Loading...
Note: File does not exist in v3.1.
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_engine_heartbeat.h"
32#include "gt/intel_gt.h"
33#include "gt/intel_gt_requests.h"
34
35#include "i915_drv.h"
36#include "i915_globals.h"
37#include "i915_sw_fence_work.h"
38#include "i915_trace.h"
39#include "i915_vma.h"
40
41static struct i915_global_vma {
42 struct i915_global base;
43 struct kmem_cache *slab_vmas;
44} global;
45
46struct i915_vma *i915_vma_alloc(void)
47{
48 return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
49}
50
51void i915_vma_free(struct i915_vma *vma)
52{
53 return kmem_cache_free(global.slab_vmas, vma);
54}
55
56#if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
57
58#include <linux/stackdepot.h>
59
60static void vma_print_allocator(struct i915_vma *vma, const char *reason)
61{
62 unsigned long *entries;
63 unsigned int nr_entries;
64 char buf[512];
65
66 if (!vma->node.stack) {
67 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
68 vma->node.start, vma->node.size, reason);
69 return;
70 }
71
72 nr_entries = stack_depot_fetch(vma->node.stack, &entries);
73 stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
74 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
75 vma->node.start, vma->node.size, reason, buf);
76}
77
78#else
79
80static void vma_print_allocator(struct i915_vma *vma, const char *reason)
81{
82}
83
84#endif
85
86static inline struct i915_vma *active_to_vma(struct i915_active *ref)
87{
88 return container_of(ref, typeof(struct i915_vma), active);
89}
90
91static int __i915_vma_active(struct i915_active *ref)
92{
93 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
94}
95
96__i915_active_call
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);
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 *
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 * 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(vm));
278 GEM_BUG_ON(!atomic_read(&vm->open));
279
280 spin_lock(&obj->vma.lock);
281 vma = 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_vma *vma;
295 struct drm_i915_gem_object *pinned;
296 struct i915_sw_dma_fence_cb cb;
297 enum i915_cache_level cache_level;
298 unsigned int flags;
299};
300
301static int __vma_bind(struct dma_fence_work *work)
302{
303 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
304 struct i915_vma *vma = vw->vma;
305 int err;
306
307 err = vma->ops->bind_vma(vma->vm, vma, vw->cache_level, vw->flags);
308 if (err)
309 atomic_or(I915_VMA_ERROR, &vma->flags);
310
311 return err;
312}
313
314static void __vma_release(struct dma_fence_work *work)
315{
316 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
317
318 if (vw->pinned)
319 __i915_gem_object_unpin_pages(vw->pinned);
320}
321
322static const struct dma_fence_work_ops bind_ops = {
323 .name = "bind",
324 .work = __vma_bind,
325 .release = __vma_release,
326};
327
328struct i915_vma_work *i915_vma_work(void)
329{
330 struct i915_vma_work *vw;
331
332 vw = kzalloc(sizeof(*vw), GFP_KERNEL);
333 if (!vw)
334 return NULL;
335
336 dma_fence_work_init(&vw->base, &bind_ops);
337 vw->base.dma.error = -EAGAIN; /* disable the worker by default */
338
339 return vw;
340}
341
342int i915_vma_wait_for_bind(struct i915_vma *vma)
343{
344 int err = 0;
345
346 if (rcu_access_pointer(vma->active.excl.fence)) {
347 struct dma_fence *fence;
348
349 rcu_read_lock();
350 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
351 rcu_read_unlock();
352 if (fence) {
353 err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
354 dma_fence_put(fence);
355 }
356 }
357
358 return err;
359}
360
361/**
362 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
363 * @vma: VMA to map
364 * @cache_level: mapping cache level
365 * @flags: flags like global or local mapping
366 * @work: preallocated worker for allocating and binding the PTE
367 *
368 * DMA addresses are taken from the scatter-gather table of this object (or of
369 * this VMA in case of non-default GGTT views) and PTE entries set up.
370 * Note that DMA addresses are also the only part of the SG table we care about.
