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_object_frontbuffer.h"
41 #include "i915_gem_ttm.h"
42 #include "i915_memcpy.h"
43 #include "i915_trace.h"
44
45 static struct pool slab_objects;
46
47 static const struct drm_gem_object_funcs i915_gem_object_funcs;
48
i915_gem_get_pat_index(struct drm_i915_private * i915,enum i915_cache_level level)49 unsigned int i915_gem_get_pat_index(struct drm_i915_private *i915,
50 enum i915_cache_level level)
51 {
52 if (drm_WARN_ON(&i915->drm, level >= I915_MAX_CACHE_LEVEL))
53 return 0;
54
55 return INTEL_INFO(i915)->cachelevel_to_pat[level];
56 }
57
i915_gem_object_has_cache_level(const struct drm_i915_gem_object * obj,enum i915_cache_level lvl)58 bool i915_gem_object_has_cache_level(const struct drm_i915_gem_object *obj,
59 enum i915_cache_level lvl)
60 {
61 /*
62 * In case the pat_index is set by user space, this kernel mode
63 * driver should leave the coherency to be managed by user space,
64 * simply return true here.
65 */
66 if (obj->pat_set_by_user)
67 return true;
68
69 /*
70 * Otherwise the pat_index should have been converted from cache_level
71 * so that the following comparison is valid.
72 */
73 return obj->pat_index == i915_gem_get_pat_index(obj_to_i915(obj), lvl);
74 }
75
i915_gem_object_alloc(void)76 struct drm_i915_gem_object *i915_gem_object_alloc(void)
77 {
78 struct drm_i915_gem_object *obj;
79
80 #ifdef __linux__
81 obj = kmem_cache_zalloc(slab_objects, GFP_KERNEL);
82 #else
83 obj = pool_get(&slab_objects, PR_WAITOK | PR_ZERO);
84 #endif
85 if (!obj)
86 return NULL;
87 obj->base.funcs = &i915_gem_object_funcs;
88
89 return obj;
90 }
91
i915_gem_object_free(struct drm_i915_gem_object * obj)92 void i915_gem_object_free(struct drm_i915_gem_object *obj)
93 {
94 #ifdef __linux__
95 return kmem_cache_free(slab_objects, obj);
96 #else
97 pool_put(&slab_objects, obj);
98 #endif
99 }
100
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)101 void i915_gem_object_init(struct drm_i915_gem_object *obj,
102 const struct drm_i915_gem_object_ops *ops,
103 struct lock_class_key *key, unsigned flags)
104 {
105 /*
106 * A gem object is embedded both in a struct ttm_buffer_object :/ and
107 * in a drm_i915_gem_object. Make sure they are aliased.
108 */
109 BUILD_BUG_ON(offsetof(typeof(*obj), base) !=
110 offsetof(typeof(*obj), __do_not_access.base));
111
112 mtx_init(&obj->vma.lock, IPL_NONE);
113 INIT_LIST_HEAD(&obj->vma.list);
114
115 INIT_LIST_HEAD(&obj->mm.link);
116
117 #ifdef CONFIG_PROC_FS
118 INIT_LIST_HEAD(&obj->client_link);
119 #endif
120
121 INIT_LIST_HEAD(&obj->lut_list);
122 mtx_init(&obj->lut_lock, IPL_NONE);
123
124 mtx_init(&obj->mmo.lock, IPL_NONE);
125 obj->mmo.offsets = RB_ROOT;
126
127 init_rcu_head(&obj->rcu);
128
129 obj->ops = ops;
130 GEM_BUG_ON(flags & ~I915_BO_ALLOC_FLAGS);
131 obj->flags = flags;
132
133 obj->mm.madv = I915_MADV_WILLNEED;
134 INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN);
135 rw_init(&obj->mm.get_page.lock, "mmget");
136 INIT_RADIX_TREE(&obj->mm.get_dma_page.radix, GFP_KERNEL | __GFP_NOWARN);
137 rw_init(&obj->mm.get_dma_page.lock, "mmgetd");
138 }
139
140 /**
141 * __i915_gem_object_fini - Clean up a GEM object initialization
142 * @obj: The gem object to cleanup
143 *
144 * This function cleans up gem object fields that are set up by
145 * drm_gem_private_object_init() and i915_gem_object_init().
146 * It's primarily intended as a helper for backends that need to
147 * clean up the gem object in separate steps.
