• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2017 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  */
24 
25 #include <linux/highmem.h>
26 #include <linux/sched/mm.h>
27 
28 #include <drm/drm_cache.h>
29 
30 #include "display/intel_frontbuffer.h"
31 #include "pxp/intel_pxp.h"
32 
33 #include "i915_drv.h"
34 #include "i915_file_private.h"
35 #include "i915_gem_clflush.h"
36 #include "i915_gem_context.h"
37 #include "i915_gem_dmabuf.h"
38 #include "i915_gem_mman.h"
39 #include "i915_gem_object.h"
40 #include "i915_gem_ttm.h"
41 #include "i915_memcpy.h"
42 #include "i915_trace.h"
43 
44 static struct kmem_cache *slab_objects;
45 
46 static const struct drm_gem_object_funcs i915_gem_object_funcs;
47 
i915_gem_object_alloc(void)48 struct drm_i915_gem_object *i915_gem_object_alloc(void)
49 {
50 	struct drm_i915_gem_object *obj;
51 
52 	obj = kmem_cache_zalloc(slab_objects, GFP_KERNEL);
53 	if (!obj)
54 		return NULL;
55 	obj->base.funcs = &i915_gem_object_funcs;
56 
57 	return obj;
58 }
59 
i915_gem_object_free(struct drm_i915_gem_object * obj)60 void i915_gem_object_free(struct drm_i915_gem_object *obj)
61 {
62 	return kmem_cache_free(slab_objects, obj);
63 }
64 
i915_gem_object_init(struct drm_i915_gem_object * obj,const struct drm_i915_gem_object_ops * ops,struct lock_class_key * key,unsigned flags)65 void i915_gem_object_init(struct drm_i915_gem_object *obj,
66 			  const struct drm_i915_gem_object_ops *ops,
67 			  struct lock_class_key *key, unsigned flags)
68 {
69 	/*
70 	 * A gem object is embedded both in a struct ttm_buffer_object :/ and
71 	 * in a drm_i915_gem_object. Make sure they are aliased.
72 	 */
73 	BUILD_BUG_ON(offsetof(typeof(*obj), base) !=
74 		     offsetof(typeof(*obj), __do_not_access.base));
75 
76 	spin_lock_init(&obj->vma.lock);
77 	INIT_LIST_HEAD(&obj->vma.list);
78 
79 	INIT_LIST_HEAD(&obj->mm.link);
80 
81 	INIT_LIST_HEAD(&obj->lut_list);
82 	spin_lock_init(&obj->lut_lock);
83 
84 	spin_lock_init(&obj->mmo.lock);
85 	obj->mmo.offsets = RB_ROOT;
86 
87 	init_rcu_head(&obj->rcu);
88 
89 	obj->ops = ops;
90 	GEM_BUG_ON(flags & ~I915_BO_ALLOC_FLAGS);
91 	obj->flags = flags;
92 
93 	obj->mm.madv = I915_MADV_WILLNEED;
94 	INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN);
95 	mutex_init(&obj->mm.get_page.lock);
96 	INIT_RADIX_TREE(&obj->mm.get_dma_page.radix, GFP_KERNEL | __GFP_NOWARN);
97 	mutex_init(&obj->mm.get_dma_page.lock);
98 }
99 
100 /**
101  * __i915_gem_object_fini - Clean up a GEM object initialization
102  * @obj: The gem object to cleanup
103  *
104  * This function cleans up gem object fields that are set up by
105  * drm_gem_private_object_init() and i915_gem_object_init().
106  * It's primarily intended as a helper for backends that need to
107  * clean up the gem object in separate steps.