371 */
372int i915_vma_bind(struct i915_vma *vma,
373 enum i915_cache_level cache_level,
374 u32 flags,
375 struct i915_vma_work *work)
376{
377 u32 bind_flags;
378 u32 vma_flags;
379 int ret;
380
381 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
382 GEM_BUG_ON(vma->size > vma->node.size);
383
384 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
385 vma->node.size,
386 vma->vm->total)))
387 return -ENODEV;
388
389 if (GEM_DEBUG_WARN_ON(!flags))
390 return -EINVAL;
391
392 bind_flags = flags;
393 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
394
395 vma_flags = atomic_read(&vma->flags);
396 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
397
398 bind_flags &= ~vma_flags;
399 if (bind_flags == 0)
400 return 0;
401
402 GEM_BUG_ON(!vma->pages);
403
404 trace_i915_vma_bind(vma, bind_flags);
405 if (work && bind_flags & vma->vm->bind_async_flags) {
406 struct dma_fence *prev;
407
408 work->vma = vma;
409 work->cache_level = cache_level;
410 work->flags = bind_flags;
411
412 /*
413 * Note we only want to chain up to the migration fence on
414 * the pages (not the object itself). As we don't track that,
415 * yet, we have to use the exclusive fence instead.
416 *
417 * Also note that we do not want to track the async vma as
418 * part of the obj->resv->excl_fence as it only affects
419 * execution and not content or object's backing store lifetime.
420 */
421 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
422 if (prev) {
423 __i915_sw_fence_await_dma_fence(&work->base.chain,
424 prev,
425 &work->cb);
426 dma_fence_put(prev);
427 }
428
429 work->base.dma.error = 0; /* enable the queue_work() */
430
431 if (vma->obj) {
432 __i915_gem_object_pin_pages(vma->obj);
433 work->pinned = vma->obj;
434 }
435 } else {
436 ret = vma->ops->bind_vma(vma->vm, vma, cache_level, bind_flags);
437 if (ret)
438 return ret;
439 }
440
441 atomic_or(bind_flags, &vma->flags);
442 return 0;
443}
444
445void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
446{
447 void __iomem *ptr;
448 int err;
449
450 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
451 err = -ENODEV;
452 goto err;
453 }
454
455 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
456 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
457
458 ptr = READ_ONCE(vma->iomap);
459 if (ptr == NULL) {
460 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
461 vma->node.start,
462 vma->node.size);
463 if (ptr == NULL) {
464 err = -ENOMEM;
465 goto err;
466 }
467
468 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
469 io_mapping_unmap(ptr);
470 ptr = vma->iomap;
471 }
472 }
473
474 __i915_vma_pin(vma);
475
476 err = i915_vma_pin_fence(vma);
477 if (err)
478 goto err_unpin;
479
480 i915_vma_set_ggtt_write(vma);
481
482 /* NB Access through the GTT requires the device to be awake. */
483 return ptr;
484
485err_unpin:
486 __i915_vma_unpin(vma);
487err:
488 return IO_ERR_PTR(err);
489}
490
491void i915_vma_flush_writes(struct i915_vma *vma)
492{
493 if (i915_vma_unset_ggtt_write(vma))
494 intel_gt_flush_ggtt_writes(vma->vm->gt);
495}
496
497void i915_vma_unpin_iomap(struct i915_vma *vma)
498{
499 GEM_BUG_ON(vma->iomap == NULL);
500
501 i915_vma_flush_writes(vma);
502
503 i915_vma_unpin_fence(vma);
504 i915_vma_unpin(vma);
505}
506
507void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
508{
509 struct i915_vma *vma;
510 struct drm_i915_gem_object *obj;
511
512 vma = fetch_and_zero(p_vma);
513 if (!vma)
514 return;
515
516 obj = vma->obj;
517 GEM_BUG_ON(!obj);
518
519 i915_vma_unpin(vma);
520
521 if (flags & I915_VMA_RELEASE_MAP)
522 i915_gem_object_unpin_map(obj);
523
524 i915_gem_object_put(obj);
525}
526
527bool i915_vma_misplaced(const struct i915_vma *vma,
528 u64 size, u64 alignment, u64 flags)
529{
530 if (!drm_mm_node_allocated(&vma->node))
531 return false;
532
533 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
534 return true;
535
536 if (vma->node.size < size)
537 return true;
538
539 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
540 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
541 return true;
542
543 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
544 return true;
545
546 if (flags & PIN_OFFSET_BIAS &&
547 vma->node.start < (flags & PIN_OFFSET_MASK))
548 return true;
549