148 */
__i915_gem_object_fini(struct drm_i915_gem_object * obj)149 void __i915_gem_object_fini(struct drm_i915_gem_object *obj)
150 {
151 mutex_destroy(&obj->mm.get_page.lock);
152 mutex_destroy(&obj->mm.get_dma_page.lock);
153 dma_resv_fini(&obj->base._resv);
154 }
155
156 /**
157 * i915_gem_object_set_cache_coherency - Mark up the object's coherency levels
158 * for a given cache_level
159 * @obj: #drm_i915_gem_object
160 * @cache_level: cache level
161 */
i915_gem_object_set_cache_coherency(struct drm_i915_gem_object * obj,unsigned int cache_level)162 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj,
163 unsigned int cache_level)
164 {
165 struct drm_i915_private *i915 = to_i915(obj->base.dev);
166
167 obj->pat_index = i915_gem_get_pat_index(i915, cache_level);
168
169 if (cache_level != I915_CACHE_NONE)
170 obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ |
171 I915_BO_CACHE_COHERENT_FOR_WRITE);
172 else if (HAS_LLC(i915))
173 obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ;
174 else
175 obj->cache_coherent = 0;
176
177 obj->cache_dirty =
178 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE) &&
179 !IS_DGFX(i915);
180 }
181
182 /**
183 * i915_gem_object_set_pat_index - set PAT index to be used in PTE encode
184 * @obj: #drm_i915_gem_object
185 * @pat_index: PAT index
186 *
187 * This is a clone of i915_gem_object_set_cache_coherency taking pat index
188 * instead of cache_level as its second argument.
189 */
i915_gem_object_set_pat_index(struct drm_i915_gem_object * obj,unsigned int pat_index)190 void i915_gem_object_set_pat_index(struct drm_i915_gem_object *obj,
191 unsigned int pat_index)
192 {
193 struct drm_i915_private *i915 = to_i915(obj->base.dev);
194
195 if (obj->pat_index == pat_index)
196 return;
197
198 obj->pat_index = pat_index;
199
200 if (pat_index != i915_gem_get_pat_index(i915, I915_CACHE_NONE))
201 obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ |
202 I915_BO_CACHE_COHERENT_FOR_WRITE);
203 else if (HAS_LLC(i915))
204 obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ;
205 else
206 obj->cache_coherent = 0;
207
208 obj->cache_dirty =
209 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE) &&
210 !IS_DGFX(i915);
211 }
212
i915_gem_object_can_bypass_llc(struct drm_i915_gem_object * obj)213 bool i915_gem_object_can_bypass_llc(struct drm_i915_gem_object *obj)
214 {
215 struct drm_i915_private *i915 = to_i915(obj->base.dev);
216
217 /*
218 * This is purely from a security perspective, so we simply don't care
219 * about non-userspace objects being able to bypass the LLC.
220 */
221 if (!(obj->flags & I915_BO_ALLOC_USER))
222 return false;
223
224 /*
225 * Always flush cache for UMD objects at creation time.
226 */
227 if (obj->pat_set_by_user)
228 return true;
229
230 /*
231 * EHL and JSL add the 'Bypass LLC' MOCS entry, which should make it
232 * possible for userspace to bypass the GTT caching bits set by the
233 * kernel, as per the given object cache_level. This is troublesome
234 * since the heavy flush we apply when first gathering the pages is
235 * skipped if the kernel thinks the object is coherent with the GPU. As
236 * a result it might be possible to bypass the cache and read the
237 * contents of the page directly, which could be stale data. If it's
238 * just a case of userspace shooting themselves in the foot then so be
239 * it, but since i915 takes the stance of always zeroing memory before
240 * handing it to userspace, we need to prevent this.
241 */
242 return (IS_JASPERLAKE(i915) || IS_ELKHARTLAKE(i915));
243 }
244
i915_gem_close_object(struct drm_gem_object * gem,struct drm_file * file)245 static void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
246 {
247 struct drm_i915_gem_object *obj = to_intel_bo(gem);
248 struct drm_i915_file_private *fpriv = file->driver_priv;
249 struct i915_lut_handle bookmark = {};
250 struct i915_mmap_offset *mmo, *mn;
251 struct i915_lut_handle *lut, *ln;
252 DRM_LIST_HEAD(close);
253
254 spin_lock(&obj->lut_lock);
255 list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) {
256 struct i915_gem_context *ctx = lut->ctx;
257
258 if (ctx && ctx->file_priv == fpriv) {
259 i915_gem_context_get(ctx);
260 list_move(&lut->obj_link, &close);
261 }
262
263 /* Break long locks, and carefully continue on from this spot */
264 if (&ln->obj_link != &obj->lut_list) {
265 list_add_tail(&bookmark.obj_link, &ln->obj_link);
266 if (cond_resched_lock(&obj->lut_lock))
267 list_safe_reset_next(&bookmark, ln, obj_link);
268 __list_del_entry(&bookmark.obj_link);
269 }
270 }
271 spin_unlock(&obj->lut_lock);
272
273 spin_lock(&obj->mmo.lock);
274 rbtree_postorder_for_each_entry_safe(mmo, mn, &obj->mmo.offsets, offset)
275 drm_vma_node_revoke(&mmo->vma_node, file);
276 spin_unlock(&obj->mmo.lock);
277
278 list_for_each_entry_safe(lut, ln, &close, obj_link) {
279 struct i915_gem_context *ctx = lut->ctx;
280 struct i915_vma *vma;
281
282 /*
283 * We allow the process to have multiple handles to the same
284 * vma, in the same fd namespace, by virtue of flink/open.
