xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision ee8a99bd)
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].handle = 0;
68 			continue;
69 		}
70 
71 		p->relocs[i].gobj = drm_gem_object_lookup(ddev, 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.written = !!r->write_domain;
82 
83 		/* the first reloc of an UVD job is the
84 		   msg and that must be in VRAM */
85 		if (p->ring == R600_RING_TYPE_UVD_INDEX && i == 0) {
86 			/* TODO: is this still needed for NI+ ? */
87 			p->relocs[i].lobj.domain =
88 				RADEON_GEM_DOMAIN_VRAM;
89 
90 			p->relocs[i].lobj.alt_domain =
91 				RADEON_GEM_DOMAIN_VRAM;
92 
93 		} else {
94 			uint32_t domain = r->write_domain ?
95 				r->write_domain : r->read_domains;
96 
97 			p->relocs[i].lobj.domain = domain;
98 			if (domain == RADEON_GEM_DOMAIN_VRAM)
99 				domain |= RADEON_GEM_DOMAIN_GTT;
100 			p->relocs[i].lobj.alt_domain = domain;
101 		}
102 
103 		p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
104 		p->relocs[i].handle = r->handle;
105 
106 		radeon_bo_list_add_object(&p->relocs[i].lobj,
107 					  &p->validated);
108 	}
109 	return radeon_bo_list_validate(&p->ticket, &p->validated, p->ring);
110 }
111 
112 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
113 {
114 	p->priority = priority;
115 
116 	switch (ring) {
117 	default:
118 		DRM_ERROR("unknown ring id: %d\n", ring);
119 		return -EINVAL;
120 	case RADEON_CS_RING_GFX:
121 		p->ring = RADEON_RING_TYPE_GFX_INDEX;
122 		break;
123 	case RADEON_CS_RING_COMPUTE:
124 		if (p->rdev->family >= CHIP_TAHITI) {
125 			if (p->priority > 0)
126 				p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
127 			else
128 				p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
129 		} else
130 			p->ring = RADEON_RING_TYPE_GFX_INDEX;
131 		break;
132 	case RADEON_CS_RING_DMA:
133 		if (p->rdev->family >= CHIP_CAYMAN) {
134 			if (p->priority > 0)
135 				p->ring = R600_RING_TYPE_DMA_INDEX;
136 			else
137 				p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
138 		} else if (p->rdev->family >= CHIP_R600) {
139 			p->ring = R600_RING_TYPE_DMA_INDEX;
140 		} else {
141 			return -EINVAL;
142 		}
143 		break;
144 	case RADEON_CS_RING_UVD:
145 		p->ring = R600_RING_TYPE_UVD_INDEX;
146 		break;
147 	}
148 	return 0;
149 }
150 
151 static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
152 {
153 	int i;
154 
155 	for (i = 0; i < p->nrelocs; i++) {
156 		if (!p->relocs[i].robj)
157 			continue;
158 
159 		radeon_ib_sync_to(&p->ib, p->relocs[i].robj->tbo.sync_obj);
160 	}
161 }
162 
163 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
164 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
165 {
166 	struct drm_radeon_cs *cs = data;
167 	uint64_t *chunk_array_ptr;
168 	unsigned size, i;
169 	u32 ring = RADEON_CS_RING_GFX;
170 	s32 priority = 0;
171 
172 	if (!cs->num_chunks) {
173 		return 0;
174 	}
175 	/* get chunks */
176 	INIT_LIST_HEAD(&p->validated);
177 	p->idx = 0;
178 	p->ib.sa_bo = NULL;
179 	p->ib.semaphore = NULL;
180 	p->const_ib.sa_bo = NULL;
181 	p->const_ib.semaphore = NULL;
182 	p->chunk_ib_idx = -1;
183 	p->chunk_relocs_idx = -1;
184 	p->chunk_flags_idx = -1;
185 	p->chunk_const_ib_idx = -1;
186 	p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
187 	if (p->chunks_array == NULL) {
188 		return -ENOMEM;
189 	}
190 	chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
191 	if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
192 			       sizeof(uint64_t)*cs->num_chunks)) {
193 		return -EFAULT;
194 	}
195 	p->cs_flags = 0;
196 	p->nchunks = cs->num_chunks;
197 	p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
198 	if (p->chunks == NULL) {
199 		return -ENOMEM;
200 	}
201 	for (i = 0; i < p->nchunks; i++) {
202 		struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
203 		struct drm_radeon_cs_chunk user_chunk;
204 		uint32_t __user *cdata;
205 
