Loading...
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_globals.h"
38#include "i915_sw_fence_work.h"
39#include "i915_trace.h"
40#include "i915_vma.h"
41
42static struct i915_global_vma {
43 struct i915_global base;
44 struct kmem_cache *slab_vmas;
45} global;
46
47struct i915_vma *i915_vma_alloc(void)
48{
49 return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
50}
51
52void i915_vma_free(struct i915_vma *vma)
53{
54 return kmem_cache_free(global.slab_vmas, vma);
55}
56
57#if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
58
59#include <linux/stackdepot.h>
60
61static void vma_print_allocator(struct i915_vma *vma, const char *reason)
62{
63 unsigned long *entries;
64 unsigned int nr_entries;
65 char buf[512];
66
67 if (!vma->node.stack) {
68 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
69 vma->node.start, vma->node.size, reason);
70 return;
71 }
72
73 nr_entries = stack_depot_fetch(vma->node.stack, &entries);
74 stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
75 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
76 vma->node.start, vma->node.size, reason, buf);
77}
78
79#else
80
81static void vma_print_allocator(struct i915_vma *vma, const char *reason)
82{
83}
84
85#endif
86
87static inline struct i915_vma *active_to_vma(struct i915_active *ref)
88{
89 return container_of(ref, typeof(struct i915_vma), active);
90}
91
92static int __i915_vma_active(struct i915_active *ref)
93{
94 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
95}
96
97static void __i915_vma_retire(struct i915_active *ref)
98{
99 i915_vma_put(active_to_vma(ref));
100}
101
102static struct i915_vma *
103vma_create(struct drm_i915_gem_object *obj,
104 struct i915_address_space *vm,
105 const struct i915_ggtt_view *view)
106{
107 struct i915_vma *pos = ERR_PTR(-E2BIG);
108 struct i915_vma *vma;
109 struct rb_node *rb, **p;
110
111 /* The aliasing_ppgtt should never be used directly! */
112 GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
113
114 vma = i915_vma_alloc();
115 if (vma == NULL)
116 return ERR_PTR(-ENOMEM);
117
118 kref_init(&vma->ref);
119 mutex_init(&vma->pages_mutex);
120 vma->vm = i915_vm_get(vm);
121 vma->ops = &vm->vma_ops;
122 vma->obj = obj;
123 vma->resv = obj->base.resv;
124 vma->size = obj->base.size;
125 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
126
127 i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0);
128
129 /* Declare ourselves safe for use inside shrinkers */
130 if (IS_ENABLED(CONFIG_LOCKDEP)) {
131 fs_reclaim_acquire(GFP_KERNEL);
132 might_lock(&vma->active.mutex);
133 fs_reclaim_release(GFP_KERNEL);
134 }
135
136 INIT_LIST_HEAD(&vma->closed_link);
137
138 if (view && view->type != I915_GGTT_VIEW_NORMAL) {
139 vma->ggtt_view = *view;
140 if (view->type == I915_GGTT_VIEW_PARTIAL) {
141 GEM_BUG_ON(range_overflows_t(u64,
142 view->partial.offset,
143 view->partial.size,
144 obj->base.size >> PAGE_SHIFT));
145 vma->size = view->partial.size;
146 vma->size <<= PAGE_SHIFT;
147 GEM_BUG_ON(vma->size > obj->base.size);
148 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
149 vma->size = intel_rotation_info_size(&view->rotated);
150 vma->size <<= PAGE_SHIFT;
151 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
152 vma->size = intel_remapped_info_size(&view->remapped);
153 vma->size <<= PAGE_SHIFT;
154 }
155 }
156
157 if (unlikely(vma->size > vm->total))
158 goto err_vma;
159
160 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
161
162 spin_lock(&obj->vma.lock);
163
164 if (i915_is_ggtt(vm)) {
165 if (unlikely(overflows_type(vma->size, u32)))
166 goto err_unlock;
167
168 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
169 i915_gem_object_get_tiling(obj),
170 i915_gem_object_get_stride(obj));
171 if (unlikely(vma->fence_size < vma->size || /* overflow */
172 vma->fence_size > vm->total))
173 goto err_unlock;
174
175 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
176
177 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
178 i915_gem_object_get_tiling(obj),
179 i915_gem_object_get_stride(obj));
180 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
181
182 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
183 }
184
185 rb = NULL;
186 p = &obj->vma.tree.rb_node;
187 while (*p) {
188 long cmp;
189
190 rb = *p;
191 pos = rb_entry(rb, struct i915_vma, obj_node);
192
193 /*
194 * If the view already exists in the tree, another thread
195 * already created a matching vma, so return the older instance
196 * and dispose of ours.
197 */
198 cmp = i915_vma_compare(pos, vm, view);
199 if (cmp < 0)
200 p = &rb->rb_right;
201 else if (cmp > 0)
202 p = &rb->rb_left;
203 else
204 goto err_unlock;
205 }
206 rb_link_node(&vma->obj_node, rb, p);
207 rb_insert_color(&vma->obj_node, &obj->vma.tree);
208
209 if (i915_vma_is_ggtt(vma))
210 /*
211 * We put the GGTT vma at the start of the vma-list, followed
212 * by the ppGGTT vma. This allows us to break early when
213 * iterating over only the GGTT vma for an object, see
214 * for_each_ggtt_vma()
215 */
216 list_add(&vma->obj_link, &obj->vma.list);
217 else
218 list_add_tail(&vma->obj_link, &obj->vma.list);
219
220 spin_unlock(&obj->vma.lock);
221
222 return vma;
223
224err_unlock:
225 spin_unlock(&obj->vma.lock);
226err_vma:
227 i915_vm_put(vm);
228 i915_vma_free(vma);
229 return pos;
230}
231
232static struct i915_vma *
233i915_vma_lookup(struct drm_i915_gem_object *obj,
234 struct i915_address_space *vm,
235 const struct i915_ggtt_view *view)
236{
237 struct rb_node *rb;
238
239 rb = obj->vma.tree.rb_node;
240 while (rb) {
241 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
242 long cmp;
243
244 cmp = i915_vma_compare(vma, vm, view);
245 if (cmp == 0)
246 return vma;
247
248 if (cmp < 0)
249 rb = rb->rb_right;
250 else
251 rb = rb->rb_left;
252 }
253
254 return NULL;
255}
256
257/**
258 * i915_vma_instance - return the singleton instance of the VMA
259 * @obj: parent &struct drm_i915_gem_object to be mapped
260 * @vm: address space in which the mapping is located
261 * @view: additional mapping requirements
262 *
263 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
264 * the same @view characteristics. If a match is not found, one is created.
265 * Once created, the VMA is kept until either the object is freed, or the
266 * address space is closed.
267 *
268 * Returns the vma, or an error pointer.
269 */
270struct i915_vma *
271i915_vma_instance(struct drm_i915_gem_object *obj,
272 struct i915_address_space *vm,
273 const struct i915_ggtt_view *view)
274{
275 struct i915_vma *vma;
276
277 GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm));
278 GEM_BUG_ON(!atomic_read(&vm->open));
279
280 spin_lock(&obj->vma.lock);
281 vma = i915_vma_lookup(obj, vm, view);
282 spin_unlock(&obj->vma.lock);
283
284 /* vma_create() will resolve the race if another creates the vma */
285 if (unlikely(!vma))
286 vma = vma_create(obj, vm, view);
287
288 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
289 return vma;
290}
291
292struct i915_vma_work {
293 struct dma_fence_work base;
294 struct i915_address_space *vm;
295 struct i915_vm_pt_stash stash;
296 struct i915_vma *vma;
297 struct drm_i915_gem_object *pinned;
298 struct i915_sw_dma_fence_cb cb;
299 enum i915_cache_level cache_level;
300 unsigned int flags;
301};
302
303static int __vma_bind(struct dma_fence_work *work)
304{
305 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
306 struct i915_vma *vma = vw->vma;
307
308 vma->ops->bind_vma(vw->vm, &vw->stash,
309 vma, vw->cache_level, vw->flags);
310 return 0;
311}
312
313static void __vma_release(struct dma_fence_work *work)
314{
315 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
316
317 if (vw->pinned) {
318 __i915_gem_object_unpin_pages(vw->pinned);
319 i915_gem_object_put(vw->pinned);
320 }
321
322 i915_vm_free_pt_stash(vw->vm, &vw->stash);
323 i915_vm_put(vw->vm);
324}
325
326static const struct dma_fence_work_ops bind_ops = {
327 .name = "bind",
328 .work = __vma_bind,
329 .release = __vma_release,
330};
331
332struct i915_vma_work *i915_vma_work(void)
333{
334 struct i915_vma_work *vw;
335
336 vw = kzalloc(sizeof(*vw), GFP_KERNEL);
337 if (!vw)
338 return NULL;
339
340 dma_fence_work_init(&vw->base, &bind_ops);
341 vw->base.dma.error = -EAGAIN; /* disable the worker by default */
342
343 return vw;
344}
345
346int i915_vma_wait_for_bind(struct i915_vma *vma)
347{
348 int err = 0;
349
350 if (rcu_access_pointer(vma->active.excl.fence)) {
351 struct dma_fence *fence;
352
353 rcu_read_lock();
354 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
355 rcu_read_unlock();
356 if (fence) {
357 err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
358 dma_fence_put(fence);
359 }
360 }
361
362 return err;
363}
364
365/**
366 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
367 * @vma: VMA to map
368 * @cache_level: mapping cache level
369 * @flags: flags like global or local mapping
370 * @work: preallocated worker for allocating and binding the PTE
371 *
372 * DMA addresses are taken from the scatter-gather table of this object (or of
373 * this VMA in case of non-default GGTT views) and PTE entries set up.