285 */
286
287 mutex_lock(&ctx->lut_mutex);
288 vma = radix_tree_delete(&ctx->handles_vma, lut->handle);
289 if (vma) {
290 GEM_BUG_ON(vma->obj != obj);
291 GEM_BUG_ON(!atomic_read(&vma->open_count));
292 i915_vma_close(vma);
293 }
294 mutex_unlock(&ctx->lut_mutex);
295
296 i915_gem_context_put(lut->ctx);
297 i915_lut_handle_free(lut);
298 i915_gem_object_put(obj);
299 }
300 }
301
__i915_gem_free_object_rcu(struct rcu_head * head)302 void __i915_gem_free_object_rcu(struct rcu_head *head)
303 {
304 struct drm_i915_gem_object *obj =
305 container_of(head, typeof(*obj), rcu);
306 struct drm_i915_private *i915 = to_i915(obj->base.dev);
307
308 /* We need to keep this alive for RCU read access from fdinfo. */
309 if (obj->mm.n_placements > 1)
310 kfree(obj->mm.placements);
311
312 #ifdef __OpenBSD__
313 if (obj->base.uao)
314 uao_detach(obj->base.uao);
315 #endif
316
317 i915_gem_object_free(obj);
318
319 GEM_BUG_ON(!atomic_read(&i915->mm.free_count));
320 atomic_dec(&i915->mm.free_count);
321 }
322
__i915_gem_object_free_mmaps(struct drm_i915_gem_object * obj)323 static void __i915_gem_object_free_mmaps(struct drm_i915_gem_object *obj)
324 {
325 /* Skip serialisation and waking the device if known to be not used. */
326
327 if (obj->userfault_count && !IS_DGFX(to_i915(obj->base.dev)))
328 i915_gem_object_release_mmap_gtt(obj);
329
330 if (!RB_EMPTY_ROOT(&obj->mmo.offsets)) {
331 struct i915_mmap_offset *mmo, *mn;
332
333 i915_gem_object_release_mmap_offset(obj);
334
335 rbtree_postorder_for_each_entry_safe(mmo, mn,
336 &obj->mmo.offsets,
337 offset) {
338 drm_vma_offset_remove(obj->base.dev->vma_offset_manager,
339 &mmo->vma_node);
340 kfree(mmo);
341 }
342 obj->mmo.offsets = RB_ROOT;
343 }
344 }
345
346 /**
347 * __i915_gem_object_pages_fini - Clean up pages use of a gem object
348 * @obj: The gem object to clean up
349 *
350 * This function cleans up usage of the object mm.pages member. It
351 * is intended for backends that need to clean up a gem object in
352 * separate steps and needs to be called when the object is idle before
353 * the object's backing memory is freed.
354 */
__i915_gem_object_pages_fini(struct drm_i915_gem_object * obj)355 void __i915_gem_object_pages_fini(struct drm_i915_gem_object *obj)
356 {
357 assert_object_held_shared(obj);
358
359 if (!list_empty(&obj->vma.list)) {
360 struct i915_vma *vma;
361
362 spin_lock(&obj->vma.lock);
363 while ((vma = list_first_entry_or_null(&obj->vma.list,
364 struct i915_vma,
365 obj_link))) {
366 GEM_BUG_ON(vma->obj != obj);
367 spin_unlock(&obj->vma.lock);
368
369 i915_vma_destroy(vma);
370
371 spin_lock(&obj->vma.lock);
372 }
373 spin_unlock(&obj->vma.lock);
374 }
375
376 __i915_gem_object_free_mmaps(obj);
377
378 atomic_set(&obj->mm.pages_pin_count, 0);
379
380 /*
381 * dma_buf_unmap_attachment() requires reservation to be
382 * locked. The imported GEM shouldn't share reservation lock
383 * and ttm_bo_cleanup_memtype_use() shouldn't be invoked for
384 * dma-buf, so it's safe to take the lock.
