1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2014-2016 Intel Corporation
5  */
6 
7 #include "display/intel_frontbuffer.h"
8 
9 #include "i915_drv.h"
10 #include "i915_gem_clflush.h"
11 #include "i915_gem_gtt.h"
12 #include "i915_gem_ioctls.h"
13 #include "i915_gem_object.h"
14 #include "i915_vma.h"
15 #include "i915_gem_lmem.h"
16 #include "i915_gem_mman.h"
17 
18 static void __i915_gem_object_flush_for_display(struct drm_i915_gem_object *obj)
19 {
20 	/*
21 	 * We manually flush the CPU domain so that we can override and
22 	 * force the flush for the display, and perform it asyncrhonously.
23 	 */
24 	i915_gem_object_flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);
25 	if (obj->cache_dirty)
26 		i915_gem_clflush_object(obj, I915_CLFLUSH_FORCE);
27 	obj->write_domain = 0;
28 }
29 
30 void i915_gem_object_flush_if_display(struct drm_i915_gem_object *obj)
31 {
32 	if (!i915_gem_object_is_framebuffer(obj))
33 		return;
34 
35 	i915_gem_object_lock(obj, NULL);
36 	__i915_gem_object_flush_for_display(obj);
37 	i915_gem_object_unlock(obj);
38 }
39 
40 void i915_gem_object_flush_if_display_locked(struct drm_i915_gem_object *obj)
41 {
42 	if (i915_gem_object_is_framebuffer(obj))
43 		__i915_gem_object_flush_for_display(obj);
44 }
45 
46 /**
47  * Moves a single object to the WC read, and possibly write domain.
48  * @obj: object to act on
49  * @write: ask for write access or read only
50  *
51  * This function returns when the move is complete, including waiting on
52  * flushes to occur.
53  */
54 int
55 i915_gem_object_set_to_wc_domain(struct drm_i915_gem_object *obj, bool write)
56 {
57 	int ret;
58 
59 	assert_object_held(obj);
60 
61 	ret = i915_gem_object_wait(obj,
62 				   I915_WAIT_INTERRUPTIBLE |
63 				   (write ? I915_WAIT_ALL : 0),
64 				   MAX_SCHEDULE_TIMEOUT);
65 	if (ret)
66 		return ret;
67 
68 	if (obj->write_domain == I915_GEM_DOMAIN_WC)
69 		return 0;
70 
71 	/* Flush and acquire obj->pages so that we are coherent through
72 	 * direct access in memory with previous cached writes through
73 	 * shmemfs and that our cache domain tracking remains valid.
74 	 * For example, if the obj->filp was moved to swap without us
75 	 * being notified and releasing the pages, we would mistakenly
76 	 * continue to assume that the obj remained out of the CPU cached
77 	 * domain.
78 	 */
79 	ret = i915_gem_object_pin_pages(obj);
80 	if (ret)
81 		return ret;
82 
83 	i915_gem_object_flush_write_domain(obj, ~I915_GEM_DOMAIN_WC);
84 
85 	/* Serialise direct access to this object with the barriers for
86 	 * coherent writes from the GPU, by effectively invalidating the
87 	 * WC domain upon first access.
88 	 */
89 	if ((obj->read_domains & I915_GEM_DOMAIN_WC) == 0)
90 		mb();
91 
92 	/* It should now be out of any other write domains, and we can update
93 	 * the domain values for our changes.
94 	 */
95 	GEM_BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_WC) != 0);
96 	obj->read_domains |= I915_GEM_DOMAIN_WC;
97 	if (write) {
98 		obj->read_domains = I915_GEM_DOMAIN_WC;
99 		obj->write_domain = I915_GEM_DOMAIN_WC;
100 		obj->mm.dirty = true;
101 	}
102 
103 	i915_gem_object_unpin_pages(obj);
104 	return 0;
105 }
106 
107 /**
108  * Moves a single object to the GTT read, and possibly write domain.
