1 /*
2  * Copyright © 2008,2010 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eric Anholt <eric@anholt.net>
25  *    Chris Wilson <chris@chris-wilson.co.uk>
26  *
27  */
28 
29 #include "drmP.h"
30 #include "drm.h"
31 #include "i915_drm.h"
32 #include "i915_drv.h"
33 #include "i915_trace.h"
34 #include "intel_drv.h"
35 #include <linux/dma_remapping.h>
36 
37 struct change_domains {
38 	uint32_t invalidate_domains;
39 	uint32_t flush_domains;
40 	uint32_t flush_rings;
41 	uint32_t flips;
42 };
43 
44 /*
45  * Set the next domain for the specified object. This
46  * may not actually perform the necessary flushing/invaliding though,
47  * as that may want to be batched with other set_domain operations
48  *
49  * This is (we hope) the only really tricky part of gem. The goal
50  * is fairly simple -- track which caches hold bits of the object
51  * and make sure they remain coherent. A few concrete examples may
52  * help to explain how it works. For shorthand, we use the notation
53  * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
54  * a pair of read and write domain masks.
55  *
56  * Case 1: the batch buffer
57  *
58  *	1. Allocated
59  *	2. Written by CPU
60  *	3. Mapped to GTT
61  *	4. Read by GPU
62  *	5. Unmapped from GTT
63  *	6. Freed
64  *
65  *	Let's take these a step at a time
66  *
67  *	1. Allocated
68  *		Pages allocated from the kernel may still have
69  *		cache contents, so we set them to (CPU, CPU) always.
70  *	2. Written by CPU (using pwrite)
71  *		The pwrite function calls set_domain (CPU, CPU) and
72  *		this function does nothing (as nothing changes)
73  *	3. Mapped by GTT
74  *		This function asserts that the object is not
75  *		currently in any GPU-based read or write domains
76  *	4. Read by GPU
77  *		i915_gem_execbuffer calls set_domain (COMMAND, 0).
78  *		As write_domain is zero, this function adds in the
79  *		current read domains (CPU+COMMAND, 0).
80  *		flush_domains is set to CPU.
81  *		invalidate_domains is set to COMMAND
82  *		clflush is run to get data out of the CPU caches
83  *		then i915_dev_set_domain calls i915_gem_flush to
84  *		emit an MI_FLUSH and drm_agp_chipset_flush
85  *	5. Unmapped from GTT
86  *		i915_gem_object_unbind calls set_domain (CPU, CPU)
87  *		flush_domains and invalidate_domains end up both zero
88  *		so no flushing/invalidating happens
89  *	6. Freed
90  *		yay, done
91  *
92  * Case 2: The shared render buffer
93  *
94  *	1. Allocated
95  *	2. Mapped to GTT
96  *	3. Read/written by GPU
97  *	4. set_domain to (CPU,CPU)
98  *	5. Read/written by CPU
99  *	6. Read/written by GPU
100  *
101  *	1. Allocated
102  *		Same as last example, (CPU, CPU)
103  *	2. Mapped to GTT
104  *		Nothing changes (assertions find that it is not in the GPU)
105  *	3. Read/written by GPU
106  *		execbuffer calls set_domain (RENDER, RENDER)
107  *		flush_domains gets CPU
108  *		invalidate_domains gets GPU
109  *		clflush (obj)
110  *		MI_FLUSH and drm_agp_chipset_flush
111  *	4. set_domain (CPU, CPU)
112  *		flush_domains gets GPU
113  *		invalidate_domains gets CPU
114  *		wait_rendering (obj) to make sure all drawing is complete.
115  *		This will include an MI_FLUSH to get the data from GPU
116  *		to memory
117  *		clflush (obj) to invalidate the CPU cache
118  *		Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
119  *	5. Read/written by CPU
120  *		cache lines are loaded and dirtied
121  *	6. Read written by GPU
122  *		Same as last GPU access
123  *
124  * Case 3: The constant buffer
125  *
126  *	1. Allocated
127  *	2. Written by CPU
128  *	3. Read by GPU
129  *	4. Updated (written) by CPU again
130  *	5. Read by GPU
131  *
132  *	1. Allocated
133  *		(CPU, CPU)
134  *	2. Written by CPU
135  *		(CPU, CPU)
136  *	3. Read by GPU
137  *		(CPU+RENDER, 0)
138  *		flush_domains = CPU
139  *		invalidate_domains = RENDER
140  *		clflush (obj)
141  *		MI_FLUSH
142  *		drm_agp_chipset_flush
143  *	4. Updated (written) by CPU again
144  *		(CPU, CPU)
145  *		flush_domains = 0 (no previous write domain)
146  *		invalidate_domains = 0 (no new read domains)
147  *	5. Read by GPU
148  *		(CPU+RENDER, 0)
149  *		flush_domains = CPU
150  *		invalidate_domains = RENDER
151  *		clflush (obj)
152  *		MI_FLUSH
153  *		drm_agp_chipset_flush
154  */
155 static void
i915_gem_object_set_to_gpu_domain(struct drm_i915_gem_object * obj,struct intel_ring_buffer * ring,struct change_domains * cd)156 i915_gem_object_set_to_gpu_domain(struct drm_i915_gem_object *obj,
157 				  struct intel_ring_buffer *ring,
158 				  struct change_domains *cd)
159 {
160 	uint32_t invalidate_domains = 0, flush_domains = 0;
161 
162 	/*
163 	 * If the object isn't moving to a new write domain,
164 	 * let the object stay in multiple read domains
165 	 */
166 	if (obj->base.pending_write_domain == 0)
167 		obj->base.pending_read_domains |= obj->base.read_domains;
168 
169 	/*
170 	 * Flush the current write domain if
171 	 * the new read domains don't match. Invalidate
172 	 * any read domains which differ from the old
173 	 * write domain
174 	 */
175 	if (obj->base.write_domain &&
176 	    (((obj->base.write_domain != obj->base.pending_read_domains ||
177 	       obj->ring != ring)) ||
178 	     (obj->fenced_gpu_access && !obj->pending_fenced_gpu_access))) {
179 		flush_domains |= obj->base.write_domain;
180 		invalidate_domains |=
181 			obj->base.pending_read_domains & ~obj->base.write_domain;
182 	}
183 	/*
184 	 * Invalidate any read caches which may have
185 	 * stale data. That is, any new read domains.
186 	 */
187 	invalidate_domains |= obj->base.pending_read_domains & ~obj->base.read_domains;
188 	if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU)
189 		i915_gem_clflush_object(obj);
190 
191 	if (obj->base.pending_write_domain)
192 		cd->flips |= atomic_read(&obj->pending_flip);
193 
194 	/* The actual obj->write_domain will be updated with
195 	 * pending_write_domain after we emit the accumulated flush for all
196 	 * of our domain changes in execbuffers (which clears objects'
197 	 * write_domains).  So if we have a current write domain that we
198 	 * aren't changing, set pending_write_domain to that.
