1 /*
2 * Copyright © 2016 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/sched/mm.h>
26 #include <drm/drm_gem.h>
27
28 #include "display/intel_frontbuffer.h"
29
30 #include "gem/i915_gem_lmem.h"
31 #include "gt/intel_engine.h"
32 #include "gt/intel_engine_heartbeat.h"
33 #include "gt/intel_gt.h"
34 #include "gt/intel_gt_requests.h"
35
36 #include "i915_drv.h"
37 #include "i915_sw_fence_work.h"
38 #include "i915_trace.h"
39 #include "i915_vma.h"
40
41 static struct kmem_cache *slab_vmas;
42
i915_vma_alloc(void)43 struct i915_vma *i915_vma_alloc(void)
44 {
45 return kmem_cache_zalloc(slab_vmas, GFP_KERNEL);
46 }
47
i915_vma_free(struct i915_vma * vma)48 void i915_vma_free(struct i915_vma *vma)
49 {
50 return kmem_cache_free(slab_vmas, vma);
51 }
52
53 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
54
55 #include <linux/stackdepot.h>
56
vma_print_allocator(struct i915_vma * vma,const char * reason)57 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
58 {
59 unsigned long *entries;
60 unsigned int nr_entries;
61 char buf[512];
62
63 if (!vma->node.stack) {
64 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
65 vma->node.start, vma->node.size, reason);
66 return;
67 }
68
69 nr_entries = stack_depot_fetch(vma->node.stack, &entries);
70 stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
71 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
72 vma->node.start, vma->node.size, reason, buf);
73 }
74
75 #else
76
vma_print_allocator(struct i915_vma * vma,const char * reason)77 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
78 {
79 }
80
81 #endif
82
active_to_vma(struct i915_active * ref)83 static inline struct i915_vma *active_to_vma(struct i915_active *ref)
84 {
85 return container_of(ref, typeof(struct i915_vma), active);
86 }
87
__i915_vma_active(struct i915_active * ref)88 static int __i915_vma_active(struct i915_active *ref)
89 {
90 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
91 }
92
__i915_vma_retire(struct i915_active * ref)93 static void __i915_vma_retire(struct i915_active *ref)
94 {
95 i915_vma_put(active_to_vma(ref));
96 }
97
98 static struct i915_vma *
vma_create(struct drm_i915_gem_object * obj,struct i915_address_space * vm,const struct i915_ggtt_view * view)99 vma_create(struct drm_i915_gem_object *obj,
100 struct i915_address_space *vm,
101 const struct i915_ggtt_view *view)
102 {
103 struct i915_vma *pos = ERR_PTR(-E2BIG);
104 struct i915_vma *vma;
105 struct rb_node *rb, **p;
106
107 /* The aliasing_ppgtt should never be used directly! */
108 GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
109
110 vma = i915_vma_alloc();
111 if (vma == NULL)
112 return ERR_PTR(-ENOMEM);
113
114 kref_init(&vma->ref);
115 mutex_init(&vma->pages_mutex);
116 vma->vm = i915_vm_get(vm);
117 vma->ops = &vm->vma_ops;
118 vma->obj = obj;
119 vma->resv = obj->base.resv;
120 vma->size = obj->base.size;
121 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
122
123 i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0);
124
125 /* Declare ourselves safe for use inside shrinkers */
126 if (IS_ENABLED(CONFIG_LOCKDEP)) {
127 fs_reclaim_acquire(GFP_KERNEL);
128 might_lock(&vma->active.mutex);
129 fs_reclaim_release(GFP_KERNEL);
130 }
131
132 INIT_LIST_HEAD(&vma->closed_link);
133
134 if (view && view->type != I915_GGTT_VIEW_NORMAL) {
135 vma->ggtt_view = *view;
136 if (view->type == I915_GGTT_VIEW_PARTIAL) {
137 GEM_BUG_ON(range_overflows_t(u64,
138 view->partial.offset,
139 view->partial.size,
140 obj->base.size >> PAGE_SHIFT));
141 vma->size = view->partial.size;
142 vma->size <<= PAGE_SHIFT;
143 GEM_BUG_ON(vma->size > obj->base.size);
144 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
145 vma->size = intel_rotation_info_size(&view->rotated);
146 vma->size <<= PAGE_SHIFT;
147 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
148 vma->size = intel_remapped_info_size(&view->remapped);
149 vma->size <<= PAGE_SHIFT;
150 }
151 }
152
153 if (unlikely(vma->size > vm->total))
154 goto err_vma;
155
156 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
157
158 spin_lock(&obj->vma.lock);
159
160 if (i915_is_ggtt(vm)) {
161 if (unlikely(overflows_type(vma->size, u32)))
162 goto err_unlock;
163
164 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
165 i915_gem_object_get_tiling(obj),
166 i915_gem_object_get_stride(obj));
167 if (unlikely(vma->fence_size < vma->size || /* overflow */
168 vma->fence_size > vm->total))
169 goto err_unlock;
170
171 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
172
173 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
174 i915_gem_object_get_tiling(obj),
175 i915_gem_object_get_stride(obj));
176 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
177
178 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
179 }
180
181 rb = NULL;
182 p = &obj->vma.tree.rb_node;
183 while (*p) {
184 long cmp;
185
186 rb = *p;
187 pos = rb_entry(rb, struct i915_vma, obj_node);
188
189 /*
190 * If the view already exists in the tree, another thread
191 * already created a matching vma, so return the older instance
192 * and dispose of ours.