206 		chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
207 		if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
208 				       sizeof(struct drm_radeon_cs_chunk))) {
209 			return -EFAULT;
210 		}
211 		p->chunks[i].length_dw = user_chunk.length_dw;
212 		p->chunks[i].kdata = NULL;
213 		p->chunks[i].chunk_id = user_chunk.chunk_id;
214 		p->chunks[i].user_ptr = (void __user *)(unsigned long)user_chunk.chunk_data;
215 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
216 			p->chunk_relocs_idx = i;
217 		}
218 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
219 			p->chunk_ib_idx = i;
220 			/* zero length IB isn't useful */
221 			if (p->chunks[i].length_dw == 0)
222 				return -EINVAL;
223 		}
224 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
225 			p->chunk_const_ib_idx = i;
226 			/* zero length CONST IB isn't useful */
227 			if (p->chunks[i].length_dw == 0)
228 				return -EINVAL;
229 		}
230 		if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
231 			p->chunk_flags_idx = i;
232 			/* zero length flags aren't useful */
233 			if (p->chunks[i].length_dw == 0)
234 				return -EINVAL;
235 		}
236 
237 		cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
238 		if ((p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) ||
239 		    (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS)) {
240 			size = p->chunks[i].length_dw * sizeof(uint32_t);
241 			p->chunks[i].kdata = kmalloc(size, GFP_KERNEL);
242 			if (p->chunks[i].kdata == NULL) {
243 				return -ENOMEM;
244 			}
245 			if (DRM_COPY_FROM_USER(p->chunks[i].kdata,
246 					       p->chunks[i].user_ptr, size)) {
247 				return -EFAULT;
248 			}
249 			if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
250 				p->cs_flags = p->chunks[i].kdata[0];
251 				if (p->chunks[i].length_dw > 1)
252 					ring = p->chunks[i].kdata[1];
253 				if (p->chunks[i].length_dw > 2)
254 					priority = (s32)p->chunks[i].kdata[2];
255 			}
256 		}
257 	}
258 
259 	/* these are KMS only */
260 	if (p->rdev) {
261 		if ((p->cs_flags & RADEON_CS_USE_VM) &&
262 		    !p->rdev->vm_manager.enabled) {
263 			DRM_ERROR("VM not active on asic!\n");
264 			return -EINVAL;
265 		}
266 
267 		if (radeon_cs_get_ring(p, ring, priority))
268 			return -EINVAL;
269 
270 		/* we only support VM on some SI+ rings */
271 		if ((p->rdev->asic->ring[p->ring].cs_parse == NULL) &&
272 		   ((p->cs_flags & RADEON_CS_USE_VM) == 0)) {
273 			DRM_ERROR("Ring %d requires VM!\n", p->ring);
274 			return -EINVAL;
275 		}
276 	}
277 
278 	/* deal with non-vm */
279 	if ((p->chunk_ib_idx != -1) &&
280 	    ((p->cs_flags & RADEON_CS_USE_VM) == 0) &&
281 	    (p->chunks[p->chunk_ib_idx].chunk_id == RADEON_CHUNK_ID_IB)) {
282 		if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
283 			DRM_ERROR("cs IB too big: %d\n",
284 				  p->chunks[p->chunk_ib_idx].length_dw);
285 			return -EINVAL;
286 		}
287 		if (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) {
288 			p->chunks[p->chunk_ib_idx].kpage[0] = kmalloc(PAGE_SIZE, GFP_KERNEL);
289 			p->chunks[p->chunk_ib_idx].kpage[1] = kmalloc(PAGE_SIZE, GFP_KERNEL);
290 			if (p->chunks[p->chunk_ib_idx].kpage[0] == NULL ||
291 			    p->chunks[p->chunk_ib_idx].kpage[1] == NULL) {
292 				kfree(p->chunks[p->chunk_ib_idx].kpage[0]);
293 				kfree(p->chunks[p->chunk_ib_idx].kpage[1]);
294 				p->chunks[p->chunk_ib_idx].kpage[0] = NULL;
295 				p->chunks[p->chunk_ib_idx].kpage[1] = NULL;
296 				return -ENOMEM;
297 			}
298 		}
299 		p->chunks[p->chunk_ib_idx].kpage_idx[0] = -1;
300 		p->chunks[p->chunk_ib_idx].kpage_idx[1] = -1;
301 		p->chunks[p->chunk_ib_idx].last_copied_page = -1;
302 		p->chunks[p->chunk_ib_idx].last_page_index =
303 			((p->chunks[p->chunk_ib_idx].length_dw * 4) - 1) / PAGE_SIZE;
304 	}
305 
306 	return 0;
307 }
308 
309 /**
310  * cs_parser_fini() - clean parser states
311  * @parser:	parser structure holding parsing context.