374 * Note that DMA addresses are also the only part of the SG table we care about.
375 */
376int i915_vma_bind(struct i915_vma *vma,
377 enum i915_cache_level cache_level,
378 u32 flags,
379 struct i915_vma_work *work)
380{
381 u32 bind_flags;
382 u32 vma_flags;
383
384 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
385 GEM_BUG_ON(vma->size > vma->node.size);
386
387 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
388 vma->node.size,
389 vma->vm->total)))
390 return -ENODEV;
391
392 if (GEM_DEBUG_WARN_ON(!flags))
393 return -EINVAL;
394
395 bind_flags = flags;
396 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
397
398 vma_flags = atomic_read(&vma->flags);
399 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
400
401 bind_flags &= ~vma_flags;
402 if (bind_flags == 0)
403 return 0;
404
405 GEM_BUG_ON(!vma->pages);
406
407 trace_i915_vma_bind(vma, bind_flags);
408 if (work && bind_flags & vma->vm->bind_async_flags) {
409 struct dma_fence *prev;
410
411 work->vma = vma;
412 work->cache_level = cache_level;
413 work->flags = bind_flags;
414
415 /*
416 * Note we only want to chain up to the migration fence on
417 * the pages (not the object itself). As we don't track that,
418 * yet, we have to use the exclusive fence instead.
419 *
420 * Also note that we do not want to track the async vma as
421 * part of the obj->resv->excl_fence as it only affects
422 * execution and not content or object's backing store lifetime.
423 */
424 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
425 if (prev) {
426 __i915_sw_fence_await_dma_fence(&work->base.chain,
427 prev,
428 &work->cb);
429 dma_fence_put(prev);
430 }
431
432 work->base.dma.error = 0; /* enable the queue_work() */
433
434 if (vma->obj) {
435 __i915_gem_object_pin_pages(vma->obj);
436 work->pinned = i915_gem_object_get(vma->obj);
437 }
438 } else {
439 vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags);
440 }
441
442 atomic_or(bind_flags, &vma->flags);
443 return 0;
444}
445
446void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
447{
448 void __iomem *ptr;
449 int err;
450
451 if (!i915_gem_object_is_lmem(vma->obj)) {
452 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
453 err = -ENODEV;
454 goto err;
455 }
456 }
457
458 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
459 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
460
461 ptr = READ_ONCE(vma->iomap);
462 if (ptr == NULL) {
463 /*
464 * TODO: consider just using i915_gem_object_pin_map() for lmem
465 * instead, which already supports mapping non-contiguous chunks
466 * of pages, that way we can also drop the
467 * I915_BO_ALLOC_CONTIGUOUS when allocating the object.
468 */
469 if (i915_gem_object_is_lmem(vma->obj))
470 ptr = i915_gem_object_lmem_io_map(vma->obj, 0,
471 vma->obj->base.size);
472 else
473 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
474 vma->node.start,
475 vma->node.size);
476 if (ptr == NULL) {
477 err = -ENOMEM;
478 goto err;
479 }
480
481 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
482 io_mapping_unmap(ptr);
483 ptr = vma->iomap;
484 }
485 }
486
487 __i915_vma_pin(vma);
488
489 err = i915_vma_pin_fence(vma);
490 if (err)
491 goto err_unpin;
492
493 i915_vma_set_ggtt_write(vma);
494
495 /* NB Access through the GTT requires the device to be awake. */
496 return ptr;
497
498err_unpin:
499 __i915_vma_unpin(vma);
500err:
501 return IO_ERR_PTR(err);
502}
503
504void i915_vma_flush_writes(struct i915_vma *vma)
505{
506 if (i915_vma_unset_ggtt_write(vma))
507 intel_gt_flush_ggtt_writes(vma->vm->gt);
508}
509
510void i915_vma_unpin_iomap(struct i915_vma *vma)
511{
512 GEM_BUG_ON(vma->iomap == NULL);
513
514 i915_vma_flush_writes(vma);
515
516 i915_vma_unpin_fence(vma);
517 i915_vma_unpin(vma);
518}
519
520void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
521{
522 struct i915_vma *vma;
523 struct drm_i915_gem_object *obj;
524
525 vma = fetch_and_zero(p_vma);
526 if (!vma)
527 return;
528
529 obj = vma->obj;
530 GEM_BUG_ON(!obj);
531
532 i915_vma_unpin(vma);
533
534 if (flags & I915_VMA_RELEASE_MAP)
535 i915_gem_object_unpin_map(obj);
536
537 i915_gem_object_put(obj);
538}
539
540bool i915_vma_misplaced(const struct i915_vma *vma,
541 u64 size, u64 alignment, u64 flags)
542{
543 if (!drm_mm_node_allocated(&vma->node))
544 return false;
545
546 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
547 return true;
548
549 if (vma->node.size < size)
550 return true;
551
552 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
553 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
554 return true;
555
556 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
557 return true;
558
559 if (flags & PIN_OFFSET_BIAS &&
560 vma->node.start < (flags & PIN_OFFSET_MASK))
561 return true;
562
563 if (flags & PIN_OFFSET_FIXED &&
564 vma->node.start != (flags & PIN_OFFSET_MASK))
565 return true;
566
567 return false;
568}
569
570void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
571{
572 bool mappable, fenceable;
573
574 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
575 GEM_BUG_ON(!vma->fence_size);
576
577 fenceable = (vma->node.size >= vma->fence_size &&
578 IS_ALIGNED(vma->node.start, vma->fence_alignment));
579
580 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
581
582 if (mappable && fenceable)
583 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
584 else
585 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
586}
587
588bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
589{
590 struct drm_mm_node *node = &vma->node;
591 struct drm_mm_node *other;
592
593 /*
594 * On some machines we have to be careful when putting differing types
595 * of snoopable memory together to avoid the prefetcher crossing memory
596 * domains and dying. During vm initialisation, we decide whether or not
597 * these constraints apply and set the drm_mm.color_adjust
598 * appropriately.
599 */
600 if (!i915_vm_has_cache_coloring(vma->vm))
601 return true;
602
603 /* Only valid to be called on an already inserted vma */
604 GEM_BUG_ON(!drm_mm_node_allocated(node));
605 GEM_BUG_ON(list_empty(&node->node_list));
606
607 other = list_prev_entry(node, node_list);
608 if (i915_node_color_differs(other, color) &&
609 !drm_mm_hole_follows(other))
610 return false;
611
612 other = list_next_entry(node, node_list);
613 if (i915_node_color_differs(other, color) &&
614 !drm_mm_hole_follows(node))
615 return false;
616
617 return true;
618}
619
620/**
621 * i915_vma_insert - finds a slot for the vma in its address space
622 * @vma: the vma
623 * @size: requested size in bytes (can be larger than the VMA)
624 * @alignment: required alignment
625 * @flags: mask of PIN_* flags to use
626 *
627 * First we try to allocate some free space that meets the requirements for
628 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
629 * preferrably the oldest idle entry to make room for the new VMA.
630 *
631 * Returns:
632 * 0 on success, negative error code otherwise.
633 */
634static int
635i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
636{
637 unsigned long color;
638 u64 start, end;
639 int ret;
640
641 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
642 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
643
644 size = max(size, vma->size);
645 alignment = max(alignment, vma->display_alignment);
646 if (flags & PIN_MAPPABLE) {
647 size = max_t(typeof(size), size, vma->fence_size);
648 alignment = max_t(typeof(alignment),
649 alignment, vma->fence_alignment);
650 }
651
652 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
653 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
654 GEM_BUG_ON(!is_power_of_2(alignment));
655
656 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
657 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
658
659 end = vma->vm->total;
660 if (flags & PIN_MAPPABLE)
661 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
662 if (flags & PIN_ZONE_4G)
663 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
664 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
665
666 /* If binding the object/GGTT view requires more space than the entire
667 * aperture has, reject it early before evicting everything in a vain
668 * attempt to find space.