385 */
386 if (obj->base.import_attach)
387 i915_gem_object_lock(obj, NULL);
388
389 __i915_gem_object_put_pages(obj);
390
391 if (obj->base.import_attach)
392 i915_gem_object_unlock(obj);
393
394 GEM_BUG_ON(i915_gem_object_has_pages(obj));
395 }
396
__i915_gem_free_object(struct drm_i915_gem_object * obj)397 void __i915_gem_free_object(struct drm_i915_gem_object *obj)
398 {
399 trace_i915_gem_object_destroy(obj);
400
401 GEM_BUG_ON(!list_empty(&obj->lut_list));
402
403 bitmap_free(obj->bit_17);
404
405 if (obj->base.import_attach)
406 drm_prime_gem_destroy(&obj->base, NULL);
407
408 drm_gem_free_mmap_offset(&obj->base);
409
410 if (obj->ops->release)
411 obj->ops->release(obj);
412
413 if (obj->shares_resv_from)
414 i915_vm_resv_put(obj->shares_resv_from);
415
416 __i915_gem_object_fini(obj);
417 }
418
__i915_gem_free_objects(struct drm_i915_private * i915,struct llist_node * freed)419 static void __i915_gem_free_objects(struct drm_i915_private *i915,
420 struct llist_node *freed)
421 {
422 struct drm_i915_gem_object *obj, *on;
423
424 llist_for_each_entry_safe(obj, on, freed, freed) {
425 might_sleep();
426 if (obj->ops->delayed_free) {
427 obj->ops->delayed_free(obj);
428 continue;
429 }
430
431 __i915_gem_object_pages_fini(obj);
432 __i915_gem_free_object(obj);
433
434 /* But keep the pointer alive for RCU-protected lookups */
435 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
436 cond_resched();
437 }
438 }
439
i915_gem_flush_free_objects(struct drm_i915_private * i915)440 void i915_gem_flush_free_objects(struct drm_i915_private *i915)
441 {
442 struct llist_node *freed = llist_del_all(&i915->mm.free_list);
443
444 if (unlikely(freed))
445 __i915_gem_free_objects(i915, freed);
446 }
447
__i915_gem_free_work(struct work_struct * work)448 static void __i915_gem_free_work(struct work_struct *work)
449 {
450 struct drm_i915_private *i915 =
451 container_of(work, struct drm_i915_private, mm.free_work);
452
453 i915_gem_flush_free_objects(i915);
454 }
455
i915_gem_free_object(struct drm_gem_object * gem_obj)456 static void i915_gem_free_object(struct drm_gem_object *gem_obj)
457 {
458 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
459 struct drm_i915_private *i915 = to_i915(obj->base.dev);
460
461 GEM_BUG_ON(i915_gem_object_is_framebuffer(obj));
462
463 i915_drm_client_remove_object(obj);
464
465 /*
466 * Before we free the object, make sure any pure RCU-only
467 * read-side critical sections are complete, e.g.
468 * i915_gem_busy_ioctl(). For the corresponding synchronized
469 * lookup see i915_gem_object_lookup_rcu().
470 */
471 atomic_inc(&i915->mm.free_count);
472
473 /*
474 * Since we require blocking on struct_mutex to unbind the freed
475 * object from the GPU before releasing resources back to the
476 * system, we can not do that directly from the RCU callback (which may
477 * be a softirq context), but must instead then defer that work onto a
478 * kthread. We use the RCU callback rather than move the freed object
479 * directly onto the work queue so that we can mix between using the
480 * worker and performing frees directly from subsequent allocations for
481 * crude but effective memory throttling.