312  * @error:	error number
313  *
314  * If error is set than unvalidate buffer, otherwise just free memory
315  * used by parsing context.
316  **/
317 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
318 {
319 	unsigned i;
320 
321 	if (!error) {
322 		ttm_eu_fence_buffer_objects(&parser->ticket,
323 					    &parser->validated,
324 					    parser->ib.fence);
325 	} else if (backoff) {
326 		ttm_eu_backoff_reservation(&parser->ticket,
327 					   &parser->validated);
328 	}
329 
330 	if (parser->relocs != NULL) {
331 		for (i = 0; i < parser->nrelocs; i++) {
332 			if (parser->relocs[i].gobj)
333 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
334 		}
335 	}
336 	kfree(parser->track);
337 	kfree(parser->relocs);
338 	kfree(parser->relocs_ptr);
339 	for (i = 0; i < parser->nchunks; i++) {
340 		kfree(parser->chunks[i].kdata);
341 		if ((parser->rdev->flags & RADEON_IS_AGP)) {
342 			kfree(parser->chunks[i].kpage[0]);
343 			kfree(parser->chunks[i].kpage[1]);
344 		}
345 	}
346 	kfree(parser->chunks);
347 	kfree(parser->chunks_array);
348 	radeon_ib_free(parser->rdev, &parser->ib);
349 	radeon_ib_free(parser->rdev, &parser->const_ib);
350 }
351 
352 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
353 			      struct radeon_cs_parser *parser)
354 {
355 	struct radeon_cs_chunk *ib_chunk;
356 	int r;
357 
358 	if (parser->chunk_ib_idx == -1)
359 		return 0;
360 
361 	if (parser->cs_flags & RADEON_CS_USE_VM)
362 		return 0;
363 
364 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
365 	/* Copy the packet into the IB, the parser will read from the
366 	 * input memory (cached) and write to the IB (which can be
367 	 * uncached).