193 */
194 cmp = i915_vma_compare(pos, vm, view);
195 if (cmp < 0)
196 p = &rb->rb_right;
197 else if (cmp > 0)
198 p = &rb->rb_left;
199 else
200 goto err_unlock;
201 }
202 rb_link_node(&vma->obj_node, rb, p);
203 rb_insert_color(&vma->obj_node, &obj->vma.tree);
204
205 if (i915_vma_is_ggtt(vma))
206 /*
207 * We put the GGTT vma at the start of the vma-list, followed
208 * by the ppGGTT vma. This allows us to break early when
209 * iterating over only the GGTT vma for an object, see
210 * for_each_ggtt_vma()
211 */
212 list_add(&vma->obj_link, &obj->vma.list);
213 else
214 list_add_tail(&vma->obj_link, &obj->vma.list);
215
216 spin_unlock(&obj->vma.lock);
217
218 return vma;
219
220 err_unlock:
221 spin_unlock(&obj->vma.lock);
222 err_vma:
223 i915_vm_put(vm);
224 i915_vma_free(vma);
225 return pos;
226 }
227
228 static struct i915_vma *
i915_vma_lookup(struct drm_i915_gem_object * obj,struct i915_address_space * vm,const struct i915_ggtt_view * view)229 i915_vma_lookup(struct drm_i915_gem_object *obj,
230 struct i915_address_space *vm,
231 const struct i915_ggtt_view *view)
232 {
233 struct rb_node *rb;
234
235 rb = obj->vma.tree.rb_node;
236 while (rb) {
237 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
238 long cmp;
239
240 cmp = i915_vma_compare(vma, vm, view);
241 if (cmp == 0)
242 return vma;
243
244 if (cmp < 0)
245 rb = rb->rb_right;
246 else
247 rb = rb->rb_left;
248 }
249
250 return NULL;
251 }
252
253 /**
254 * i915_vma_instance - return the singleton instance of the VMA
255 * @obj: parent &struct drm_i915_gem_object to be mapped
256 * @vm: address space in which the mapping is located
257 * @view: additional mapping requirements
258 *
259 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
260 * the same @view characteristics. If a match is not found, one is created.
261 * Once created, the VMA is kept until either the object is freed, or the
262 * address space is closed.
263 *
264 * Returns the vma, or an error pointer.
265 */
266 struct i915_vma *
i915_vma_instance(struct drm_i915_gem_object * obj,struct i915_address_space * vm,const struct i915_ggtt_view * view)267 i915_vma_instance(struct drm_i915_gem_object *obj,
268 struct i915_address_space *vm,
269 const struct i915_ggtt_view *view)
270 {
271 struct i915_vma *vma;
272
273 GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm));
274 GEM_BUG_ON(!atomic_read(&vm->open));
275
276 spin_lock(&obj->vma.lock);
277 vma = i915_vma_lookup(obj, vm, view);
278 spin_unlock(&obj->vma.lock);
279
280 /* vma_create() will resolve the race if another creates the vma */
281 if (unlikely(!vma))
282 vma = vma_create(obj, vm, view);
283
284 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
285 return vma;
286 }
287
288 struct i915_vma_work {
289 struct dma_fence_work base;
290 struct i915_address_space *vm;
291 struct i915_vm_pt_stash stash;
292 struct i915_vma *vma;
293 struct drm_i915_gem_object *pinned;
294 struct i915_sw_dma_fence_cb cb;
295 enum i915_cache_level cache_level;
296 unsigned int flags;
297 };
298
__vma_bind(struct dma_fence_work * work)299 static void __vma_bind(struct dma_fence_work *work)
300 {
301 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
302 struct i915_vma *vma = vw->vma;
303
304 vma->ops->bind_vma(vw->vm, &vw->stash,
305 vma, vw->cache_level, vw->flags);
306 }
307
__vma_release(struct dma_fence_work * work)308 static void __vma_release(struct dma_fence_work *work)
309 {
310 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
311
312 if (vw->pinned) {
313 __i915_gem_object_unpin_pages(vw->pinned);
314 i915_gem_object_put(vw->pinned);
315 }
316
317 i915_vm_free_pt_stash(vw->vm, &vw->stash);
318 i915_vm_put(vw->vm);
319 }
320
321 static const struct dma_fence_work_ops bind_ops = {
322 .name = "bind",
323 .work = __vma_bind,
324 .release = __vma_release,
325 };
326
i915_vma_work(void)327 struct i915_vma_work *i915_vma_work(void)
328 {
329 struct i915_vma_work *vw;
330
331 vw = kzalloc(sizeof(*vw), GFP_KERNEL);
332 if (!vw)
333 return NULL;
334
335 dma_fence_work_init(&vw->base, &bind_ops);
336 vw->base.dma.error = -EAGAIN; /* disable the worker by default */
337
338 return vw;
339 }
340
i915_vma_wait_for_bind(struct i915_vma * vma)341 int i915_vma_wait_for_bind(struct i915_vma *vma)
342 {
343 int err = 0;
344
345 if (rcu_access_pointer(vma->active.excl.fence)) {
346 struct dma_fence *fence;
347
348 rcu_read_lock();
349 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
350 rcu_read_unlock();
351 if (fence) {
352 err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
353 dma_fence_put(fence);
354 }
355 }
356
357 return err;
358 }
359
360 /**
361 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
362 * @vma: VMA to map
363 * @cache_level: mapping cache level
364 * @flags: flags like global or local mapping
365 * @work: preallocated worker for allocating and binding the PTE
366 *
367 * DMA addresses are taken from the scatter-gather table of this object (or of
368 * this VMA in case of non-default GGTT views) and PTE entries set up.