550 if (flags & PIN_OFFSET_FIXED &&
551 vma->node.start != (flags & PIN_OFFSET_MASK))
552 return true;
553
554 return false;
555}
556
557void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
558{
559 bool mappable, fenceable;
560
561 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
562 GEM_BUG_ON(!vma->fence_size);
563
564 fenceable = (vma->node.size >= vma->fence_size &&
565 IS_ALIGNED(vma->node.start, vma->fence_alignment));
566
567 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
568
569 if (mappable && fenceable)
570 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
571 else
572 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
573}
574
575bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
576{
577 struct drm_mm_node *node = &vma->node;
578 struct drm_mm_node *other;
579
580 /*
581 * On some machines we have to be careful when putting differing types
582 * of snoopable memory together to avoid the prefetcher crossing memory
583 * domains and dying. During vm initialisation, we decide whether or not
584 * these constraints apply and set the drm_mm.color_adjust
585 * appropriately.
586 */
587 if (!i915_vm_has_cache_coloring(vma->vm))
588 return true;
589
590 /* Only valid to be called on an already inserted vma */
591 GEM_BUG_ON(!drm_mm_node_allocated(node));
592 GEM_BUG_ON(list_empty(&node->node_list));
593
594 other = list_prev_entry(node, node_list);
595 if (i915_node_color_differs(other, color) &&
596 !drm_mm_hole_follows(other))
597 return false;
598
599 other = list_next_entry(node, node_list);
600 if (i915_node_color_differs(other, color) &&
601 !drm_mm_hole_follows(node))
602 return false;
603
604 return true;
605}
606
607/**
608 * i915_vma_insert - finds a slot for the vma in its address space
609 * @vma: the vma
610 * @size: requested size in bytes (can be larger than the VMA)
611 * @alignment: required alignment
612 * @flags: mask of PIN_* flags to use
613 *
614 * First we try to allocate some free space that meets the requirements for
615 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
616 * preferrably the oldest idle entry to make room for the new VMA.
617 *
618 * Returns:
619 * 0 on success, negative error code otherwise.
620 */
621static int
622i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
623{
624 unsigned long color;
625 u64 start, end;
626 int ret;
627
628 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
629 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
630
631 size = max(size, vma->size);
632 alignment = max(alignment, vma->display_alignment);
633 if (flags & PIN_MAPPABLE) {
634 size = max_t(typeof(size), size, vma->fence_size);
635 alignment = max_t(typeof(alignment),
636 alignment, vma->fence_alignment);
637 }
638
639 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
640 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
641 GEM_BUG_ON(!is_power_of_2(alignment));
642
643 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
644 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
645
646 end = vma->vm->total;
647 if (flags & PIN_MAPPABLE)
648 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
649 if (flags & PIN_ZONE_4G)
650 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
651 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
652
653 /* If binding the object/GGTT view requires more space than the entire
654 * aperture has, reject it early before evicting everything in a vain
655 * attempt to find space.
656 */
657 if (size > end) {
658 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
659 size, flags & PIN_MAPPABLE ? "mappable" : "total",
660 end);
661 return -ENOSPC;
662 }
663
664 color = 0;
665 if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
666 color = vma->obj->cache_level;
667
668 if (flags & PIN_OFFSET_FIXED) {
669 u64 offset = flags & PIN_OFFSET_MASK;
670 if (!IS_ALIGNED(offset, alignment) ||
671 range_overflows(offset, size, end))
672 return -EINVAL;
673
674 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
675 size, offset, color,
676 flags);
677 if (ret)
678 return ret;
679 } else {
680 /*
681 * We only support huge gtt pages through the 48b PPGTT,
682 * however we also don't want to force any alignment for
683 * objects which need to be tightly packed into the low 32bits.