109  * @obj: object to act on
110  * @write: ask for write access or read only
111  *
112  * This function returns when the move is complete, including waiting on
113  * flushes to occur.
114  */
115 int
116 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
117 {
118 	int ret;
119 
120 	assert_object_held(obj);
121 
122 	ret = i915_gem_object_wait(obj,
123 				   I915_WAIT_INTERRUPTIBLE |
124 				   (write ? I915_WAIT_ALL : 0),
125 				   MAX_SCHEDULE_TIMEOUT);
126 	if (ret)
127 		return ret;
128 
129 	if (obj->write_domain == I915_GEM_DOMAIN_GTT)
130 		return 0;
131 
132 	/* Flush and acquire obj->pages so that we are coherent through
133 	 * direct access in memory with previous cached writes through
134 	 * shmemfs and that our cache domain tracking remains valid.
135 	 * For example, if the obj->filp was moved to swap without us
136 	 * being notified and releasing the pages, we would mistakenly
137 	 * continue to assume that the obj remained out of the CPU cached
138 	 * domain.
139 	 */
140 	ret = i915_gem_object_pin_pages(obj);
141 	if (ret)
142 		return ret;
143 
144 	i915_gem_object_flush_write_domain(obj, ~I915_GEM_DOMAIN_GTT);
145 
146 	/* Serialise direct access to this object with the barriers for
147 	 * coherent writes from the GPU, by effectively invalidating the
148 	 * GTT domain upon first access.
149 	 */
150 	if ((obj->read_domains & I915_GEM_DOMAIN_GTT) == 0)
151 		mb();
152 
153 	/* It should now be out of any other write domains, and we can update
154 	 * the domain values for our changes.
155 	 */
156 	GEM_BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
157 	obj->read_domains |= I915_GEM_DOMAIN_GTT;
158 	if (write) {
159 		struct i915_vma *vma;
160 
161 		obj->read_domains = I915_GEM_DOMAIN_GTT;
162 		obj->write_domain = I915_GEM_DOMAIN_GTT;
163 		obj->mm.dirty = true;
164 
165 		spin_lock(&obj->vma.lock);
166 		for_each_ggtt_vma(vma, obj)
167 			if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
168 				i915_vma_set_ggtt_write(vma);
169 		spin_unlock(&obj->vma.lock);
170 	}
171 
172 	i915_gem_object_unpin_pages(obj);
173 	return 0;
174 }
175 
176 /**
177  * Changes the cache-level of an object across all VMA.
178  * @obj: object to act on
179  * @cache_level: new cache level to set for the object
180  *
181  * After this function returns, the object will be in the new cache-level
182  * across all GTT and the contents of the backing storage will be coherent,
183  * with respect to the new cache-level. In order to keep the backing storage
184  * coherent for all users, we only allow a single cache level to be set
185  * globally on the object and prevent it from being changed whilst the
186  * hardware is reading from the object. That is if the object is currently
187  * on the scanout it will be set to uncached (or equivalent display
188  * cache coherency) and all non-MOCS GPU access will also be uncached so
189  * that all direct access to the scanout remains coherent.