199 	 */
200 	if (flush_domains == 0 && obj->base.pending_write_domain == 0)
201 		obj->base.pending_write_domain = obj->base.write_domain;
202 
203 	cd->invalidate_domains |= invalidate_domains;
204 	cd->flush_domains |= flush_domains;
205 	if (flush_domains & I915_GEM_GPU_DOMAINS)
206 		cd->flush_rings |= intel_ring_flag(obj->ring);
207 	if (invalidate_domains & I915_GEM_GPU_DOMAINS)
208 		cd->flush_rings |= intel_ring_flag(ring);
209 }
210 
211 struct eb_objects {
212 	int and;
213 	struct hlist_head buckets[0];
214 };
215 
216 static struct eb_objects *
eb_create(int size)217 eb_create(int size)
218 {
219 	struct eb_objects *eb;
220 	int count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
221 	while (count > size)
222 		count >>= 1;
223 	eb = kzalloc(count*sizeof(struct hlist_head) +
224 		     sizeof(struct eb_objects),
225 		     GFP_KERNEL);
226 	if (eb == NULL)
227 		return eb;
228 
229 	eb->and = count - 1;
230 	return eb;
231 }
232 
233 static void
eb_reset(struct eb_objects * eb)234 eb_reset(struct eb_objects *eb)
235 {
236 	memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
237 }
238 
239 static void
eb_add_object(struct eb_objects * eb,struct drm_i915_gem_object * obj)240 eb_add_object(struct eb_objects *eb, struct drm_i915_gem_object *obj)
241 {
242 	hlist_add_head(&obj->exec_node,
243 		       &eb->buckets[obj->exec_handle & eb->and]);
244 }
245 
246 static struct drm_i915_gem_object *
eb_get_object(struct eb_objects * eb,unsigned long handle)247 eb_get_object(struct eb_objects *eb, unsigned long handle)
248 {
249 	struct hlist_head *head;
250 	struct hlist_node *node;
251 	struct drm_i915_gem_object *obj;
252 
253 	head = &eb->buckets[handle & eb->and];
254 	hlist_for_each(node, head) {
255 		obj = hlist_entry(node, struct drm_i915_gem_object, exec_node);
256 		if (obj->exec_handle == handle)
257 			return obj;
258 	}
259 
260 	return NULL;
261 }
262 
263 static void
eb_destroy(struct eb_objects * eb)264 eb_destroy(struct eb_objects *eb)
265 {
266 	kfree(eb);
267 }
268 
269 static int
i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object * obj,struct eb_objects * eb,struct drm_i915_gem_relocation_entry * reloc)270 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
271 				   struct eb_objects *eb,
272 				   struct drm_i915_gem_relocation_entry *reloc)
273 {
274 	struct drm_device *dev = obj->base.dev;
275 	struct drm_gem_object *target_obj;
276 	uint32_t target_offset;
277 	int ret = -EINVAL;
278 
279 	/* we've already hold a reference to all valid objects */
280 	target_obj = &eb_get_object(eb, reloc->target_handle)->base;
281 	if (unlikely(target_obj == NULL))
282 		return -ENOENT;
283 
284 	target_offset = to_intel_bo(target_obj)->gtt_offset;
285 
286 	/* The target buffer should have appeared before us in the
287 	 * exec_object list, so it should have a GTT space bound by now.
288 	 */
289 	if (unlikely(target_offset == 0)) {
290 		DRM_DEBUG("No GTT space found for object %d\n",
291 			  reloc->target_handle);
292 		return ret;
293 	}
294 
295 	/* Validate that the target is in a valid r/w GPU domain */
296 	if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
297 		DRM_DEBUG("reloc with multiple write domains: "
298 			  "obj %p target %d offset %d "
299 			  "read %08x write %08x",
300 			  obj, reloc->target_handle,
301 			  (int) reloc->offset,
302 			  reloc->read_domains,
303 			  reloc->write_domain);
304 		return ret;
305 	}
306 	if (unlikely((reloc->write_domain | reloc->read_domains)
307 		     & ~I915_GEM_GPU_DOMAINS)) {
308 		DRM_DEBUG("reloc with read/write non-GPU domains: "
309 			  "obj %p target %d offset %d "
310 			  "read %08x write %08x",
311 			  obj, reloc->target_handle,
312 			  (int) reloc->offset,
313 			  reloc->read_domains,
314 			  reloc->write_domain);
315 		return ret;
316 	}
317 	if (unlikely(reloc->write_domain && target_obj->pending_write_domain &&
318 		     reloc->write_domain != target_obj->pending_write_domain)) {
319 		DRM_DEBUG("Write domain conflict: "
320 			  "obj %p target %d offset %d "
321 			  "new %08x old %08x\n",
322 			  obj, reloc->target_handle,
323 			  (int) reloc->offset,
324 			  reloc->write_domain,
325 			  target_obj->pending_write_domain);
326 		return ret;
327 	}
328 
329 	target_obj->pending_read_domains |= reloc->read_domains;
330 	target_obj->pending_write_domain |= reloc->write_domain;
331 
332 	/* If the relocation already has the right value in it, no
333 	 * more work needs to be done.