684 *
685 * Note that we assume that GGTT are limited to 4GiB for the
686 * forseeable future. See also i915_ggtt_offset().
687 */
688 if (upper_32_bits(end - 1) &&
689 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
690 /*
691 * We can't mix 64K and 4K PTEs in the same page-table
692 * (2M block), and so to avoid the ugliness and
693 * complexity of coloring we opt for just aligning 64K
694 * objects to 2M.
695 */
696 u64 page_alignment =
697 rounddown_pow_of_two(vma->page_sizes.sg |
698 I915_GTT_PAGE_SIZE_2M);
699
700 /*
701 * Check we don't expand for the limited Global GTT
702 * (mappable aperture is even more precious!). This
703 * also checks that we exclude the aliasing-ppgtt.
704 */
705 GEM_BUG_ON(i915_vma_is_ggtt(vma));
706
707 alignment = max(alignment, page_alignment);
708
709 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
710 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
711 }
712
713 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
714 size, alignment, color,
715 start, end, flags);
716 if (ret)
717 return ret;
718
719 GEM_BUG_ON(vma->node.start < start);
720 GEM_BUG_ON(vma->node.start + vma->node.size > end);
721 }
722 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
723 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
724
725 list_add_tail(&vma->vm_link, &vma->vm->bound_list);
726
727 return 0;
728}
729
730static void
731i915_vma_detach(struct i915_vma *vma)
732{
733 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
734 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
735
736 /*
737 * And finally now the object is completely decoupled from this
738 * vma, we can drop its hold on the backing storage and allow
739 * it to be reaped by the shrinker.
740 */
741 list_del(&vma->vm_link);
742}
743
744static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
745{
746 unsigned int bound;
747 bool pinned = true;
748
749 bound = atomic_read(&vma->flags);
750 do {
751 if (unlikely(flags & ~bound))
752 return false;
753
754 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
755 return false;
756
757 if (!(bound & I915_VMA_PIN_MASK))
758 goto unpinned;
759
760 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
761 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
762
763 return true;
764
765unpinned:
766 /*
767 * If pin_count==0, but we are bound, check under the lock to avoid
768 * racing with a concurrent i915_vma_unbind().
769 */
770 mutex_lock(&vma->vm->mutex);
771 do {
772 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
773 pinned = false;
774 break;
775 }
776
777 if (unlikely(flags & ~bound)) {
778 pinned = false;
779 break;
780 }
781 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
782 mutex_unlock(&vma->vm->mutex);
783
784 return pinned;
785}
786
787static int vma_get_pages(struct i915_vma *vma)
788{
789 int err = 0;
790
791 if (atomic_add_unless(&vma->pages_count, 1, 0))
792 return 0;
793
794 /* Allocations ahoy! */
795 if (mutex_lock_interruptible(&vma->pages_mutex))
796 return -EINTR;
797
798 if (!atomic_read(&vma->pages_count)) {
799 if (vma->obj) {
800 err = i915_gem_object_pin_pages(vma->obj);
801 if (err)
802 goto unlock;
803 }
804
805 err = vma->ops->set_pages(vma);
806 if (err) {
807 if (vma->obj)
808 i915_gem_object_unpin_pages(vma->obj);
809 goto unlock;
810 }
811 }
812 atomic_inc(&vma->pages_count);
813
814unlock:
815 mutex_unlock(&vma->pages_mutex);
816
817 return err;
818}
819
820static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
821{
822 /* We allocate under vma_get_pages, so beware the shrinker */
823 mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
824 GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
825 if (atomic_sub_return(count, &vma->pages_count) == 0) {
826 vma->ops->clear_pages(vma);
827 GEM_BUG_ON(vma->pages);
828 if (vma->obj)
829 i915_gem_object_unpin_pages(vma->obj);
830 }
831 mutex_unlock(&vma->pages_mutex);
832}
833
834static void vma_put_pages(struct i915_vma *vma)
835{
836 if (atomic_add_unless(&vma->pages_count, -1, 1))
837 return;
838
839 __vma_put_pages(vma, 1);
840}
841
842static void vma_unbind_pages(struct i915_vma *vma)
843{
844 unsigned int count;
845
846 lockdep_assert_held(&vma->vm->mutex);
847
848 /* The upper portion of pages_count is the number of bindings */
849 count = atomic_read(&vma->pages_count);
850 count >>= I915_VMA_PAGES_BIAS;
851 GEM_BUG_ON(!count);
852
853 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
854}
855
856int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
857{
858 struct i915_vma_work *work = NULL;
859 intel_wakeref_t wakeref = 0;
860 unsigned int bound;
861 int err;
862
863 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
864 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
865
866 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
867
868 /* First try and grab the pin without rebinding the vma */
869 if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
870 return 0;
871
872 err = vma_get_pages(vma);
873 if (err)
874 return err;
875
876 if (flags & vma->vm->bind_async_flags) {
877 work = i915_vma_work();
878 if (!work) {
879 err = -ENOMEM;
880 goto err_pages;
881 }
882 }
883
884 if (flags & PIN_GLOBAL)
885 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
886
887 /*
888 * Differentiate between user/kernel vma inside the aliasing-ppgtt.
