xref: /dragonfly/sys/dev/drm/radeon/radeon_cs.c (revision d4ef6694)
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  * $FreeBSD: head/sys/dev/drm2/radeon/radeon_cs.c 254885 2013-08-25 19:37:15Z dumbbell $
28  */
29 
30 #include <drm/drmP.h>
31 #include <uapi_drm/radeon_drm.h>
32 #include "radeon_reg.h"
33 #include "radeon.h"
34 
35 void r100_cs_dump_packet(struct radeon_cs_parser *p,
36 			 struct radeon_cs_packet *pkt);
37 
38 static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
39 {
40 	struct drm_device *ddev = p->rdev->ddev;
41 	struct radeon_cs_chunk *chunk;
42 	unsigned i, j;
43 	bool duplicate;
44 
45 	if (p->chunk_relocs_idx == -1) {
46 		return 0;
47 	}
48 	chunk = &p->chunks[p->chunk_relocs_idx];
49 	p->dma_reloc_idx = 0;
50 	/* FIXME: we assume that each relocs use 4 dwords */
51 	p->nrelocs = chunk->length_dw / 4;
52 	p->relocs_ptr = kmalloc(p->nrelocs * sizeof(void *), M_DRM,
53 				M_ZERO | M_WAITOK);
54 	if (p->relocs_ptr == NULL) {
55 		return -ENOMEM;
56 	}
57 	p->relocs = kmalloc(p->nrelocs * sizeof(struct radeon_cs_reloc),
58 			    M_DRM, M_ZERO | M_WAITOK);
59 	if (p->relocs == NULL) {
60 		return -ENOMEM;
61 	}
62 	for (i = 0; i < p->nrelocs; i++) {
63 		struct drm_radeon_cs_reloc *r;
64 
65 		duplicate = false;
66 		r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
67 		for (j = 0; j < i; j++) {
68 			if (r->handle == p->relocs[j].handle) {
69 				p->relocs_ptr[i] = &p->relocs[j];
70 				duplicate = true;
71 				break;
72 			}
73 		}
74 		if (!duplicate) {
75 			p->relocs[i].gobj = drm_gem_object_lookup(ddev,
76 								  p->filp,
77 								  r->handle);
78 			if (p->relocs[i].gobj == NULL) {
79 				DRM_ERROR("gem object lookup failed 0x%x\n",
80 					  r->handle);
81 				return -ENOENT;
82 			}
83 			p->relocs_ptr[i] = &p->relocs[i];
84 			p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
85 			p->relocs[i].lobj.bo = p->relocs[i].robj;
86 			p->relocs[i].lobj.wdomain = r->write_domain;
87 			p->relocs[i].lobj.rdomain = r->read_domains;
88 			p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
89 			p->relocs[i].handle = r->handle;
90 			p->relocs[i].flags = r->flags;
91 			radeon_bo_list_add_object(&p->relocs[i].lobj,
92 						  &p->validated);
93 
94 		} else
95 			p->relocs[i].handle = 0;
96 	}
97 	return radeon_bo_list_validate(&p->validated);
98 }
99 
100 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
101 {
102 	p->priority = priority;
103 
104 	switch (ring) {
105 	default:
106 		DRM_ERROR("unknown ring id: %d\n", ring);
107 		return -EINVAL;
108 	case RADEON_CS_RING_GFX:
109 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
110 		break;
111 	case RADEON_CS_RING_COMPUTE:
112 		if (p->rdev->family >= CHIP_TAHITI) {
113 			if (p->priority > 0)
114 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
115 			else
116 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
117 		} else
118 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
119 		break;
120 	case RADEON_CS_RING_DMA:
121 		if (p->rdev->family >= CHIP_CAYMAN) {
122 			if (p->priority > 0)
123 				p->ring = R600_RING_TYPE_DMA_INDEX;
124 			else
125 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
126 		} else if (p->rdev->family >= CHIP_R600) {
127 			p->ring = R600_RING_TYPE_DMA_INDEX;
128 		} else {
129 			return -EINVAL;
130 		}
131 		break;
132 	}
133 	return 0;
134 }
135 
136 static void radeon_cs_sync_to(struct radeon_cs_parser *p,
137 			      struct radeon_fence *fence)
138 {
139 	struct radeon_fence *other;
140 
141 	if (!fence)
142 		return;
143 
144 	other = p->ib.sync_to[fence->ring];
145 	p->ib.sync_to[fence->ring] = radeon_fence_later(fence, other);
146 }
147 
148 static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