369 * Note that DMA addresses are also the only part of the SG table we care about.
370 */
i915_vma_bind(struct i915_vma * vma,enum i915_cache_level cache_level,u32 flags,struct i915_vma_work * work)371 int i915_vma_bind(struct i915_vma *vma,
372 enum i915_cache_level cache_level,
373 u32 flags,
374 struct i915_vma_work *work)
375 {
376 u32 bind_flags;
377 u32 vma_flags;
378
379 lockdep_assert_held(&vma->vm->mutex);
380 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
381 GEM_BUG_ON(vma->size > vma->node.size);
382
383 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
384 vma->node.size,
385 vma->vm->total)))
386 return -ENODEV;
387
388 if (GEM_DEBUG_WARN_ON(!flags))
389 return -EINVAL;
390
391 bind_flags = flags;
392 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
393
394 vma_flags = atomic_read(&vma->flags);
395 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
396
397 bind_flags &= ~vma_flags;
398 if (bind_flags == 0)
399 return 0;
400
401 GEM_BUG_ON(!vma->pages);
402
403 trace_i915_vma_bind(vma, bind_flags);
404 if (work && bind_flags & vma->vm->bind_async_flags) {
405 struct dma_fence *prev;
406
407 work->vma = vma;
408 work->cache_level = cache_level;
409 work->flags = bind_flags;
410
411 /*
412 * Note we only want to chain up to the migration fence on
413 * the pages (not the object itself). As we don't track that,
414 * yet, we have to use the exclusive fence instead.
415 *
416 * Also note that we do not want to track the async vma as
417 * part of the obj->resv->excl_fence as it only affects
418 * execution and not content or object's backing store lifetime.
419 */
420 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
421 if (prev) {
422 __i915_sw_fence_await_dma_fence(&work->base.chain,
423 prev,
424 &work->cb);
425 dma_fence_put(prev);
426 }
427
428 work->base.dma.error = 0; /* enable the queue_work() */
429
430 if (vma->obj) {
431 __i915_gem_object_pin_pages(vma->obj);
432 work->pinned = i915_gem_object_get(vma->obj);
433 }
434 } else {
435 vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags);
436 }
437
438 if (vma->obj)
439 set_bit(I915_BO_WAS_BOUND_BIT, &vma->obj->flags);
440
441 atomic_or(bind_flags, &vma->flags);
442 return 0;
443 }
444
i915_vma_pin_iomap(struct i915_vma * vma)445 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
446 {
447 void __iomem *ptr;
448 int err;
449
450 if (!i915_gem_object_is_lmem(vma->obj)) {
451 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
452 err = -ENODEV;
453 goto err;
454 }
455 }
456
457 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
458 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
459
460 ptr = READ_ONCE(vma->iomap);
461 if (ptr == NULL) {
462 /*
463 * TODO: consider just using i915_gem_object_pin_map() for lmem
464 * instead, which already supports mapping non-contiguous chunks
465 * of pages, that way we can also drop the
466 * I915_BO_ALLOC_CONTIGUOUS when allocating the object.
467 */
468 if (i915_gem_object_is_lmem(vma->obj))
469 ptr = i915_gem_object_lmem_io_map(vma->obj, 0,
470 vma->obj->base.size);
471 else
472 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
473 vma->node.start,
474 vma->node.size);
475 if (ptr == NULL) {
476 err = -ENOMEM;
477 goto err;
478 }
479
480 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
481 io_mapping_unmap(ptr);
482 ptr = vma->iomap;
483 }
484 }
485
486 __i915_vma_pin(vma);
487
488 err = i915_vma_pin_fence(vma);
489 if (err)
490 goto err_unpin;
491
492 i915_vma_set_ggtt_write(vma);
493
494 /* NB Access through the GTT requires the device to be awake. */
495 return ptr;
496
497 err_unpin:
498 __i915_vma_unpin(vma);
499 err:
500 return IO_ERR_PTR(err);
501 }
502
i915_vma_flush_writes(struct i915_vma * vma)503 void i915_vma_flush_writes(struct i915_vma *vma)
504 {
505 if (i915_vma_unset_ggtt_write(vma))
506 intel_gt_flush_ggtt_writes(vma->vm->gt);
507 }
508
i915_vma_unpin_iomap(struct i915_vma * vma)509 void i915_vma_unpin_iomap(struct i915_vma *vma)
510 {
511 GEM_BUG_ON(vma->iomap == NULL);
512
513 i915_vma_flush_writes(vma);
514
515 i915_vma_unpin_fence(vma);
516 i915_vma_unpin(vma);
517 }
518
i915_vma_unpin_and_release(struct i915_vma ** p_vma,unsigned int flags)519 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
520 {
521 struct i915_vma *vma;
522 struct drm_i915_gem_object *obj;
523
524 vma = fetch_and_zero(p_vma);
525 if (!vma)
526 return;
527
528 obj = vma->obj;
529 GEM_BUG_ON(!obj);
530
531 i915_vma_unpin(vma);
532
533 if (flags & I915_VMA_RELEASE_MAP)
534 i915_gem_object_unpin_map(obj);
535
536 i915_gem_object_put(obj);
537 }
538
i915_vma_misplaced(const struct i915_vma * vma,u64 size,u64 alignment,u64 flags)539 bool i915_vma_misplaced(const struct i915_vma *vma,
540 u64 size, u64 alignment, u64 flags)
541 {
542 if (!drm_mm_node_allocated(&vma->node))
543 return false;
544
545 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
546 return true;
547
548 if (vma->node.size < size)
549 return true;
550
551 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
552 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
553 return true;
554
555 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
556 return true;
557
558 if (flags & PIN_OFFSET_BIAS &&
559 vma->node.start < (flags & PIN_OFFSET_MASK))
560 return true;
561
562 if (flags & PIN_OFFSET_FIXED &&
563 vma->node.start != (flags & PIN_OFFSET_MASK))
564 return true;
565
566 return false;
567 }
568
__i915_vma_set_map_and_fenceable(struct i915_vma * vma)569 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
570 {
571 bool mappable, fenceable;
572
573 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
574 GEM_BUG_ON(!vma->fence_size);
575
576 fenceable = (vma->node.size >= vma->fence_size &&
577 IS_ALIGNED(vma->node.start, vma->fence_alignment));
578
579 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
580
581 if (mappable && fenceable)
582 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
583 else
584 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
585 }
586
i915_gem_valid_gtt_space(struct i915_vma * vma,unsigned long color)587 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
588 {
589 struct drm_mm_node *node = &vma->node;
590 struct drm_mm_node *other;
591
592 /*
593 * On some machines we have to be careful when putting differing types
594 * of snoopable memory together to avoid the prefetcher crossing memory
595 * domains and dying. During vm initialisation, we decide whether or not
596 * these constraints apply and set the drm_mm.color_adjust
597 * appropriately.