889 *
890 * We conflate the Global GTT with the user's vma when using the
891 * aliasing-ppgtt, but it is still vitally important to try and
892 * keep the use cases distinct. For example, userptr objects are
893 * not allowed inside the Global GTT as that will cause lock
894 * inversions when we have to evict them the mmu_notifier callbacks -
895 * but they are allowed to be part of the user ppGTT which can never
896 * be mapped. As such we try to give the distinct users of the same
897 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
898 * and i915_ppgtt separate].
899 *
900 * NB this may cause us to mask real lock inversions -- while the
901 * code is safe today, lockdep may not be able to spot future
902 * transgressions.
903 */
904 err = mutex_lock_interruptible_nested(&vma->vm->mutex,
905 !(flags & PIN_GLOBAL));
906 if (err)
907 goto err_fence;
908
909 /* No more allocations allowed now we hold vm->mutex */
910
911 if (unlikely(i915_vma_is_closed(vma))) {
912 err = -ENOENT;
913 goto err_unlock;
914 }
915
916 bound = atomic_read(&vma->flags);
917 if (unlikely(bound & I915_VMA_ERROR)) {
918 err = -ENOMEM;
919 goto err_unlock;
920 }
921
922 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
923 err = -EAGAIN; /* pins are meant to be fairly temporary */
924 goto err_unlock;
925 }
926
927 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
928 __i915_vma_pin(vma);
929 goto err_unlock;
930 }
931
932 err = i915_active_acquire(&vma->active);
933 if (err)
934 goto err_unlock;
935
936 if (!(bound & I915_VMA_BIND_MASK)) {
937 err = i915_vma_insert(vma, size, alignment, flags);
938 if (err)
939 goto err_active;
940
941 if (i915_is_ggtt(vma->vm))
942 __i915_vma_set_map_and_fenceable(vma);
943 }
944
945 GEM_BUG_ON(!vma->pages);
946 err = i915_vma_bind(vma,
947 vma->obj ? vma->obj->cache_level : 0,
948 flags, work);
949 if (err)
950 goto err_remove;
951
952 /* There should only be at most 2 active bindings (user, global) */
953 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
954 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
955 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
956
957 __i915_vma_pin(vma);
958 GEM_BUG_ON(!i915_vma_is_pinned(vma));
959 GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
960 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
961
962err_remove:
963 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
964 i915_vma_detach(vma);
965 drm_mm_remove_node(&vma->node);
966 }
967err_active:
968 i915_active_release(&vma->active);
969err_unlock:
970 mutex_unlock(&vma->vm->mutex);
971err_fence:
972 if (work)
973 dma_fence_work_commit_imm(&work->base);
974 if (wakeref)
975 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
976err_pages:
977 vma_put_pages(vma);
978 return err;
979}
980
981static void flush_idle_contexts(struct intel_gt *gt)
982{
983 struct intel_engine_cs *engine;
984 enum intel_engine_id id;
985
986 for_each_engine(engine, gt, id)
987 intel_engine_flush_barriers(engine);
988
989 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
990}
991
992int i915_ggtt_pin(struct i915_vma *vma, u32 align, unsigned int flags)
993{
994 struct i915_address_space *vm = vma->vm;
995 int err;
996
997 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
998
999 do {
1000 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1001 if (err != -ENOSPC) {
1002 if (!err) {
1003 err = i915_vma_wait_for_bind(vma);
