xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_cs.c (revision 39b6f3aa)
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->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)
318 {
319 	unsigned i;
320 
321 	if (!error) {
322 		ttm_eu_fence_buffer_objects(&parser->validated,
323 					    parser->ib.fence);
324 	} else {
325 		ttm_eu_backoff_reservation(&parser->validated);
326 	}
327 
328 	if (parser->relocs != NULL) {
329 		for (i = 0; i < parser->nrelocs; i++) {
330 			if (parser->relocs[i].gobj)
331 				drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
332 		}
333 	}
334 	kfree(parser->track);
335 	kfree(parser->relocs);
336 	kfree(parser->relocs_ptr);
337 	for (i = 0; i < parser->nchunks; i++) {
338 		kfree(parser->chunks[i].kdata);
339 		if ((parser->rdev->flags & RADEON_IS_AGP)) {
340 			kfree(parser->chunks[i].kpage[0]);
341 			kfree(parser->chunks[i].kpage[1]);
342 		}
343 	}
344 	kfree(parser->chunks);
345 	kfree(parser->chunks_array);
346 	radeon_ib_free(parser->rdev, &parser->ib);
347 	radeon_ib_free(parser->rdev, &parser->const_ib);
348 }
349 
350 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
351 			      struct radeon_cs_parser *parser)
352 {
353 	struct radeon_cs_chunk *ib_chunk;
354 	int r;
355 
356 	if (parser->chunk_ib_idx == -1)
357 		return 0;
358 
359 	if (parser->cs_flags & RADEON_CS_USE_VM)
360 		return 0;
361 
362 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
363 	/* Copy the packet into the IB, the parser will read from the
364 	 * input memory (cached) and write to the IB (which can be
365 	 * uncached).
366 	 */
367 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
368 			   NULL, ib_chunk->length_dw * 4);
369 	if (r) {
370 		DRM_ERROR("Failed to get ib !\n");
371 		return r;
372 	}
373 	parser->ib.length_dw = ib_chunk->length_dw;
374 	r = radeon_cs_parse(rdev, parser->ring, parser);
375 	if (r || parser->parser_error) {
376 		DRM_ERROR("Invalid command stream !\n");
377 		return r;
378 	}
379 	r = radeon_cs_finish_pages(parser);
380 	if (r) {
381 		DRM_ERROR("Invalid command stream !\n");
382 		return r;
383 	}
384 	radeon_cs_sync_rings(parser);
385 	r = radeon_ib_schedule(rdev, &parser->ib, NULL);
386 	if (r) {
387 		DRM_ERROR("Failed to schedule IB !\n");
388 	}
389 	return r;
390 }
391 
392 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
393 				   struct radeon_vm *vm)
394 {
395 	struct radeon_device *rdev = parser->rdev;
396 	struct radeon_bo_list *lobj;
397 	struct radeon_bo *bo;
398 	int r;
399 
400 	r = radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo, &rdev->ring_tmp_bo.bo->tbo.mem);
401 	if (r) {
402 		return r;
403 	}
404 	list_for_each_entry(lobj, &parser->validated, tv.head) {
405 		bo = lobj->bo;
406 		r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
407 		if (r) {
408 			return r;
409 		}
410 	}
411 	return 0;
412 }
413 
414 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
415 				 struct radeon_cs_parser *parser)
416 {
417 	struct radeon_cs_chunk *ib_chunk;
418 	struct radeon_fpriv *fpriv = parser->filp->driver_priv;
419 	struct radeon_vm *vm = &fpriv->vm;
420 	int r;
421 
422 	if (parser->chunk_ib_idx == -1)
423 		return 0;
424 	if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
425 		return 0;
426 
427 	if ((rdev->family >= CHIP_TAHITI) &&
428 	    (parser->chunk_const_ib_idx != -1)) {
429 		ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
430 		if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
431 			DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
432 			return -EINVAL;
433 		}
434 		r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
435 				   vm, ib_chunk->length_dw * 4);
436 		if (r) {
437 			DRM_ERROR("Failed to get const ib !\n");
438 			return r;
439 		}
440 		parser->const_ib.is_const_ib = true;
441 		parser->const_ib.length_dw = ib_chunk->length_dw;
442 		/* Copy the packet into the IB */
443 		if (DRM_COPY_FROM_USER(parser->const_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->const_ib);