334 	 */
335 	if (target_offset == reloc->presumed_offset)
336 		return 0;
337 
338 	/* Check that the relocation address is valid... */
339 	if (unlikely(reloc->offset > obj->base.size - 4)) {
340 		DRM_DEBUG("Relocation beyond object bounds: "
341 			  "obj %p target %d offset %d size %d.\n",
342 			  obj, reloc->target_handle,
343 			  (int) reloc->offset,
344 			  (int) obj->base.size);
345 		return ret;
346 	}
347 	if (unlikely(reloc->offset & 3)) {
348 		DRM_DEBUG("Relocation not 4-byte aligned: "
349 			  "obj %p target %d offset %d.\n",
350 			  obj, reloc->target_handle,
351 			  (int) reloc->offset);
352 		return ret;
353 	}
354 
355 	reloc->delta += target_offset;
356 	if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) {
357 		uint32_t page_offset = reloc->offset & ~PAGE_MASK;
358 		char *vaddr;
359 
360 		vaddr = kmap_atomic(obj->pages[reloc->offset >> PAGE_SHIFT]);
361 		*(uint32_t *)(vaddr + page_offset) = reloc->delta;
362 		kunmap_atomic(vaddr);
363 	} else {
364 		struct drm_i915_private *dev_priv = dev->dev_private;
365 		uint32_t __iomem *reloc_entry;
366 		void __iomem *reloc_page;
367 
368 		/* We can't wait for rendering with pagefaults disabled */
369 		if (obj->active && in_atomic())
370 			return -EFAULT;
371 
372 		ret = i915_gem_object_set_to_gtt_domain(obj, 1);
373 		if (ret)
374 			return ret;
375 
376 		/* Map the page containing the relocation we're going to perform.  */
377 		reloc->offset += obj->gtt_offset;
378 		reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
379 						      reloc->offset & PAGE_MASK);
380 		reloc_entry = (uint32_t __iomem *)
381 			(reloc_page + (reloc->offset & ~PAGE_MASK));
382 		iowrite32(reloc->delta, reloc_entry);
383 		io_mapping_unmap_atomic(reloc_page);
384 	}
385 
386 	/* and update the user's relocation entry */
387 	reloc->presumed_offset = target_offset;
388 
389 	return 0;
390 }
391 
392 static int
i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object * obj,struct eb_objects * eb)393 i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
394 				    struct eb_objects *eb)
395 {
396 	struct drm_i915_gem_relocation_entry __user *user_relocs;
397 	struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
398 	int i, ret;
399 
400 	user_relocs = (void __user *)(uintptr_t)entry->relocs_ptr;
401 	for (i = 0; i < entry->relocation_count; i++) {
402 		struct drm_i915_gem_relocation_entry reloc;
403 
404 		if (__copy_from_user_inatomic(&reloc,
405 					      user_relocs+i,
406 					      sizeof(reloc)))
407 			return -EFAULT;
408 
409 		ret = i915_gem_execbuffer_relocate_entry(obj, eb, &reloc);
410 		if (ret)
411 			return ret;
412 
413 		if (__copy_to_user_inatomic(&user_relocs[i].presumed_offset,
414 					    &reloc.presumed_offset,
415 					    sizeof(reloc.presumed_offset)))
416 			return -EFAULT;
417 	}
418 
419 	return 0;
420 }
421 
422 static int
i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object * obj,struct eb_objects * eb,struct drm_i915_gem_relocation_entry * relocs)423 i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj,
424 					 struct eb_objects *eb,
425 					 struct drm_i915_gem_relocation_entry *relocs)
426 {
427 	const struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
428 	int i, ret;
429 
430 	for (i = 0; i < entry->relocation_count; i++) {
431 		ret = i915_gem_execbuffer_relocate_entry(obj, eb, &relocs[i]);
432 		if (ret)
433 			return ret;
434 	}
435 
436 	return 0;
437 }
438 
439 static int
i915_gem_execbuffer_relocate(struct drm_device * dev,struct eb_objects * eb,struct list_head * objects)440 i915_gem_execbuffer_relocate(struct drm_device *dev,
441 			     struct eb_objects *eb,
442 			     struct list_head *objects)
443 {
444 	struct drm_i915_gem_object *obj;
445 	int ret = 0;
446 
447 	/* This is the fast path and we cannot handle a pagefault whilst
448 	 * holding the struct mutex lest the user pass in the relocations
449 	 * contained within a mmaped bo. For in such a case we, the page
450 	 * fault handler would call i915_gem_fault() and we would try to
451 	 * acquire the struct mutex again. Obviously this is bad and so
452 	 * lockdep complains vehemently.
453 	 */
454 	pagefault_disable();
455 	list_for_each_entry(obj, objects, exec_list) {
456 		ret = i915_gem_execbuffer_relocate_object(obj, eb);
457 		if (ret)
458 			break;
459 	}
460 	pagefault_enable();
461 
462 	return ret;
463 }
464 
465 #define  __EXEC_OBJECT_HAS_FENCE (1<<31)
466 
467 static int
pin_and_fence_object(struct drm_i915_gem_object * obj,struct intel_ring_buffer * ring)468 pin_and_fence_object(struct drm_i915_gem_object *obj,
469 		     struct intel_ring_buffer *ring)
470 {
471 	struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
472 	bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
473 	bool need_fence, need_mappable;
474 	int ret;
475 
476 	need_fence =
477 		has_fenced_gpu_access &&
478 		entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
479 		obj->tiling_mode != I915_TILING_NONE;
480 	need_mappable =
481 		entry->relocation_count ? true : need_fence;
482 
483 	ret = i915_gem_object_pin(obj, entry->alignment, need_mappable);
484 	if (ret)
485 		return ret;
486 
487 	if (has_fenced_gpu_access) {
488 		if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
489 			if (obj->tiling_mode) {
490 				ret = i915_gem_object_get_fence(obj, ring);
491 				if (ret)
492 					goto err_unpin;
493 
494 				entry->flags |= __EXEC_OBJECT_HAS_FENCE;
495 				i915_gem_object_pin_fence(obj);
496 			} else {
497 				ret = i915_gem_object_put_fence(obj);
498 				if (ret)
499 					goto err_unpin;
500 			}
501 			obj->pending_fenced_gpu_access = true;
502 		}
503 	}
504 
505 	entry->offset = obj->gtt_offset;
506 	return 0;
507 
508 err_unpin:
509 	i915_gem_object_unpin(obj);
510 	return ret;
511 }
512 
513 static int
i915_gem_execbuffer_reserve(struct intel_ring_buffer * ring,struct drm_file * file,struct list_head * objects)514 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
515 			    struct drm_file *file,
516 			    struct list_head *objects)
517 {
518 	drm_i915_private_t *dev_priv = ring->dev->dev_private;
519 	struct drm_i915_gem_object *obj;
520 	int ret, retry;
521 	bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
522 	struct list_head ordered_objects;
523 
524 	INIT_LIST_HEAD(&ordered_objects);
525 	while (!list_empty(objects)) {
526 		struct drm_i915_gem_exec_object2 *entry;
527 		bool need_fence, need_mappable;
528 
529 		obj = list_first_entry(objects,
530 				       struct drm_i915_gem_object,
531 				       exec_list);
532 		entry = obj->exec_entry;
533 
534 		need_fence =
535 			has_fenced_gpu_access &&
536 			entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
537 			obj->tiling_mode != I915_TILING_NONE;
538 		need_mappable =
539 			entry->relocation_count ? true : need_fence;
540 
541 		if (need_mappable)
542 			list_move(&obj->exec_list, &ordered_objects);
543 		else
544 			list_move_tail(&obj->exec_list, &ordered_objects);
545 
546 		obj->base.pending_read_domains = 0;
547 		obj->base.pending_write_domain = 0;
548 	}
549 	list_splice(&ordered_objects, objects);
550 
551 	/* Attempt to pin all of the buffers into the GTT.
552 	 * This is done in 3 phases:
553 	 *
554 	 * 1a. Unbind all objects that do not match the GTT constraints for
555 	 *     the execbuffer (fenceable, mappable, alignment etc).
556 	 * 1b. Increment pin count for already bound objects.
557 	 * 2.  Bind new objects.
558 	 * 3.  Decrement pin count.
