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 <linux/dma-fence-array.h>
27#include <drm/drm_gem.h>
28
29#include "display/intel_frontbuffer.h"
30#include "gem/i915_gem_lmem.h"
31#include "gem/i915_gem_tiling.h"
32#include "gt/intel_engine.h"
33#include "gt/intel_engine_heartbeat.h"
34#include "gt/intel_gt.h"
35#include "gt/intel_gt_requests.h"
36
37#include "i915_drv.h"
38#include "i915_gem_evict.h"
39#include "i915_sw_fence_work.h"
40#include "i915_trace.h"
41#include "i915_vma.h"
42#include "i915_vma_resource.h"
43
44static inline void assert_vma_held_evict(const struct i915_vma *vma)
45{
46 /*
47 * We may be forced to unbind when the vm is dead, to clean it up.
48 * This is the only exception to the requirement of the object lock
49 * being held.
50 */
51 if (kref_read(&vma->vm->ref))
52 assert_object_held_shared(vma->obj);
53}
54
55static struct kmem_cache *slab_vmas;
56
57static struct i915_vma *i915_vma_alloc(void)
58{
59 return kmem_cache_zalloc(slab_vmas, GFP_KERNEL);
60}
61
62static void i915_vma_free(struct i915_vma *vma)
63{
64 return kmem_cache_free(slab_vmas, vma);
65}
66
67#if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
68
69#include <linux/stackdepot.h>
70
71static void vma_print_allocator(struct i915_vma *vma, const char *reason)
72{
73 char buf[512];
74
75 if (!vma->node.stack) {
76 drm_dbg(&to_i915(vma->obj->base.dev)->drm,
77 "vma.node [%08llx + %08llx] %s: unknown owner\n",
78 vma->node.start, vma->node.size, reason);
79 return;
80 }
81
82 stack_depot_snprint(vma->node.stack, buf, sizeof(buf), 0);
83 drm_dbg(&to_i915(vma->obj->base.dev)->drm,
84 "vma.node [%08llx + %08llx] %s: inserted at %s\n",
85 vma->node.start, vma->node.size, reason, buf);
86}
87
88#else
89
90static void vma_print_allocator(struct i915_vma *vma, const char *reason)
91{
92}
93
94#endif
95
96static inline struct i915_vma *active_to_vma(struct i915_active *ref)
97{
98 return container_of(ref, typeof(struct i915_vma), active);
99}
100
101static int __i915_vma_active(struct i915_active *ref)
102{
103 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
104}
105
106static void __i915_vma_retire(struct i915_active *ref)
107{
108 i915_vma_put(active_to_vma(ref));
109}
110
111static struct i915_vma *
112vma_create(struct drm_i915_gem_object *obj,
113 struct i915_address_space *vm,
114 const struct i915_gtt_view *view)
115{
116 struct i915_vma *pos = ERR_PTR(-E2BIG);
117 struct i915_vma *vma;
118 struct rb_node *rb, **p;
119 int err;
120
121 /* The aliasing_ppgtt should never be used directly! */
122 GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
123
124 vma = i915_vma_alloc();
125 if (vma == NULL)
126 return ERR_PTR(-ENOMEM);
127
128 vma->ops = &vm->vma_ops;
129 vma->obj = obj;
130 vma->size = obj->base.size;
131 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
132
133 i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0);
134
135 /* Declare ourselves safe for use inside shrinkers */
136 if (IS_ENABLED(CONFIG_LOCKDEP)) {
137 fs_reclaim_acquire(GFP_KERNEL);
138 might_lock(&vma->active.mutex);
139 fs_reclaim_release(GFP_KERNEL);
140 }
141
142 INIT_LIST_HEAD(&vma->closed_link);
143 INIT_LIST_HEAD(&vma->obj_link);
144 RB_CLEAR_NODE(&vma->obj_node);
145
146 if (view && view->type != I915_GTT_VIEW_NORMAL) {
147 vma->gtt_view = *view;
148 if (view->type == I915_GTT_VIEW_PARTIAL) {
149 GEM_BUG_ON(range_overflows_t(u64,
150 view->partial.offset,
151 view->partial.size,
152 obj->base.size >> PAGE_SHIFT));
153 vma->size = view->partial.size;
154 vma->size <<= PAGE_SHIFT;
155 GEM_BUG_ON(vma->size > obj->base.size);
156 } else if (view->type == I915_GTT_VIEW_ROTATED) {
157 vma->size = intel_rotation_info_size(&view->rotated);
158 vma->size <<= PAGE_SHIFT;
159 } else if (view->type == I915_GTT_VIEW_REMAPPED) {
160 vma->size = intel_remapped_info_size(&view->remapped);
161 vma->size <<= PAGE_SHIFT;
162 }
163 }
164
165 if (unlikely(vma->size > vm->total))
166 goto err_vma;
167
168 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
169
170 err = mutex_lock_interruptible(&vm->mutex);
171 if (err) {
172 pos = ERR_PTR(err);
173 goto err_vma;
174 }
175
176 vma->vm = vm;
177 list_add_tail(&vma->vm_link, &vm->unbound_list);
178
179 spin_lock(&obj->vma.lock);
180 if (i915_is_ggtt(vm)) {