1004 if (err)
1005 i915_vma_unpin(vma);
1006 }
1007 return err;
1008 }
1009
1010 /* Unlike i915_vma_pin, we don't take no for an answer! */
1011 flush_idle_contexts(vm->gt);
1012 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1013 i915_gem_evict_vm(vm);
1014 mutex_unlock(&vm->mutex);
1015 }
1016 } while (1);
1017}
1018
1019static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1020{
1021 /*
1022 * We defer actually closing, unbinding and destroying the VMA until
1023 * the next idle point, or if the object is freed in the meantime. By
1024 * postponing the unbind, we allow for it to be resurrected by the
1025 * client, avoiding the work required to rebind the VMA. This is
1026 * advantageous for DRI, where the client/server pass objects
1027 * between themselves, temporarily opening a local VMA to the
1028 * object, and then closing it again. The same object is then reused
1029 * on the next frame (or two, depending on the depth of the swap queue)
1030 * causing us to rebind the VMA once more. This ends up being a lot
1031 * of wasted work for the steady state.
1032 */
1033 GEM_BUG_ON(i915_vma_is_closed(vma));
1034 list_add(&vma->closed_link, >->closed_vma);
1035}
1036
1037void i915_vma_close(struct i915_vma *vma)
1038{
1039 struct intel_gt *gt = vma->vm->gt;
1040 unsigned long flags;
1041
1042 if (i915_vma_is_ggtt(vma))
1043 return;
1044
1045 GEM_BUG_ON(!atomic_read(&vma->open_count));
1046 if (atomic_dec_and_lock_irqsave(&vma->open_count,
1047 >->closed_lock,
1048 flags)) {
1049 __vma_close(vma, gt);
1050 spin_unlock_irqrestore(>->closed_lock, flags);
1051 }
1052}
1053
1054static void __i915_vma_remove_closed(struct i915_vma *vma)
1055{
1056 struct intel_gt *gt = vma->vm->gt;
1057
1058 spin_lock_irq(>->closed_lock);
1059 list_del_init(&vma->closed_link);
1060 spin_unlock_irq(>->closed_lock);
1061}
1062
1063void i915_vma_reopen(struct i915_vma *vma)
1064{
1065 if (i915_vma_is_closed(vma))
1066 __i915_vma_remove_closed(vma);
1067}
1068
1069void i915_vma_release(struct kref *ref)
1070{
1071 struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1072
1073 if (drm_mm_node_allocated(&vma->node)) {
1074 mutex_lock(&vma->vm->mutex);
1075 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1076 WARN_ON(__i915_vma_unbind(vma));
1077 mutex_unlock(&vma->vm->mutex);
1078 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1079 }
1080 GEM_BUG_ON(i915_vma_is_active(vma));
1081
1082 if (vma->obj) {
1083 struct drm_i915_gem_object *obj = vma->obj;
1084
1085 spin_lock(&obj->vma.lock);
1086 list_del(&vma->obj_link);
1087 if (!RB_EMPTY_NODE(&vma->obj_node))
1088 rb_erase(&vma->obj_node, &obj->vma.tree);
1089 spin_unlock(&obj->vma.lock);
1090 }
1091
1092 __i915_vma_remove_closed(vma);
1093 i915_vm_put(vma->vm);
1094
1095 i915_active_fini(&vma->active);
1096 i915_vma_free(vma);
1097}
1098
1099void i915_vma_parked(struct intel_gt *gt)
1100{
1101 struct i915_vma *vma, *next;
1102 LIST_HEAD(closed);
1103
1104 spin_lock_irq(>->closed_lock);
1105 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
1106 struct drm_i915_gem_object *obj = vma->obj;
1107 struct i915_address_space *vm = vma->vm;
1108
1109 /* XXX All to avoid keeping a reference on i915_vma itself */
1110
1111 if (!kref_get_unless_zero(&obj->base.refcount))
1112 continue;
1113
1114 if (!i915_vm_tryopen(vm)) {
1115 i915_gem_object_put(obj);
1116 continue;
1117 }
1118
1119 list_move(&vma->closed_link, &closed);