108  */
__i915_gem_object_fini(struct drm_i915_gem_object * obj)109 void __i915_gem_object_fini(struct drm_i915_gem_object *obj)
110 {
111 	mutex_destroy(&obj->mm.get_page.lock);
112 	mutex_destroy(&obj->mm.get_dma_page.lock);
113 	dma_resv_fini(&obj->base._resv);
114 }
115 
116 /**
117  * i915_gem_object_set_cache_coherency - Mark up the object's coherency levels
118  * for a given cache_level
119  * @obj: #drm_i915_gem_object
120  * @cache_level: cache level
121  */
i915_gem_object_set_cache_coherency(struct drm_i915_gem_object * obj,unsigned int cache_level)122 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj,
123 					 unsigned int cache_level)
124 {
125 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
126 
127 	obj->cache_level = cache_level;
128 
129 	if (cache_level != I915_CACHE_NONE)
130 		obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ |
131 				       I915_BO_CACHE_COHERENT_FOR_WRITE);
132 	else if (HAS_LLC(i915))
133 		obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ;
134 	else
135 		obj->cache_coherent = 0;
136 
137 	obj->cache_dirty =
138 		!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE) &&
139 		!IS_DGFX(i915);
140 }
141 
i915_gem_object_can_bypass_llc(struct drm_i915_gem_object * obj)142 bool i915_gem_object_can_bypass_llc(struct drm_i915_gem_object *obj)
143 {
144 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
145 
146 	/*
147 	 * This is purely from a security perspective, so we simply don't care
148 	 * about non-userspace objects being able to bypass the LLC.
149 	 */
150 	if (!(obj->flags & I915_BO_ALLOC_USER))
151 		return false;
152 
153 	/*
154 	 * EHL and JSL add the 'Bypass LLC' MOCS entry, which should make it
155 	 * possible for userspace to bypass the GTT caching bits set by the
156 	 * kernel, as per the given object cache_level. This is troublesome
157 	 * since the heavy flush we apply when first gathering the pages is
158 	 * skipped if the kernel thinks the object is coherent with the GPU. As
159 	 * a result it might be possible to bypass the cache and read the
160 	 * contents of the page directly, which could be stale data. If it's
161 	 * just a case of userspace shooting themselves in the foot then so be
162 	 * it, but since i915 takes the stance of always zeroing memory before
163 	 * handing it to userspace, we need to prevent this.
164 	 */
165 	return IS_JSL_EHL(i915);
166 }
167 
i915_gem_close_object(struct drm_gem_object * gem,struct drm_file * file)168 static void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
169 {
170 	struct drm_i915_gem_object *obj = to_intel_bo(gem);
171 	struct drm_i915_file_private *fpriv = file->driver_priv;
172 	struct i915_lut_handle bookmark = {};
173 	struct i915_mmap_offset *mmo, *mn;
174 	struct i915_lut_handle *lut, *ln;
175 	LIST_HEAD(close);
176 
177 	spin_lock(&obj->lut_lock);
178 	list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) {
179 		struct i915_gem_context *ctx = lut->ctx;
180 
181 		if (ctx && ctx->file_priv == fpriv) {
182 			i915_gem_context_get(ctx);
183 			list_move(&lut->obj_link, &close);
184 		}
185 
186 		/* Break long locks, and carefully continue on from this spot */
187 		if (&ln->obj_link != &obj->lut_list) {
188 			list_add_tail(&bookmark.obj_link, &ln->obj_link);
189 			if (cond_resched_lock(&obj->lut_lock))
190 				list_safe_reset_next(&bookmark, ln, obj_link);
191 			__list_del_entry(&bookmark.obj_link);
192 		}
193 	}
194 	spin_unlock(&obj->lut_lock);
195 
196 	spin_lock(&obj->mmo.lock);
197 	rbtree_postorder_for_each_entry_safe(mmo, mn, &obj->mmo.offsets, offset)
198 		drm_vma_node_revoke(&mmo->vma_node, file);
199 	spin_unlock(&obj->mmo.lock);
200 
201 	list_for_each_entry_safe(lut, ln, &close, obj_link) {
202 		struct i915_gem_context *ctx = lut->ctx;
203 		struct i915_vma *vma;
204 
205 		/*
206 		 * We allow the process to have multiple handles to the same
207 		 * vma, in the same fd namespace, by virtue of flink/open.