598 */
599 if (!i915_vm_has_cache_coloring(vma->vm))
600 return true;
601
602 /* Only valid to be called on an already inserted vma */
603 GEM_BUG_ON(!drm_mm_node_allocated(node));
604 GEM_BUG_ON(list_empty(&node->node_list));
605
606 other = list_prev_entry(node, node_list);
607 if (i915_node_color_differs(other, color) &&
608 !drm_mm_hole_follows(other))
609 return false;
610
611 other = list_next_entry(node, node_list);
612 if (i915_node_color_differs(other, color) &&
613 !drm_mm_hole_follows(node))
614 return false;
615
616 return true;
617 }
618
619 /**
620 * i915_vma_insert - finds a slot for the vma in its address space
621 * @vma: the vma
622 * @size: requested size in bytes (can be larger than the VMA)
623 * @alignment: required alignment
624 * @flags: mask of PIN_* flags to use
625 *
626 * First we try to allocate some free space that meets the requirements for
627 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
628 * preferrably the oldest idle entry to make room for the new VMA.
629 *
630 * Returns:
631 * 0 on success, negative error code otherwise.
632 */
633 static int
i915_vma_insert(struct i915_vma * vma,u64 size,u64 alignment,u64 flags)634 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
635 {
636 unsigned long color;
637 u64 start, end;
638 int ret;
639
640 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
641 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
642
643 size = max(size, vma->size);
644 alignment = max(alignment, vma->display_alignment);
645 if (flags & PIN_MAPPABLE) {
646 size = max_t(typeof(size), size, vma->fence_size);
647 alignment = max_t(typeof(alignment),
648 alignment, vma->fence_alignment);
649 }
650
651 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
652 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
653 GEM_BUG_ON(!is_power_of_2(alignment));
654
655 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
656 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
657
658 end = vma->vm->total;
659 if (flags & PIN_MAPPABLE)
660 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
661 if (flags & PIN_ZONE_4G)
662 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
663 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
664
665 /* If binding the object/GGTT view requires more space than the entire
666 * aperture has, reject it early before evicting everything in a vain
667 * attempt to find space.
668 */
669 if (size > end) {
670 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
671 size, flags & PIN_MAPPABLE ? "mappable" : "total",
672 end);
673 return -ENOSPC;
674 }
675
676 color = 0;
677 if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
678 color = vma->obj->cache_level;
679
680 if (flags & PIN_OFFSET_FIXED) {
681 u64 offset = flags & PIN_OFFSET_MASK;
682 if (!IS_ALIGNED(offset, alignment) ||
683 range_overflows(offset, size, end))
684 return -EINVAL;
685
686 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
687 size, offset, color,
688 flags);
689 if (ret)
690 return ret;
691 } else {
692 /*
693 * We only support huge gtt pages through the 48b PPGTT,
694 * however we also don't want to force any alignment for
695 * objects which need to be tightly packed into the low 32bits.
696 *
697 * Note that we assume that GGTT are limited to 4GiB for the
698 * forseeable future. See also i915_ggtt_offset().
699 */
700 if (upper_32_bits(end - 1) &&
701 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
702 /*
703 * We can't mix 64K and 4K PTEs in the same page-table
704 * (2M block), and so to avoid the ugliness and
705 * complexity of coloring we opt for just aligning 64K
706 * objects to 2M.
707 */
708 u64 page_alignment =
709 rounddown_pow_of_two(vma->page_sizes.sg |
710 I915_GTT_PAGE_SIZE_2M);
711
712 /*
713 * Check we don't expand for the limited Global GTT
714 * (mappable aperture is even more precious!). This
715 * also checks that we exclude the aliasing-ppgtt.
716 */
717 GEM_BUG_ON(i915_vma_is_ggtt(vma));
718
719 alignment = max(alignment, page_alignment);
720
721 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
722 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
723 }
724
725 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
726 size, alignment, color,
727 start, end, flags);
728 if (ret)
729 return ret;
730
731 GEM_BUG_ON(vma->node.start < start);
732 GEM_BUG_ON(vma->node.start + vma->node.size > end);
733 }
734 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
735 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
736
737 list_add_tail(&vma->vm_link, &vma->vm->bound_list);
738
739 return 0;
740 }
741
742 static void
i915_vma_detach(struct i915_vma * vma)743 i915_vma_detach(struct i915_vma *vma)
744 {
745 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
746 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
747
748 /*
749 * And finally now the object is completely decoupled from this
750 * vma, we can drop its hold on the backing storage and allow
751 * it to be reaped by the shrinker.