190  */
191 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
192 				    enum i915_cache_level cache_level)
193 {
194 	int ret;
195 
196 	if (obj->cache_level == cache_level)
197 		return 0;
198 
199 	ret = i915_gem_object_wait(obj,
200 				   I915_WAIT_INTERRUPTIBLE |
201 				   I915_WAIT_ALL,
202 				   MAX_SCHEDULE_TIMEOUT);
203 	if (ret)
204 		return ret;
205 
206 	/* Always invalidate stale cachelines */
207 	if (obj->cache_level != cache_level) {
208 		i915_gem_object_set_cache_coherency(obj, cache_level);
209 		obj->cache_dirty = true;
210 	}
211 
212 	/* The cache-level will be applied when each vma is rebound. */
213 	return i915_gem_object_unbind(obj,
214 				      I915_GEM_OBJECT_UNBIND_ACTIVE |
215 				      I915_GEM_OBJECT_UNBIND_BARRIER);
216 }
217 
218 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
219 			       struct drm_file *file)
220 {
221 	struct drm_i915_gem_caching *args = data;
222 	struct drm_i915_gem_object *obj;
223 	int err = 0;
224 
225 	rcu_read_lock();
226 	obj = i915_gem_object_lookup_rcu(file, args->handle);
227 	if (!obj) {
228 		err = -ENOENT;
229 		goto out;
230 	}
231 
232 	switch (obj->cache_level) {
233 	case I915_CACHE_LLC:
234 	case I915_CACHE_L3_LLC:
235 		args->caching = I915_CACHING_CACHED;
236 		break;
237 
238 	case I915_CACHE_WT:
239 		args->caching = I915_CACHING_DISPLAY;
240 		break;
241 
242 	default:
243 		args->caching = I915_CACHING_NONE;
244 		break;
245 	}
246 out:
247 	rcu_read_unlock();
248 	return err;
249 }
250 
251 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
252 			       struct drm_file *file)
253 {
254 	struct drm_i915_private *i915 = to_i915(dev);
255 	struct drm_i915_gem_caching *args = data;
256 	struct drm_i915_gem_object *obj;
257 	enum i915_cache_level level;
258 	int ret = 0;
259 
260 	switch (args->caching) {
261 	case I915_CACHING_NONE:
262 		level = I915_CACHE_NONE;
263 		break;
264 	case I915_CACHING_CACHED:
265 		/*
266 		 * Due to a HW issue on BXT A stepping, GPU stores via a
267 		 * snooped mapping may leave stale data in a corresponding CPU
268 		 * cacheline, whereas normally such cachelines would get
269 		 * invalidated.
270 		 */
271 		if (!HAS_LLC(i915) && !HAS_SNOOP(i915))
272 			return -ENODEV;
273 
274 		level = I915_CACHE_LLC;
275 		break;
276 	case I915_CACHING_DISPLAY:
277 		level = HAS_WT(i915) ? I915_CACHE_WT : I915_CACHE_NONE;
278 		break;
279 	default:
280 		return -EINVAL;
281 	}
282 
283 	obj = i915_gem_object_lookup(file, args->handle);
284 	if (!obj)
285 		return -ENOENT;
286 
287 	/*
288 	 * The caching mode of proxy object is handled by its generator, and
289 	 * not allowed to be changed by userspace.
290 	 */
291 	if (i915_gem_object_is_proxy(obj)) {
292 		ret = -ENXIO;
293 		goto out;
294 	}
295 
296 	ret = i915_gem_object_lock_interruptible(obj, NULL);
297 	if (ret)
298 		goto out;
299 
300 	ret = i915_gem_object_set_cache_level(obj, level);
301 	i915_gem_object_unlock(obj);
302 
303 out:
304 	i915_gem_object_put(obj);
305 	return ret;
306 }
307 
308 /*
309  * Prepare buffer for display plane (scanout, cursors, etc). Can be called from
310  * an uninterruptible phase (modesetting) and allows any flushes to be pipelined
311  * (for pageflips). We only flush the caches while preparing the buffer for
312  * display, the callers are responsible for frontbuffer flush.
313  */
314 struct i915_vma *
315 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
316 				     u32 alignment,
317 				     const struct i915_ggtt_view *view,
318 				     unsigned int flags)
319 {
320 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
321 	struct i915_gem_ww_ctx ww;
322 	struct i915_vma *vma;
323 	int ret;
324 
325 	/* Frame buffer must be in LMEM (no migration yet) */
326 	if (HAS_LMEM(i915) && !i915_gem_object_is_lmem(obj))
327 		return ERR_PTR(-EINVAL);
328 
329 	i915_gem_ww_ctx_init(&ww, true);
330 retry:
331 	ret = i915_gem_object_lock(obj, &ww);
332 	if (ret)
333 		goto err;
334 	/*
335 	 * The display engine is not coherent with the LLC cache on gen6.  As
336 	 * a result, we make sure that the pinning that is about to occur is
337 	 * done with uncached PTEs. This is lowest common denominator for all
338 	 * chipsets.