208 		 */
209 
210 		mutex_lock(&ctx->lut_mutex);
211 		vma = radix_tree_delete(&ctx->handles_vma, lut->handle);
212 		if (vma) {
213 			GEM_BUG_ON(vma->obj != obj);
214 			GEM_BUG_ON(!atomic_read(&vma->open_count));
215 			i915_vma_close(vma);
216 		}
217 		mutex_unlock(&ctx->lut_mutex);
218 
219 		i915_gem_context_put(lut->ctx);
220 		i915_lut_handle_free(lut);
221 		i915_gem_object_put(obj);
222 	}
223 }
224 
__i915_gem_free_object_rcu(struct rcu_head * head)225 void __i915_gem_free_object_rcu(struct rcu_head *head)
226 {
227 	struct drm_i915_gem_object *obj =
228 		container_of(head, typeof(*obj), rcu);
229 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
230 
231 	i915_gem_object_free(obj);
232 
233 	GEM_BUG_ON(!atomic_read(&i915->mm.free_count));
234 	atomic_dec(&i915->mm.free_count);
235 }
236 
__i915_gem_object_free_mmaps(struct drm_i915_gem_object * obj)237 static void __i915_gem_object_free_mmaps(struct drm_i915_gem_object *obj)
238 {
239 	/* Skip serialisation and waking the device if known to be not used. */
240 
241 	if (obj->userfault_count && !IS_DGFX(to_i915(obj->base.dev)))
242 		i915_gem_object_release_mmap_gtt(obj);
243 
244 	if (!RB_EMPTY_ROOT(&obj->mmo.offsets)) {
245 		struct i915_mmap_offset *mmo, *mn;
246 
247 		i915_gem_object_release_mmap_offset(obj);
248 
249 		rbtree_postorder_for_each_entry_safe(mmo, mn,
250 						     &obj->mmo.offsets,
251 						     offset) {
252 			drm_vma_offset_remove(obj->base.dev->vma_offset_manager,
253 					      &mmo->vma_node);
254 			kfree(mmo);
255 		}
256 		obj->mmo.offsets = RB_ROOT;
257 	}
258 }
259 
260 /**
261  * __i915_gem_object_pages_fini - Clean up pages use of a gem object
262  * @obj: The gem object to clean up
263  *
264  * This function cleans up usage of the object mm.pages member. It
265  * is intended for backends that need to clean up a gem object in
266  * separate steps and needs to be called when the object is idle before
267  * the object's backing memory is freed.
268  */
__i915_gem_object_pages_fini(struct drm_i915_gem_object * obj)269 void __i915_gem_object_pages_fini(struct drm_i915_gem_object *obj)
270 {
271 	assert_object_held_shared(obj);
272 
273 	if (!list_empty(&obj->vma.list)) {
274 		struct i915_vma *vma;
275 
276 		spin_lock(&obj->vma.lock);
277 		while ((vma = list_first_entry_or_null(&obj->vma.list,
278 						       struct i915_vma,
279 						       obj_link))) {
280 			GEM_BUG_ON(vma->obj != obj);
281 			spin_unlock(&obj->vma.lock);
282 
283 			i915_vma_destroy(vma);
284 
285 			spin_lock(&obj->vma.lock);
286 		}
287 		spin_unlock(&obj->vma.lock);
288 	}
289 
290 	__i915_gem_object_free_mmaps(obj);
291 
292 	atomic_set(&obj->mm.pages_pin_count, 0);
293 	__i915_gem_object_put_pages(obj);
294 	GEM_BUG_ON(i915_gem_object_has_pages(obj));
295 }
296 
__i915_gem_free_object(struct drm_i915_gem_object * obj)297 void __i915_gem_free_object(struct drm_i915_gem_object *obj)
298 {
299 	trace_i915_gem_object_destroy(obj);
300 
301 	GEM_BUG_ON(!list_empty(&obj->lut_list));
302 
303 	bitmap_free(obj->bit_17);
304 
305 	if (obj->base.import_attach)
306 		drm_prime_gem_destroy(&obj->base, NULL);
307 
308 	drm_gem_free_mmap_offset(&obj->base);
309 
310 	if (obj->ops->release)
311 		obj->ops->release(obj);
312 
313 	if (obj->mm.n_placements > 1)
314 		kfree(obj->mm.placements);
315 
316 	if (obj->shares_resv_from)
317 		i915_vm_resv_put(obj->shares_resv_from);
318 
319 	__i915_gem_object_fini(obj);
320 }
321 
__i915_gem_free_objects(struct drm_i915_private * i915,struct llist_node * freed)322 static void __i915_gem_free_objects(struct drm_i915_private *i915,
323 				    struct llist_node *freed)
324 {
325 	struct drm_i915_gem_object *obj, *on;
326 
327 	llist_for_each_entry_safe(obj, on, freed, freed) {
328 		might_sleep();
329 		if (obj->ops->delayed_free) {
330 			obj->ops->delayed_free(obj);
331 			continue;
332 		}
333 
334 		__i915_gem_object_pages_fini(obj);
335 		__i915_gem_free_object(obj);
336 
337 		/* But keep the pointer alive for RCU-protected lookups */
338 		call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
339 		cond_resched();
340 	}
341 }
342 
i915_gem_flush_free_objects(struct drm_i915_private * i915)343 void i915_gem_flush_free_objects(struct drm_i915_private *i915)
344 {
345 	struct llist_node *freed = llist_del_all(&i915->mm.free_list);