368 	 */
369 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
370 			   NULL, ib_chunk->length_dw * 4);
371 	if (r) {
372 		DRM_ERROR("Failed to get ib !\n");
373 		return r;
374 	}
375 	parser->ib.length_dw = ib_chunk->length_dw;
376 	r = radeon_cs_parse(rdev, parser->ring, parser);
377 	if (r || parser->parser_error) {
378 		DRM_ERROR("Invalid command stream !\n");
379 		return r;
380 	}
381 	r = radeon_cs_finish_pages(parser);
382 	if (r) {
383 		DRM_ERROR("Invalid command stream !\n");
384 		return r;
385 	}
386 	radeon_cs_sync_rings(parser);
387 	r = radeon_ib_schedule(rdev, &parser->ib, NULL);
388 	if (r) {
389 		DRM_ERROR("Failed to schedule IB !\n");
390 	}
391 	return r;
392 }
393 
394 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
395 				   struct radeon_vm *vm)
396 {
397 	struct radeon_device *rdev = parser->rdev;
398 	struct radeon_bo_list *lobj;
399 	struct radeon_bo *bo;
400 	int r;
401 
402 	r = radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo, &rdev->ring_tmp_bo.bo->tbo.mem);
403 	if (r) {
404 		return r;
405 	}
406 	list_for_each_entry(lobj, &parser->validated, tv.head) {
407 		bo = lobj->bo;
408 		r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
409 		if (r) {
410 			return r;
411 		}
412 	}
413 	return 0;
414 }
415 
416 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
417 				 struct radeon_cs_parser *parser)
418 {
419 	struct radeon_cs_chunk *ib_chunk;
420 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
421 	struct radeon_vm *vm = &fpriv->vm;
422 	int r;
423 
424 	if (parser->chunk_ib_idx == -1)
425 		return 0;
426 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
427 		return 0;
428 
429 	if ((rdev->family >= CHIP_TAHITI) &&
430 	    (parser->chunk_const_ib_idx != -1)) {
431 		ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
432 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
433 			DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
434 			return -EINVAL;
435 		}
436 		r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
437 				   vm, ib_chunk->length_dw * 4);
438 		if (r) {
439 			DRM_ERROR("Failed to get const ib !\n");
440 			return r;
441 		}
442 		parser->const_ib.is_const_ib = true;
443 		parser->const_ib.length_dw = ib_chunk->length_dw;
444 		/* Copy the packet into the IB */
445 		if (DRM_COPY_FROM_USER(parser->const_ib.ptr, ib_chunk->user_ptr,
446 				       ib_chunk->length_dw * 4)) {
447 			return -EFAULT;
448 		}
449 		r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
450 		if (r) {
451 			return r;
452 		}
453 	}
454 
455 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
456 	if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
457 		DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
458 		return -EINVAL;
459 	}
460 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
461 			   vm, ib_chunk->length_dw * 4);
462 	if (r) {
463 		DRM_ERROR("Failed to get ib !\n");
464 		return r;
465 	}
466 	parser->ib.length_dw = ib_chunk->length_dw;
467 	/* Copy the packet into the IB */
468 	if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
469 			       ib_chunk->length_dw * 4)) {
470 		return -EFAULT;
471 	}
472 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
473 	if (r) {
474 		return r;
475 	}
476 
477 	mutex_lock(&rdev->vm_manager.lock);
478 	mutex_lock(&vm->mutex);
479 	r = radeon_vm_alloc_pt(rdev, vm);
480 	if (r) {
481 		goto out;
482 	}
483 	r = radeon_bo_vm_update_pte(parser, vm);
484 	if (r) {
485 		goto out;
486 	}
487 	radeon_cs_sync_rings(parser);
488 	radeon_ib_sync_to(&parser->ib, vm->fence);
489 	radeon_ib_sync_to(&parser->ib, radeon_vm_grab_id(
490 		rdev, vm, parser->ring));
491 
492 	if ((rdev->family >= CHIP_TAHITI) &&
493 	    (parser->chunk_const_ib_idx != -1)) {
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 	if (!r) {
500 		radeon_vm_fence(rdev, vm, parser->ib.fence);
501 	}
502 
503 out:
504 	radeon_vm_add_to_lru(rdev, vm);
505 	mutex_unlock(&vm->mutex);
506 	mutex_unlock(&rdev->vm_manager.lock);
507 	return r;
508 }
509 
510 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
511 {
512 	if (r == -EDEADLK) {
513 		r = radeon_gpu_reset(rdev);
514 		if (!r)
515 			r = -EAGAIN;
516 	}