559 	 *
560 	 * This avoid unnecessary unbinding of later objects in order to makr
561 	 * room for the earlier objects *unless* we need to defragment.
562 	 */
563 	retry = 0;
564 	do {
565 		ret = 0;
566 
567 		/* Unbind any ill-fitting objects or pin. */
568 		list_for_each_entry(obj, objects, exec_list) {
569 			struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
570 			bool need_fence, need_mappable;
571 
572 			if (!obj->gtt_space)
573 				continue;
574 
575 			need_fence =
576 				has_fenced_gpu_access &&
577 				entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
578 				obj->tiling_mode != I915_TILING_NONE;
579 			need_mappable =
580 				entry->relocation_count ? true : need_fence;
581 
582 			if ((entry->alignment && obj->gtt_offset & (entry->alignment - 1)) ||
583 			    (need_mappable && !obj->map_and_fenceable))
584 				ret = i915_gem_object_unbind(obj);
585 			else
586 				ret = pin_and_fence_object(obj, ring);
587 			if (ret)
588 				goto err;
589 		}
590 
591 		/* Bind fresh objects */
592 		list_for_each_entry(obj, objects, exec_list) {
593 			if (obj->gtt_space)
594 				continue;
595 
596 			ret = pin_and_fence_object(obj, ring);
597 			if (ret) {
598 				int ret_ignore;
599 
600 				/* This can potentially raise a harmless
601 				 * -EINVAL if we failed to bind in the above
602 				 * call. It cannot raise -EINTR since we know
603 				 * that the bo is freshly bound and so will
604 				 * not need to be flushed or waited upon.
605 				 */
606 				ret_ignore = i915_gem_object_unbind(obj);
607 				(void)ret_ignore;
608 				WARN_ON(obj->gtt_space);
609 				break;
610 			}
611 		}
612 
613 		/* Decrement pin count for bound objects */
614 		list_for_each_entry(obj, objects, exec_list) {
615 			struct drm_i915_gem_exec_object2 *entry;
616 
617 			if (!obj->gtt_space)
618 				continue;
619 
620 			entry = obj->exec_entry;
621 			if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
622 				i915_gem_object_unpin_fence(obj);
623 				entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
624 			}
625 
626 			i915_gem_object_unpin(obj);
627 
628 			/* ... and ensure ppgtt mapping exist if needed. */
629 			if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) {
630 				i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
631 						       obj, obj->cache_level);
632 
633 				obj->has_aliasing_ppgtt_mapping = 1;
634 			}
635 		}
636 
637 		if (ret != -ENOSPC || retry > 1)
638 			return ret;
639 
640 		/* First attempt, just clear anything that is purgeable.
641 		 * Second attempt, clear the entire GTT.
642 		 */
643 		ret = i915_gem_evict_everything(ring->dev, retry == 0);
644 		if (ret)
645 			return ret;
646 
647 		retry++;
648 	} while (1);
649 
650 err:
651 	list_for_each_entry_continue_reverse(obj, objects, exec_list) {
652 		struct drm_i915_gem_exec_object2 *entry;
653 
654 		if (!obj->gtt_space)
655 			continue;
656 
657 		entry = obj->exec_entry;
658 		if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
659 			i915_gem_object_unpin_fence(obj);
660 			entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
661 		}
662 
663 		i915_gem_object_unpin(obj);
664 	}
665 
666 	return ret;
667 }
668 
669 static int
i915_gem_execbuffer_relocate_slow(struct drm_device * dev,struct drm_file * file,struct intel_ring_buffer * ring,struct list_head * objects,struct eb_objects * eb,struct drm_i915_gem_exec_object2 * exec,int count)670 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
671 				  struct drm_file *file,
672 				  struct intel_ring_buffer *ring,
673 				  struct list_head *objects,
674 				  struct eb_objects *eb,
675 				  struct drm_i915_gem_exec_object2 *exec,
676 				  int count)
677 {
678 	struct drm_i915_gem_relocation_entry *reloc;
679 	struct drm_i915_gem_object *obj;
680 	int *reloc_offset;
681 	int i, total, ret;
682 
683 	/* We may process another execbuffer during the unlock... */
684 	while (!list_empty(objects)) {
685 		obj = list_first_entry(objects,
686 				       struct drm_i915_gem_object,
687 				       exec_list);
688 		list_del_init(&obj->exec_list);
689 		drm_gem_object_unreference(&obj->base);
690 	}
691 
692 	mutex_unlock(&dev->struct_mutex);
693 
694 	total = 0;
695 	for (i = 0; i < count; i++)
696 		total += exec[i].relocation_count;
697 
698 	reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
699 	reloc = drm_malloc_ab(total, sizeof(*reloc));
700 	if (reloc == NULL || reloc_offset == NULL) {
701 		drm_free_large(reloc);
702 		drm_free_large(reloc_offset);
703 		mutex_lock(&dev->struct_mutex);
704 		return -ENOMEM;
705 	}
706 
707 	total = 0;
708 	for (i = 0; i < count; i++) {
709 		struct drm_i915_gem_relocation_entry __user *user_relocs;
710 		u64 invalid_offset = (u64)-1;
711 		int j;
712 
713 		user_relocs = (void __user *)(uintptr_t)exec[i].relocs_ptr;
714 
715 		if (copy_from_user(reloc+total, user_relocs,
716 				   exec[i].relocation_count * sizeof(*reloc))) {
717 			ret = -EFAULT;
718 			mutex_lock(&dev->struct_mutex);
719 			goto err;
720 		}
721 
722 		/* As we do not update the known relocation offsets after
723 		 * relocating (due to the complexities in lock handling),
724 		 * we need to mark them as invalid now so that we force the
725 		 * relocation processing next time. Just in case the target
726 		 * object is evicted and then rebound into its old
727 		 * presumed_offset before the next execbuffer - if that
728 		 * happened we would make the mistake of assuming that the
729 		 * relocations were valid.