181 if (unlikely(overflows_type(vma->size, u32)))
182 goto err_unlock;
183
184 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
185 i915_gem_object_get_tiling(obj),
186 i915_gem_object_get_stride(obj));
187 if (unlikely(vma->fence_size < vma->size || /* overflow */
188 vma->fence_size > vm->total))
189 goto err_unlock;
190
191 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
192
193 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
194 i915_gem_object_get_tiling(obj),
195 i915_gem_object_get_stride(obj));
196 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
197
198 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
199 }
200
201 rb = NULL;
202 p = &obj->vma.tree.rb_node;
203 while (*p) {
204 long cmp;
205
206 rb = *p;
207 pos = rb_entry(rb, struct i915_vma, obj_node);
208
209 /*
210 * If the view already exists in the tree, another thread
211 * already created a matching vma, so return the older instance
212 * and dispose of ours.
213 */
214 cmp = i915_vma_compare(pos, vm, view);
215 if (cmp < 0)
216 p = &rb->rb_right;
217 else if (cmp > 0)
218 p = &rb->rb_left;
219 else
220 goto err_unlock;
221 }
222 rb_link_node(&vma->obj_node, rb, p);
223 rb_insert_color(&vma->obj_node, &obj->vma.tree);
224
225 if (i915_vma_is_ggtt(vma))
226 /*
227 * We put the GGTT vma at the start of the vma-list, followed
228 * by the ppGGTT vma. This allows us to break early when
229 * iterating over only the GGTT vma for an object, see
230 * for_each_ggtt_vma()
231 */
232 list_add(&vma->obj_link, &obj->vma.list);
233 else
234 list_add_tail(&vma->obj_link, &obj->vma.list);
235
236 spin_unlock(&obj->vma.lock);
237 mutex_unlock(&vm->mutex);
238
239 return vma;
240
241err_unlock:
242 spin_unlock(&obj->vma.lock);
243 list_del_init(&vma->vm_link);
244 mutex_unlock(&vm->mutex);
245err_vma:
246 i915_vma_free(vma);
247 return pos;
248}
249
250static struct i915_vma *
251i915_vma_lookup(struct drm_i915_gem_object *obj,
252 struct i915_address_space *vm,
253 const struct i915_gtt_view *view)
254{
255 struct rb_node *rb;
256
257 rb = obj->vma.tree.rb_node;
258 while (rb) {
259 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
260 long cmp;
261
262 cmp = i915_vma_compare(vma, vm, view);
263 if (cmp == 0)
264 return vma;
265
266 if (cmp < 0)
267 rb = rb->rb_right;
268 else
269 rb = rb->rb_left;
270 }
271
272 return NULL;
273}
274
275/**
276 * i915_vma_instance - return the singleton instance of the VMA
277 * @obj: parent &struct drm_i915_gem_object to be mapped
278 * @vm: address space in which the mapping is located
279 * @view: additional mapping requirements
280 *
281 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
282 * the same @view characteristics. If a match is not found, one is created.
283 * Once created, the VMA is kept until either the object is freed, or the
284 * address space is closed.
285 *
286 * Returns the vma, or an error pointer.
287 */
288struct i915_vma *
289i915_vma_instance(struct drm_i915_gem_object *obj,
290 struct i915_address_space *vm,
291 const struct i915_gtt_view *view)
292{
293 struct i915_vma *vma;
294
295 GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm));
296 GEM_BUG_ON(!kref_read(&vm->ref));
297
298 spin_lock(&obj->vma.lock);
299 vma = i915_vma_lookup(obj, vm, view);
300 spin_unlock(&obj->vma.lock);
301
302 /* vma_create() will resolve the race if another creates the vma */
303 if (unlikely(!vma))
304 vma = vma_create(obj, vm, view);
305
306 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
307 return vma;
308}
309
310struct i915_vma_work {
311 struct dma_fence_work base;
312 struct i915_address_space *vm;
313 struct i915_vm_pt_stash stash;
314 struct i915_vma_resource *vma_res;
315 struct drm_i915_gem_object *obj;
316 struct i915_sw_dma_fence_cb cb;
317 enum i915_cache_level cache_level;
318 unsigned int flags;
319};
320
321static void __vma_bind(struct dma_fence_work *work)
322{
323 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
324 struct i915_vma_resource *vma_res = vw->vma_res;
325
326 /*
327 * We are about the bind the object, which must mean we have already
328 * signaled the work to potentially clear/move the pages underneath. If
329 * something went wrong at that stage then the object should have
330 * unknown_state set, in which case we need to skip the bind.