669 */
670 if (size > end) {
671 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
672 size, flags & PIN_MAPPABLE ? "mappable" : "total",
673 end);
674 return -ENOSPC;
675 }
676
677 color = 0;
678 if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
679 color = vma->obj->cache_level;
680
681 if (flags & PIN_OFFSET_FIXED) {
682 u64 offset = flags & PIN_OFFSET_MASK;
683 if (!IS_ALIGNED(offset, alignment) ||
684 range_overflows(offset, size, end))
685 return -EINVAL;
686
687 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
688 size, offset, color,
689 flags);
690 if (ret)
691 return ret;
692 } else {
693 /*
694 * We only support huge gtt pages through the 48b PPGTT,
695 * however we also don't want to force any alignment for
696 * objects which need to be tightly packed into the low 32bits.
697 *
698 * Note that we assume that GGTT are limited to 4GiB for the
699 * forseeable future. See also i915_ggtt_offset().
700 */
701 if (upper_32_bits(end - 1) &&
702 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
703 /*
704 * We can't mix 64K and 4K PTEs in the same page-table
705 * (2M block), and so to avoid the ugliness and
706 * complexity of coloring we opt for just aligning 64K
707 * objects to 2M.
708 */
709 u64 page_alignment =
710 rounddown_pow_of_two(vma->page_sizes.sg |
711 I915_GTT_PAGE_SIZE_2M);
712
713 /*
714 * Check we don't expand for the limited Global GTT
715 * (mappable aperture is even more precious!). This
716 * also checks that we exclude the aliasing-ppgtt.
717 */
718 GEM_BUG_ON(i915_vma_is_ggtt(vma));
719
720 alignment = max(alignment, page_alignment);
721
722 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
723 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
724 }
725
726 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
727 size, alignment, color,
728 start, end, flags);
729 if (ret)
730 return ret;
731
732 GEM_BUG_ON(vma->node.start < start);
733 GEM_BUG_ON(vma->node.start + vma->node.size > end);
734 }
735 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
736 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
737
738 list_add_tail(&vma->vm_link, &vma->vm->bound_list);
739
740 return 0;
741}
742
743static void
744i915_vma_detach(struct i915_vma *vma)
745{
746 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
747 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
748
749 /*
750 * And finally now the object is completely decoupled from this
751 * vma, we can drop its hold on the backing storage and allow
752 * it to be reaped by the shrinker.
753 */
754 list_del(&vma->vm_link);
755}
756
757static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
758{
759 unsigned int bound;
760 bool pinned = true;
761
762 bound = atomic_read(&vma->flags);
763 do {
764 if (unlikely(flags & ~bound))
765 return false;
766
767 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
768 return false;
769
770 if (!(bound & I915_VMA_PIN_MASK))
771 goto unpinned;
772
773 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
774 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
775
776 return true;
777
778unpinned:
779 /*
780 * If pin_count==0, but we are bound, check under the lock to avoid
781 * racing with a concurrent i915_vma_unbind().
782 */
783 mutex_lock(&vma->vm->mutex);
784 do {
785 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
786 pinned = false;
787 break;
788 }
789
790 if (unlikely(flags & ~bound)) {
791 pinned = false;
792 break;
793 }
794 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
795 mutex_unlock(&vma->vm->mutex);
796
797 return pinned;
798}
799
800static int vma_get_pages(struct i915_vma *vma)
801{
802 int err = 0;
803 bool pinned_pages = false;
804
805 if (atomic_add_unless(&vma->pages_count, 1, 0))
806 return 0;
807
808 if (vma->obj) {
809 err = i915_gem_object_pin_pages(vma->obj);
810 if (err)
811 return err;
812 pinned_pages = true;
813 }
814
815 /* Allocations ahoy! */
816 if (mutex_lock_interruptible(&vma->pages_mutex)) {
817 err = -EINTR;
818 goto unpin;
819 }
820
821 if (!atomic_read(&vma->pages_count)) {
822 err = vma->ops->set_pages(vma);
823 if (err)
824 goto unlock;
825 pinned_pages = false;
826 }
827 atomic_inc(&vma->pages_count);
828
829unlock:
830 mutex_unlock(&vma->pages_mutex);
831unpin:
832 if (pinned_pages)
833 __i915_gem_object_unpin_pages(vma->obj);
834
835 return err;
836}
837
838static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
839{
840 /* We allocate under vma_get_pages, so beware the shrinker */
841 mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
842 GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
843 if (atomic_sub_return(count, &vma->pages_count) == 0) {
844 vma->ops->clear_pages(vma);
845 GEM_BUG_ON(vma->pages);
846 if (vma->obj)
847 i915_gem_object_unpin_pages(vma->obj);
848 }
849 mutex_unlock(&vma->pages_mutex);
850}
851
852static void vma_put_pages(struct i915_vma *vma)
853{
854 if (atomic_add_unless(&vma->pages_count, -1, 1))
855 return;
856
857 __vma_put_pages(vma, 1);
858}
859
860static void vma_unbind_pages(struct i915_vma *vma)
861{
862 unsigned int count;
863
864 lockdep_assert_held(&vma->vm->mutex);
865
866 /* The upper portion of pages_count is the number of bindings */
867 count = atomic_read(&vma->pages_count);
868 count >>= I915_VMA_PAGES_BIAS;
869 GEM_BUG_ON(!count);
870
871 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
872}
873
874int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
875 u64 size, u64 alignment, u64 flags)
876{
877 struct i915_vma_work *work = NULL;
878 intel_wakeref_t wakeref = 0;
879 unsigned int bound;
880 int err;
881
882#ifdef CONFIG_PROVE_LOCKING
883 if (debug_locks && !WARN_ON(!ww) && vma->resv)
884 assert_vma_held(vma);
885#endif
886
887 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
888 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
889
890 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
891
892 /* First try and grab the pin without rebinding the vma */
893 if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
894 return 0;
895
896 err = vma_get_pages(vma);
897 if (err)
898 return err;
899
900 if (flags & PIN_GLOBAL)
901 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
902
903 if (flags & vma->vm->bind_async_flags) {
904 /* lock VM */
905 err = i915_vm_lock_objects(vma->vm, ww);
906 if (err)
907 goto err_rpm;
908
909 work = i915_vma_work();
910 if (!work) {
911 err = -ENOMEM;
912 goto err_rpm;
913 }
914
915 work->vm = i915_vm_get(vma->vm);
916
917 /* Allocate enough page directories to used PTE */
918 if (vma->vm->allocate_va_range) {
919 err = i915_vm_alloc_pt_stash(vma->vm,
920 &work->stash,
921 vma->size);
922 if (err)
923 goto err_fence;
924
925 err = i915_vm_map_pt_stash(vma->vm, &work->stash);
926 if (err)
927 goto err_fence;
928 }
929 }
930
931 /*
932 * Differentiate between user/kernel vma inside the aliasing-ppgtt.
933 *
934 * We conflate the Global GTT with the user's vma when using the
935 * aliasing-ppgtt, but it is still vitally important to try and
936 * keep the use cases distinct. For example, userptr objects are
937 * not allowed inside the Global GTT as that will cause lock
938 * inversions when we have to evict them the mmu_notifier callbacks -
939 * but they are allowed to be part of the user ppGTT which can never
940 * be mapped. As such we try to give the distinct users of the same
941 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
942 * and i915_ppgtt separate].
943 *
944 * NB this may cause us to mask real lock inversions -- while the
945 * code is safe today, lockdep may not be able to spot future
946 * transgressions.