730 		 */
731 		for (j = 0; j < exec[i].relocation_count; j++) {
732 			if (copy_to_user(&user_relocs[j].presumed_offset,
733 					 &invalid_offset,
734 					 sizeof(invalid_offset))) {
735 				ret = -EFAULT;
736 				mutex_lock(&dev->struct_mutex);
737 				goto err;
738 			}
739 		}
740 
741 		reloc_offset[i] = total;
742 		total += exec[i].relocation_count;
743 	}
744 
745 	ret = i915_mutex_lock_interruptible(dev);
746 	if (ret) {
747 		mutex_lock(&dev->struct_mutex);
748 		goto err;
749 	}
750 
751 	/* reacquire the objects */
752 	eb_reset(eb);
753 	for (i = 0; i < count; i++) {
754 		obj = to_intel_bo(drm_gem_object_lookup(dev, file,
755 							exec[i].handle));
756 		if (&obj->base == NULL) {
757 			DRM_DEBUG("Invalid object handle %d at index %d\n",
758 				   exec[i].handle, i);
759 			ret = -ENOENT;
760 			goto err;
761 		}
762 
763 		list_add_tail(&obj->exec_list, objects);
764 		obj->exec_handle = exec[i].handle;
765 		obj->exec_entry = &exec[i];
766 		eb_add_object(eb, obj);
767 	}
768 
769 	ret = i915_gem_execbuffer_reserve(ring, file, objects);
770 	if (ret)
771 		goto err;
772 
773 	list_for_each_entry(obj, objects, exec_list) {
774 		int offset = obj->exec_entry - exec;
775 		ret = i915_gem_execbuffer_relocate_object_slow(obj, eb,
776 							       reloc + reloc_offset[offset]);
777 		if (ret)
778 			goto err;
779 	}
780 
781 	/* Leave the user relocations as are, this is the painfully slow path,
782 	 * and we want to avoid the complication of dropping the lock whilst
783 	 * having buffers reserved in the aperture and so causing spurious
784 	 * ENOSPC for random operations.
785 	 */
786 
787 err:
788 	drm_free_large(reloc);
789 	drm_free_large(reloc_offset);
790 	return ret;
791 }
792 
793 static int
i915_gem_execbuffer_flush(struct drm_device * dev,uint32_t invalidate_domains,uint32_t flush_domains,uint32_t flush_rings)794 i915_gem_execbuffer_flush(struct drm_device *dev,
795 			  uint32_t invalidate_domains,
796 			  uint32_t flush_domains,
797 			  uint32_t flush_rings)
798 {
799 	drm_i915_private_t *dev_priv = dev->dev_private;
800 	int i, ret;
801 
802 	if (flush_domains & I915_GEM_DOMAIN_CPU)
803 		intel_gtt_chipset_flush();
804 
805 	if (flush_domains & I915_GEM_DOMAIN_GTT)
806 		wmb();
807 
808 	if ((flush_domains | invalidate_domains) & I915_GEM_GPU_DOMAINS) {
809 		for (i = 0; i < I915_NUM_RINGS; i++)
810 			if (flush_rings & (1 << i)) {
811 				ret = i915_gem_flush_ring(&dev_priv->ring[i],
812 							  invalidate_domains,
813 							  flush_domains);
814 				if (ret)
815 					return ret;
816 			}
817 	}
818 
819 	return 0;
820 }
821 
822 static bool
intel_enable_semaphores(struct drm_device * dev)823 intel_enable_semaphores(struct drm_device *dev)
824 {
825 	if (INTEL_INFO(dev)->gen < 6)
826 		return 0;
827 
828 	if (i915_semaphores >= 0)
829 		return i915_semaphores;
830 
831 	/* Disable semaphores on SNB */
832 	if (INTEL_INFO(dev)->gen == 6)
833 		return 0;
834 
835 	return 1;
836 }
837 
838 static int
i915_gem_execbuffer_sync_rings(struct drm_i915_gem_object * obj,struct intel_ring_buffer * to)839 i915_gem_execbuffer_sync_rings(struct drm_i915_gem_object *obj,
840 			       struct intel_ring_buffer *to)
841 {
842 	struct intel_ring_buffer *from = obj->ring;
843 	u32 seqno;
844 	int ret, idx;
845 
846 	if (from == NULL || to == from)
847 		return 0;
848 
849 	/* XXX gpu semaphores are implicated in various hard hangs on SNB */
850 	if (!intel_enable_semaphores(obj->base.dev))
851 		return i915_gem_object_wait_rendering(obj);
852 
853 	idx = intel_ring_sync_index(from, to);
854 
855 	seqno = obj->last_rendering_seqno;
856 	if (seqno <= from->sync_seqno[idx])
857 		return 0;
858 
859 	if (seqno == from->outstanding_lazy_request) {
860 		struct drm_i915_gem_request *request;
861 
862 		request = kzalloc(sizeof(*request), GFP_KERNEL);
863 		if (request == NULL)
864 			return -ENOMEM;
865 
866 		ret = i915_add_request(from, NULL, request);
867 		if (ret) {
868 			kfree(request);
869 			return ret;
870 		}
871 
872 		seqno = request->seqno;
873 	}
874 
875 	from->sync_seqno[idx] = seqno;
876 
877 	return to->sync_to(to, from, seqno - 1);
878 }
879 
880 static int
i915_gem_execbuffer_wait_for_flips(struct intel_ring_buffer * ring,u32 flips)881 i915_gem_execbuffer_wait_for_flips(struct intel_ring_buffer *ring, u32 flips)
882 {
883 	u32 plane, flip_mask;
884 	int ret;
885 
886 	/* Check for any pending flips. As we only maintain a flip queue depth
887 	 * of 1, we can simply insert a WAIT for the next display flip prior
888 	 * to executing the batch and avoid stalling the CPU.
889 	 */
890 
891 	for (plane = 0; flips >> plane; plane++) {
892 		if (((flips >> plane) & 1) == 0)
893 			continue;
894 
895 		if (plane)
896 			flip_mask = MI_WAIT_FOR_PLANE_B_FLIP;
897 		else
898 			flip_mask = MI_WAIT_FOR_PLANE_A_FLIP;
899 
900 		ret = intel_ring_begin(ring, 2);
901 		if (ret)
902 			return ret;
903 
904 		intel_ring_emit(ring, MI_WAIT_FOR_EVENT | flip_mask);
905 		intel_ring_emit(ring, MI_NOOP);
906 		intel_ring_advance(ring);
907 	}
908 
909 	return 0;
910 }
911 
912 
913 static int
i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer * ring,struct list_head * objects)914 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
915 				struct list_head *objects)
916 {
917 	struct drm_i915_gem_object *obj;
918 	struct change_domains cd;
919 	int ret;
920 
921 	memset(&cd, 0, sizeof(cd));
922 	list_for_each_entry(obj, objects, exec_list)
923 		i915_gem_object_set_to_gpu_domain(obj, ring, &cd);
924 
925 	if (cd.invalidate_domains | cd.flush_domains) {
926 		ret = i915_gem_execbuffer_flush(ring->dev,
927 						cd.invalidate_domains,
928 						cd.flush_domains,
929 						cd.flush_rings);
930 		if (ret)
931 			return ret;
932 	}
933 
934 	if (cd.flips) {
935 		ret = i915_gem_execbuffer_wait_for_flips(ring, cd.flips);
936 		if (ret)
937 			return ret;
938 	}
939 
940 	list_for_each_entry(obj, objects, exec_list) {
941 		ret = i915_gem_execbuffer_sync_rings(obj, ring);
942 		if (ret)
943 			return ret;
944 	}
945 
946 	return 0;
947 }
948 
949 static bool
i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 * exec)950 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
951 {
952 	return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
953 }
954 
955 static int
validate_exec_list(struct drm_i915_gem_exec_object2 * exec,int count)956 validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
957 		   int count)
958 {
959 	int i;
960 	int relocs_total = 0;
961 	int relocs_max = INT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
962 
963 	for (i = 0; i < count; i++) {
964 		char __user *ptr = (char __user *)(uintptr_t)exec[i].relocs_ptr;
965 		int length; /* limited by fault_in_pages_readable() */
966 
967 		/* First check for malicious input causing overflow in
968 		 * the worst case where we need to allocate the entire
969 		 * relocation tree as a single array.