346 
347 	if (unlikely(freed))
348 		__i915_gem_free_objects(i915, freed);
349 }
350 
__i915_gem_free_work(struct work_struct * work)351 static void __i915_gem_free_work(struct work_struct *work)
352 {
353 	struct drm_i915_private *i915 =
354 		container_of(work, struct drm_i915_private, mm.free_work);
355 
356 	i915_gem_flush_free_objects(i915);
357 }
358 
i915_gem_free_object(struct drm_gem_object * gem_obj)359 static void i915_gem_free_object(struct drm_gem_object *gem_obj)
360 {
361 	struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
362 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
363 
364 	GEM_BUG_ON(i915_gem_object_is_framebuffer(obj));
365 
366 	/*
367 	 * Before we free the object, make sure any pure RCU-only
368 	 * read-side critical sections are complete, e.g.
369 	 * i915_gem_busy_ioctl(). For the corresponding synchronized
370 	 * lookup see i915_gem_object_lookup_rcu().
371 	 */
372 	atomic_inc(&i915->mm.free_count);
373 
374 	/*
375 	 * Since we require blocking on struct_mutex to unbind the freed
376 	 * object from the GPU before releasing resources back to the
377 	 * system, we can not do that directly from the RCU callback (which may
378 	 * be a softirq context), but must instead then defer that work onto a
379 	 * kthread. We use the RCU callback rather than move the freed object
380 	 * directly onto the work queue so that we can mix between using the
381 	 * worker and performing frees directly from subsequent allocations for
382 	 * crude but effective memory throttling.
383 	 */
384 
385 	if (llist_add(&obj->freed, &i915->mm.free_list))
386 		queue_work(i915->wq, &i915->mm.free_work);
387 }
388 
__i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object * obj,enum fb_op_origin origin)389 void __i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object *obj,
390 					 enum fb_op_origin origin)
391 {
392 	struct intel_frontbuffer *front;
393 
394 	front = __intel_frontbuffer_get(obj);
395 	if (front) {
396 		intel_frontbuffer_flush(front, origin);
397 		intel_frontbuffer_put(front);
398 	}
399 }
400 
__i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object * obj,enum fb_op_origin origin)401 void __i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object *obj,
402 					      enum fb_op_origin origin)
403 {
404 	struct intel_frontbuffer *front;
405 
406 	front = __intel_frontbuffer_get(obj);
407 	if (front) {
408 		intel_frontbuffer_invalidate(front, origin);
409 		intel_frontbuffer_put(front);
410 	}
411 }
412 
413 static void
i915_gem_object_read_from_page_kmap(struct drm_i915_gem_object * obj,u64 offset,void * dst,int size)414 i915_gem_object_read_from_page_kmap(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size)
415 {
416 	void *src_map;
417 	void *src_ptr;
418 
419 	src_map = kmap_atomic(i915_gem_object_get_page(obj, offset >> PAGE_SHIFT));
420 
421 	src_ptr = src_map + offset_in_page(offset);
422 	if (!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ))
423 		drm_clflush_virt_range(src_ptr, size);
424 	memcpy(dst, src_ptr, size);
425 
426 	kunmap_atomic(src_map);
427 }
428 
429 static void
i915_gem_object_read_from_page_iomap(struct drm_i915_gem_object * obj,u64 offset,void * dst,int size)430 i915_gem_object_read_from_page_iomap(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size)
431 {
432 	void __iomem *src_map;
433 	void __iomem *src_ptr;
434 	dma_addr_t dma = i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT);
435 
436 	src_map = io_mapping_map_wc(&obj->mm.region->iomap,
437 				    dma - obj->mm.region->region.start,
438 				    PAGE_SIZE);
439 
440 	src_ptr = src_map + offset_in_page(offset);
441 	if (!i915_memcpy_from_wc(dst, (void __force *)src_ptr, size))
442 		memcpy_fromio(dst, src_ptr, size);
443 
444 	io_mapping_unmap(src_map);
445 }
446 
447 /**
448  * i915_gem_object_read_from_page - read data from the page of a GEM object
449  * @obj: GEM object to read from
450  * @offset: offset within the object
451  * @dst: buffer to store the read data
452  * @size: size to read
453  *
454  * Reads data from @obj at the specified offset. The requested region to read
455  * from can't cross a page boundary. The caller must ensure that @obj pages
456  * are pinned and that @obj is synced wrt. any related writes.