482 */
483
484 if (llist_add(&obj->freed, &i915->mm.free_list))
485 queue_work(i915->wq, &i915->mm.free_work);
486 }
487
__i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object * obj,enum fb_op_origin origin)488 void __i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object *obj,
489 enum fb_op_origin origin)
490 {
491 struct intel_frontbuffer *front;
492
493 front = i915_gem_object_get_frontbuffer(obj);
494 if (front) {
495 intel_frontbuffer_flush(front, origin);
496 intel_frontbuffer_put(front);
497 }
498 }
499
__i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object * obj,enum fb_op_origin origin)500 void __i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object *obj,
501 enum fb_op_origin origin)
502 {
503 struct intel_frontbuffer *front;
504
505 front = i915_gem_object_get_frontbuffer(obj);
506 if (front) {
507 intel_frontbuffer_invalidate(front, origin);
508 intel_frontbuffer_put(front);
509 }
510 }
511
512 static void
i915_gem_object_read_from_page_kmap(struct drm_i915_gem_object * obj,u64 offset,void * dst,int size)513 i915_gem_object_read_from_page_kmap(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size)
514 {
515 pgoff_t idx = offset >> PAGE_SHIFT;
516 void *src_ptr;
517
518 src_ptr = kmap_local_page(i915_gem_object_get_page(obj, idx))
519 + offset_in_page(offset);
520 if (!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ))
521 drm_clflush_virt_range(src_ptr, size);
522 memcpy(dst, src_ptr, size);
523
524 kunmap_local(src_ptr);
525 }
526
527 static void
i915_gem_object_read_from_page_iomap(struct drm_i915_gem_object * obj,u64 offset,void * dst,int size)528 i915_gem_object_read_from_page_iomap(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size)
529 {
530 pgoff_t idx = offset >> PAGE_SHIFT;
531 dma_addr_t dma = i915_gem_object_get_dma_address(obj, idx);
532 void __iomem *src_map;
533 void __iomem *src_ptr;
534
535 src_map = io_mapping_map_wc(&obj->mm.region->iomap,
536 dma - obj->mm.region->region.start,
537 PAGE_SIZE);
538
539 src_ptr = src_map + offset_in_page(offset);
540 if (!i915_memcpy_from_wc(dst, (void __force *)src_ptr, size))
541 memcpy_fromio(dst, src_ptr, size);
542
543 io_mapping_unmap(src_map);
544 }
545
object_has_mappable_iomem(struct drm_i915_gem_object * obj)546 static bool object_has_mappable_iomem(struct drm_i915_gem_object *obj)
547 {
548 GEM_BUG_ON(!i915_gem_object_has_iomem(obj));
549
550 if (IS_DGFX(to_i915(obj->base.dev)))
551 return i915_ttm_resource_mappable(i915_gem_to_ttm(obj)->resource);
552
553 return true;
554 }
555
556 /**
557 * i915_gem_object_read_from_page - read data from the page of a GEM object
558 * @obj: GEM object to read from
559 * @offset: offset within the object
560 * @dst: buffer to store the read data
561 * @size: size to read
562 *
563 * Reads data from @obj at the specified offset. The requested region to read
564 * from can't cross a page boundary. The caller must ensure that @obj pages
565 * are pinned and that @obj is synced wrt. any related writes.
566 *
567 * Return: %0 on success or -ENODEV if the type of @obj's backing store is
568 * unsupported.
569 */
i915_gem_object_read_from_page(struct drm_i915_gem_object * obj,u64 offset,void * dst,int size)570 int i915_gem_object_read_from_page(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size)
571 {
572 GEM_BUG_ON(overflows_type(offset >> PAGE_SHIFT, pgoff_t));
573 GEM_BUG_ON(offset >= obj->base.size);
574 GEM_BUG_ON(offset_in_page(offset) > PAGE_SIZE - size);
575 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
576
577 if (i915_gem_object_has_struct_page(obj))
578 i915_gem_object_read_from_page_kmap(obj, offset, dst, size);
579 else if (i915_gem_object_has_iomem(obj) && object_has_mappable_iomem(obj))
580 i915_gem_object_read_from_page_iomap(obj, offset, dst, size);
581 else
582 return -ENODEV;
583
584 return 0;
585 }
586
587 /**
588 * i915_gem_object_evictable - Whether object is likely evictable after unbind.
589 * @obj: The object to check
590 *
591 * This function checks whether the object is likely unvictable after unbind.
592 * If the object is not locked when checking, the result is only advisory.
593 * If the object is locked when checking, and the function returns true,
594 * then an eviction should indeed be possible. But since unlocked vma
595 * unpinning and unbinding is currently possible, the object can actually
596 * become evictable even if this function returns false.
597 *
598 * Return: true if the object may be evictable. False otherwise.
599 */
i915_gem_object_evictable(struct drm_i915_gem_object * obj)600 bool i915_gem_object_evictable(struct drm_i915_gem_object *obj)
601 {
602 struct i915_vma *vma;
603 int pin_count = atomic_read(&obj->mm.pages_pin_count);
604
605 if (!pin_count)
606 return true;
607
608 spin_lock(&obj->vma.lock);
609 list_for_each_entry(vma, &obj->vma.list, obj_link) {
610 if (i915_vma_is_pinned(vma)) {
611 spin_unlock(&obj->vma.lock);
612 return false;
613 }
614 if (atomic_read(&vma->pages_count))
615 pin_count--;
616 }
617 spin_unlock(&obj->vma.lock);
618 GEM_WARN_ON(pin_count < 0);
619
620 return pin_count == 0;
621 }
622
623 /**
624 * i915_gem_object_migratable - Whether the object is migratable out of the
625 * current region.
626 * @obj: Pointer to the object.