970 		 */
971 		if (exec[i].relocation_count > relocs_max - relocs_total)
972 			return -EINVAL;
973 		relocs_total += exec[i].relocation_count;
974 
975 		length = exec[i].relocation_count *
976 			sizeof(struct drm_i915_gem_relocation_entry);
977 		if (!access_ok(VERIFY_READ, ptr, length))
978 			return -EFAULT;
979 
980 		/* we may also need to update the presumed offsets */
981 		if (!access_ok(VERIFY_WRITE, ptr, length))
982 			return -EFAULT;
983 
984 		if (fault_in_pages_readable(ptr, length))
985 			return -EFAULT;
986 	}
987 
988 	return 0;
989 }
990 
991 static void
i915_gem_execbuffer_move_to_active(struct list_head * objects,struct intel_ring_buffer * ring,u32 seqno)992 i915_gem_execbuffer_move_to_active(struct list_head *objects,
993 				   struct intel_ring_buffer *ring,
994 				   u32 seqno)
995 {
996 	struct drm_i915_gem_object *obj;
997 
998 	list_for_each_entry(obj, objects, exec_list) {
999 		  u32 old_read = obj->base.read_domains;
1000 		  u32 old_write = obj->base.write_domain;
1001 
1002 
1003 		obj->base.read_domains = obj->base.pending_read_domains;
1004 		obj->base.write_domain = obj->base.pending_write_domain;
1005 		obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
1006 
1007 		i915_gem_object_move_to_active(obj, ring, seqno);
1008 		if (obj->base.write_domain) {
1009 			obj->dirty = 1;
1010 			obj->pending_gpu_write = true;
1011 			list_move_tail(&obj->gpu_write_list,
1012 				       &ring->gpu_write_list);
1013 			intel_mark_busy(ring->dev, obj);
1014 		}
1015 
1016 		trace_i915_gem_object_change_domain(obj, old_read, old_write);
1017 	}
1018 }
1019 
1020 static void
i915_gem_execbuffer_retire_commands(struct drm_device * dev,struct drm_file * file,struct intel_ring_buffer * ring)1021 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
1022 				    struct drm_file *file,
1023 				    struct intel_ring_buffer *ring)
1024 {
1025 	struct drm_i915_gem_request *request;
1026 	u32 invalidate;
1027 
1028 	/*
1029 	 * Ensure that the commands in the batch buffer are
1030 	 * finished before the interrupt fires.
1031 	 *
1032 	 * The sampler always gets flushed on i965 (sigh).
1033 	 */
1034 	invalidate = I915_GEM_DOMAIN_COMMAND;
1035 	if (INTEL_INFO(dev)->gen >= 4)
1036 		invalidate |= I915_GEM_DOMAIN_SAMPLER;
1037 	if (ring->flush(ring, invalidate, 0)) {
1038 		i915_gem_next_request_seqno(ring);
1039 		return;
1040 	}
1041 
1042 	/* Add a breadcrumb for the completion of the batch buffer */
1043 	request = kzalloc(sizeof(*request), GFP_KERNEL);
1044 	if (request == NULL || i915_add_request(ring, file, request)) {
1045 		i915_gem_next_request_seqno(ring);
1046 		kfree(request);
1047 	}
1048 }
1049 
1050 static int
i915_reset_gen7_sol_offsets(struct drm_device * dev,struct intel_ring_buffer * ring)1051 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1052 			    struct intel_ring_buffer *ring)
1053 {
1054 	drm_i915_private_t *dev_priv = dev->dev_private;
1055 	int ret, i;
1056 
1057 	if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS])
1058 		return 0;
1059 
1060 	ret = intel_ring_begin(ring, 4 * 3);
1061 	if (ret)
1062 		return ret;
1063 
1064 	for (i = 0; i < 4; i++) {
1065 		intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1066 		intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
1067 		intel_ring_emit(ring, 0);
1068 	}
1069 
1070 	intel_ring_advance(ring);
1071 
1072 	return 0;
1073 }
1074 
1075 static int
i915_gem_do_execbuffer(struct drm_device * dev,void * data,struct drm_file * file,struct drm_i915_gem_execbuffer2 * args,struct drm_i915_gem_exec_object2 * exec)1076 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1077 		       struct drm_file *file,
1078 		       struct drm_i915_gem_execbuffer2 *args,
1079 		       struct drm_i915_gem_exec_object2 *exec)
1080 {
1081 	drm_i915_private_t *dev_priv = dev->dev_private;
1082 	struct list_head objects;
1083 	struct eb_objects *eb;
1084 	struct drm_i915_gem_object *batch_obj;
1085 	struct drm_clip_rect *cliprects = NULL;
1086 	struct intel_ring_buffer *ring;
1087 	u32 exec_start, exec_len;
1088 	u32 seqno;
1089 	u32 mask;
1090 	int ret, mode, i;
1091 
1092 	if (!i915_gem_check_execbuffer(args)) {
1093 		DRM_DEBUG("execbuf with invalid offset/length\n");
1094 		return -EINVAL;
1095 	}
1096 
1097 	ret = validate_exec_list(exec, args->buffer_count);
1098 	if (ret)
1099 		return ret;
1100 
1101 	switch (args->flags & I915_EXEC_RING_MASK) {
1102 	case I915_EXEC_DEFAULT:
1103 	case I915_EXEC_RENDER:
1104 		ring = &dev_priv->ring[RCS];
1105 		break;
1106 	case I915_EXEC_BSD:
1107 		if (!HAS_BSD(dev)) {
1108 			DRM_DEBUG("execbuf with invalid ring (BSD)\n");
1109 			return -EINVAL;
1110 		}
1111 		ring = &dev_priv->ring[VCS];
1112 		break;
1113 	case I915_EXEC_BLT:
1114 		if (!HAS_BLT(dev)) {
1115 			DRM_DEBUG("execbuf with invalid ring (BLT)\n");