149 {
150 	int i;
151 
152 	for (i = 0; i < p->nrelocs; i++) {
153 		if (!p->relocs[i].robj)
154 			continue;
155 
156 		radeon_cs_sync_to(p, p->relocs[i].robj->tbo.sync_obj);
157 	}
158 }
159 
160 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
161 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
162 {
163 	struct drm_radeon_cs *cs = data;
164 	uint64_t *chunk_array_ptr;
165 	unsigned size, i;
166 	u32 ring = RADEON_CS_RING_GFX;
167 	s32 priority = 0;
168 
169 	if (!cs->num_chunks) {
170 		return 0;
171 	}
172 	/* get chunks */
173 	INIT_LIST_HEAD(&p->validated);
174 	p->idx = 0;
175 	p->ib.sa_bo = NULL;
176 	p->ib.semaphore = NULL;
177 	p->const_ib.sa_bo = NULL;
178 	p->const_ib.semaphore = NULL;
179 	p->chunk_ib_idx = -1;
180 	p->chunk_relocs_idx = -1;
181 	p->chunk_flags_idx = -1;
182 	p->chunk_const_ib_idx = -1;
183 	p->chunks_array = kmalloc(cs->num_chunks * sizeof(uint64_t),
184 				  M_DRM, M_ZERO | M_WAITOK);
185 	if (p->chunks_array == NULL) {
186 		return -ENOMEM;
187 	}
188 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
189 	if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
190 			       sizeof(uint64_t)*cs->num_chunks)) {
191 		return -EFAULT;
192 	}
193 	p->cs_flags = 0;
194 	p->nchunks = cs->num_chunks;
195 	p->chunks = kmalloc(p->nchunks * sizeof(struct radeon_cs_chunk),
196 			    M_DRM, M_ZERO | M_WAITOK);
197 	if (p->chunks == NULL) {
198 		return -ENOMEM;
199 	}
200 	for (i = 0; i < p->nchunks; i++) {
201 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
202 		struct drm_radeon_cs_chunk user_chunk;
203 		uint32_t __user *cdata;
204 
205 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
206 		if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
207 				       sizeof(struct drm_radeon_cs_chunk))) {
208 			return -EFAULT;
209 		}
210 		p->chunks[i].length_dw = user_chunk.length_dw;
211 		p->chunks[i].kdata = NULL;
212 		p->chunks[i].chunk_id = user_chunk.chunk_id;
213 
214 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
215 			p->chunk_relocs_idx = i;
216 		}
217 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
218 			p->chunk_ib_idx = i;
219 			/* zero length IB isn't useful */
220 			if (p->chunks[i].length_dw == 0)
221 				return -EINVAL;
222 		}
223 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
224 			p->chunk_const_ib_idx = i;
225 			/* zero length CONST IB isn't useful */
226 			if (p->chunks[i].length_dw == 0)
227 				return -EINVAL;
228 		}
229 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
230 			p->chunk_flags_idx = i;
231 			/* zero length flags aren't useful */
232 			if (p->chunks[i].length_dw == 0)
233 				return -EINVAL;
234 		}
235 
236 		p->chunks[i].length_dw = user_chunk.length_dw;
237 		p->chunks[i].user_ptr = (void __user *)(unsigned long)user_chunk.chunk_data;
238 
239 		cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
240 		if ((p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) ||
241 		    (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS)) {
242 			size = p->chunks[i].length_dw * sizeof(uint32_t);
243 			p->chunks[i].kdata = kmalloc(size, M_DRM,
244 						     M_WAITOK);
245 			if (p->chunks[i].kdata == NULL) {
246 				return -ENOMEM;
247 			}
248 			if (DRM_COPY_FROM_USER(p->chunks[i].kdata,
249 					       p->chunks[i].user_ptr, size)) {
250 				return -EFAULT;
251 			}
252 			if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
253 				p->cs_flags = p->chunks[i].kdata[0];
254 				if (p->chunks[i].length_dw > 1)
255 					ring = p->chunks[i].kdata[1];
256 				if (p->chunks[i].length_dw > 2)
257 					priority = (s32)p->chunks[i].kdata[2];
258 			}
259 		}
260 	}
261 
262 	/* these are KMS only */