627 *
628 * Return: Whether the object is allowed to be resident in other
629 * regions than the current while pages are present.
630 */
i915_gem_object_migratable(struct drm_i915_gem_object * obj)631 bool i915_gem_object_migratable(struct drm_i915_gem_object *obj)
632 {
633 struct intel_memory_region *mr = READ_ONCE(obj->mm.region);
634
635 if (!mr)
636 return false;
637
638 return obj->mm.n_placements > 1;
639 }
640
641 /**
642 * i915_gem_object_has_struct_page - Whether the object is page-backed
643 * @obj: The object to query.
644 *
645 * This function should only be called while the object is locked or pinned,
646 * otherwise the page backing may change under the caller.
647 *
648 * Return: True if page-backed, false otherwise.
649 */
i915_gem_object_has_struct_page(const struct drm_i915_gem_object * obj)650 bool i915_gem_object_has_struct_page(const struct drm_i915_gem_object *obj)
651 {
652 #ifdef CONFIG_LOCKDEP
653 if (IS_DGFX(to_i915(obj->base.dev)) &&
654 i915_gem_object_evictable((void __force *)obj))
655 assert_object_held_shared(obj);
656 #endif
657 return obj->mem_flags & I915_BO_FLAG_STRUCT_PAGE;
658 }
659
660 /**
661 * i915_gem_object_has_iomem - Whether the object is iomem-backed
662 * @obj: The object to query.
663 *
664 * This function should only be called while the object is locked or pinned,
665 * otherwise the iomem backing may change under the caller.
666 *
667 * Return: True if iomem-backed, false otherwise.
668 */
i915_gem_object_has_iomem(const struct drm_i915_gem_object * obj)669 bool i915_gem_object_has_iomem(const struct drm_i915_gem_object *obj)
670 {
671 #ifdef CONFIG_LOCKDEP
672 if (IS_DGFX(to_i915(obj->base.dev)) &&
673 i915_gem_object_evictable((void __force *)obj))
674 assert_object_held_shared(obj);
675 #endif
676 return obj->mem_flags & I915_BO_FLAG_IOMEM;
677 }
678
679 /**
680 * i915_gem_object_can_migrate - Whether an object likely can be migrated
681 *
682 * @obj: The object to migrate
683 * @id: The region intended to migrate to
684 *
685 * Check whether the object backend supports migration to the
686 * given region. Note that pinning may affect the ability to migrate as
687 * returned by this function.
688 *
689 * This function is primarily intended as a helper for checking the
690 * possibility to migrate objects and might be slightly less permissive
691 * than i915_gem_object_migrate() when it comes to objects with the
692 * I915_BO_ALLOC_USER flag set.
693 *
694 * Return: true if migration is possible, false otherwise.
695 */
i915_gem_object_can_migrate(struct drm_i915_gem_object * obj,enum intel_region_id id)696 bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj,
697 enum intel_region_id id)
698 {
699 struct drm_i915_private *i915 = to_i915(obj->base.dev);
700 unsigned int num_allowed = obj->mm.n_placements;
701 struct intel_memory_region *mr;
702 unsigned int i;
703
704 GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
705 GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
706
707 mr = i915->mm.regions[id];
708 if (!mr)
709 return false;
710
711 if (!IS_ALIGNED(obj->base.size, mr->min_page_size))
712 return false;
713
714 if (obj->mm.region == mr)
715 return true;
716
717 if (!i915_gem_object_evictable(obj))
718 return false;
719
720 if (!obj->ops->migrate)
721 return false;
722
723 if (!(obj->flags & I915_BO_ALLOC_USER))
724 return true;
725
726 if (num_allowed == 0)
727 return false;
728
729 for (i = 0; i < num_allowed; ++i) {
730 if (mr == obj->mm.placements[i])
731 return true;
732 }
733
734 return false;
735 }
736
737 /**
738 * i915_gem_object_migrate - Migrate an object to the desired region id
739 * @obj: The object to migrate.
740 * @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may
741 * not be successful in evicting other objects to make room for this object.
742 * @id: The region id to migrate to.
743 *
744 * Attempt to migrate the object to the desired memory region. The
745 * object backend must support migration and the object may not be
746 * pinned, (explicitly pinned pages or pinned vmas). The object must
747 * be locked.
748 * On successful completion, the object will have pages pointing to
749 * memory in the new region, but an async migration task may not have
750 * completed yet, and to accomplish that, i915_gem_object_wait_migration()
751 * must be called.
752 *
753 * Note: the @ww parameter is not used yet, but included to make sure
754 * callers put some effort into obtaining a valid ww ctx if one is
755 * available.