448 		if (r) {
449 			return r;
450 		}
451 	}
452 
453 	ib_chunk = &parser->chunks[parser->chunk_ib_idx];
454 	if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
455 		DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
456 		return -EINVAL;
457 	}
458 	r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
459 			   vm, ib_chunk->length_dw * 4);
460 	if (r) {
461 		DRM_ERROR("Failed to get ib !\n");
462 		return r;
463 	}
464 	parser->ib.length_dw = ib_chunk->length_dw;
465 	/* Copy the packet into the IB */
466 	if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
467 			       ib_chunk->length_dw * 4)) {
468 		return -EFAULT;
469 	}
470 	r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
471 	if (r) {
472 		return r;
473 	}
474 
475 	mutex_lock(&rdev->vm_manager.lock);
476 	mutex_lock(&vm->mutex);
477 	r = radeon_vm_alloc_pt(rdev, vm);
478 	if (r) {
479 		goto out;
480 	}
481 	r = radeon_bo_vm_update_pte(parser, vm);
482 	if (r) {
483 		goto out;
484 	}
485 	radeon_cs_sync_rings(parser);
486 	radeon_ib_sync_to(&parser->ib, vm->fence);
487 	radeon_ib_sync_to(&parser->ib, radeon_vm_grab_id(
488 		rdev, vm, parser->ring));
489 
490 	if ((rdev->family >= CHIP_TAHITI) &&
491 	    (parser->chunk_const_ib_idx != -1)) {
492 		r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib);
493 	} else {
494 		r = radeon_ib_schedule(rdev, &parser->ib, NULL);
495 	}
496 
497 	if (!r) {
498 		radeon_vm_fence(rdev, vm, parser->ib.fence);
499 	}
500 
501 out:
502 	radeon_vm_add_to_lru(rdev, vm);
503 	mutex_unlock(&vm->mutex);
504 	mutex_unlock(&rdev->vm_manager.lock);
505 	return r;
506 }
507 
508 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
509 {
510 	if (r == -EDEADLK) {
511 		r = radeon_gpu_reset(rdev);
512 		if (!r)
513 			r = -EAGAIN;
514 	}
515 	return r;
516 }
517 
518 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
519 {
520 	struct radeon_device *rdev = dev->dev_private;
521 	struct radeon_cs_parser parser;
522 	int r;
523 
524 	down_read(&rdev->exclusive_lock);
525 	if (!rdev->accel_working) {
526 		up_read(&rdev->exclusive_lock);
527 		return -EBUSY;
528 	}
529 	/* initialize parser */
530 	memset(&parser, 0, sizeof(struct radeon_cs_parser));
531 	parser.filp = filp;
532 	parser.rdev = rdev;
533 	parser.dev = rdev->dev;
534 	parser.family = rdev->family;
535 	r = radeon_cs_parser_init(&parser, data);
536 	if (r) {
537 		DRM_ERROR("Failed to initialize parser !\n");
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 	r = radeon_cs_parser_relocs(&parser);
544 	if (r) {
545 		if (r != -ERESTARTSYS)
546 			DRM_ERROR("Failed to parse relocation %d!\n", r);
547 		radeon_cs_parser_fini(&parser, r);
548 		up_read(&rdev->exclusive_lock);
549 		r = radeon_cs_handle_lockup(rdev, r);
550 		return r;
551 	}
552 
553 	if (parser.ring == R600_RING_TYPE_UVD_INDEX)
554 		radeon_uvd_note_usage(rdev);
555 
556 	r = radeon_cs_ib_chunk(rdev, &parser);
557 	if (r) {
558 		goto out;
559 	}
560 	r = radeon_cs_ib_vm_chunk(rdev, &parser);
561 	if (r) {
562 		goto out;
563 	}
564 out:
565 	radeon_cs_parser_fini(&parser, r);
566 	up_read(&rdev->exclusive_lock);
567 	r = radeon_cs_handle_lockup(rdev, r);
568 	return r;
569 }
570 
571 int radeon_cs_finish_pages(struct radeon_cs_parser *p)
572 {
573 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
574 	int i;
575 	int size = PAGE_SIZE;
576 
577 	for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
578 		if (i == ibc->last_page_index) {
579 			size = (ibc->length_dw * 4) % PAGE_SIZE;
580 			if (size == 0)
581 				size = PAGE_SIZE;
582 		}
583 
584 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
585 				       ibc->user_ptr + (i * PAGE_SIZE),
586 				       size))
587 			return -EFAULT;
588 	}
589 	return 0;
590 }
591 
592 static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
593 {
594 	int new_page;
595 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
596 	int i;
597 	int size = PAGE_SIZE;
598 	bool copy1 = (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) ?