339 	 *
340 	 * However for gen6+, we could do better by using the GFDT bit instead
341 	 * of uncaching, which would allow us to flush all the LLC-cached data
342 	 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
343 	 */
344 	ret = i915_gem_object_set_cache_level(obj,
345 					      HAS_WT(i915) ?
346 					      I915_CACHE_WT : I915_CACHE_NONE);
347 	if (ret)
348 		goto err;
349 
350 	/*
351 	 * As the user may map the buffer once pinned in the display plane
352 	 * (e.g. libkms for the bootup splash), we have to ensure that we
353 	 * always use map_and_fenceable for all scanout buffers. However,
354 	 * it may simply be too big to fit into mappable, in which case
355 	 * put it anyway and hope that userspace can cope (but always first
356 	 * try to preserve the existing ABI).
357 	 */
358 	vma = ERR_PTR(-ENOSPC);
359 	if ((flags & PIN_MAPPABLE) == 0 &&
360 	    (!view || view->type == I915_GGTT_VIEW_NORMAL))
361 		vma = i915_gem_object_ggtt_pin_ww(obj, &ww, view, 0, alignment,
362 						  flags | PIN_MAPPABLE |
363 						  PIN_NONBLOCK);
364 	if (IS_ERR(vma) && vma != ERR_PTR(-EDEADLK))
365 		vma = i915_gem_object_ggtt_pin_ww(obj, &ww, view, 0,
366 						  alignment, flags);
367 	if (IS_ERR(vma)) {
368 		ret = PTR_ERR(vma);
369 		goto err;
370 	}
371 
372 	vma->display_alignment = max_t(u64, vma->display_alignment, alignment);
373 
374 	i915_gem_object_flush_if_display_locked(obj);
375 
376 err:
377 	if (ret == -EDEADLK) {
378 		ret = i915_gem_ww_ctx_backoff(&ww);
379 		if (!ret)
380 			goto retry;
381 	}
382 	i915_gem_ww_ctx_fini(&ww);
383 
384 	if (ret)
385 		return ERR_PTR(ret);
386 
387 	return vma;
388 }
389 
390 /**
391  * Moves a single object to the CPU read, and possibly write domain.
392  * @obj: object to act on
393  * @write: requesting write or read-only access
394  *
395  * This function returns when the move is complete, including waiting on
396  * flushes to occur.
397  */
398 int
399 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
400 {
401 	int ret;
402 
403 	assert_object_held(obj);
404 
405 	ret = i915_gem_object_wait(obj,
406 				   I915_WAIT_INTERRUPTIBLE |
407 				   (write ? I915_WAIT_ALL : 0),
408 				   MAX_SCHEDULE_TIMEOUT);
409 	if (ret)
410 		return ret;
411 
412 	i915_gem_object_flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);
413 
414 	/* Flush the CPU cache if it's still invalid. */
415 	if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0) {
416 		i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC);
417 		obj->read_domains |= I915_GEM_DOMAIN_CPU;
418 	}
419 
420 	/* It should now be out of any other write domains, and we can update
421 	 * the domain values for our changes.
422 	 */
423 	GEM_BUG_ON(obj->write_domain & ~I915_GEM_DOMAIN_CPU);
424 
425 	/* If we're writing through the CPU, then the GPU read domains will
426 	 * need to be invalidated at next use.
427 	 */
428 	if (write)
429 		__start_cpu_write(obj);
430 
431 	return 0;
432 }
433 
434 /**
435  * Called when user space prepares to use an object with the CPU, either
436  * through the mmap ioctl's mapping or a GTT mapping.