331 */
332 if (i915_gem_object_has_unknown_state(vw->obj))
333 return;
334
335 vma_res->ops->bind_vma(vma_res->vm, &vw->stash,
336 vma_res, vw->cache_level, vw->flags);
337}
338
339static void __vma_release(struct dma_fence_work *work)
340{
341 struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
342
343 if (vw->obj)
344 i915_gem_object_put(vw->obj);
345
346 i915_vm_free_pt_stash(vw->vm, &vw->stash);
347 if (vw->vma_res)
348 i915_vma_resource_put(vw->vma_res);
349}
350
351static const struct dma_fence_work_ops bind_ops = {
352 .name = "bind",
353 .work = __vma_bind,
354 .release = __vma_release,
355};
356
357struct i915_vma_work *i915_vma_work(void)
358{
359 struct i915_vma_work *vw;
360
361 vw = kzalloc(sizeof(*vw), GFP_KERNEL);
362 if (!vw)
363 return NULL;
364
365 dma_fence_work_init(&vw->base, &bind_ops);
366 vw->base.dma.error = -EAGAIN; /* disable the worker by default */
367
368 return vw;
369}
370
371int i915_vma_wait_for_bind(struct i915_vma *vma)
372{
373 int err = 0;
374
375 if (rcu_access_pointer(vma->active.excl.fence)) {
376 struct dma_fence *fence;
377
378 rcu_read_lock();
379 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
380 rcu_read_unlock();
381 if (fence) {
382 err = dma_fence_wait(fence, true);
383 dma_fence_put(fence);
384 }
385 }
386
387 return err;
388}
389
390#if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
391static int i915_vma_verify_bind_complete(struct i915_vma *vma)
392{
393 struct dma_fence *fence = i915_active_fence_get(&vma->active.excl);
394 int err;
395
396 if (!fence)
397 return 0;
398
399 if (dma_fence_is_signaled(fence))
400 err = fence->error;
401 else
402 err = -EBUSY;
403
404 dma_fence_put(fence);
405
406 return err;
407}
408#else
409#define i915_vma_verify_bind_complete(_vma) 0
410#endif
411
412I915_SELFTEST_EXPORT void
413i915_vma_resource_init_from_vma(struct i915_vma_resource *vma_res,
414 struct i915_vma *vma)
415{
416 struct drm_i915_gem_object *obj = vma->obj;
417
418 i915_vma_resource_init(vma_res, vma->vm, vma->pages, &vma->page_sizes,
419 obj->mm.rsgt, i915_gem_object_is_readonly(obj),
420 i915_gem_object_is_lmem(obj), obj->mm.region,
421 vma->ops, vma->private, vma->node.start,
422 vma->node.size, vma->size);
423}
424
425/**
426 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
427 * @vma: VMA to map
428 * @cache_level: mapping cache level
429 * @flags: flags like global or local mapping
430 * @work: preallocated worker for allocating and binding the PTE
431 * @vma_res: pointer to a preallocated vma resource. The resource is either
432 * consumed or freed.
433 *
434 * DMA addresses are taken from the scatter-gather table of this object (or of
435 * this VMA in case of non-default GGTT views) and PTE entries set up.
436 * Note that DMA addresses are also the only part of the SG table we care about.