752 */
753 list_del(&vma->vm_link);
754 }
755
try_qad_pin(struct i915_vma * vma,unsigned int flags)756 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
757 {
758 unsigned int bound;
759 bool pinned = true;
760
761 bound = atomic_read(&vma->flags);
762 do {
763 if (unlikely(flags & ~bound))
764 return false;
765
766 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
767 return false;
768
769 if (!(bound & I915_VMA_PIN_MASK))
770 goto unpinned;
771
772 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
773 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
774
775 return true;
776
777 unpinned:
778 /*
779 * If pin_count==0, but we are bound, check under the lock to avoid
780 * racing with a concurrent i915_vma_unbind().
781 */
782 mutex_lock(&vma->vm->mutex);
783 do {
784 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
785 pinned = false;
786 break;
787 }
788
789 if (unlikely(flags & ~bound)) {
790 pinned = false;
791 break;
792 }
793 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
794 mutex_unlock(&vma->vm->mutex);
795
796 return pinned;
797 }
798
vma_get_pages(struct i915_vma * vma)799 static int vma_get_pages(struct i915_vma *vma)
800 {
801 int err = 0;
802 bool pinned_pages = false;
803
804 if (atomic_add_unless(&vma->pages_count, 1, 0))
805 return 0;
806
807 if (vma->obj) {
808 err = i915_gem_object_pin_pages(vma->obj);
809 if (err)
810 return err;
811 pinned_pages = true;
812 }
813
814 /* Allocations ahoy! */
815 if (mutex_lock_interruptible(&vma->pages_mutex)) {
816 err = -EINTR;
817 goto unpin;
818 }
819
820 if (!atomic_read(&vma->pages_count)) {
821 err = vma->ops->set_pages(vma);
822 if (err)
823 goto unlock;
824 pinned_pages = false;
825 }
826 atomic_inc(&vma->pages_count);
827
828 unlock:
829 mutex_unlock(&vma->pages_mutex);
830 unpin:
831 if (pinned_pages)
832 __i915_gem_object_unpin_pages(vma->obj);
833
834 return err;
835 }
836
__vma_put_pages(struct i915_vma * vma,unsigned int count)837 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
838 {
839 /* We allocate under vma_get_pages, so beware the shrinker */
840 mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
841 GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
842 if (atomic_sub_return(count, &vma->pages_count) == 0) {
843 vma->ops->clear_pages(vma);
844 GEM_BUG_ON(vma->pages);
845 if (vma->obj)
846 i915_gem_object_unpin_pages(vma->obj);
847 }
848 mutex_unlock(&vma->pages_mutex);
849 }
850
vma_put_pages(struct i915_vma * vma)851 static void vma_put_pages(struct i915_vma *vma)
852 {
853 if (atomic_add_unless(&vma->pages_count, -1, 1))
854 return;
855
856 __vma_put_pages(vma, 1);
857 }
858
vma_unbind_pages(struct i915_vma * vma)859 static void vma_unbind_pages(struct i915_vma *vma)
860 {
861 unsigned int count;
862
863 lockdep_assert_held(&vma->vm->mutex);
864
865 /* The upper portion of pages_count is the number of bindings */
866 count = atomic_read(&vma->pages_count);
867 count >>= I915_VMA_PAGES_BIAS;
868 GEM_BUG_ON(!count);
869
870 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
871 }
872
i915_vma_pin_ww(struct i915_vma * vma,struct i915_gem_ww_ctx * ww,u64 size,u64 alignment,u64 flags)873 int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
874 u64 size, u64 alignment, u64 flags)
875 {
876 struct i915_vma_work *work = NULL;
877 intel_wakeref_t wakeref = 0;
878 unsigned int bound;
879 int err;
880
881 #ifdef CONFIG_PROVE_LOCKING
882 if (debug_locks && !WARN_ON(!ww) && vma->resv)
883 assert_vma_held(vma);
884 #endif
885
886 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
887 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
888
889 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
890
891 /* First try and grab the pin without rebinding the vma */
892 if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
893 return 0;
894
895 err = vma_get_pages(vma);
896 if (err)
897 return err;
898
899 if (flags & PIN_GLOBAL)
900 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
901
902 if (flags & vma->vm->bind_async_flags) {
903 /* lock VM */
904 err = i915_vm_lock_objects(vma->vm, ww);
905 if (err)
906 goto err_rpm;
907
908 work = i915_vma_work();
909 if (!work) {
910 err = -ENOMEM;
911 goto err_rpm;
912 }
913
914 work->vm = i915_vm_get(vma->vm);
915
916 /* Allocate enough page directories to used PTE */
917 if (vma->vm->allocate_va_range) {
918 err = i915_vm_alloc_pt_stash(vma->vm,
919 &work->stash,
920 vma->size);
921 if (err)
922 goto err_fence;
923
924 err = i915_vm_map_pt_stash(vma->vm, &work->stash);
925 if (err)
926 goto err_fence;
927 }
928 }
929
930 /*
931 * Differentiate between user/kernel vma inside the aliasing-ppgtt.