756 *
757 * Return: 0 on success. Negative error code on failure. In particular may
758 * return -ENXIO on lack of region space, -EDEADLK for deadlock avoidance
759 * if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and
760 * -EBUSY if the object is pinned.
761 */
i915_gem_object_migrate(struct drm_i915_gem_object * obj,struct i915_gem_ww_ctx * ww,enum intel_region_id id)762 int i915_gem_object_migrate(struct drm_i915_gem_object *obj,
763 struct i915_gem_ww_ctx *ww,
764 enum intel_region_id id)
765 {
766 return __i915_gem_object_migrate(obj, ww, id, obj->flags);
767 }
768
769 /**
770 * __i915_gem_object_migrate - Migrate an object to the desired region id, with
771 * control of the extra flags
772 * @obj: The object to migrate.
773 * @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may
774 * not be successful in evicting other objects to make room for this object.
775 * @id: The region id to migrate to.
776 * @flags: The object flags. Normally just obj->flags.
777 *
778 * Attempt to migrate the object to the desired memory region. The
779 * object backend must support migration and the object may not be
780 * pinned, (explicitly pinned pages or pinned vmas). The object must
781 * be locked.
782 * On successful completion, the object will have pages pointing to
783 * memory in the new region, but an async migration task may not have
784 * completed yet, and to accomplish that, i915_gem_object_wait_migration()
785 * must be called.
786 *
787 * Note: the @ww parameter is not used yet, but included to make sure
788 * callers put some effort into obtaining a valid ww ctx if one is
789 * available.
790 *
791 * Return: 0 on success. Negative error code on failure. In particular may
792 * return -ENXIO on lack of region space, -EDEADLK for deadlock avoidance
793 * if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and
794 * -EBUSY if the object is pinned.
795 */
__i915_gem_object_migrate(struct drm_i915_gem_object * obj,struct i915_gem_ww_ctx * ww,enum intel_region_id id,unsigned int flags)796 int __i915_gem_object_migrate(struct drm_i915_gem_object *obj,
797 struct i915_gem_ww_ctx *ww,
798 enum intel_region_id id,
799 unsigned int flags)
800 {
801 struct drm_i915_private *i915 = to_i915(obj->base.dev);
802 struct intel_memory_region *mr;
803
804 GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
805 GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
806 assert_object_held(obj);
807
808 mr = i915->mm.regions[id];
809 GEM_BUG_ON(!mr);
810
811 if (!i915_gem_object_can_migrate(obj, id))
812 return -EINVAL;
813
814 if (!obj->ops->migrate) {
815 if (GEM_WARN_ON(obj->mm.region != mr))
816 return -EINVAL;
817 return 0;
818 }
819
820 return obj->ops->migrate(obj, mr, flags);
821 }
822
823 /**
824 * i915_gem_object_placement_possible - Check whether the object can be
825 * placed at certain memory type
826 * @obj: Pointer to the object
827 * @type: The memory type to check
828 *
829 * Return: True if the object can be placed in @type. False otherwise.
830 */
i915_gem_object_placement_possible(struct drm_i915_gem_object * obj,enum intel_memory_type type)831 bool i915_gem_object_placement_possible(struct drm_i915_gem_object *obj,
832 enum intel_memory_type type)
833 {
834 unsigned int i;
835
836 if (!obj->mm.n_placements) {
837 switch (type) {
838 case INTEL_MEMORY_LOCAL:
839 return i915_gem_object_has_iomem(obj);
840 case INTEL_MEMORY_SYSTEM:
841 return i915_gem_object_has_pages(obj);
842 default:
843 /* Ignore stolen for now */
844 GEM_BUG_ON(1);
845 return false;
846 }
847 }
848
849 for (i = 0; i < obj->mm.n_placements; i++) {
850 if (obj->mm.placements[i]->type == type)
851 return true;
852 }
853
854 return false;
855 }
856
857 /**
858 * i915_gem_object_needs_ccs_pages - Check whether the object requires extra
859 * pages when placed in system-memory, in order to save and later restore the
860 * flat-CCS aux state when the object is moved between local-memory and
861 * system-memory
862 * @obj: Pointer to the object
863 *
864 * Return: True if the object needs extra ccs pages. False otherwise.