457  *
458  * Return: %0 on success or -ENODEV if the type of @obj's backing store is
459  * unsupported.
460  */
i915_gem_object_read_from_page(struct drm_i915_gem_object * obj,u64 offset,void * dst,int size)461 int i915_gem_object_read_from_page(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size)
462 {
463 	GEM_BUG_ON(offset >= obj->base.size);
464 	GEM_BUG_ON(offset_in_page(offset) > PAGE_SIZE - size);
465 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
466 
467 	if (i915_gem_object_has_struct_page(obj))
468 		i915_gem_object_read_from_page_kmap(obj, offset, dst, size);
469 	else if (i915_gem_object_has_iomem(obj))
470 		i915_gem_object_read_from_page_iomap(obj, offset, dst, size);
471 	else
472 		return -ENODEV;
473 
474 	return 0;
475 }
476 
477 /**
478  * i915_gem_object_evictable - Whether object is likely evictable after unbind.
479  * @obj: The object to check
480  *
481  * This function checks whether the object is likely unvictable after unbind.
482  * If the object is not locked when checking, the result is only advisory.
483  * If the object is locked when checking, and the function returns true,
484  * then an eviction should indeed be possible. But since unlocked vma
485  * unpinning and unbinding is currently possible, the object can actually
486  * become evictable even if this function returns false.
487  *
488  * Return: true if the object may be evictable. False otherwise.
489  */
i915_gem_object_evictable(struct drm_i915_gem_object * obj)490 bool i915_gem_object_evictable(struct drm_i915_gem_object *obj)
491 {
492 	struct i915_vma *vma;
493 	int pin_count = atomic_read(&obj->mm.pages_pin_count);
494 
495 	if (!pin_count)
496 		return true;
497 
498 	spin_lock(&obj->vma.lock);
499 	list_for_each_entry(vma, &obj->vma.list, obj_link) {
500 		if (i915_vma_is_pinned(vma)) {
501 			spin_unlock(&obj->vma.lock);
502 			return false;
503 		}
504 		if (atomic_read(&vma->pages_count))
505 			pin_count--;
506 	}
507 	spin_unlock(&obj->vma.lock);
508 	GEM_WARN_ON(pin_count < 0);
509 
510 	return pin_count == 0;
511 }
512 
513 /**
514  * i915_gem_object_migratable - Whether the object is migratable out of the
515  * current region.
516  * @obj: Pointer to the object.
517  *
518  * Return: Whether the object is allowed to be resident in other
519  * regions than the current while pages are present.
520  */
i915_gem_object_migratable(struct drm_i915_gem_object * obj)521 bool i915_gem_object_migratable(struct drm_i915_gem_object *obj)
522 {
523 	struct intel_memory_region *mr = READ_ONCE(obj->mm.region);
524 
525 	if (!mr)
526 		return false;
527 
528 	return obj->mm.n_placements > 1;
529 }
530 
531 /**
532  * i915_gem_object_has_struct_page - Whether the object is page-backed
533  * @obj: The object to query.
534  *
535  * This function should only be called while the object is locked or pinned,
536  * otherwise the page backing may change under the caller.
537  *
538  * Return: True if page-backed, false otherwise.
539  */
i915_gem_object_has_struct_page(const struct drm_i915_gem_object * obj)540 bool i915_gem_object_has_struct_page(const struct drm_i915_gem_object *obj)
541 {
542 #ifdef CONFIG_LOCKDEP
543 	if (IS_DGFX(to_i915(obj->base.dev)) &&
544 	    i915_gem_object_evictable((void __force *)obj))
545 		assert_object_held_shared(obj);
546 #endif
547 	return obj->mem_flags & I915_BO_FLAG_STRUCT_PAGE;
548 }
549 
550 /**
551  * i915_gem_object_has_iomem - Whether the object is iomem-backed
552  * @obj: The object to query.