932 *
933 * We conflate the Global GTT with the user's vma when using the
934 * aliasing-ppgtt, but it is still vitally important to try and
935 * keep the use cases distinct. For example, userptr objects are
936 * not allowed inside the Global GTT as that will cause lock
937 * inversions when we have to evict them the mmu_notifier callbacks -
938 * but they are allowed to be part of the user ppGTT which can never
939 * be mapped. As such we try to give the distinct users of the same
940 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
941 * and i915_ppgtt separate].
942 *
943 * NB this may cause us to mask real lock inversions -- while the
944 * code is safe today, lockdep may not be able to spot future
945 * transgressions.
946 */
947 err = mutex_lock_interruptible_nested(&vma->vm->mutex,
948 !(flags & PIN_GLOBAL));
949 if (err)
950 goto err_fence;
951
952 /* No more allocations allowed now we hold vm->mutex */
953
954 if (unlikely(i915_vma_is_closed(vma))) {
955 err = -ENOENT;
956 goto err_unlock;
957 }
958
959 bound = atomic_read(&vma->flags);
960 if (unlikely(bound & I915_VMA_ERROR)) {
961 err = -ENOMEM;
962 goto err_unlock;
963 }
964
965 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
966 err = -EAGAIN; /* pins are meant to be fairly temporary */
967 goto err_unlock;
968 }
969
970 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
971 __i915_vma_pin(vma);
972 goto err_unlock;
973 }
974
975 err = i915_active_acquire(&vma->active);
976 if (err)
977 goto err_unlock;
978
979 if (!(bound & I915_VMA_BIND_MASK)) {
980 err = i915_vma_insert(vma, size, alignment, flags);
981 if (err)
982 goto err_active;
983
984 if (i915_is_ggtt(vma->vm))
985 __i915_vma_set_map_and_fenceable(vma);
986 }
987
988 GEM_BUG_ON(!vma->pages);
989 err = i915_vma_bind(vma,
990 vma->obj ? vma->obj->cache_level : 0,
991 flags, work);
992 if (err)
993 goto err_remove;
994
995 /* There should only be at most 2 active bindings (user, global) */
996 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
997 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
998 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
999
1000 __i915_vma_pin(vma);
1001 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1002 GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
1003 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
1004
1005 err_remove:
1006 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
1007 i915_vma_detach(vma);
1008 drm_mm_remove_node(&vma->node);
1009 }
1010 err_active:
1011 i915_active_release(&vma->active);
1012 err_unlock:
1013 mutex_unlock(&vma->vm->mutex);
1014 err_fence:
1015 if (work)
1016 dma_fence_work_commit_imm(&work->base);
1017 err_rpm:
1018 if (wakeref)
1019 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
1020 vma_put_pages(vma);
1021 return err;
1022 }
1023
flush_idle_contexts(struct intel_gt * gt)1024 static void flush_idle_contexts(struct intel_gt *gt)
1025 {
1026 struct intel_engine_cs *engine;
1027 enum intel_engine_id id;
1028
1029 for_each_engine(engine, gt, id)
1030 intel_engine_flush_barriers(engine);
1031
1032 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1033 }
1034
i915_ggtt_pin(struct i915_vma * vma,struct i915_gem_ww_ctx * ww,u32 align,unsigned int flags)1035 int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1036 u32 align, unsigned int flags)
1037 {
1038 struct i915_address_space *vm = vma->vm;
1039 int err;
1040
1041 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1042
1043 #ifdef CONFIG_LOCKDEP
1044 WARN_ON(!ww && vma->resv && dma_resv_held(vma->resv));
1045 #endif
1046
1047 do {
1048 if (ww)
1049 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);
1050 else
1051 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1052 if (err != -ENOSPC) {
1053 if (!err) {
1054 err = i915_vma_wait_for_bind(vma);
1055 if (err)
1056 i915_vma_unpin(vma);
1057 }
1058 return err;
1059 }
1060
1061 /* Unlike i915_vma_pin, we don't take no for an answer! */
1062 flush_idle_contexts(vm->gt);
1063 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1064 i915_gem_evict_vm(vm);
1065 mutex_unlock(&vm->mutex);
1066 }
1067 } while (1);
1068 }
1069
__vma_close(struct i915_vma * vma,struct intel_gt * gt)1070 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1071 {
1072 /*
1073 * We defer actually closing, unbinding and destroying the VMA until
1074 * the next idle point, or if the object is freed in the meantime. By
1075 * postponing the unbind, we allow for it to be resurrected by the
1076 * client, avoiding the work required to rebind the VMA. This is
1077 * advantageous for DRI, where the client/server pass objects
1078 * between themselves, temporarily opening a local VMA to the
1079 * object, and then closing it again. The same object is then reused
1080 * on the next frame (or two, depending on the depth of the swap queue)
1081 * causing us to rebind the VMA once more. This ends up being a lot
1082 * of wasted work for the steady state.