599 		false : true;
600 
601 	for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
602 		if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
603 				       ibc->user_ptr + (i * PAGE_SIZE),
604 				       PAGE_SIZE)) {
605 			p->parser_error = -EFAULT;
606 			return 0;
607 		}
608 	}
609 
610 	if (pg_idx == ibc->last_page_index) {
611 		size = (ibc->length_dw * 4) % PAGE_SIZE;
612 		if (size == 0)
613 			size = PAGE_SIZE;
614 	}
615 
616 	new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
617 	if (copy1)
618 		ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
619 
620 	if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
621 			       ibc->user_ptr + (pg_idx * PAGE_SIZE),
622 			       size)) {
623 		p->parser_error = -EFAULT;
624 		return 0;
625 	}
626 
627 	/* copy to IB for non single case */
628 	if (!copy1)
629 		memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
630 
631 	ibc->last_copied_page = pg_idx;
632 	ibc->kpage_idx[new_page] = pg_idx;
633 
634 	return new_page;
635 }
636 
637 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
638 {
639 	struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
640 	u32 pg_idx, pg_offset;
641 	u32 idx_value = 0;
642 	int new_page;
643 
644 	pg_idx = (idx * 4) / PAGE_SIZE;
645 	pg_offset = (idx * 4) % PAGE_SIZE;
646 
647 	if (ibc->kpage_idx[0] == pg_idx)
648 		return ibc->kpage[0][pg_offset/4];
649 	if (ibc->kpage_idx[1] == pg_idx)
650 		return ibc->kpage[1][pg_offset/4];
651 
652 	new_page = radeon_cs_update_pages(p, pg_idx);
653 	if (new_page < 0) {
654 		p->parser_error = new_page;
655 		return 0;
656 	}
657 
658 	idx_value = ibc->kpage[new_page][pg_offset/4];
659 	return idx_value;
660 }
661 
662 /**
663  * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
664  * @parser:	parser structure holding parsing context.
665  * @pkt:	where to store packet information
666  *
667  * Assume that chunk_ib_index is properly set. Will return -EINVAL
668  * if packet is bigger than remaining ib size. or if packets is unknown.
669  **/
670 int radeon_cs_packet_parse(struct radeon_cs_parser *p,
671 			   struct radeon_cs_packet *pkt,
672 			   unsigned idx)
673 {
674 	struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
675 	struct radeon_device *rdev = p->rdev;
676 	uint32_t header;
677 
678 	if (idx >= ib_chunk->length_dw) {
679 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
680 			  idx, ib_chunk->length_dw);
681 		return -EINVAL;
682 	}
683 	header = radeon_get_ib_value(p, idx);
684 	pkt->idx = idx;
685 	pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
686 	pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
687 	pkt->one_reg_wr = 0;
688 	switch (pkt->type) {
689 	case RADEON_PACKET_TYPE0:
690 		if (rdev->family < CHIP_R600) {
691 			pkt->reg = R100_CP_PACKET0_GET_REG(header);
692 			pkt->one_reg_wr =
693 				RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
694 		} else
695 			pkt->reg = R600_CP_PACKET0_GET_REG(header);
696 		break;
697 	case RADEON_PACKET_TYPE3:
698 		pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
699 		break;
700 	case RADEON_PACKET_TYPE2:
701 		pkt->count = -1;
702 		break;
703 	default:
704 		DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
705 		return -EINVAL;
706 	}
707 	if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
708 		DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
709 			  pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
710 		return -EINVAL;
711 	}
712 	return 0;
713 }
714 
715 /**
716  * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
717  * @p:		structure holding the parser context.