553  *
554  * This function should only be called while the object is locked or pinned,
555  * otherwise the iomem backing may change under the caller.
556  *
557  * Return: True if iomem-backed, false otherwise.
558  */
i915_gem_object_has_iomem(const struct drm_i915_gem_object * obj)559 bool i915_gem_object_has_iomem(const struct drm_i915_gem_object *obj)
560 {
561 #ifdef CONFIG_LOCKDEP
562 	if (IS_DGFX(to_i915(obj->base.dev)) &&
563 	    i915_gem_object_evictable((void __force *)obj))
564 		assert_object_held_shared(obj);
565 #endif
566 	return obj->mem_flags & I915_BO_FLAG_IOMEM;
567 }
568 
569 /**
570  * i915_gem_object_can_migrate - Whether an object likely can be migrated
571  *
572  * @obj: The object to migrate
573  * @id: The region intended to migrate to
574  *
575  * Check whether the object backend supports migration to the
576  * given region. Note that pinning may affect the ability to migrate as
577  * returned by this function.
578  *
579  * This function is primarily intended as a helper for checking the
580  * possibility to migrate objects and might be slightly less permissive
581  * than i915_gem_object_migrate() when it comes to objects with the
582  * I915_BO_ALLOC_USER flag set.
583  *
584  * Return: true if migration is possible, false otherwise.
585  */
i915_gem_object_can_migrate(struct drm_i915_gem_object * obj,enum intel_region_id id)586 bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj,
587 				 enum intel_region_id id)
588 {
589 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
590 	unsigned int num_allowed = obj->mm.n_placements;
591 	struct intel_memory_region *mr;
592 	unsigned int i;
593 
594 	GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
595 	GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
596 
597 	mr = i915->mm.regions[id];
598 	if (!mr)
599 		return false;
600 
601 	if (!IS_ALIGNED(obj->base.size, mr->min_page_size))
602 		return false;
603 
604 	if (obj->mm.region == mr)
605 		return true;
606 
607 	if (!i915_gem_object_evictable(obj))
608 		return false;
609 
610 	if (!obj->ops->migrate)
611 		return false;
612 
613 	if (!(obj->flags & I915_BO_ALLOC_USER))
614 		return true;
615 
616 	if (num_allowed == 0)
617 		return false;
618 
619 	for (i = 0; i < num_allowed; ++i) {
620 		if (mr == obj->mm.placements[i])
621 			return true;
622 	}
623 
624 	return false;
625 }
626 
627 /**
628  * i915_gem_object_migrate - Migrate an object to the desired region id
629  * @obj: The object to migrate.
630  * @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may
631  * not be successful in evicting other objects to make room for this object.
632  * @id: The region id to migrate to.
633  *
634  * Attempt to migrate the object to the desired memory region. The
635  * object backend must support migration and the object may not be
636  * pinned, (explicitly pinned pages or pinned vmas). The object must
637  * be locked.
638  * On successful completion, the object will have pages pointing to
639  * memory in the new region, but an async migration task may not have
640  * completed yet, and to accomplish that, i915_gem_object_wait_migration()
641  * must be called.
642  *
643  * Note: the @ww parameter is not used yet, but included to make sure
644  * callers put some effort into obtaining a valid ww ctx if one is
645  * available.
646  *
647  * Return: 0 on success. Negative error code on failure. In particular may
648  * return -ENXIO on lack of region space, -EDEADLK for deadlock avoidance
649  * if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and
650  * -EBUSY if the object is pinned.
651  */
i915_gem_object_migrate(struct drm_i915_gem_object * obj,struct i915_gem_ww_ctx * ww,enum intel_region_id id)652 int i915_gem_object_migrate(struct drm_i915_gem_object *obj,
653 			    struct i915_gem_ww_ctx *ww,
654 			    enum intel_region_id id)
655 {
656 	return __i915_gem_object_migrate(obj, ww, id, obj->flags);
657 }
658 
659 /**
660  * __i915_gem_object_migrate - Migrate an object to the desired region id, with
661  * control of the extra flags
662  * @obj: The object to migrate.