947 */
948 err = mutex_lock_interruptible_nested(&vma->vm->mutex,
949 !(flags & PIN_GLOBAL));
950 if (err)
951 goto err_fence;
952
953 /* No more allocations allowed now we hold vm->mutex */
954
955 if (unlikely(i915_vma_is_closed(vma))) {
956 err = -ENOENT;
957 goto err_unlock;
958 }
959
960 bound = atomic_read(&vma->flags);
961 if (unlikely(bound & I915_VMA_ERROR)) {
962 err = -ENOMEM;
963 goto err_unlock;
964 }
965
966 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
967 err = -EAGAIN; /* pins are meant to be fairly temporary */
968 goto err_unlock;
969 }
970
971 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
972 __i915_vma_pin(vma);
973 goto err_unlock;
974 }
975
976 err = i915_active_acquire(&vma->active);
977 if (err)
978 goto err_unlock;
979
980 if (!(bound & I915_VMA_BIND_MASK)) {
981 err = i915_vma_insert(vma, size, alignment, flags);
982 if (err)
983 goto err_active;
984
985 if (i915_is_ggtt(vma->vm))
986 __i915_vma_set_map_and_fenceable(vma);
987 }
988
989 GEM_BUG_ON(!vma->pages);
990 err = i915_vma_bind(vma,
991 vma->obj ? vma->obj->cache_level : 0,
992 flags, work);
993 if (err)
994 goto err_remove;
995
996 /* There should only be at most 2 active bindings (user, global) */
997 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
998 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
999 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
1000
1001 __i915_vma_pin(vma);
1002 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1003 GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
1004 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
1005
1006err_remove:
1007 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
1008 i915_vma_detach(vma);
1009 drm_mm_remove_node(&vma->node);
1010 }
1011err_active:
1012 i915_active_release(&vma->active);
1013err_unlock:
1014 mutex_unlock(&vma->vm->mutex);
1015err_fence:
1016 if (work)
1017 dma_fence_work_commit_imm(&work->base);
1018err_rpm:
1019 if (wakeref)
1020 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
1021 vma_put_pages(vma);
1022 return err;
1023}
1024
1025static void flush_idle_contexts(struct intel_gt *gt)
1026{
1027 struct intel_engine_cs *engine;
1028 enum intel_engine_id id;
1029
1030 for_each_engine(engine, gt, id)
1031 intel_engine_flush_barriers(engine);
1032
1033 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1034}
1035
1036int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1037 u32 align, unsigned int flags)
1038{
1039 struct i915_address_space *vm = vma->vm;
1040 int err;
1041
1042 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1043
1044#ifdef CONFIG_LOCKDEP
1045 WARN_ON(!ww && vma->resv && dma_resv_held(vma->resv));
1046#endif
1047
1048 do {
1049 if (ww)
1050 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);
1051 else
1052 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1053 if (err != -ENOSPC) {
1054 if (!err) {
1055 err = i915_vma_wait_for_bind(vma);
1056 if (err)
1057 i915_vma_unpin(vma);
1058 }
1059 return err;
1060 }
1061
1062 /* Unlike i915_vma_pin, we don't take no for an answer! */
1063 flush_idle_contexts(vm->gt);
1064 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1065 i915_gem_evict_vm(vm);
1066 mutex_unlock(&vm->mutex);
1067 }
1068 } while (1);
1069}
1070
1071static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1072{
1073 /*
1074 * We defer actually closing, unbinding and destroying the VMA until
1075 * the next idle point, or if the object is freed in the meantime. By
1076 * postponing the unbind, we allow for it to be resurrected by the
1077 * client, avoiding the work required to rebind the VMA. This is
1078 * advantageous for DRI, where the client/server pass objects
1079 * between themselves, temporarily opening a local VMA to the
1080 * object, and then closing it again. The same object is then reused
1081 * on the next frame (or two, depending on the depth of the swap queue)
1082 * causing us to rebind the VMA once more. This ends up being a lot
1083 * of wasted work for the steady state.
1084 */
1085 GEM_BUG_ON(i915_vma_is_closed(vma));
1086 list_add(&vma->closed_link, >->closed_vma);
1087}
1088
1089void i915_vma_close(struct i915_vma *vma)
1090{
1091 struct intel_gt *gt = vma->vm->gt;
1092 unsigned long flags;
1093
1094 if (i915_vma_is_ggtt(vma))
1095 return;
1096
1097 GEM_BUG_ON(!atomic_read(&vma->open_count));
1098 if (atomic_dec_and_lock_irqsave(&vma->open_count,
1099 >->closed_lock,
1100 flags)) {
1101 __vma_close(vma, gt);
1102 spin_unlock_irqrestore(>->closed_lock, flags);
1103 }
1104}
1105
1106static void __i915_vma_remove_closed(struct i915_vma *vma)
1107{
1108 struct intel_gt *gt = vma->vm->gt;
1109
1110 spin_lock_irq(>->closed_lock);
1111 list_del_init(&vma->closed_link);
1112 spin_unlock_irq(>->closed_lock);
1113}
1114
1115void i915_vma_reopen(struct i915_vma *vma)
1116{
1117 if (i915_vma_is_closed(vma))
1118 __i915_vma_remove_closed(vma);
1119}
1120
1121void i915_vma_release(struct kref *ref)
1122{
1123 struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1124
1125 if (drm_mm_node_allocated(&vma->node)) {
1126 mutex_lock(&vma->vm->mutex);
1127 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1128 WARN_ON(__i915_vma_unbind(vma));
1129 mutex_unlock(&vma->vm->mutex);
1130 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1131 }
1132 GEM_BUG_ON(i915_vma_is_active(vma));
1133
1134 if (vma->obj) {
1135 struct drm_i915_gem_object *obj = vma->obj;
1136
1137 spin_lock(&obj->vma.lock);
1138 list_del(&vma->obj_link);
1139 if (!RB_EMPTY_NODE(&vma->obj_node))
1140 rb_erase(&vma->obj_node, &obj->vma.tree);
1141 spin_unlock(&obj->vma.lock);
1142 }
1143
1144 __i915_vma_remove_closed(vma);
1145 i915_vm_put(vma->vm);
1146
1147 i915_active_fini(&vma->active);
1148 i915_vma_free(vma);
1149}
1150
1151void i915_vma_parked(struct intel_gt *gt)
1152{
1153 struct i915_vma *vma, *next;
1154 LIST_HEAD(closed);
1155
1156 spin_lock_irq(>->closed_lock);
1157 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
1158 struct drm_i915_gem_object *obj = vma->obj;
1159 struct i915_address_space *vm = vma->vm;
1160
1161 /* XXX All to avoid keeping a reference on i915_vma itself */
1162
1163 if (!kref_get_unless_zero(&obj->base.refcount))
1164 continue;
1165
1166 if (!i915_vm_tryopen(vm)) {
1167 i915_gem_object_put(obj);
1168 continue;
1169 }
1170
1171 list_move(&vma->closed_link, &closed);
1172 }
1173 spin_unlock_irq(>->closed_lock);
1174
1175 /* As the GT is held idle, no vma can be reopened as we destroy them */
1176 list_for_each_entry_safe(vma, next, &closed, closed_link) {
1177 struct drm_i915_gem_object *obj = vma->obj;
1178 struct i915_address_space *vm = vma->vm;
1179
1180 INIT_LIST_HEAD(&vma->closed_link);
1181 __i915_vma_put(vma);
1182
1183 i915_gem_object_put(obj);
1184 i915_vm_close(vm);
1185 }
1186}
1187
1188static void __i915_vma_iounmap(struct i915_vma *vma)
1189{
1190 GEM_BUG_ON(i915_vma_is_pinned(vma));
1191
1192 if (vma->iomap == NULL)
1193 return;
1194
1195 io_mapping_unmap(vma->iomap);
1196 vma->iomap = NULL;
1197}
1198
1199void i915_vma_revoke_mmap(struct i915_vma *vma)
1200{
1201 struct drm_vma_offset_node *node;
1202 u64 vma_offset;
1203
1204 if (!i915_vma_has_userfault(vma))
1205 return;
1206
1207 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1208 GEM_BUG_ON(!vma->obj->userfault_count);
1209
1210 node = &vma->mmo->vma_node;
1211 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1212 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1213 drm_vma_node_offset_addr(node) + vma_offset,
1214 vma->size,
1215 1);
1216
1217 i915_vma_unset_userfault(vma);
1218 if (!--vma->obj->userfault_count)
1219 list_del(&vma->obj->userfault_link);
1220}
1221
1222static int
1223__i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma)
1224{
1225 return __i915_request_await_exclusive(rq, &vma->active);
1226}
1227
1228int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1229{
1230 int err;
1231
1232 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1233
1234 /* Wait for the vma to be bound before we start! */
1235 err = __i915_request_await_bind(rq, vma);
1236 if (err)
1237 return err;
1238
1239 return i915_active_add_request(&vma->active, rq);
1240}
1241
1242int i915_vma_move_to_active(struct i915_vma *vma,
1243 struct i915_request *rq,
1244 unsigned int flags)
1245{
1246 struct drm_i915_gem_object *obj = vma->obj;
1247 int err;
1248
1249 assert_object_held(obj);
1250
1251 err = __i915_vma_move_to_active(vma, rq);
1252 if (unlikely(err))
1253 return err;
1254
1255 if (flags & EXEC_OBJECT_WRITE) {
1256 struct intel_frontbuffer *front;
1257
1258 front = __intel_frontbuffer_get(obj);
1259 if (unlikely(front)) {
1260 if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1261 i915_active_add_request(&front->write, rq);
1262 intel_frontbuffer_put(front);
1263 }
1264
1265 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1266 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1267 obj->read_domains = 0;
1268 } else {
1269 if (!(flags & __EXEC_OBJECT_NO_RESERVE)) {
1270 err = dma_resv_reserve_shared(vma->resv, 1);
1271 if (unlikely(err))
1272 return err;
1273 }
1274
1275 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1276 obj->write_domain = 0;
1277 }
1278
1279 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1280 i915_active_add_request(&vma->fence->active, rq);
1281
1282 obj->read_domains |= I915_GEM_GPU_DOMAINS;
1283 obj->mm.dirty = true;
1284
1285 GEM_BUG_ON(!i915_vma_is_active(vma));
1286 return 0;
1287}
1288
1289void __i915_vma_evict(struct i915_vma *vma)
1290{
1291 GEM_BUG_ON(i915_vma_is_pinned(vma));
1292
1293 if (i915_vma_is_map_and_fenceable(vma)) {
1294 /* Force a pagefault for domain tracking on next user access */
1295 i915_vma_revoke_mmap(vma);
1296
1297 /*
1298 * Check that we have flushed all writes through the GGTT
1299 * before the unbind, other due to non-strict nature of those
1300 * indirect writes they may end up referencing the GGTT PTE
1301 * after the unbind.