1083 */
1084 GEM_BUG_ON(i915_vma_is_closed(vma));
1085 list_add(&vma->closed_link, >->closed_vma);
1086 }
1087
i915_vma_close(struct i915_vma * vma)1088 void i915_vma_close(struct i915_vma *vma)
1089 {
1090 struct intel_gt *gt = vma->vm->gt;
1091 unsigned long flags;
1092
1093 if (i915_vma_is_ggtt(vma))
1094 return;
1095
1096 GEM_BUG_ON(!atomic_read(&vma->open_count));
1097 if (atomic_dec_and_lock_irqsave(&vma->open_count,
1098 >->closed_lock,
1099 flags)) {
1100 __vma_close(vma, gt);
1101 spin_unlock_irqrestore(>->closed_lock, flags);
1102 }
1103 }
1104
__i915_vma_remove_closed(struct i915_vma * vma)1105 static void __i915_vma_remove_closed(struct i915_vma *vma)
1106 {
1107 struct intel_gt *gt = vma->vm->gt;
1108
1109 spin_lock_irq(>->closed_lock);
1110 list_del_init(&vma->closed_link);
1111 spin_unlock_irq(>->closed_lock);
1112 }
1113
i915_vma_reopen(struct i915_vma * vma)1114 void i915_vma_reopen(struct i915_vma *vma)
1115 {
1116 if (i915_vma_is_closed(vma))
1117 __i915_vma_remove_closed(vma);
1118 }
1119
i915_vma_release(struct kref * ref)1120 void i915_vma_release(struct kref *ref)
1121 {
1122 struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1123
1124 if (drm_mm_node_allocated(&vma->node)) {
1125 mutex_lock(&vma->vm->mutex);
1126 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1127 WARN_ON(__i915_vma_unbind(vma));
1128 mutex_unlock(&vma->vm->mutex);
1129 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1130 }
1131 GEM_BUG_ON(i915_vma_is_active(vma));
1132
1133 if (vma->obj) {
1134 struct drm_i915_gem_object *obj = vma->obj;
1135
1136 spin_lock(&obj->vma.lock);
1137 list_del(&vma->obj_link);
1138 if (!RB_EMPTY_NODE(&vma->obj_node))
1139 rb_erase(&vma->obj_node, &obj->vma.tree);
1140 spin_unlock(&obj->vma.lock);
1141 }
1142
1143 __i915_vma_remove_closed(vma);
1144 i915_vm_put(vma->vm);
1145
1146 i915_active_fini(&vma->active);
1147 i915_vma_free(vma);
1148 }
1149
i915_vma_parked(struct intel_gt * gt)1150 void i915_vma_parked(struct intel_gt *gt)
1151 {
1152 struct i915_vma *vma, *next;
1153 LIST_HEAD(closed);
1154
1155 spin_lock_irq(>->closed_lock);
1156 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
1157 struct drm_i915_gem_object *obj = vma->obj;
1158 struct i915_address_space *vm = vma->vm;
1159
1160 /* XXX All to avoid keeping a reference on i915_vma itself */
1161
1162 if (!kref_get_unless_zero(&obj->base.refcount))
1163 continue;
1164
1165 if (!i915_vm_tryopen(vm)) {
1166 i915_gem_object_put(obj);
1167 continue;
1168 }
1169
1170 list_move(&vma->closed_link, &closed);
1171 }
1172 spin_unlock_irq(>->closed_lock);
1173
1174 /* As the GT is held idle, no vma can be reopened as we destroy them */
1175 list_for_each_entry_safe(vma, next, &closed, closed_link) {
1176 struct drm_i915_gem_object *obj = vma->obj;
1177 struct i915_address_space *vm = vma->vm;
1178
1179 INIT_LIST_HEAD(&vma->closed_link);
1180 __i915_vma_put(vma);
1181
1182 i915_gem_object_put(obj);
1183 i915_vm_close(vm);
1184 }
1185 }
1186
__i915_vma_iounmap(struct i915_vma * vma)1187 static void __i915_vma_iounmap(struct i915_vma *vma)
1188 {
1189 GEM_BUG_ON(i915_vma_is_pinned(vma));
1190
1191 if (vma->iomap == NULL)
1192 return;
1193
1194 io_mapping_unmap(vma->iomap);
1195 vma->iomap = NULL;
1196 }
1197
i915_vma_revoke_mmap(struct i915_vma * vma)1198 void i915_vma_revoke_mmap(struct i915_vma *vma)
1199 {
1200 struct drm_vma_offset_node *node;
1201 u64 vma_offset;
1202
1203 if (!i915_vma_has_userfault(vma))
1204 return;
1205
1206 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1207 GEM_BUG_ON(!vma->obj->userfault_count);
1208
1209 node = &vma->mmo->vma_node;
1210 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1211 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1212 drm_vma_node_offset_addr(node) + vma_offset,
1213 vma->size,
1214 1);
1215
1216 i915_vma_unset_userfault(vma);
1217 if (!--vma->obj->userfault_count)
1218 list_del(&vma->obj->userfault_link);
1219 }
1220
1221 static int
__i915_request_await_bind(struct i915_request * rq,struct i915_vma * vma)1222 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma)
1223 {
1224 return __i915_request_await_exclusive(rq, &vma->active);
1225 }
1226
__i915_vma_move_to_active(struct i915_vma * vma,struct i915_request * rq)1227 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1228 {
1229 int err;
1230
1231 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1232
1233 /* Wait for the vma to be bound before we start! */
1234 err = __i915_request_await_bind(rq, vma);
1235 if (err)
1236 return err;
1237
1238 return i915_active_add_request(&vma->active, rq);
1239 }
1240
i915_vma_move_to_active(struct i915_vma * vma,struct i915_request * rq,unsigned int flags)1241 int i915_vma_move_to_active(struct i915_vma *vma,
1242 struct i915_request *rq,
1243 unsigned int flags)
1244 {
1245 struct drm_i915_gem_object *obj = vma->obj;
1246 int err;
1247
1248 assert_object_held(obj);
1249
1250 err = __i915_vma_move_to_active(vma, rq);
1251 if (unlikely(err))
1252 return err;
1253
1254 if (flags & EXEC_OBJECT_WRITE) {
1255 struct intel_frontbuffer *front;
1256
1257 front = __intel_frontbuffer_get(obj);
1258 if (unlikely(front)) {
1259 if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1260 i915_active_add_request(&front->write, rq);
1261 intel_frontbuffer_put(front);
1262 }
1263
1264 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1265 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1266 obj->read_domains = 0;
1267 } else {
1268 if (!(flags & __EXEC_OBJECT_NO_RESERVE)) {
1269 err = dma_resv_reserve_shared(vma->resv, 1);
1270 if (unlikely(err))
1271 return err;
1272 }
1273
1274 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1275 obj->write_domain = 0;
1276 }
1277
1278 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1279 i915_active_add_request(&vma->fence->active, rq);
1280
1281 obj->read_domains |= I915_GEM_GPU_DOMAINS;
1282 obj->mm.dirty = true;
1283
1284 GEM_BUG_ON(!i915_vma_is_active(vma));
1285 return 0;
1286 }
1287
__i915_vma_evict(struct i915_vma * vma)1288 void __i915_vma_evict(struct i915_vma *vma)
1289 {
1290 GEM_BUG_ON(i915_vma_is_pinned(vma));
1291
1292 if (i915_vma_is_map_and_fenceable(vma)) {
1293 /* Force a pagefault for domain tracking on next user access */
1294 i915_vma_revoke_mmap(vma);
1295
1296 /*
1297 * Check that we have flushed all writes through the GGTT
1298 * before the unbind, other due to non-strict nature of those
1299 * indirect writes they may end up referencing the GGTT PTE
1300 * after the unbind.