263 	if (p->rdev) {
264 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
265 		    !p->rdev->vm_manager.enabled) {
266 			DRM_ERROR("VM not active on asic!\n");
267 			return -EINVAL;
268 		}
269 
270 		/* we only support VM on SI+ */
271 		if ((p->rdev->family >= CHIP_TAHITI) &&
272 		    ((p->cs_flags & RADEON_CS_USE_VM) == 0)) {
273 			DRM_ERROR("VM required on SI+!\n");
274 			return -EINVAL;
275 		}
276 
277 		if (radeon_cs_get_ring(p, ring, priority))
278 			return -EINVAL;
279 	}
280 
281 	/* deal with non-vm */
282 	if ((p->chunk_ib_idx != -1) &&
283 	    ((p->cs_flags & RADEON_CS_USE_VM) == 0) &&
284 	    (p->chunks[p->chunk_ib_idx].chunk_id == RADEON_CHUNK_ID_IB)) {
285 		if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
286 			DRM_ERROR("cs IB too big: %d\n",
287 				  p->chunks[p->chunk_ib_idx].length_dw);
288 			return -EINVAL;
289 		}
290 		if (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) {
291 			p->chunks[p->chunk_ib_idx].kpage[0] = kmalloc(PAGE_SIZE,
292 								      M_DRM,
293 								      M_WAITOK);
294 			p->chunks[p->chunk_ib_idx].kpage[1] = kmalloc(PAGE_SIZE,
295 								      M_DRM,
296 								      M_WAITOK);
297 			if (p->chunks[p->chunk_ib_idx].kpage[0] == NULL ||
298 			    p->chunks[p->chunk_ib_idx].kpage[1] == NULL) {
299 				drm_free(p->chunks[p->chunk_ib_idx].kpage[0],
300 					 M_DRM);
301 				drm_free(p->chunks[p->chunk_ib_idx].kpage[1],
302 					 M_DRM);
303 				p->chunks[p->chunk_ib_idx].kpage[0] = NULL;
304 				p->chunks[p->chunk_ib_idx].kpage[1] = NULL;
305 				return -ENOMEM;
306 			}
307 		}
308 		p->chunks[p->chunk_ib_idx].kpage_idx[0] = -1;
309 		p->chunks[p->chunk_ib_idx].kpage_idx[1] = -1;
310 		p->chunks[p->chunk_ib_idx].last_copied_page = -1;
311 		p->chunks[p->chunk_ib_idx].last_page_index =
312 			((p->chunks[p->chunk_ib_idx].length_dw * 4) - 1) / PAGE_SIZE;
313 	}
314 
315 	return 0;
316 }
317 
318 /**
319  * cs_parser_fini() - clean parser states
320  * @parser:	parser structure holding parsing context.
321  * @error:	error number
322  *
323  * If error is set than unvalidate buffer, otherwise just free memory
324  * used by parsing context.
325  **/
326 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
327 {
328 	unsigned i;
329 
330 	if (!error) {
331 		ttm_eu_fence_buffer_objects(&parser->validated,
332 					    parser->ib.fence);
333 	} else {
334 		ttm_eu_backoff_reservation(&parser->validated);
335 	}
336 
337 	if (parser->relocs != NULL) {
338 		for (i = 0; i < parser->nrelocs; i++) {
339 			if (parser->relocs[i].gobj)
340 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
341 		}
342 	}
343 	drm_free(parser->track, M_DRM);
344 	drm_free(parser->relocs, M_DRM);
345 	drm_free(parser->relocs_ptr, M_DRM);
346 	for (i = 0; i < parser->nchunks; i++) {
347 		drm_free(parser->chunks[i].kdata, M_DRM);
348 		if ((parser->rdev->flags & RADEON_IS_AGP)) {
349 			drm_free(parser->chunks[i].kpage[0], M_DRM);
350 			drm_free(parser->chunks[i].kpage[1], M_DRM);
351 		}
352 	}
353 	drm_free(parser->chunks, M_DRM);
354 	drm_free(parser->chunks_array, M_DRM);
355 	radeon_ib_free(parser->rdev, &parser->ib);
356 	radeon_ib_free(parser->rdev, &parser->const_ib);
357 }
358 
359 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
360 			      struct radeon_cs_parser *parser)
361 {
362 	struct radeon_cs_chunk *ib_chunk;
363 	int r;
364 
365 	if (parser->chunk_ib_idx == -1)
366 		return 0;
367 
368 	if (parser->cs_flags & RADEON_CS_USE_VM)
369 		return 0;
370 
371 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
372 	/* Copy the packet into the IB, the parser will read from the
373 	 * input memory (cached) and write to the IB (which can be
374 	 * uncached).