1302 *
1303 * Note that we may be concurrently poking at the GGTT_WRITE
1304 * bit from set-domain, as we mark all GGTT vma associated
1305 * with an object. We know this is for another vma, as we
1306 * are currently unbinding this one -- so if this vma will be
1307 * reused, it will be refaulted and have its dirty bit set
1308 * before the next write.
1309 */
1310 i915_vma_flush_writes(vma);
1311
1312 /* release the fence reg _after_ flushing */
1313 i915_vma_revoke_fence(vma);
1314
1315 __i915_vma_iounmap(vma);
1316 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1317 }
1318 GEM_BUG_ON(vma->fence);
1319 GEM_BUG_ON(i915_vma_has_userfault(vma));
1320
1321 if (likely(atomic_read(&vma->vm->open))) {
1322 trace_i915_vma_unbind(vma);
1323 vma->ops->unbind_vma(vma->vm, vma);
1324 }
1325 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1326 &vma->flags);
1327
1328 i915_vma_detach(vma);
1329 vma_unbind_pages(vma);
1330}
1331
1332int __i915_vma_unbind(struct i915_vma *vma)
1333{
1334 int ret;
1335
1336 lockdep_assert_held(&vma->vm->mutex);
1337
1338 if (!drm_mm_node_allocated(&vma->node))
1339 return 0;
1340
1341 if (i915_vma_is_pinned(vma)) {
1342 vma_print_allocator(vma, "is pinned");
1343 return -EAGAIN;
1344 }
1345
1346 /*
1347 * After confirming that no one else is pinning this vma, wait for
1348 * any laggards who may have crept in during the wait (through
1349 * a residual pin skipping the vm->mutex) to complete.
1350 */
1351 ret = i915_vma_sync(vma);
1352 if (ret)
1353 return ret;
1354
1355 GEM_BUG_ON(i915_vma_is_active(vma));
1356 __i915_vma_evict(vma);
1357
1358 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1359 return 0;
1360}
1361
1362int i915_vma_unbind(struct i915_vma *vma)
1363{
1364 struct i915_address_space *vm = vma->vm;
1365 intel_wakeref_t wakeref = 0;
1366 int err;
1367
1368 /* Optimistic wait before taking the mutex */
1369 err = i915_vma_sync(vma);
1370 if (err)
1371 return err;
1372
1373 if (!drm_mm_node_allocated(&vma->node))
1374 return 0;
1375
1376 if (i915_vma_is_pinned(vma)) {
1377 vma_print_allocator(vma, "is pinned");
1378 return -EAGAIN;
1379 }
1380
1381 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1382 /* XXX not always required: nop_clear_range */
1383 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1384
1385 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1386 if (err)
1387 goto out_rpm;
1388
1389 err = __i915_vma_unbind(vma);
1390 mutex_unlock(&vm->mutex);
1391
1392out_rpm:
1393 if (wakeref)
1394 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1395 return err;
1396}
1397
1398struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1399{
1400 i915_gem_object_make_unshrinkable(vma->obj);
1401 return vma;
1402}
1403
1404void i915_vma_make_shrinkable(struct i915_vma *vma)
1405{
1406 i915_gem_object_make_shrinkable(vma->obj);
1407}
1408
1409void i915_vma_make_purgeable(struct i915_vma *vma)
1410{
1411 i915_gem_object_make_purgeable(vma->obj);
1412}
1413
1414#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1415#include "selftests/i915_vma.c"
1416#endif
1417
1418static void i915_global_vma_shrink(void)
1419{
1420 kmem_cache_shrink(global.slab_vmas);
1421}
1422
1423static void i915_global_vma_exit(void)
1424{
1425 kmem_cache_destroy(global.slab_vmas);
1426}
1427
1428static struct i915_global_vma global = { {
1429 .shrink = i915_global_vma_shrink,
1430 .exit = i915_global_vma_exit,
1431} };
1432
1433int __init i915_global_vma_init(void)
1434{
1435 global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1436 if (!global.slab_vmas)
1437 return -ENOMEM;
1438
1439 i915_global_register(&global.base);
1440 return 0;
1441}
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 "gt/intel_engine.h"
31#include "gt/intel_gt.h"
32
33#include "i915_drv.h"
34#include "i915_globals.h"
35#include "i915_trace.h"
36#include "i915_vma.h"
37
38static struct i915_global_vma {
39 struct i915_global base;
40 struct kmem_cache *slab_vmas;
41} global;
42
43struct i915_vma *i915_vma_alloc(void)
44{
45 return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
46}
47
48void i915_vma_free(struct i915_vma *vma)
49{
50 return kmem_cache_free(global.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
57static 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
77static void vma_print_allocator(struct i915_vma *vma, const char *reason)
78{
79}
80
81#endif
82
83static inline struct i915_vma *active_to_vma(struct i915_active *ref)
84{
85 return container_of(ref, typeof(struct i915_vma), active);
86}
87
88static int __i915_vma_active(struct i915_active *ref)
89{
90 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
91}
92
93static void __i915_vma_retire(struct i915_active *ref)
94{
95 i915_vma_put(active_to_vma(ref));
96}
97
98static struct i915_vma *
99vma_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 *vma;
104 struct rb_node *rb, **p;
105
106 /* The aliasing_ppgtt should never be used directly! */
107 GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm);
108
109 vma = i915_vma_alloc();
110 if (vma == NULL)
111 return ERR_PTR(-ENOMEM);
112
113 vma->vm = vm;
114 vma->ops = &vm->vma_ops;
115 vma->obj = obj;
116 vma->resv = obj->base.resv;
117 vma->size = obj->base.size;
118 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
119
120 i915_active_init(vm->i915, &vma->active,
121 __i915_vma_active, __i915_vma_retire);
122
123 /* Declare ourselves safe for use inside shrinkers */
124 if (IS_ENABLED(CONFIG_LOCKDEP)) {
125 fs_reclaim_acquire(GFP_KERNEL);
126 might_lock(&vma->active.mutex);
127 fs_reclaim_release(GFP_KERNEL);
128 }
129
130 INIT_LIST_HEAD(&vma->closed_link);
131
132 if (view && view->type != I915_GGTT_VIEW_NORMAL) {
133 vma->ggtt_view = *view;
134 if (view->type == I915_GGTT_VIEW_PARTIAL) {
135 GEM_BUG_ON(range_overflows_t(u64,
136 view->partial.offset,
137 view->partial.size,
138 obj->base.size >> PAGE_SHIFT));
139 vma->size = view->partial.size;
140 vma->size <<= PAGE_SHIFT;
141 GEM_BUG_ON(vma->size > obj->base.size);
142 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
143 vma->size = intel_rotation_info_size(&view->rotated);
144 vma->size <<= PAGE_SHIFT;
145 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
146 vma->size = intel_remapped_info_size(&view->remapped);
147 vma->size <<= PAGE_SHIFT;
148 }
149 }
150
151 if (unlikely(vma->size > vm->total))
152 goto err_vma;
153
154 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
155
156 if (i915_is_ggtt(vm)) {
157 if (unlikely(overflows_type(vma->size, u32)))
158 goto err_vma;
159
160 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
161 i915_gem_object_get_tiling(obj),
162 i915_gem_object_get_stride(obj));
163 if (unlikely(vma->fence_size < vma->size || /* overflow */
164 vma->fence_size > vm->total))
165 goto err_vma;
166
167 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
168
169 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
170 i915_gem_object_get_tiling(obj),
171 i915_gem_object_get_stride(obj));
172 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
173
174 vma->flags |= I915_VMA_GGTT;
175 }
176
177 spin_lock(&obj->vma.lock);
178
179 rb = NULL;
180 p = &obj->vma.tree.rb_node;
181 while (*p) {
182 struct i915_vma *pos;
183 long cmp;
184
185 rb = *p;
186 pos = rb_entry(rb, struct i915_vma, obj_node);
187
188 /*
189 * If the view already exists in the tree, another thread
190 * already created a matching vma, so return the older instance
191 * and dispose of ours.