437  * @dev: drm device
438  * @data: ioctl data blob
439  * @file: drm file
440  */
441 int
442 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
443 			  struct drm_file *file)
444 {
445 	struct drm_i915_gem_set_domain *args = data;
446 	struct drm_i915_gem_object *obj;
447 	u32 read_domains = args->read_domains;
448 	u32 write_domain = args->write_domain;
449 	int err;
450 
451 	/* Only handle setting domains to types used by the CPU. */
452 	if ((write_domain | read_domains) & I915_GEM_GPU_DOMAINS)
453 		return -EINVAL;
454 
455 	/*
456 	 * Having something in the write domain implies it's in the read
457 	 * domain, and only that read domain.  Enforce that in the request.
458 	 */
459 	if (write_domain && read_domains != write_domain)
460 		return -EINVAL;
461 
462 	if (!read_domains)
463 		return 0;
464 
465 	obj = i915_gem_object_lookup(file, args->handle);
466 	if (!obj)
467 		return -ENOENT;
468 
469 	/*
470 	 * Try to flush the object off the GPU without holding the lock.
471 	 * We will repeat the flush holding the lock in the normal manner
472 	 * to catch cases where we are gazumped.
473 	 */
474 	err = i915_gem_object_wait(obj,
475 				   I915_WAIT_INTERRUPTIBLE |
476 				   I915_WAIT_PRIORITY |
477 				   (write_domain ? I915_WAIT_ALL : 0),
478 				   MAX_SCHEDULE_TIMEOUT);
479 	if (err)
480 		goto out;
481 
482 	/*
483 	 * Proxy objects do not control access to the backing storage, ergo
484 	 * they cannot be used as a means to manipulate the cache domain
485 	 * tracking for that backing storage. The proxy object is always
486 	 * considered to be outside of any cache domain.
487 	 */
488 	if (i915_gem_object_is_proxy(obj)) {
489 		err = -ENXIO;
490 		goto out;
491 	}
492 
493 	/*
494 	 * Flush and acquire obj->pages so that we are coherent through
495 	 * direct access in memory with previous cached writes through
496 	 * shmemfs and that our cache domain tracking remains valid.
497 	 * For example, if the obj->filp was moved to swap without us
498 	 * being notified and releasing the pages, we would mistakenly
499 	 * continue to assume that the obj remained out of the CPU cached
500 	 * domain.
501 	 */
502 	err = i915_gem_object_pin_pages(obj);
503 	if (err)
504 		goto out;
505 
506 	/*
507 	 * Already in the desired write domain? Nothing for us to do!
508 	 *
509 	 * We apply a little bit of cunning here to catch a broader set of
510 	 * no-ops. If obj->write_domain is set, we must be in the same
511 	 * obj->read_domains, and only that domain. Therefore, if that
512 	 * obj->write_domain matches the request read_domains, we are
513 	 * already in the same read/write domain and can skip the operation,
514 	 * without having to further check the requested write_domain.
515 	 */
516 	if (READ_ONCE(obj->write_domain) == read_domains)
517 		goto out_unpin;
518 
519 	err = i915_gem_object_lock_interruptible(obj, NULL);
520 	if (err)
521 		goto out_unpin;
522 
523 	if (read_domains & I915_GEM_DOMAIN_WC)
524 		err = i915_gem_object_set_to_wc_domain(obj, write_domain);
525 	else if (read_domains & I915_GEM_DOMAIN_GTT)
526 		err = i915_gem_object_set_to_gtt_domain(obj, write_domain);
527 	else
528 		err = i915_gem_object_set_to_cpu_domain(obj, write_domain);
529 
530 	i915_gem_object_unlock(obj);
531 
532 	if (write_domain)
533 		i915_gem_object_invalidate_frontbuffer(obj, ORIGIN_CPU);
534 
535 out_unpin:
536 	i915_gem_object_unpin_pages(obj);
537 out:
538 	i915_gem_object_put(obj);
539 	return err;
540 }
541 
542 /*
543  * Pins the specified object's pages and synchronizes the object with
544  * GPU accesses. Sets needs_clflush to non-zero if the caller should
545  * flush the object from the CPU cache.