1116 			return -EINVAL;
1117 		}
1118 		ring = &dev_priv->ring[BCS];
1119 		break;
1120 	default:
1121 		DRM_DEBUG("execbuf with unknown ring: %d\n",
1122 			  (int)(args->flags & I915_EXEC_RING_MASK));
1123 		return -EINVAL;
1124 	}
1125 
1126 	mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1127 	mask = I915_EXEC_CONSTANTS_MASK;
1128 	switch (mode) {
1129 	case I915_EXEC_CONSTANTS_REL_GENERAL:
1130 	case I915_EXEC_CONSTANTS_ABSOLUTE:
1131 	case I915_EXEC_CONSTANTS_REL_SURFACE:
1132 		if (ring == &dev_priv->ring[RCS] &&
1133 		    mode != dev_priv->relative_constants_mode) {
1134 			if (INTEL_INFO(dev)->gen < 4)
1135 				return -EINVAL;
1136 
1137 			if (INTEL_INFO(dev)->gen > 5 &&
1138 			    mode == I915_EXEC_CONSTANTS_REL_SURFACE)
1139 				return -EINVAL;
1140 
1141 			/* The HW changed the meaning on this bit on gen6 */
1142 			if (INTEL_INFO(dev)->gen >= 6)
1143 				mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1144 		}
1145 		break;
1146 	default:
1147 		DRM_DEBUG("execbuf with unknown constants: %d\n", mode);
1148 		return -EINVAL;
1149 	}
1150 
1151 	if (args->buffer_count < 1) {
1152 		DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1153 		return -EINVAL;
1154 	}
1155 
1156 	if (args->num_cliprects != 0) {
1157 		if (ring != &dev_priv->ring[RCS]) {
1158 			DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1159 			return -EINVAL;
1160 		}
1161 
1162 		if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1163 			DRM_DEBUG("execbuf with %u cliprects\n",
1164 				  args->num_cliprects);
1165 			return -EINVAL;
1166 		}
1167 		cliprects = kmalloc(args->num_cliprects * sizeof(*cliprects),
1168 				    GFP_KERNEL);
1169 		if (cliprects == NULL) {
1170 			ret = -ENOMEM;
1171 			goto pre_mutex_err;
1172 		}
1173 
1174 		if (copy_from_user(cliprects,
1175 				     (struct drm_clip_rect __user *)(uintptr_t)
1176 				     args->cliprects_ptr,
1177 				     sizeof(*cliprects)*args->num_cliprects)) {
1178 			ret = -EFAULT;
1179 			goto pre_mutex_err;
1180 		}
1181 	}
1182 
1183 	ret = i915_mutex_lock_interruptible(dev);
1184 	if (ret)
1185 		goto pre_mutex_err;
1186 
1187 	if (dev_priv->mm.suspended) {
1188 		mutex_unlock(&dev->struct_mutex);
1189 		ret = -EBUSY;
1190 		goto pre_mutex_err;
1191 	}
1192 
1193 	eb = eb_create(args->buffer_count);
1194 	if (eb == NULL) {
1195 		mutex_unlock(&dev->struct_mutex);
1196 		ret = -ENOMEM;
1197 		goto pre_mutex_err;
1198 	}
1199 
1200 	/* Look up object handles */
1201 	INIT_LIST_HEAD(&objects);
1202 	for (i = 0; i < args->buffer_count; i++) {
1203 		struct drm_i915_gem_object *obj;
1204 
1205 		obj = to_intel_bo(drm_gem_object_lookup(dev, file,
1206 							exec[i].handle));
1207 		if (&obj->base == NULL) {
1208 			DRM_DEBUG("Invalid object handle %d at index %d\n",
1209 				   exec[i].handle, i);
1210 			/* prevent error path from reading uninitialized data */
1211 			ret = -ENOENT;
1212 			goto err;
1213 		}
1214 
1215 		if (!list_empty(&obj->exec_list)) {
1216 			DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
1217 				   obj, exec[i].handle, i);
1218 			ret = -EINVAL;
1219 			goto err;
1220 		}
1221 
1222 		list_add_tail(&obj->exec_list, &objects);
1223 		obj->exec_handle = exec[i].handle;
1224 		obj->exec_entry = &exec[i];
1225 		eb_add_object(eb, obj);
1226 	}
1227 
1228 	/* take note of the batch buffer before we might reorder the lists */
1229 	batch_obj = list_entry(objects.prev,
1230 			       struct drm_i915_gem_object,
1231 			       exec_list);
1232 
1233 	/* Move the objects en-masse into the GTT, evicting if necessary. */
1234 	ret = i915_gem_execbuffer_reserve(ring, file, &objects);
1235 	if (ret)
1236 		goto err;
1237 
1238 	/* The objects are in their final locations, apply the relocations. */
1239 	ret = i915_gem_execbuffer_relocate(dev, eb, &objects);
1240 	if (ret) {
1241 		if (ret == -EFAULT) {
1242 			ret = i915_gem_execbuffer_relocate_slow(dev, file, ring,
1243 								&objects, eb,
1244 								exec,
1245 								args->buffer_count);
1246 			BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1247 		}
1248 		if (ret)
1249 			goto err;
1250 	}
1251 
1252 	/* Set the pending read domains for the batch buffer to COMMAND */
1253 	if (batch_obj->base.pending_write_domain) {
1254 		DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1255 		ret = -EINVAL;
1256 		goto err;
1257 	}
1258 	batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1259 
1260 	ret = i915_gem_execbuffer_move_to_gpu(ring, &objects);
1261 	if (ret)
1262 		goto err;
1263 
1264 	seqno = i915_gem_next_request_seqno(ring);
1265 	for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++) {
1266 		if (seqno < ring->sync_seqno[i]) {
1267 			/* The GPU can not handle its semaphore value wrapping,
1268 			 * so every billion or so execbuffers, we need to stall
1269 			 * the GPU in order to reset the counters.