192 */
193 cmp = i915_vma_compare(pos, vm, view);
194 if (cmp == 0) {
195 spin_unlock(&obj->vma.lock);
196 i915_vma_free(vma);
197 return pos;
198 }
199
200 if (cmp < 0)
201 p = &rb->rb_right;
202 else
203 p = &rb->rb_left;
204 }
205 rb_link_node(&vma->obj_node, rb, p);
206 rb_insert_color(&vma->obj_node, &obj->vma.tree);
207
208 if (i915_vma_is_ggtt(vma))
209 /*
210 * We put the GGTT vma at the start of the vma-list, followed
211 * by the ppGGTT vma. This allows us to break early when
212 * iterating over only the GGTT vma for an object, see
213 * for_each_ggtt_vma()
214 */
215 list_add(&vma->obj_link, &obj->vma.list);
216 else
217 list_add_tail(&vma->obj_link, &obj->vma.list);
218
219 spin_unlock(&obj->vma.lock);
220
221 mutex_lock(&vm->mutex);
222 list_add(&vma->vm_link, &vm->unbound_list);
223 mutex_unlock(&vm->mutex);
224
225 return vma;
226
227err_vma:
228 i915_vma_free(vma);
229 return ERR_PTR(-E2BIG);
230}
231
232static struct i915_vma *
233vma_lookup(struct drm_i915_gem_object *obj,
234 struct i915_address_space *vm,
235 const struct i915_ggtt_view *view)
236{
237 struct rb_node *rb;
238
239 rb = obj->vma.tree.rb_node;
240 while (rb) {
241 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
242 long cmp;
243
244 cmp = i915_vma_compare(vma, vm, view);
245 if (cmp == 0)
246 return vma;
247
248 if (cmp < 0)
249 rb = rb->rb_right;
250 else
251 rb = rb->rb_left;
252 }
253
254 return NULL;
255}
256
257/**
258 * i915_vma_instance - return the singleton instance of the VMA
259 * @obj: parent &struct drm_i915_gem_object to be mapped
260 * @vm: address space in which the mapping is located
261 * @view: additional mapping requirements
262 *
263 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
264 * the same @view characteristics. If a match is not found, one is created.
265 * Once created, the VMA is kept until either the object is freed, or the
266 * address space is closed.
267 *
268 * Must be called with struct_mutex held.
269 *
270 * Returns the vma, or an error pointer.
271 */
272struct i915_vma *
273i915_vma_instance(struct drm_i915_gem_object *obj,
274 struct i915_address_space *vm,
275 const struct i915_ggtt_view *view)
276{
277 struct i915_vma *vma;
278
279 GEM_BUG_ON(view && !i915_is_ggtt(vm));
280 GEM_BUG_ON(vm->closed);
281
282 spin_lock(&obj->vma.lock);
283 vma = vma_lookup(obj, vm, view);
284 spin_unlock(&obj->vma.lock);
285
286 /* vma_create() will resolve the race if another creates the vma */
287 if (unlikely(!vma))
288 vma = vma_create(obj, vm, view);
289
290 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
291 return vma;
292}
293
294/**
295 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
296 * @vma: VMA to map
297 * @cache_level: mapping cache level
298 * @flags: flags like global or local mapping
299 *
300 * DMA addresses are taken from the scatter-gather table of this object (or of
301 * this VMA in case of non-default GGTT views) and PTE entries set up.
302 * Note that DMA addresses are also the only part of the SG table we care about.
303 */
304int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
305 u32 flags)
306{
307 u32 bind_flags;
308 u32 vma_flags;
309 int ret;
310
311 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
312 GEM_BUG_ON(vma->size > vma->node.size);
313
314 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
315 vma->node.size,
316 vma->vm->total)))
317 return -ENODEV;
318
319 if (GEM_DEBUG_WARN_ON(!flags))
320 return -EINVAL;
321
322 bind_flags = 0;
323 if (flags & PIN_GLOBAL)
324 bind_flags |= I915_VMA_GLOBAL_BIND;
325 if (flags & PIN_USER)
326 bind_flags |= I915_VMA_LOCAL_BIND;
327
328 vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
329 if (flags & PIN_UPDATE)
330 bind_flags |= vma_flags;
331 else
332 bind_flags &= ~vma_flags;
333 if (bind_flags == 0)
334 return 0;
335
336 GEM_BUG_ON(!vma->pages);
337
338 trace_i915_vma_bind(vma, bind_flags);
339 ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
340 if (ret)
341 return ret;
342
343 vma->flags |= bind_flags;
344 return 0;
345}
346
347void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
348{
349 void __iomem *ptr;
350 int err;
351
352 /* Access through the GTT requires the device to be awake. */
353 assert_rpm_wakelock_held(&vma->vm->i915->runtime_pm);
354
355 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
356 if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
357 err = -ENODEV;
358 goto err;
359 }
360
361 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
362 GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
363
364 ptr = vma->iomap;
365 if (ptr == NULL) {
366 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
367 vma->node.start,
368 vma->node.size);
369 if (ptr == NULL) {
370 err = -ENOMEM;
371 goto err;
372 }
373
374 vma->iomap = ptr;
375 }
376
377 __i915_vma_pin(vma);
378
379 err = i915_vma_pin_fence(vma);
380 if (err)
381 goto err_unpin;
382
383 i915_vma_set_ggtt_write(vma);
384 return ptr;
385
386err_unpin:
387 __i915_vma_unpin(vma);
388err:
389 return IO_ERR_PTR(err);
390}
391
392void i915_vma_flush_writes(struct i915_vma *vma)
393{
394 if (!i915_vma_has_ggtt_write(vma))
395 return;
396
397 intel_gt_flush_ggtt_writes(vma->vm->gt);
398
399 i915_vma_unset_ggtt_write(vma);
400}
401
402void i915_vma_unpin_iomap(struct i915_vma *vma)
403{
404 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
405
406 GEM_BUG_ON(vma->iomap == NULL);
407
408 i915_vma_flush_writes(vma);
409
410 i915_vma_unpin_fence(vma);
411 i915_vma_unpin(vma);
412}
413
414void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
415{
416 struct i915_vma *vma;
417 struct drm_i915_gem_object *obj;
418
419 vma = fetch_and_zero(p_vma);
420 if (!vma)
421 return;
422
423 obj = vma->obj;
424 GEM_BUG_ON(!obj);
425
426 i915_vma_unpin(vma);
427 i915_vma_close(vma);
428
429 if (flags & I915_VMA_RELEASE_MAP)
430 i915_gem_object_unpin_map(obj);
431
432 i915_gem_object_put(obj);
433}
434
435bool i915_vma_misplaced(const struct i915_vma *vma,
436 u64 size, u64 alignment, u64 flags)
437{
438 if (!drm_mm_node_allocated(&vma->node))
439 return false;
440
441 if (vma->node.size < size)
442 return true;
443
444 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
445 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
446 return true;
447
448 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
449 return true;
450
451 if (flags & PIN_OFFSET_BIAS &&
452 vma->node.start < (flags & PIN_OFFSET_MASK))
453 return true;
454
455 if (flags & PIN_OFFSET_FIXED &&
456 vma->node.start != (flags & PIN_OFFSET_MASK))
457 return true;
458
459 return false;
460}
461
462void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
463{
464 bool mappable, fenceable;
465
466 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
467 GEM_BUG_ON(!vma->fence_size);
468
469 fenceable = (vma->node.size >= vma->fence_size &&
470 IS_ALIGNED(vma->node.start, vma->fence_alignment));
471
472 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
473
474 if (mappable && fenceable)
475 vma->flags |= I915_VMA_CAN_FENCE;
476 else
477 vma->flags &= ~I915_VMA_CAN_FENCE;
478}
479
480static bool color_differs(struct drm_mm_node *node, unsigned long color)
481{
482 return node->allocated && node->color != color;
483}
484
485bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
486{
487 struct drm_mm_node *node = &vma->node;
488 struct drm_mm_node *other;
489
490 /*
491 * On some machines we have to be careful when putting differing types
492 * of snoopable memory together to avoid the prefetcher crossing memory
493 * domains and dying. During vm initialisation, we decide whether or not
494 * these constraints apply and set the drm_mm.color_adjust
495 * appropriately.