517 	return r;
518 }
519 
520 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
521 {
522 	struct radeon_device *rdev = dev->dev_private;
523 	struct radeon_cs_parser parser;
524 	int r;
525 
526 	down_read(&rdev->exclusive_lock);
527 	if (!rdev->accel_working) {
528 		up_read(&rdev->exclusive_lock);
529 		return -EBUSY;
530 	}
531 	/* initialize parser */
532 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
533 	parser.filp = filp;
534 	parser.rdev = rdev;
535 	parser.dev = rdev->dev;
536 	parser.family = rdev->family;
537 	r = radeon_cs_parser_init(&parser, data);
538 	if (r) {
539 		DRM_ERROR("Failed to initialize parser !\n");
540 		radeon_cs_parser_fini(&parser, r, false);
541 		up_read(&rdev->exclusive_lock);
542 		r = radeon_cs_handle_lockup(rdev, r);
543 		return r;
544 	}
545 	r = radeon_cs_parser_relocs(&parser);
546 	if (r) {
547 		if (r != -ERESTARTSYS)
548 			DRM_ERROR("Failed to parse relocation %d!\n", r);
549 		radeon_cs_parser_fini(&parser, r, false);
550 		up_read(&rdev->exclusive_lock);
551 		r = radeon_cs_handle_lockup(rdev, r);
552 		return r;
553 	}
554 
555 	/* XXX pick SD/HD/MVC */
556 	if (parser.ring == R600_RING_TYPE_UVD_INDEX)
557 		radeon_uvd_note_usage(rdev);
558 
559 	r = radeon_cs_ib_chunk(rdev, &parser);
560 	if (r) {
561 		goto out;
562 	}
563 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
564 	if (r) {
565 		goto out;
566 	}
567 out:
568 	radeon_cs_parser_fini(&parser, r, true);
569 	up_read(&rdev->exclusive_lock);
570 	r = radeon_cs_handle_lockup(rdev, r);
571 	return r;
572 }
573 
574 int radeon_cs_finish_pages(struct radeon_cs_parser *p)
575 {
576 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
577 	int i;
578 	int size = PAGE_SIZE;
579 
580 	for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
581 		if (i == ibc->last_page_index) {
582 			size = (ibc->length_dw * 4) % PAGE_SIZE;
583 			if (size == 0)
584 				size = PAGE_SIZE;
585 		}
586 
587 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
588 				       ibc->user_ptr + (i * PAGE_SIZE),
589 				       size))
590 			return -EFAULT;
591 	}
592 	return 0;
593 }
594 
595 static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
596 {
597 	int new_page;
598 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
599 	int i;
600 	int size = PAGE_SIZE;
601 	bool copy1 = (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) ?
602 		false : true;
603 
604 	for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
605 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
606 				       ibc->user_ptr + (i * PAGE_SIZE),
607 				       PAGE_SIZE)) {
608 			p->parser_error = -EFAULT;
609 			return 0;
610 		}
611 	}
612 
613 	if (pg_idx == ibc->last_page_index) {
614 		size = (ibc->length_dw * 4) % PAGE_SIZE;
615 		if (size == 0)
616 			size = PAGE_SIZE;
617 	}
618 
619 	new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
620 	if (copy1)
621 		ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
622 
623 	if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
624 			       ibc->user_ptr + (pg_idx * PAGE_SIZE),
625 			       size)) {
626 		p->parser_error = -EFAULT;
627 		return 0;
628 	}
629 
630 	/* copy to IB for non single case */
631 	if (!copy1)
632 		memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
633 
634 	ibc->last_copied_page = pg_idx;
635 	ibc->kpage_idx[new_page] = pg_idx;
636 
637 	return new_page;
638 }
639 
640 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
641 {
642 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
643 	u32 pg_idx, pg_offset;
644 	u32 idx_value = 0;
645 	int new_page;
646 
647 	pg_idx = (idx * 4) / PAGE_SIZE;
648 	pg_offset = (idx * 4) % PAGE_SIZE;
649 
650 	if (ibc->kpage_idx[0] == pg_idx)
651 		return ibc->kpage[0][pg_offset/4];
652 	if (ibc->kpage_idx[1] == pg_idx)
653 		return ibc->kpage[1][pg_offset/4];
654 
655 	new_page = radeon_cs_update_pages(p, pg_idx);
656 	if (new_page < 0) {
657 		p->parser_error = new_page;
658 		return 0;
659 	}
660 
661 	idx_value = ibc->kpage[new_page][pg_offset/4];
662 	return idx_value;
663 }
664 
665 /**
666  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
667  * @parser:	parser structure holding parsing context.
