497 	return r;
498 }
499 
500 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
501 {
502 	if (r == -EDEADLK) {
503 		r = radeon_gpu_reset(rdev);
504 		if (!r)
505 			r = -EAGAIN;
506 	}
507 	return r;
508 }
509 
510 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
511 {
512 	struct radeon_device *rdev = dev->dev_private;
513 	struct radeon_cs_parser parser;
514 	int r;
515 
516 	down_read(&rdev->exclusive_lock);
517 	if (!rdev->accel_working) {
518 		up_read(&rdev->exclusive_lock);
519 		return -EBUSY;
520 	}
521 	/* initialize parser */
522 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
523 	parser.filp = filp;
524 	parser.rdev = rdev;
525 	parser.dev = rdev->dev;
526 	parser.family = rdev->family;
527 	r = radeon_cs_parser_init(&parser, data);
528 	if (r) {
529 		DRM_ERROR("Failed to initialize parser !\n");
530 		radeon_cs_parser_fini(&parser, r);
531 		up_read(&rdev->exclusive_lock);
532 		r = radeon_cs_handle_lockup(rdev, r);
533 		return r;
534 	}
535 	r = radeon_cs_parser_relocs(&parser);
536 	if (r) {
537 		if (r != -ERESTARTSYS)
538 			DRM_ERROR("Failed to parse relocation %d!\n", r);
539 		radeon_cs_parser_fini(&parser, r);
540 		up_read(&rdev->exclusive_lock);
541 		r = radeon_cs_handle_lockup(rdev, r);
542 		return r;
543 	}
544 	r = radeon_cs_ib_chunk(rdev, &parser);
545 	if (r) {
546 		goto out;
547 	}
548 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
549 	if (r) {
550 		goto out;
551 	}
552 out:
553 	radeon_cs_parser_fini(&parser, r);
554 	up_read(&rdev->exclusive_lock);
555 	r = radeon_cs_handle_lockup(rdev, r);
556 	return r;
557 }
558 
559 int radeon_cs_finish_pages(struct radeon_cs_parser *p)
560 {
561 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
562 	int i;
563 	int size = PAGE_SIZE;
564 
565 	for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
566 		if (i == ibc->last_page_index) {
567 			size = (ibc->length_dw * 4) % PAGE_SIZE;
568 			if (size == 0)
569 				size = PAGE_SIZE;
570 		}
571 
572 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
573 				       ibc->user_ptr + (i * PAGE_SIZE),
574 				       size))
575 			return -EFAULT;
576 	}
577 	return 0;
578 }
579 
580 static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
581 {
582 	int new_page;
583 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
584 	int i;
585 	int size = PAGE_SIZE;
586 	bool copy1 = (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) ?
587 		false : true;
588 
589 	for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
590 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
591 				       ibc->user_ptr + (i * PAGE_SIZE),
592 				       PAGE_SIZE)) {
593 			p->parser_error = -EFAULT;
594 			return 0;
595 		}
596 	}
597 
598 	if (pg_idx == ibc->last_page_index) {
599 		size = (ibc->length_dw * 4) % PAGE_SIZE;
600 		if (size == 0)
601 			size = PAGE_SIZE;
602 	}
603 
604 	new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
605 	if (copy1)
606 		ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
607 
608 	if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
609 			       ibc->user_ptr + (pg_idx * PAGE_SIZE),
610 			       size)) {
611 		p->parser_error = -EFAULT;
612 		return 0;
613 	}
614 
615 	/* copy to IB for non single case */
616 	if (!copy1)
617 		memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
618 
619 	ibc->last_copied_page = pg_idx;
620 	ibc->kpage_idx[new_page] = pg_idx;
621 
622 	return new_page;
623 }
624 
625 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
626 {
627 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
628 	u32 pg_idx, pg_offset;
629 	u32 idx_value = 0;
630 	int new_page;
631 
632 	pg_idx = (idx * 4) / PAGE_SIZE;
633 	pg_offset = (idx * 4) % PAGE_SIZE;
634 
635 	if (ibc->kpage_idx[0] == pg_idx)
636 		return ibc->kpage[0][pg_offset/4];
637 	if (ibc->kpage_idx[1] == pg_idx)
638 		return ibc->kpage[1][pg_offset/4];
639 
640 	new_page = radeon_cs_update_pages(p, pg_idx);
641 	if (new_page < 0) {
642 		p->parser_error = new_page;
643 		return 0;
644 	}
645 
646 	idx_value = ibc->kpage[new_page][pg_offset/4];
647 	return idx_value;
648 }
649