496 */
497 if (vma->vm->mm.color_adjust == NULL)
498 return true;
499
500 /* Only valid to be called on an already inserted vma */
501 GEM_BUG_ON(!drm_mm_node_allocated(node));
502 GEM_BUG_ON(list_empty(&node->node_list));
503
504 other = list_prev_entry(node, node_list);
505 if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
506 return false;
507
508 other = list_next_entry(node, node_list);
509 if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
510 return false;
511
512 return true;
513}
514
515static void assert_bind_count(const struct drm_i915_gem_object *obj)
516{
517 /*
518 * Combine the assertion that the object is bound and that we have
519 * pinned its pages. But we should never have bound the object
520 * more than we have pinned its pages. (For complete accuracy, we
521 * assume that no else is pinning the pages, but as a rough assertion
522 * that we will not run into problems later, this will do!)
523 */
524 GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < atomic_read(&obj->bind_count));
525}
526
527/**
528 * i915_vma_insert - finds a slot for the vma in its address space
529 * @vma: the vma
530 * @size: requested size in bytes (can be larger than the VMA)
531 * @alignment: required alignment
532 * @flags: mask of PIN_* flags to use
533 *
534 * First we try to allocate some free space that meets the requirements for
535 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
536 * preferrably the oldest idle entry to make room for the new VMA.
537 *
538 * Returns:
539 * 0 on success, negative error code otherwise.
540 */
541static int
542i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
543{
544 struct drm_i915_private *dev_priv = vma->vm->i915;
545 unsigned int cache_level;
546 u64 start, end;
547 int ret;
548
549 GEM_BUG_ON(i915_vma_is_closed(vma));
550 GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
551 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
552
553 size = max(size, vma->size);
554 alignment = max(alignment, vma->display_alignment);
555 if (flags & PIN_MAPPABLE) {
556 size = max_t(typeof(size), size, vma->fence_size);
557 alignment = max_t(typeof(alignment),
558 alignment, vma->fence_alignment);
559 }
560
561 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
562 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
563 GEM_BUG_ON(!is_power_of_2(alignment));
564
565 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
566 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
567
568 end = vma->vm->total;
569 if (flags & PIN_MAPPABLE)
570 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
571 if (flags & PIN_ZONE_4G)
572 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
573 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
574
575 /* If binding the object/GGTT view requires more space than the entire
576 * aperture has, reject it early before evicting everything in a vain
577 * attempt to find space.
578 */
579 if (size > end) {
580 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
581 size, flags & PIN_MAPPABLE ? "mappable" : "total",
582 end);
583 return -ENOSPC;
584 }
585
586 if (vma->obj) {
587 ret = i915_gem_object_pin_pages(vma->obj);
588 if (ret)
589 return ret;
590
591 cache_level = vma->obj->cache_level;
592 } else {
593 cache_level = 0;
594 }
595
596 GEM_BUG_ON(vma->pages);
597
598 ret = vma->ops->set_pages(vma);
599 if (ret)
600 goto err_unpin;
601
602 if (flags & PIN_OFFSET_FIXED) {
603 u64 offset = flags & PIN_OFFSET_MASK;
604 if (!IS_ALIGNED(offset, alignment) ||
605 range_overflows(offset, size, end)) {
606 ret = -EINVAL;
607 goto err_clear;
608 }
609
610 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
611 size, offset, cache_level,
612 flags);
613 if (ret)
614 goto err_clear;
615 } else {
616 /*
617 * We only support huge gtt pages through the 48b PPGTT,
618 * however we also don't want to force any alignment for
619 * objects which need to be tightly packed into the low 32bits.
620 *
621 * Note that we assume that GGTT are limited to 4GiB for the
622 * forseeable future. See also i915_ggtt_offset().
623 */
624 if (upper_32_bits(end - 1) &&
625 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
626 /*
627 * We can't mix 64K and 4K PTEs in the same page-table
628 * (2M block), and so to avoid the ugliness and
629 * complexity of coloring we opt for just aligning 64K
630 * objects to 2M.
631 */
632 u64 page_alignment =
633 rounddown_pow_of_two(vma->page_sizes.sg |
634 I915_GTT_PAGE_SIZE_2M);
635
636 /*
637 * Check we don't expand for the limited Global GTT
638 * (mappable aperture is even more precious!). This
639 * also checks that we exclude the aliasing-ppgtt.
640 */
641 GEM_BUG_ON(i915_vma_is_ggtt(vma));
642
643 alignment = max(alignment, page_alignment);
644
645 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
646 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
647 }
648
649 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
650 size, alignment, cache_level,
651 start, end, flags);
652 if (ret)
653 goto err_clear;
654
655 GEM_BUG_ON(vma->node.start < start);
656 GEM_BUG_ON(vma->node.start + vma->node.size > end);
657 }
658 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
659 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, cache_level));
660
661 mutex_lock(&vma->vm->mutex);
662 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
663 mutex_unlock(&vma->vm->mutex);
664
665 if (vma->obj) {
666 atomic_inc(&vma->obj->bind_count);
667 assert_bind_count(vma->obj);
668 }
669
670 return 0;
671
672err_clear:
673 vma->ops->clear_pages(vma);
674err_unpin:
675 if (vma->obj)
676 i915_gem_object_unpin_pages(vma->obj);
677 return ret;
678}
679
680static void
681i915_vma_remove(struct i915_vma *vma)
682{
683 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
684 GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
685
686 vma->ops->clear_pages(vma);
687
688 mutex_lock(&vma->vm->mutex);
689 drm_mm_remove_node(&vma->node);
690 list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
691 mutex_unlock(&vma->vm->mutex);
692
693 /*
694 * Since the unbound list is global, only move to that list if
695 * no more VMAs exist.
696 */
697 if (vma->obj) {
698 struct drm_i915_gem_object *obj = vma->obj;
699
700 atomic_dec(&obj->bind_count);
701
702 /*
703 * And finally now the object is completely decoupled from this
704 * vma, we can drop its hold on the backing storage and allow
705 * it to be reaped by the shrinker.
706 */
707 i915_gem_object_unpin_pages(obj);
708 assert_bind_count(obj);
709 }
710}
711
712int __i915_vma_do_pin(struct i915_vma *vma,
713 u64 size, u64 alignment, u64 flags)
714{
715 const unsigned int bound = vma->flags;
716 int ret;
717
718 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
719 GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
720 GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
721
722 if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
723 ret = -EBUSY;
724 goto err_unpin;
725 }
726
727 if ((bound & I915_VMA_BIND_MASK) == 0) {
728 ret = i915_vma_insert(vma, size, alignment, flags);
729 if (ret)
730 goto err_unpin;
731 }
732 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
733
734 ret = i915_vma_bind(vma, vma->obj ? vma->obj->cache_level : 0, flags);
735 if (ret)
736 goto err_remove;
737
738 GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
739
740 if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
741 __i915_vma_set_map_and_fenceable(vma);
742
743 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
744 return 0;
745
746err_remove:
747 if ((bound & I915_VMA_BIND_MASK) == 0) {
748 i915_vma_remove(vma);
749 GEM_BUG_ON(vma->pages);
750 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
751 }
752err_unpin:
753 __i915_vma_unpin(vma);
754 return ret;
755}
756
757void i915_vma_close(struct i915_vma *vma)
758{
759 struct drm_i915_private *i915 = vma->vm->i915;
760 unsigned long flags;
761
762 GEM_BUG_ON(i915_vma_is_closed(vma));
763
764 /*
765 * We defer actually closing, unbinding and destroying the VMA until
766 * the next idle point, or if the object is freed in the meantime. By
767 * postponing the unbind, we allow for it to be resurrected by the
768 * client, avoiding the work required to rebind the VMA. This is
769 * advantageous for DRI, where the client/server pass objects
770 * between themselves, temporarily opening a local VMA to the
771 * object, and then closing it again. The same object is then reused
772 * on the next frame (or two, depending on the depth of the swap queue)
773 * causing us to rebind the VMA once more. This ends up being a lot
774 * of wasted work for the steady state.