437 */
438int i915_vma_bind(struct i915_vma *vma,
439 enum i915_cache_level cache_level,
440 u32 flags,
441 struct i915_vma_work *work,
442 struct i915_vma_resource *vma_res)
443{
444 u32 bind_flags;
445 u32 vma_flags;
446 int ret;
447
448 lockdep_assert_held(&vma->vm->mutex);
449 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
450 GEM_BUG_ON(vma->size > vma->node.size);
451
452 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
453 vma->node.size,
454 vma->vm->total))) {
455 i915_vma_resource_free(vma_res);
456 return -ENODEV;
457 }
458
459 if (GEM_DEBUG_WARN_ON(!flags)) {
460 i915_vma_resource_free(vma_res);
461 return -EINVAL;
462 }
463
464 bind_flags = flags;
465 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
466
467 vma_flags = atomic_read(&vma->flags);
468 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
469
470 bind_flags &= ~vma_flags;
471 if (bind_flags == 0) {
472 i915_vma_resource_free(vma_res);
473 return 0;
474 }
475
476 GEM_BUG_ON(!atomic_read(&vma->pages_count));
477
478 /* Wait for or await async unbinds touching our range */
479 if (work && bind_flags & vma->vm->bind_async_flags)
480 ret = i915_vma_resource_bind_dep_await(vma->vm,
481 &work->base.chain,
482 vma->node.start,
483 vma->node.size,
484 true,
485 GFP_NOWAIT |
486 __GFP_RETRY_MAYFAIL |
487 __GFP_NOWARN);
488 else
489 ret = i915_vma_resource_bind_dep_sync(vma->vm, vma->node.start,
490 vma->node.size, true);
491 if (ret) {
492 i915_vma_resource_free(vma_res);
493 return ret;
494 }
495
496 if (vma->resource || !vma_res) {
497 /* Rebinding with an additional I915_VMA_*_BIND */
498 GEM_WARN_ON(!vma_flags);
499 i915_vma_resource_free(vma_res);
500 } else {
501 i915_vma_resource_init_from_vma(vma_res, vma);
502 vma->resource = vma_res;
503 }
504 trace_i915_vma_bind(vma, bind_flags);
505 if (work && bind_flags & vma->vm->bind_async_flags) {
506 struct dma_fence *prev;
507
508 work->vma_res = i915_vma_resource_get(vma->resource);
509 work->cache_level = cache_level;
510 work->flags = bind_flags;
511
512 /*
513 * Note we only want to chain up to the migration fence on
514 * the pages (not the object itself). As we don't track that,
515 * yet, we have to use the exclusive fence instead.
516 *
517 * Also note that we do not want to track the async vma as
518 * part of the obj->resv->excl_fence as it only affects
519 * execution and not content or object's backing store lifetime.
520 */
521 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
522 if (prev) {
523 __i915_sw_fence_await_dma_fence(&work->base.chain,
524 prev,
525 &work->cb);
526 dma_fence_put(prev);
527 }
528
529 work->base.dma.error = 0; /* enable the queue_work() */
530 work->obj = i915_gem_object_get(vma->obj);
531 } else {
532 ret = i915_gem_object_wait_moving_fence(vma->obj, true);
533 if (ret) {
534 i915_vma_resource_free(vma->resource);
535 vma->resource = NULL;
536
537 return ret;
538 }
539 vma->ops->bind_vma(vma->vm, NULL, vma->resource, cache_level,
540 bind_flags);
541 }
542
543 atomic_or(bind_flags, &vma->flags);
544 return 0;
545}
546
547void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
548{
549 void __iomem *ptr;
550 int err;
551
552 if (WARN_ON_ONCE(vma->obj->flags & I915_BO_ALLOC_GPU_ONLY))
553 return IOMEM_ERR_PTR(-EINVAL);
554
555 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
556 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
557 GEM_BUG_ON(i915_vma_verify_bind_complete(vma));
558
559 ptr = READ_ONCE(vma->iomap);
560 if (ptr == NULL) {
561 /*
562 * TODO: consider just using i915_gem_object_pin_map() for lmem
563 * instead, which already supports mapping non-contiguous chunks
564 * of pages, that way we can also drop the
565 * I915_BO_ALLOC_CONTIGUOUS when allocating the object.