375 	 */
376 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
377 			   NULL, ib_chunk->length_dw * 4);
378 	if (r) {
379 		DRM_ERROR("Failed to get ib !\n");
380 		return r;
381 	}
382 	parser->ib.length_dw = ib_chunk->length_dw;
383 	r = radeon_cs_parse(rdev, parser->ring, parser);
384 	if (r || parser->parser_error) {
385 		DRM_ERROR("Invalid command stream !\n");
386 		return r;
387 	}
388 	r = radeon_cs_finish_pages(parser);
389 	if (r) {
390 		DRM_ERROR("Invalid command stream !\n");
391 		return r;
392 	}
393 	radeon_cs_sync_rings(parser);
394 	r = radeon_ib_schedule(rdev, &parser->ib, NULL);
395 	if (r) {
396 		DRM_ERROR("Failed to schedule IB !\n");
397 	}
398 	return r;
399 }
400 
401 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
402 				   struct radeon_vm *vm)
403 {
404 	struct radeon_device *rdev = parser->rdev;
405 	struct radeon_bo_list *lobj;
406 	struct radeon_bo *bo;
407 	int r;
408 
409 	r = radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo, &rdev->ring_tmp_bo.bo->tbo.mem);
410 	if (r) {
411 		return r;
412 	}
413 	list_for_each_entry(lobj, &parser->validated, tv.head) {
414 		bo = lobj->bo;
415 		r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
416 		if (r) {
417 			return r;
418 		}
419 	}
420 	return 0;
421 }
422 
423 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
424 				 struct radeon_cs_parser *parser)
425 {
426 	struct radeon_cs_chunk *ib_chunk;
427 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
428 	struct radeon_vm *vm = &fpriv->vm;
429 	int r;
430 
431 	if (parser->chunk_ib_idx == -1)
432 		return 0;
433 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
434 		return 0;
435 
436 	if ((rdev->family >= CHIP_TAHITI) &&
437 	    (parser->chunk_const_ib_idx != -1)) {
438 		ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
439 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
440 			DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
441 			return -EINVAL;
442 		}
443 		r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
444 				   vm, ib_chunk->length_dw * 4);
445 		if (r) {
446 			DRM_ERROR("Failed to get const ib !\n");
447 			return r;
448 		}
449 		parser->const_ib.is_const_ib = true;
450 		parser->const_ib.length_dw = ib_chunk->length_dw;
451 		/* Copy the packet into the IB */
452 		if (DRM_COPY_FROM_USER(parser->const_ib.ptr, ib_chunk->user_ptr,
453 				       ib_chunk->length_dw * 4)) {
454 			return -EFAULT;
455 		}
456 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
457 		if (r) {
458 			return r;
459 		}
460 	}
461 
462 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
463 	if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
464 		DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
465 		return -EINVAL;
466 	}
467 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
468 			   vm, ib_chunk->length_dw * 4);
469 	if (r) {
470 		DRM_ERROR("Failed to get ib !\n");
471 		return r;
472 	}
473 	parser->ib.length_dw = ib_chunk->length_dw;
474 	/* Copy the packet into the IB */
475 	if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
476 			       ib_chunk->length_dw * 4)) {
477 		return -EFAULT;
478 	}
479 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
480 	if (r) {
481 		return r;
482 	}
483 
484 	lockmgr(&rdev->vm_manager.lock, LK_EXCLUSIVE);
485 	lockmgr(&vm->mutex, LK_EXCLUSIVE);
486 	r = radeon_vm_alloc_pt(rdev, vm);
487 	if (r) {
488 		goto out;
489 	}
490 	r = radeon_bo_vm_update_pte(parser, vm);
491 	if (r) {
492 		goto out;
493 	}
494 	radeon_cs_sync_rings(parser);
495 	radeon_cs_sync_to(parser, vm->fence);
496 	radeon_cs_sync_to(parser, radeon_vm_grab_id(rdev, vm, parser->ring));
497 
498 	if ((rdev->family >= CHIP_TAHITI) &&
499 	    (parser->chunk_const_ib_idx != -1)) {
500 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib);
501 	} else {
502 		r = radeon_ib_schedule(rdev, &parser->ib, NULL);
503 	}
504 
505 	if (!r) {
506 		radeon_vm_fence(rdev, vm, parser->ib.fence);
507 	}
508 
509 out:
510 	radeon_vm_add_to_lru(rdev, vm);
511 	lockmgr(&vm->mutex, LK_RELEASE);
512 	lockmgr(&rdev->vm_manager.lock, LK_RELEASE);
513 	return r;
514 }
515 