775 */
776 spin_lock_irqsave(&i915->gt.closed_lock, flags);
777 list_add(&vma->closed_link, &i915->gt.closed_vma);
778 spin_unlock_irqrestore(&i915->gt.closed_lock, flags);
779}
780
781static void __i915_vma_remove_closed(struct i915_vma *vma)
782{
783 struct drm_i915_private *i915 = vma->vm->i915;
784
785 if (!i915_vma_is_closed(vma))
786 return;
787
788 spin_lock_irq(&i915->gt.closed_lock);
789 list_del_init(&vma->closed_link);
790 spin_unlock_irq(&i915->gt.closed_lock);
791}
792
793void i915_vma_reopen(struct i915_vma *vma)
794{
795 __i915_vma_remove_closed(vma);
796}
797
798static void __i915_vma_destroy(struct i915_vma *vma)
799{
800 GEM_BUG_ON(vma->node.allocated);
801 GEM_BUG_ON(vma->fence);
802
803 mutex_lock(&vma->vm->mutex);
804 list_del(&vma->vm_link);
805 mutex_unlock(&vma->vm->mutex);
806
807 if (vma->obj) {
808 struct drm_i915_gem_object *obj = vma->obj;
809
810 spin_lock(&obj->vma.lock);
811 list_del(&vma->obj_link);
812 rb_erase(&vma->obj_node, &vma->obj->vma.tree);
813 spin_unlock(&obj->vma.lock);
814 }
815
816 i915_active_fini(&vma->active);
817
818 i915_vma_free(vma);
819}
820
821void i915_vma_destroy(struct i915_vma *vma)
822{
823 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
824
825 GEM_BUG_ON(i915_vma_is_pinned(vma));
826
827 __i915_vma_remove_closed(vma);
828
829 WARN_ON(i915_vma_unbind(vma));
830 GEM_BUG_ON(i915_vma_is_active(vma));
831
832 __i915_vma_destroy(vma);
833}
834
835void i915_vma_parked(struct drm_i915_private *i915)
836{
837 struct i915_vma *vma, *next;
838
839 spin_lock_irq(&i915->gt.closed_lock);
840 list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
841 list_del_init(&vma->closed_link);
842 spin_unlock_irq(&i915->gt.closed_lock);
843
844 i915_vma_destroy(vma);
845
846 spin_lock_irq(&i915->gt.closed_lock);
847 }
848 spin_unlock_irq(&i915->gt.closed_lock);
849}
850
851static void __i915_vma_iounmap(struct i915_vma *vma)
852{
853 GEM_BUG_ON(i915_vma_is_pinned(vma));
854
855 if (vma->iomap == NULL)
856 return;
857
858 io_mapping_unmap(vma->iomap);
859 vma->iomap = NULL;
860}
861
862void i915_vma_revoke_mmap(struct i915_vma *vma)
863{
864 struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
865 u64 vma_offset;
866
867 lockdep_assert_held(&vma->vm->mutex);
868
869 if (!i915_vma_has_userfault(vma))
870 return;
871
872 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
873 GEM_BUG_ON(!vma->obj->userfault_count);
874
875 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
876 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
877 drm_vma_node_offset_addr(node) + vma_offset,
878 vma->size,
879 1);
880
881 i915_vma_unset_userfault(vma);
882 if (!--vma->obj->userfault_count)
883 list_del(&vma->obj->userfault_link);
884}
885
886int i915_vma_move_to_active(struct i915_vma *vma,
887 struct i915_request *rq,
888 unsigned int flags)
889{
890 struct drm_i915_gem_object *obj = vma->obj;
891 int err;
892
893 assert_vma_held(vma);
894 assert_object_held(obj);
895 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
896
897 /*
898 * Add a reference if we're newly entering the active list.
899 * The order in which we add operations to the retirement queue is
900 * vital here: mark_active adds to the start of the callback list,
901 * such that subsequent callbacks are called first. Therefore we
902 * add the active reference first and queue for it to be dropped
903 * *last*.
904 */
905 err = i915_active_ref(&vma->active, rq->timeline, rq);
906 if (unlikely(err))
907 return err;
908
909 if (flags & EXEC_OBJECT_WRITE) {
910 if (intel_frontbuffer_invalidate(obj->frontbuffer, ORIGIN_CS))
911 i915_active_ref(&obj->frontbuffer->write,
912 rq->timeline,
913 rq);
914
915 dma_resv_add_excl_fence(vma->resv, &rq->fence);
916 obj->write_domain = I915_GEM_DOMAIN_RENDER;
917 obj->read_domains = 0;
918 } else {
919 err = dma_resv_reserve_shared(vma->resv, 1);
920 if (unlikely(err))
921 return err;
922
923 dma_resv_add_shared_fence(vma->resv, &rq->fence);
924 obj->write_domain = 0;
925 }
926 obj->read_domains |= I915_GEM_GPU_DOMAINS;
927 obj->mm.dirty = true;
928
929 GEM_BUG_ON(!i915_vma_is_active(vma));
930 return 0;
931}
932
933int i915_vma_unbind(struct i915_vma *vma)
934{
935 int ret;
936
937 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
938
939 /*
940 * First wait upon any activity as retiring the request may
941 * have side-effects such as unpinning or even unbinding this vma.
942 */
943 might_sleep();
944 if (i915_vma_is_active(vma)) {
945 /*
946 * When a closed VMA is retired, it is unbound - eek.
947 * In order to prevent it from being recursively closed,
948 * take a pin on the vma so that the second unbind is
949 * aborted.
950 *
951 * Even more scary is that the retire callback may free
952 * the object (last active vma). To prevent the explosion
953 * we defer the actual object free to a worker that can
954 * only proceed once it acquires the struct_mutex (which
955 * we currently hold, therefore it cannot free this object
956 * before we are finished).
957 */
958 __i915_vma_pin(vma);
959 ret = i915_active_wait(&vma->active);
960 __i915_vma_unpin(vma);
961 if (ret)
962 return ret;
963 }
964 GEM_BUG_ON(i915_vma_is_active(vma));
965
966 if (i915_vma_is_pinned(vma)) {
967 vma_print_allocator(vma, "is pinned");
968 return -EBUSY;
969 }
970
971 if (!drm_mm_node_allocated(&vma->node))
972 return 0;
973
974 if (i915_vma_is_map_and_fenceable(vma)) {
975 /*
976 * Check that we have flushed all writes through the GGTT
977 * before the unbind, other due to non-strict nature of those
978 * indirect writes they may end up referencing the GGTT PTE
979 * after the unbind.
980 */
981 i915_vma_flush_writes(vma);
982 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
983
984 /* release the fence reg _after_ flushing */
985 mutex_lock(&vma->vm->mutex);
986 ret = i915_vma_revoke_fence(vma);
987 mutex_unlock(&vma->vm->mutex);
988 if (ret)
989 return ret;
990
991 /* Force a pagefault for domain tracking on next user access */
992 mutex_lock(&vma->vm->mutex);
993 i915_vma_revoke_mmap(vma);
994 mutex_unlock(&vma->vm->mutex);
995
996 __i915_vma_iounmap(vma);
997 vma->flags &= ~I915_VMA_CAN_FENCE;
998 }
999 GEM_BUG_ON(vma->fence);
1000 GEM_BUG_ON(i915_vma_has_userfault(vma));
1001
1002 if (likely(!vma->vm->closed)) {
1003 trace_i915_vma_unbind(vma);
1004 vma->ops->unbind_vma(vma);
1005 }
1006 vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
1007
1008 i915_vma_remove(vma);
1009
1010 return 0;
1011}
1012
1013struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1014{
1015 i915_gem_object_make_unshrinkable(vma->obj);
1016 return vma;
1017}
1018
1019void i915_vma_make_shrinkable(struct i915_vma *vma)
1020{
1021 i915_gem_object_make_shrinkable(vma->obj);
1022}
1023
1024void i915_vma_make_purgeable(struct i915_vma *vma)
1025{
1026 i915_gem_object_make_purgeable(vma->obj);
1027}
1028
1029#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1030#include "selftests/i915_vma.c"
1031#endif
1032
1033static void i915_global_vma_shrink(void)
1034{
1035 kmem_cache_shrink(global.slab_vmas);
1036}
1037
1038static void i915_global_vma_exit(void)
1039{
1040 kmem_cache_destroy(global.slab_vmas);
1041}
1042
1043static struct i915_global_vma global = { {
1044 .shrink = i915_global_vma_shrink,
1045 .exit = i915_global_vma_exit,
1046} };
1047
1048int __init i915_global_vma_init(void)
1049{
1050 global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1051 if (!global.slab_vmas)
1052 return -ENOMEM;
1053
1054 i915_global_register(&global.base);
1055 return 0;
1056}