663  * @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may
664  * not be successful in evicting other objects to make room for this object.
665  * @id: The region id to migrate to.
666  * @flags: The object flags. Normally just obj->flags.
667  *
668  * Attempt to migrate the object to the desired memory region. The
669  * object backend must support migration and the object may not be
670  * pinned, (explicitly pinned pages or pinned vmas). The object must
671  * be locked.
672  * On successful completion, the object will have pages pointing to
673  * memory in the new region, but an async migration task may not have
674  * completed yet, and to accomplish that, i915_gem_object_wait_migration()
675  * must be called.
676  *
677  * Note: the @ww parameter is not used yet, but included to make sure
678  * callers put some effort into obtaining a valid ww ctx if one is
679  * available.
680  *
681  * Return: 0 on success. Negative error code on failure. In particular may
682  * return -ENXIO on lack of region space, -EDEADLK for deadlock avoidance
683  * if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and
684  * -EBUSY if the object is pinned.
685  */
__i915_gem_object_migrate(struct drm_i915_gem_object * obj,struct i915_gem_ww_ctx * ww,enum intel_region_id id,unsigned int flags)686 int __i915_gem_object_migrate(struct drm_i915_gem_object *obj,
687 			      struct i915_gem_ww_ctx *ww,
688 			      enum intel_region_id id,
689 			      unsigned int flags)
690 {
691 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
692 	struct intel_memory_region *mr;
693 
694 	GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
695 	GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
696 	assert_object_held(obj);
697 
698 	mr = i915->mm.regions[id];
699 	GEM_BUG_ON(!mr);
700 
701 	if (!i915_gem_object_can_migrate(obj, id))
702 		return -EINVAL;
703 
704 	if (!obj->ops->migrate) {
705 		if (GEM_WARN_ON(obj->mm.region != mr))
706 			return -EINVAL;
707 		return 0;
708 	}
709 
710 	return obj->ops->migrate(obj, mr, flags);
711 }
712 
713 /**
714  * i915_gem_object_placement_possible - Check whether the object can be
715  * placed at certain memory type
716  * @obj: Pointer to the object
717  * @type: The memory type to check
718  *
719  * Return: True if the object can be placed in @type. False otherwise.
720  */
i915_gem_object_placement_possible(struct drm_i915_gem_object * obj,enum intel_memory_type type)721 bool i915_gem_object_placement_possible(struct drm_i915_gem_object *obj,
722 					enum intel_memory_type type)
723 {
724 	unsigned int i;
725 
726 	if (!obj->mm.n_placements) {
727 		switch (type) {
728 		case INTEL_MEMORY_LOCAL:
729 			return i915_gem_object_has_iomem(obj);
730 		case INTEL_MEMORY_SYSTEM:
731 			return i915_gem_object_has_pages(obj);
732 		default:
733 			/* Ignore stolen for now */
734 			GEM_BUG_ON(1);
735 			return false;
736 		}
737 	}
738 
739 	for (i = 0; i < obj->mm.n_placements; i++) {
740 		if (obj->mm.placements[i]->type == type)
741 			return true;
742 	}
743 
744 	return false;
745 }
746 
747 /**
748  * i915_gem_object_needs_ccs_pages - Check whether the object requires extra
749  * pages when placed in system-memory, in order to save and later restore the
750  * flat-CCS aux state when the object is moved between local-memory and
751  * system-memory
752  * @obj: Pointer to the object
753  *
754  * Return: True if the object needs extra ccs pages. False otherwise.