1301 *
1302 * Note that we may be concurrently poking at the GGTT_WRITE
1303 * bit from set-domain, as we mark all GGTT vma associated
1304 * with an object. We know this is for another vma, as we
1305 * are currently unbinding this one -- so if this vma will be
1306 * reused, it will be refaulted and have its dirty bit set
1307 * before the next write.
1308 */
1309 i915_vma_flush_writes(vma);
1310
1311 /* release the fence reg _after_ flushing */
1312 i915_vma_revoke_fence(vma);
1313
1314 __i915_vma_iounmap(vma);
1315 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1316 }
1317 GEM_BUG_ON(vma->fence);
1318 GEM_BUG_ON(i915_vma_has_userfault(vma));
1319
1320 if (likely(atomic_read(&vma->vm->open))) {
1321 trace_i915_vma_unbind(vma);
1322 vma->ops->unbind_vma(vma->vm, vma);
1323 }
1324 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1325 &vma->flags);
1326
1327 i915_vma_detach(vma);
1328 vma_unbind_pages(vma);
1329 }
1330
__i915_vma_unbind(struct i915_vma * vma)1331 int __i915_vma_unbind(struct i915_vma *vma)
1332 {
1333 int ret;
1334
1335 lockdep_assert_held(&vma->vm->mutex);
1336
1337 if (!drm_mm_node_allocated(&vma->node))
1338 return 0;
1339
1340 if (i915_vma_is_pinned(vma)) {
1341 vma_print_allocator(vma, "is pinned");
1342 return -EAGAIN;
1343 }
1344
1345 /*
1346 * After confirming that no one else is pinning this vma, wait for
1347 * any laggards who may have crept in during the wait (through
1348 * a residual pin skipping the vm->mutex) to complete.
1349 */
1350 ret = i915_vma_sync(vma);
1351 if (ret)
1352 return ret;
1353
1354 GEM_BUG_ON(i915_vma_is_active(vma));
1355 __i915_vma_evict(vma);
1356
1357 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1358 return 0;
1359 }
1360
i915_vma_unbind(struct i915_vma * vma)1361 int i915_vma_unbind(struct i915_vma *vma)
1362 {
1363 struct i915_address_space *vm = vma->vm;
1364 intel_wakeref_t wakeref = 0;
1365 int err;
1366
1367 /* Optimistic wait before taking the mutex */
1368 err = i915_vma_sync(vma);
1369 if (err)
1370 return err;
1371
1372 if (!drm_mm_node_allocated(&vma->node))
1373 return 0;
1374
1375 if (i915_vma_is_pinned(vma)) {
1376 vma_print_allocator(vma, "is pinned");
1377 return -EAGAIN;
1378 }
1379
1380 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1381 /* XXX not always required: nop_clear_range */
1382 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1383
1384 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1385 if (err)
1386 goto out_rpm;
1387
1388 err = __i915_vma_unbind(vma);
1389 mutex_unlock(&vm->mutex);
1390
1391 out_rpm:
1392 if (wakeref)
1393 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1394 return err;
1395 }
1396
i915_vma_make_unshrinkable(struct i915_vma * vma)1397 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1398 {
1399 i915_gem_object_make_unshrinkable(vma->obj);
1400 return vma;
1401 }
1402
i915_vma_make_shrinkable(struct i915_vma * vma)1403 void i915_vma_make_shrinkable(struct i915_vma *vma)
1404 {
1405 i915_gem_object_make_shrinkable(vma->obj);
1406 }
1407
i915_vma_make_purgeable(struct i915_vma * vma)1408 void i915_vma_make_purgeable(struct i915_vma *vma)
1409 {
1410 i915_gem_object_make_purgeable(vma->obj);
1411 }
1412
1413 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1414 #include "selftests/i915_vma.c"
1415 #endif
1416
i915_vma_module_exit(void)1417 void i915_vma_module_exit(void)
1418 {
1419 kmem_cache_destroy(slab_vmas);
1420 }
1421
i915_vma_module_init(void)1422 int __init i915_vma_module_init(void)
1423 {
1424 slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1425 if (!slab_vmas)
1426 return -ENOMEM;
1427
1428 return 0;
1429 }
1430