1270 			 */
1271 			ret = i915_gpu_idle(dev, true);
1272 			if (ret)
1273 				goto err;
1274 
1275 			BUG_ON(ring->sync_seqno[i]);
1276 		}
1277 	}
1278 
1279 	if (ring == &dev_priv->ring[RCS] &&
1280 	    mode != dev_priv->relative_constants_mode) {
1281 		ret = intel_ring_begin(ring, 4);
1282 		if (ret)
1283 				goto err;
1284 
1285 		intel_ring_emit(ring, MI_NOOP);
1286 		intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1287 		intel_ring_emit(ring, INSTPM);
1288 		intel_ring_emit(ring, mask << 16 | mode);
1289 		intel_ring_advance(ring);
1290 
1291 		dev_priv->relative_constants_mode = mode;
1292 	}
1293 
1294 	if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1295 		ret = i915_reset_gen7_sol_offsets(dev, ring);
1296 		if (ret)
1297 			goto err;
1298 	}
1299 
1300 	trace_i915_gem_ring_dispatch(ring, seqno);
1301 
1302 	exec_start = batch_obj->gtt_offset + args->batch_start_offset;
1303 	exec_len = args->batch_len;
1304 	if (cliprects) {
1305 		for (i = 0; i < args->num_cliprects; i++) {
1306 			ret = i915_emit_box(dev, &cliprects[i],
1307 					    args->DR1, args->DR4);
1308 			if (ret)
1309 				goto err;
1310 
1311 			ret = ring->dispatch_execbuffer(ring,
1312 							exec_start, exec_len);
1313 			if (ret)
1314 				goto err;
1315 		}
1316 	} else {
1317 		ret = ring->dispatch_execbuffer(ring, exec_start, exec_len);
1318 		if (ret)
1319 			goto err;
1320 	}
1321 
1322 	i915_gem_execbuffer_move_to_active(&objects, ring, seqno);
1323 	i915_gem_execbuffer_retire_commands(dev, file, ring);
1324 
1325 err:
1326 	eb_destroy(eb);
1327 	while (!list_empty(&objects)) {
1328 		struct drm_i915_gem_object *obj;
1329 
1330 		obj = list_first_entry(&objects,
1331 				       struct drm_i915_gem_object,
1332 				       exec_list);
1333 		list_del_init(&obj->exec_list);
1334 		drm_gem_object_unreference(&obj->base);
1335 	}
1336 
1337 	mutex_unlock(&dev->struct_mutex);
1338 
1339 pre_mutex_err:
1340 	kfree(cliprects);
1341 	return ret;
1342 }
1343 
1344 /*
1345  * Legacy execbuffer just creates an exec2 list from the original exec object
1346  * list array and passes it to the real function.
1347  */
1348 int
i915_gem_execbuffer(struct drm_device * dev,void * data,struct drm_file * file)1349 i915_gem_execbuffer(struct drm_device *dev, void *data,
1350 		    struct drm_file *file)
1351 {
1352 	struct drm_i915_gem_execbuffer *args = data;
1353 	struct drm_i915_gem_execbuffer2 exec2;
1354 	struct drm_i915_gem_exec_object *exec_list = NULL;
1355 	struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1356 	int ret, i;
1357 
1358 	if (args->buffer_count < 1) {
1359 		DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1360 		return -EINVAL;
1361 	}
1362 
1363 	/* Copy in the exec list from userland */
1364 	exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1365 	exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1366 	if (exec_list == NULL || exec2_list == NULL) {
1367 		DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1368 			  args->buffer_count);
1369 		drm_free_large(exec_list);
1370 		drm_free_large(exec2_list);
1371 		return -ENOMEM;
1372 	}
1373 	ret = copy_from_user(exec_list,
1374 			     (struct drm_i915_relocation_entry __user *)
1375 			     (uintptr_t) args->buffers_ptr,
1376 			     sizeof(*exec_list) * args->buffer_count);
1377 	if (ret != 0) {
1378 		DRM_DEBUG("copy %d exec entries failed %d\n",
1379 			  args->buffer_count, ret);
1380 		drm_free_large(exec_list);
1381 		drm_free_large(exec2_list);
1382 		return -EFAULT;
1383 	}
1384 
1385 	for (i = 0; i < args->buffer_count; i++) {
1386 		exec2_list[i].handle = exec_list[i].handle;
1387 		exec2_list[i].relocation_count = exec_list[i].relocation_count;
1388 		exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1389 		exec2_list[i].alignment = exec_list[i].alignment;
1390 		exec2_list[i].offset = exec_list[i].offset;
1391 		if (INTEL_INFO(dev)->gen < 4)
1392 			exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1393 		else
1394 			exec2_list[i].flags = 0;
1395 	}
1396 
1397 	exec2.buffers_ptr = args->buffers_ptr;
1398 	exec2.buffer_count = args->buffer_count;
1399 	exec2.batch_start_offset = args->batch_start_offset;
1400 	exec2.batch_len = args->batch_len;
1401 	exec2.DR1 = args->DR1;
1402 	exec2.DR4 = args->DR4;
1403 	exec2.num_cliprects = args->num_cliprects;
1404 	exec2.cliprects_ptr = args->cliprects_ptr;
1405 	exec2.flags = I915_EXEC_RENDER;
1406 
1407 	ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1408 	if (!ret) {
1409 		/* Copy the new buffer offsets back to the user's exec list. */
1410 		for (i = 0; i < args->buffer_count; i++)
1411 			exec_list[i].offset = exec2_list[i].offset;
1412 		/* ... and back out to userspace */
1413 		ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1414 				   (uintptr_t) args->buffers_ptr,
1415 				   exec_list,
1416 				   sizeof(*exec_list) * args->buffer_count);
1417 		if (ret) {
1418 			ret = -EFAULT;
1419 			DRM_DEBUG("failed to copy %d exec entries "
1420 				  "back to user (%d)\n",
1421 				  args->buffer_count, ret);
1422 		}
1423 	}
1424 
1425 	drm_free_large(exec_list);
1426 	drm_free_large(exec2_list);
1427 	return ret;
1428 }
1429 
1430 int
i915_gem_execbuffer2(struct drm_device * dev,void * data,struct drm_file * file)1431 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1432 		     struct drm_file *file)
1433 {
1434 	struct drm_i915_gem_execbuffer2 *args = data;
1435 	struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1436 	int ret;
1437 
1438 	if (args->buffer_count < 1 ||
1439 	    args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1440 		DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1441 		return -EINVAL;
1442 	}
1443 
1444 	exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1445 			     GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
1446 	if (exec2_list == NULL)
1447 		exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1448 					   args->buffer_count);
1449 	if (exec2_list == NULL) {
1450 		DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1451 			  args->buffer_count);
1452 		return -ENOMEM;
1453 	}
1454 	ret = copy_from_user(exec2_list,
1455 			     (struct drm_i915_relocation_entry __user *)
1456 			     (uintptr_t) args->buffers_ptr,
1457 			     sizeof(*exec2_list) * args->buffer_count);
1458 	if (ret != 0) {
1459 		DRM_DEBUG("copy %d exec entries failed %d\n",
1460 			  args->buffer_count, ret);
1461 		drm_free_large(exec2_list);
1462 		return -EFAULT;
1463 	}
1464 
1465 	ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1466 	if (!ret) {
1467 		/* Copy the new buffer offsets back to the user's exec list. */
1468 		ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1469 				   (uintptr_t) args->buffers_ptr,
1470 				   exec2_list,
1471 				   sizeof(*exec2_list) * args->buffer_count);
1472 		if (ret) {
1473 			ret = -EFAULT;
1474 			DRM_DEBUG("failed to copy %d exec entries "
1475 				  "back to user (%d)\n",
1476 				  args->buffer_count, ret);
1477 		}
1478 	}
1479 
1480 	drm_free_large(exec2_list);
1481 	return ret;
1482 }
1483