546  */
547 int i915_gem_object_prepare_read(struct drm_i915_gem_object *obj,
548 				 unsigned int *needs_clflush)
549 {
550 	int ret;
551 
552 	*needs_clflush = 0;
553 	if (!i915_gem_object_has_struct_page(obj))
554 		return -ENODEV;
555 
556 	assert_object_held(obj);
557 
558 	ret = i915_gem_object_wait(obj,
559 				   I915_WAIT_INTERRUPTIBLE,
560 				   MAX_SCHEDULE_TIMEOUT);
561 	if (ret)
562 		return ret;
563 
564 	ret = i915_gem_object_pin_pages(obj);
565 	if (ret)
566 		return ret;
567 
568 	if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ ||
569 	    !static_cpu_has(X86_FEATURE_CLFLUSH)) {
570 		ret = i915_gem_object_set_to_cpu_domain(obj, false);
571 		if (ret)
572 			goto err_unpin;
573 		else
574 			goto out;
575 	}
576 
577 	i915_gem_object_flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);
578 
579 	/* If we're not in the cpu read domain, set ourself into the gtt
580 	 * read domain and manually flush cachelines (if required). This
581 	 * optimizes for the case when the gpu will dirty the data
582 	 * anyway again before the next pread happens.
583 	 */
584 	if (!obj->cache_dirty &&
585 	    !(obj->read_domains & I915_GEM_DOMAIN_CPU))
586 		*needs_clflush = CLFLUSH_BEFORE;
587 
588 out:
589 	/* return with the pages pinned */
590 	return 0;
591 
592 err_unpin:
593 	i915_gem_object_unpin_pages(obj);
594 	return ret;
595 }
596 
597 int i915_gem_object_prepare_write(struct drm_i915_gem_object *obj,
598 				  unsigned int *needs_clflush)
599 {
600 	int ret;
601 
602 	*needs_clflush = 0;
603 	if (!i915_gem_object_has_struct_page(obj))
604 		return -ENODEV;
605 
606 	assert_object_held(obj);
607 
608 	ret = i915_gem_object_wait(obj,
609 				   I915_WAIT_INTERRUPTIBLE |
610 				   I915_WAIT_ALL,
611 				   MAX_SCHEDULE_TIMEOUT);
612 	if (ret)
613 		return ret;
614 
615 	ret = i915_gem_object_pin_pages(obj);
616 	if (ret)
617 		return ret;
618 
619 	if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE ||
620 	    !static_cpu_has(X86_FEATURE_CLFLUSH)) {
621 		ret = i915_gem_object_set_to_cpu_domain(obj, true);
622 		if (ret)
623 			goto err_unpin;
624 		else
625 			goto out;
626 	}
627 
628 	i915_gem_object_flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);
629 
630 	/* If we're not in the cpu write domain, set ourself into the
631 	 * gtt write domain and manually flush cachelines (as required).
632 	 * This optimizes for the case when the gpu will use the data
633 	 * right away and we therefore have to clflush anyway.
634 	 */
635 	if (!obj->cache_dirty) {
636 		*needs_clflush |= CLFLUSH_AFTER;
637 
638 		/*
639 		 * Same trick applies to invalidate partially written
640 		 * cachelines read before writing.
641 		 */
642 		if (!(obj->read_domains & I915_GEM_DOMAIN_CPU))
643 			*needs_clflush |= CLFLUSH_BEFORE;
644 	}
645 
646 out:
647 	i915_gem_object_invalidate_frontbuffer(obj, ORIGIN_CPU);
648 	obj->mm.dirty = true;
649 	/* return with the pages pinned */
650 	return 0;
651 
652 err_unpin:
653 	i915_gem_object_unpin_pages(obj);
654 	return ret;
655 }
656