516 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
517 {
518 	if (r == -EDEADLK) {
519 		r = radeon_gpu_reset(rdev);
520 		if (!r)
521 			r = -EAGAIN;
522 	}
523 	return r;
524 }
525 
526 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
527 {
528 	struct radeon_device *rdev = dev->dev_private;
529 	struct radeon_cs_parser parser;
530 	int r;
531 
532 	lockmgr(&rdev->exclusive_lock, LK_EXCLUSIVE);
533 	if (!rdev->accel_working) {
534 		lockmgr(&rdev->exclusive_lock, LK_RELEASE);
535 		return -EBUSY;
536 	}
537 	/* initialize parser */
538 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
539 	parser.filp = filp;
540 	parser.rdev = rdev;
541 	parser.dev = rdev->dev;
542 	parser.family = rdev->family;
543 	r = radeon_cs_parser_init(&parser, data);
544 	if (r) {
545 		DRM_ERROR("Failed to initialize parser !\n");
546 		radeon_cs_parser_fini(&parser, r);
547 		lockmgr(&rdev->exclusive_lock, LK_RELEASE);
548 		r = radeon_cs_handle_lockup(rdev, r);
549 		return r;
550 	}
551 	r = radeon_cs_parser_relocs(&parser);
552 	if (r) {
553 		if (r != -ERESTARTSYS)
554 			DRM_ERROR("Failed to parse relocation %d!\n", r);
555 		radeon_cs_parser_fini(&parser, r);
556 		lockmgr(&rdev->exclusive_lock, LK_RELEASE);
557 		r = radeon_cs_handle_lockup(rdev, r);
558 		return r;
559 	}
560 	r = radeon_cs_ib_chunk(rdev, &parser);
561 	if (r) {
562 		goto out;
563 	}
564 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
565 	if (r) {
566 		goto out;
567 	}
568 out:
569 	radeon_cs_parser_fini(&parser, r);
570 	lockmgr(&rdev->exclusive_lock, LK_RELEASE);
571 	r = radeon_cs_handle_lockup(rdev, r);
572 	return r;
573 }
574 
575 int radeon_cs_finish_pages(struct radeon_cs_parser *p)
576 {
577 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
578 	int i;
579 	int size = PAGE_SIZE;
580 
581 	for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
582 		if (i == ibc->last_page_index) {
583 			size = (ibc->length_dw * 4) % PAGE_SIZE;
584 			if (size == 0)
585 				size = PAGE_SIZE;
586 		}
587 
588 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
589 				       (char *)ibc->user_ptr + (i * PAGE_SIZE),
590 				       size))
591 			return -EFAULT;
592 	}
593 	return 0;
594 }
595 
596 static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
597 {
598 	int new_page;
599 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
600 	int i;
601 	int size = PAGE_SIZE;
602 	bool copy1 = (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) ?
603 		false : true;
604 
605 	for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
606 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
607 				       (char *)ibc->user_ptr + (i * PAGE_SIZE),
608 				       PAGE_SIZE)) {
609 			p->parser_error = -EFAULT;
610 			return 0;
611 		}
612 	}
613 
614 	if (pg_idx == ibc->last_page_index) {
615 		size = (ibc->length_dw * 4) % PAGE_SIZE;
616 		if (size == 0)
617 			size = PAGE_SIZE;
618 	}
619 
620 	new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
621 	if (copy1)
622 		ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
623 
624 	if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
625 			       (char *)ibc->user_ptr + (pg_idx * PAGE_SIZE),
626 			       size)) {
627 		p->parser_error = -EFAULT;
628 		return 0;
629 	}
630 
631 	/* copy to IB for non single case */
632 	if (!copy1)
633 		memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
634 
635 	ibc->last_copied_page = pg_idx;
636 	ibc->kpage_idx[new_page] = pg_idx;
637 
638 	return new_page;
639 }
640 
641 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
642 {
643 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
644 	u32 pg_idx, pg_offset;
645 	u32 idx_value = 0;
646 	int new_page;
647 
648 	pg_idx = (idx * 4) / PAGE_SIZE;
649 	pg_offset = (idx * 4) % PAGE_SIZE;
650 
651 	if (ibc->kpage_idx[0] == pg_idx)
652 		return ibc->kpage[0][pg_offset/4];
653 	if (ibc->kpage_idx[1] == pg_idx)
654 		return ibc->kpage[1][pg_offset/4];
655 
656 	new_page = radeon_cs_update_pages(p, pg_idx);
657 	if (new_page < 0) {
658 		p->parser_error = new_page;
659 		return 0;
660 	}
661 
662 	idx_value = ibc->kpage[new_page][pg_offset/4];
663 	return idx_value;
664 }
665