718  *
719  * Check if the next packet is NOP relocation packet3.
720  **/
721 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
722 {
723 	struct radeon_cs_packet p3reloc;
724 	int r;
725 
726 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
727 	if (r)
728 		return false;
729 	if (p3reloc.type != RADEON_PACKET_TYPE3)
730 		return false;
731 	if (p3reloc.opcode != RADEON_PACKET3_NOP)
732 		return false;
733 	return true;
734 }
735 
736 /**
737  * radeon_cs_dump_packet() - dump raw packet context
738  * @p:		structure holding the parser context.
739  * @pkt:	structure holding the packet.
740  *
741  * Used mostly for debugging and error reporting.
742  **/
743 void radeon_cs_dump_packet(struct radeon_cs_parser *p,
744 			   struct radeon_cs_packet *pkt)
745 {
746 	volatile uint32_t *ib;
747 	unsigned i;
748 	unsigned idx;
749 
750 	ib = p->ib.ptr;
751 	idx = pkt->idx;
752 	for (i = 0; i <= (pkt->count + 1); i++, idx++)
753 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
754 }
755 
756 /**
757  * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
758  * @parser:		parser structure holding parsing context.
759  * @data:		pointer to relocation data
760  * @offset_start:	starting offset
761  * @offset_mask:	offset mask (to align start offset on)
762  * @reloc:		reloc informations
763  *
764  * Check if next packet is relocation packet3, do bo validation and compute
765  * GPU offset using the provided start.
766  **/
767 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
768 				struct radeon_cs_reloc **cs_reloc,
769 				int nomm)
770 {
771 	struct radeon_cs_chunk *relocs_chunk;
772 	struct radeon_cs_packet p3reloc;
773 	unsigned idx;
774 	int r;
775 
776 	if (p->chunk_relocs_idx == -1) {
777 		DRM_ERROR("No relocation chunk !\n");
778 		return -EINVAL;
779 	}
780 	*cs_reloc = NULL;
781 	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
782 	r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
783 	if (r)
784 		return r;
785 	p->idx += p3reloc.count + 2;
786 	if (p3reloc.type != RADEON_PACKET_TYPE3 ||
787 	    p3reloc.opcode != RADEON_PACKET3_NOP) {
788 		DRM_ERROR("No packet3 for relocation for packet at %d.\n",
789 			  p3reloc.idx);
790 		radeon_cs_dump_packet(p, &p3reloc);
791 		return -EINVAL;
792 	}
793 	idx = radeon_get_ib_value(p, p3reloc.idx + 1);
794 	if (idx >= relocs_chunk->length_dw) {
795 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
796 			  idx, relocs_chunk->length_dw);
797 		radeon_cs_dump_packet(p, &p3reloc);
798 		return -EINVAL;
799 	}
800 	/* FIXME: we assume reloc size is 4 dwords */
801 	if (nomm) {
802 		*cs_reloc = p->relocs;
803 		(*cs_reloc)->lobj.gpu_offset =
804 			(u64)relocs_chunk->kdata[idx + 3] << 32;
805 		(*cs_reloc)->lobj.gpu_offset |= relocs_chunk->kdata[idx + 0];
806 	} else
807 		*cs_reloc = p->relocs_ptr[(idx / 4)];
808 	return 0;
809 }
810