1120 }
1121 spin_unlock_irq(>->closed_lock);
1122
1123 /* As the GT is held idle, no vma can be reopened as we destroy them */
1124 list_for_each_entry_safe(vma, next, &closed, closed_link) {
1125 struct drm_i915_gem_object *obj = vma->obj;
1126 struct i915_address_space *vm = vma->vm;
1127
1128 INIT_LIST_HEAD(&vma->closed_link);
1129 __i915_vma_put(vma);
1130
1131 i915_gem_object_put(obj);
1132 i915_vm_close(vm);
1133 }
1134}
1135
1136static void __i915_vma_iounmap(struct i915_vma *vma)
1137{
1138 GEM_BUG_ON(i915_vma_is_pinned(vma));
1139
1140 if (vma->iomap == NULL)
1141 return;
1142
1143 io_mapping_unmap(vma->iomap);
1144 vma->iomap = NULL;
1145}
1146
1147void i915_vma_revoke_mmap(struct i915_vma *vma)
1148{
1149 struct drm_vma_offset_node *node;
1150 u64 vma_offset;
1151
1152 if (!i915_vma_has_userfault(vma))
1153 return;
1154
1155 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1156 GEM_BUG_ON(!vma->obj->userfault_count);
1157
1158 node = &vma->mmo->vma_node;
1159 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1160 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1161 drm_vma_node_offset_addr(node) + vma_offset,
1162 vma->size,
1163 1);
1164
1165 i915_vma_unset_userfault(vma);
1166 if (!--vma->obj->userfault_count)
1167 list_del(&vma->obj->userfault_link);
1168}
1169
1170int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1171{
1172 int err;
1173
1174 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1175
1176 /* Wait for the vma to be bound before we start! */
1177 err = i915_request_await_active(rq, &vma->active,
1178 I915_ACTIVE_AWAIT_EXCL);
1179 if (err)
1180 return err;
1181
1182 return i915_active_add_request(&vma->active, rq);
1183}
1184
1185int i915_vma_move_to_active(struct i915_vma *vma,
1186 struct i915_request *rq,
1187 unsigned int flags)
1188{
1189 struct drm_i915_gem_object *obj = vma->obj;
1190 int err;
1191
1192 assert_object_held(obj);
1193
1194 err = __i915_vma_move_to_active(vma, rq);
1195 if (unlikely(err))
1196 return err;
1197
1198 if (flags & EXEC_OBJECT_WRITE) {
1199 struct intel_frontbuffer *front;
1200
1201 front = __intel_frontbuffer_get(obj);
1202 if (unlikely(front)) {
1203 if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1204 i915_active_add_request(&front->write, rq);
1205 intel_frontbuffer_put(front);
1206 }
1207
1208 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1209 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1210 obj->read_domains = 0;
1211 } else {
1212 err = dma_resv_reserve_shared(vma->resv, 1);
1213 if (unlikely(err))
1214 return err;
1215
1216 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1217 obj->write_domain = 0;
1218 }
1219
1220 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1221 i915_active_add_request(&vma->fence->active, rq);
1222
1223 obj->read_domains |= I915_GEM_GPU_DOMAINS;
1224 obj->mm.dirty = true;
1225
1226 GEM_BUG_ON(!i915_vma_is_active(vma));
1227 return 0;
1228}
1229
1230void __i915_vma_evict(struct i915_vma *vma)
1231{
1232 GEM_BUG_ON(i915_vma_is_pinned(vma));
1233
1234 if (i915_vma_is_map_and_fenceable(vma)) {
1235 /* Force a pagefault for domain tracking on next user access */
1236 i915_vma_revoke_mmap(vma);
1237
1238 /*
1239 * Check that we have flushed all writes through the GGTT
1240 * before the unbind, other due to non-strict nature of those
1241 * indirect writes they may end up referencing the GGTT PTE
1242 * after the unbind.