865 */
i915_gem_object_needs_ccs_pages(struct drm_i915_gem_object * obj)866 bool i915_gem_object_needs_ccs_pages(struct drm_i915_gem_object *obj)
867 {
868 bool lmem_placement = false;
869 int i;
870
871 if (!HAS_FLAT_CCS(to_i915(obj->base.dev)))
872 return false;
873
874 if (obj->flags & I915_BO_ALLOC_CCS_AUX)
875 return true;
876
877 for (i = 0; i < obj->mm.n_placements; i++) {
878 /* Compression is not allowed for the objects with smem placement */
879 if (obj->mm.placements[i]->type == INTEL_MEMORY_SYSTEM)
880 return false;
881 if (!lmem_placement &&
882 obj->mm.placements[i]->type == INTEL_MEMORY_LOCAL)
883 lmem_placement = true;
884 }
885
886 return lmem_placement;
887 }
888
i915_gem_init__objects(struct drm_i915_private * i915)889 void i915_gem_init__objects(struct drm_i915_private *i915)
890 {
891 INIT_WORK(&i915->mm.free_work, __i915_gem_free_work);
892 }
893
i915_objects_module_exit(void)894 void i915_objects_module_exit(void)
895 {
896 #ifdef __linux__
897 kmem_cache_destroy(slab_objects);
898 #else
899 pool_destroy(&slab_objects);
900 #endif
901 }
902
i915_objects_module_init(void)903 int __init i915_objects_module_init(void)
904 {
905 #ifdef __linux__
906 slab_objects = KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN);
907 if (!slab_objects)
908 return -ENOMEM;
909 #else
910 pool_init(&slab_objects, sizeof(struct drm_i915_gem_object),
911 CACHELINESIZE, IPL_NONE, 0, "drmobj", NULL);
912 #endif
913
914 return 0;
915 }
916
917 static const struct drm_gem_object_funcs i915_gem_object_funcs = {
918 .free = i915_gem_free_object,
919 .close = i915_gem_close_object,
920 .export = i915_gem_prime_export,
921 };
922
923 /**
924 * i915_gem_object_get_moving_fence - Get the object's moving fence if any
925 * @obj: The object whose moving fence to get.
926 * @fence: The resulting fence
927 *
928 * A non-signaled moving fence means that there is an async operation
929 * pending on the object that needs to be waited on before setting up
930 * any GPU- or CPU PTEs to the object's pages.
931 *
932 * Return: Negative error code or 0 for success.
933 */
i915_gem_object_get_moving_fence(struct drm_i915_gem_object * obj,struct dma_fence ** fence)934 int i915_gem_object_get_moving_fence(struct drm_i915_gem_object *obj,
935 struct dma_fence **fence)
936 {
937 return dma_resv_get_singleton(obj->base.resv, DMA_RESV_USAGE_KERNEL,
938 fence);
939 }
940
941 /**
942 * i915_gem_object_wait_moving_fence - Wait for the object's moving fence if any
943 * @obj: The object whose moving fence to wait for.
944 * @intr: Whether to wait interruptible.
945 *
946 * If the moving fence signaled without an error, it is detached from the
947 * object and put.
948 *
949 * Return: 0 if successful, -ERESTARTSYS if the wait was interrupted,
950 * negative error code if the async operation represented by the
951 * moving fence failed.
952 */
i915_gem_object_wait_moving_fence(struct drm_i915_gem_object * obj,bool intr)953 int i915_gem_object_wait_moving_fence(struct drm_i915_gem_object *obj,
954 bool intr)
955 {
956 long ret;
957
958 assert_object_held(obj);
959
960 ret = dma_resv_wait_timeout(obj->base. resv, DMA_RESV_USAGE_KERNEL,
961 intr, MAX_SCHEDULE_TIMEOUT);
962 if (!ret)
963 ret = -ETIME;
964 else if (ret > 0 && i915_gem_object_has_unknown_state(obj))
965 ret = -EIO;
966
967 return ret < 0 ? ret : 0;
968 }
969
970 /*
971 * i915_gem_object_has_unknown_state - Return true if the object backing pages are
972 * in an unknown_state. This means that userspace must NEVER be allowed to touch
973 * the pages, with either the GPU or CPU.
974 *
975 * ONLY valid to be called after ensuring that all kernel fences have signalled
976 * (in particular the fence for moving/clearing the object).
977 */
i915_gem_object_has_unknown_state(struct drm_i915_gem_object * obj)978 bool i915_gem_object_has_unknown_state(struct drm_i915_gem_object *obj)
979 {
980 /*
981 * The below barrier pairs with the dma_fence_signal() in
982 * __memcpy_work(). We should only sample the unknown_state after all
983 * the kernel fences have signalled.
984 */
985 smp_rmb();
986 return obj->mm.unknown_state;
987 }
988
989 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
990 #include "selftests/huge_gem_object.c"
991 #include "selftests/huge_pages.c"
992 #include "selftests/i915_gem_migrate.c"
993 #include "selftests/i915_gem_object.c"
994 #include "selftests/i915_gem_coherency.c"
995 #endif
996