755  */
i915_gem_object_needs_ccs_pages(struct drm_i915_gem_object * obj)756 bool i915_gem_object_needs_ccs_pages(struct drm_i915_gem_object *obj)
757 {
758 	bool lmem_placement = false;
759 	int i;
760 
761 	if (!HAS_FLAT_CCS(to_i915(obj->base.dev)))
762 		return false;
763 
764 	if (obj->flags & I915_BO_ALLOC_CCS_AUX)
765 		return true;
766 
767 	for (i = 0; i < obj->mm.n_placements; i++) {
768 		/* Compression is not allowed for the objects with smem placement */
769 		if (obj->mm.placements[i]->type == INTEL_MEMORY_SYSTEM)
770 			return false;
771 		if (!lmem_placement &&
772 		    obj->mm.placements[i]->type == INTEL_MEMORY_LOCAL)
773 			lmem_placement = true;
774 	}
775 
776 	return lmem_placement;
777 }
778 
i915_gem_init__objects(struct drm_i915_private * i915)779 void i915_gem_init__objects(struct drm_i915_private *i915)
780 {
781 	INIT_WORK(&i915->mm.free_work, __i915_gem_free_work);
782 }
783 
i915_objects_module_exit(void)784 void i915_objects_module_exit(void)
785 {
786 	kmem_cache_destroy(slab_objects);
787 }
788 
i915_objects_module_init(void)789 int __init i915_objects_module_init(void)
790 {
791 	slab_objects = KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN);
792 	if (!slab_objects)
793 		return -ENOMEM;
794 
795 	return 0;
796 }
797 
798 static const struct drm_gem_object_funcs i915_gem_object_funcs = {
799 	.free = i915_gem_free_object,
800 	.close = i915_gem_close_object,
801 	.export = i915_gem_prime_export,
802 };
803 
804 /**
805  * i915_gem_object_get_moving_fence - Get the object's moving fence if any
806  * @obj: The object whose moving fence to get.
807  * @fence: The resulting fence
808  *
809  * A non-signaled moving fence means that there is an async operation
810  * pending on the object that needs to be waited on before setting up
811  * any GPU- or CPU PTEs to the object's pages.
812  *
813  * Return: Negative error code or 0 for success.
814  */
i915_gem_object_get_moving_fence(struct drm_i915_gem_object * obj,struct dma_fence ** fence)815 int i915_gem_object_get_moving_fence(struct drm_i915_gem_object *obj,
816 				     struct dma_fence **fence)
817 {
818 	return dma_resv_get_singleton(obj->base.resv, DMA_RESV_USAGE_KERNEL,
819 				      fence);
820 }
821 
822 /**
823  * i915_gem_object_wait_moving_fence - Wait for the object's moving fence if any
824  * @obj: The object whose moving fence to wait for.
825  * @intr: Whether to wait interruptible.
826  *
827  * If the moving fence signaled without an error, it is detached from the
828  * object and put.
829  *
830  * Return: 0 if successful, -ERESTARTSYS if the wait was interrupted,
831  * negative error code if the async operation represented by the
832  * moving fence failed.
833  */
i915_gem_object_wait_moving_fence(struct drm_i915_gem_object * obj,bool intr)834 int i915_gem_object_wait_moving_fence(struct drm_i915_gem_object *obj,
835 				      bool intr)
836 {
837 	long ret;
838 
839 	assert_object_held(obj);
840 
841 	ret = dma_resv_wait_timeout(obj->base. resv, DMA_RESV_USAGE_KERNEL,
842 				    intr, MAX_SCHEDULE_TIMEOUT);
843 	if (!ret)
844 		ret = -ETIME;
845 	else if (ret > 0 && i915_gem_object_has_unknown_state(obj))
846 		ret = -EIO;
847 
848 	return ret < 0 ? ret : 0;
849 }
850 
851 /**
852  * i915_gem_object_has_unknown_state - Return true if the object backing pages are
853  * in an unknown_state. This means that userspace must NEVER be allowed to touch
854  * the pages, with either the GPU or CPU.
855  *
856  * ONLY valid to be called after ensuring that all kernel fences have signalled
857  * (in particular the fence for moving/clearing the object).
858  */
i915_gem_object_has_unknown_state(struct drm_i915_gem_object * obj)859 bool i915_gem_object_has_unknown_state(struct drm_i915_gem_object *obj)
860 {
861 	/*
862 	 * The below barrier pairs with the dma_fence_signal() in
863 	 * __memcpy_work(). We should only sample the unknown_state after all
864 	 * the kernel fences have signalled.
865 	 */
866 	smp_rmb();
867 	return obj->mm.unknown_state;
868 }
869 
870 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
871 #include "selftests/huge_gem_object.c"
872 #include "selftests/huge_pages.c"
873 #include "selftests/i915_gem_migrate.c"
874 #include "selftests/i915_gem_object.c"
875 #include "selftests/i915_gem_coherency.c"
876 #endif
877