1243 *
1244 * Note that we may be concurrently poking at the GGTT_WRITE
1245 * bit from set-domain, as we mark all GGTT vma associated
1246 * with an object. We know this is for another vma, as we
1247 * are currently unbinding this one -- so if this vma will be
1248 * reused, it will be refaulted and have its dirty bit set
1249 * before the next write.
1250 */
1251 i915_vma_flush_writes(vma);
1252
1253 /* release the fence reg _after_ flushing */
1254 i915_vma_revoke_fence(vma);
1255
1256 __i915_vma_iounmap(vma);
1257 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1258 }
1259 GEM_BUG_ON(vma->fence);
1260 GEM_BUG_ON(i915_vma_has_userfault(vma));
1261
1262 if (likely(atomic_read(&vma->vm->open))) {
1263 trace_i915_vma_unbind(vma);
1264 vma->ops->unbind_vma(vma->vm, vma);
1265 }
1266 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1267 &vma->flags);
1268
1269 i915_vma_detach(vma);
1270 vma_unbind_pages(vma);
1271}
1272
1273int __i915_vma_unbind(struct i915_vma *vma)
1274{
1275 int ret;
1276
1277 lockdep_assert_held(&vma->vm->mutex);
1278
1279 if (!drm_mm_node_allocated(&vma->node))
1280 return 0;
1281
1282 if (i915_vma_is_pinned(vma)) {
1283 vma_print_allocator(vma, "is pinned");
1284 return -EAGAIN;
1285 }
1286
1287 /*
1288 * After confirming that no one else is pinning this vma, wait for
1289 * any laggards who may have crept in during the wait (through
1290 * a residual pin skipping the vm->mutex) to complete.
1291 */
1292 ret = i915_vma_sync(vma);
1293 if (ret)
1294 return ret;
1295
1296 GEM_BUG_ON(i915_vma_is_active(vma));
1297 __i915_vma_evict(vma);
1298
1299 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1300 return 0;
1301}
1302
1303int i915_vma_unbind(struct i915_vma *vma)
1304{
1305 struct i915_address_space *vm = vma->vm;
1306 intel_wakeref_t wakeref = 0;
1307 int err;
1308
1309 /* Optimistic wait before taking the mutex */
1310 err = i915_vma_sync(vma);
1311 if (err)
1312 return err;
1313
1314 if (!drm_mm_node_allocated(&vma->node))
1315 return 0;
1316
1317 if (i915_vma_is_pinned(vma)) {
1318 vma_print_allocator(vma, "is pinned");
1319 return -EAGAIN;
1320 }
1321
1322 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1323 /* XXX not always required: nop_clear_range */
1324 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1325
1326 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1327 if (err)
1328 goto out_rpm;
1329
1330 err = __i915_vma_unbind(vma);
1331 mutex_unlock(&vm->mutex);
1332
1333out_rpm:
1334 if (wakeref)
1335 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1336 return err;
1337}
1338
1339struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1340{
1341 i915_gem_object_make_unshrinkable(vma->obj);
1342 return vma;
1343}
1344
1345void i915_vma_make_shrinkable(struct i915_vma *vma)
1346{
1347 i915_gem_object_make_shrinkable(vma->obj);
1348}
1349
1350void i915_vma_make_purgeable(struct i915_vma *vma)
1351{
1352 i915_gem_object_make_purgeable(vma->obj);
1353}
1354
1355#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1356#include "selftests/i915_vma.c"
1357#endif
1358
1359static void i915_global_vma_shrink(void)
1360{
1361 kmem_cache_shrink(global.slab_vmas);
1362}
1363
1364static void i915_global_vma_exit(void)
1365{
1366 kmem_cache_destroy(global.slab_vmas);
1367}
1368
1369static struct i915_global_vma global = { {
1370 .shrink = i915_global_vma_shrink,
1371 .exit = i915_global_vma_exit,
1372} };
1373
1374int __init i915_global_vma_init(void)
1375{
1376 global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1377 if (!global.slab_vmas)
1378 return -ENOMEM;
1379
1380 i915_global_register(&global.base);
1381 return 0;
1382}