668  * @pkt:	where to store packet information
669  *
670  * Assume that chunk_ib_index is properly set. Will return -EINVAL
671  * if packet is bigger than remaining ib size. or if packets is unknown.
672  **/
673 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
674 			   struct radeon_cs_packet *pkt,
675 			   unsigned idx)
676 {
677 	struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
678 	struct radeon_device *rdev = p->rdev;
679 	uint32_t header;
680 
681 	if (idx >= ib_chunk->length_dw) {
682 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
683 			  idx, ib_chunk->length_dw);
684 		return -EINVAL;
685 	}
686 	header = radeon_get_ib_value(p, idx);
687 	pkt->idx = idx;
688 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
689 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
690 	pkt->one_reg_wr = 0;
691 	switch (pkt->type) {
692 	case RADEON_PACKET_TYPE0:
693 		if (rdev->family < CHIP_R600) {
694 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
695 			pkt->one_reg_wr =
696 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
697 		} else
698 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
699 		break;
700 	case RADEON_PACKET_TYPE3:
701 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
702 		break;
703 	case RADEON_PACKET_TYPE2:
704 		pkt->count = -1;
705 		break;
706 	default:
707 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
708 		return -EINVAL;
709 	}
710 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
711 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
712 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
713 		return -EINVAL;
714 	}
715 	return 0;
716 }
717 
718 /**
719  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
720  * @p:		structure holding the parser context.
721  *
722  * Check if the next packet is NOP relocation packet3.
723  **/
724 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
725 {
726 	struct radeon_cs_packet p3reloc;
727 	int r;
728 
729 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
730 	if (r)
731 		return false;
732 	if (p3reloc.type != RADEON_PACKET_TYPE3)
733 		return false;
734 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
735 		return false;
736 	return true;
737 }
738 
739 /**
740  * radeon_cs_dump_packet() - dump raw packet context
741  * @p:		structure holding the parser context.
742  * @pkt:	structure holding the packet.
743  *
744  * Used mostly for debugging and error reporting.
745  **/
746 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
747 			   struct radeon_cs_packet *pkt)
748 {
749 	volatile uint32_t *ib;
750 	unsigned i;
751 	unsigned idx;
752 
753 	ib = p->ib.ptr;
754 	idx = pkt->idx;
755 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
756 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
757 }
758 
759 /**
760  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
761  * @parser:		parser structure holding parsing context.
762  * @data:		pointer to relocation data
763  * @offset_start:	starting offset
764  * @offset_mask:	offset mask (to align start offset on)
765  * @reloc:		reloc informations
766  *
767  * Check if next packet is relocation packet3, do bo validation and compute
768  * GPU offset using the provided start.
769  **/
770 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
771 				struct radeon_cs_reloc **cs_reloc,
772 				int nomm)
773 {
774 	struct radeon_cs_chunk *relocs_chunk;
775 	struct radeon_cs_packet p3reloc;
776 	unsigned idx;
777 	int r;
778 
779 	if (p->chunk_relocs_idx == -1) {
780 		DRM_ERROR("No relocation chunk !\n");
781 		return -EINVAL;
782 	}
783 	*cs_reloc = NULL;
784 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
785 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
786 	if (r)
787 		return r;
788 	p->idx += p3reloc.count + 2;
789 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
790 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
791 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
792 			  p3reloc.idx);
793 		radeon_cs_dump_packet(p, &p3reloc);
794 		return -EINVAL;
795 	}
796 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
797 	if (idx >= relocs_chunk->length_dw) {
798 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
799 			  idx, relocs_chunk->length_dw);
800 		radeon_cs_dump_packet(p, &p3reloc);
801 		return -EINVAL;
802 	}
803 	/* FIXME: we assume reloc size is 4 dwords */
804 	if (nomm) {
805 		*cs_reloc = p->relocs;
806 		(*cs_reloc)->lobj.gpu_offset =
807 			(u64)relocs_chunk->kdata[idx + 3] << 32;
808 		(*cs_reloc)->lobj.gpu_offset |= relocs_chunk->kdata[idx + 0];
809 	} else
810 		*cs_reloc = p->relocs_ptr[(idx / 4)];
811 	return 0;
812 }
813