566 */
567 if (i915_gem_object_is_lmem(vma->obj)) {
568 ptr = i915_gem_object_lmem_io_map(vma->obj, 0,
569 vma->obj->base.size);
570 } else if (i915_vma_is_map_and_fenceable(vma)) {
571 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
572 vma->node.start,
573 vma->node.size);
574 } else {
575 ptr = (void __iomem *)
576 i915_gem_object_pin_map(vma->obj, I915_MAP_WC);
577 if (IS_ERR(ptr)) {
578 err = PTR_ERR(ptr);
579 goto err;
580 }
581 ptr = page_pack_bits(ptr, 1);
582 }
583
584 if (ptr == NULL) {
585 err = -ENOMEM;
586 goto err;
587 }
588
589 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
590 if (page_unmask_bits(ptr))
591 __i915_gem_object_release_map(vma->obj);
592 else
593 io_mapping_unmap(ptr);
594 ptr = vma->iomap;
595 }
596 }
597
598 __i915_vma_pin(vma);
599
600 err = i915_vma_pin_fence(vma);
601 if (err)
602 goto err_unpin;
603
604 i915_vma_set_ggtt_write(vma);
605
606 /* NB Access through the GTT requires the device to be awake. */
607 return page_mask_bits(ptr);
608
609err_unpin:
610 __i915_vma_unpin(vma);
611err:
612 return IOMEM_ERR_PTR(err);
613}
614
615void i915_vma_flush_writes(struct i915_vma *vma)
616{
617 if (i915_vma_unset_ggtt_write(vma))
618 intel_gt_flush_ggtt_writes(vma->vm->gt);
619}
620
621void i915_vma_unpin_iomap(struct i915_vma *vma)
622{
623 GEM_BUG_ON(vma->iomap == NULL);
624
625 /* XXX We keep the mapping until __i915_vma_unbind()/evict() */
626
627 i915_vma_flush_writes(vma);
628
629 i915_vma_unpin_fence(vma);
630 i915_vma_unpin(vma);
631}
632
633void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
634{
635 struct i915_vma *vma;
636 struct drm_i915_gem_object *obj;
637
638 vma = fetch_and_zero(p_vma);
639 if (!vma)
640 return;
641
642 obj = vma->obj;
643 GEM_BUG_ON(!obj);
644
645 i915_vma_unpin(vma);
646
647 if (flags & I915_VMA_RELEASE_MAP)
648 i915_gem_object_unpin_map(obj);
649
650 i915_gem_object_put(obj);
651}
652
653bool i915_vma_misplaced(const struct i915_vma *vma,
654 u64 size, u64 alignment, u64 flags)
655{
656 if (!drm_mm_node_allocated(&vma->node))
657 return false;
658
659 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
660 return true;
661
662 if (vma->node.size < size)
663 return true;
664
665 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
666 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
667 return true;
668
669 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
670 return true;
671
672 if (flags & PIN_OFFSET_BIAS &&
673 vma->node.start < (flags & PIN_OFFSET_MASK))
674 return true;
675
676 if (flags & PIN_OFFSET_FIXED &&
677 vma->node.start != (flags & PIN_OFFSET_MASK))
678 return true;
679
680 return false;
681}
682
683void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
684{
685 bool mappable, fenceable;
686
687 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
688 GEM_BUG_ON(!vma->fence_size);
689
690 fenceable = (vma->node.size >= vma->fence_size &&
691 IS_ALIGNED(vma->node.start, vma->fence_alignment));
692
693 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
694
695 if (mappable && fenceable)
696 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
697 else
698 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
699}
700
701bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
702{
703 struct drm_mm_node *node = &vma->node;
704 struct drm_mm_node *other;
705
706 /*
707 * On some machines we have to be careful when putting differing types
708 * of snoopable memory together to avoid the prefetcher crossing memory
709 * domains and dying. During vm initialisation, we decide whether or not
710 * these constraints apply and set the drm_mm.color_adjust
711 * appropriately.
712 */
713 if (!i915_vm_has_cache_coloring(vma->vm))
714 return true;
715
716 /* Only valid to be called on an already inserted vma */
717 GEM_BUG_ON(!drm_mm_node_allocated(node));
718 GEM_BUG_ON(list_empty(&node->node_list));
719
720 other = list_prev_entry(node, node_list);
721 if (i915_node_color_differs(other, color) &&
722 !drm_mm_hole_follows(other))
723 return false;
724
725 other = list_next_entry(node, node_list);
726 if (i915_node_color_differs(other, color) &&
727 !drm_mm_hole_follows(node))
728 return false;
729
730 return true;
731}
732
733/**
734 * i915_vma_insert - finds a slot for the vma in its address space
735 * @vma: the vma
736 * @size: requested size in bytes (can be larger than the VMA)
737 * @alignment: required alignment
738 * @flags: mask of PIN_* flags to use
739 *
740 * First we try to allocate some free space that meets the requirements for
741 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
742 * preferrably the oldest idle entry to make room for the new VMA.
743 *
744 * Returns:
745 * 0 on success, negative error code otherwise.
746 */
747static int
748i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
749 u64 size, u64 alignment, u64 flags)
750{
751 unsigned long color;
752 u64 start, end;
753 int ret;
754
755 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
756 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
757
758 size = max(size, vma->size);
759 alignment = max(alignment, vma->display_alignment);
760 if (flags & PIN_MAPPABLE) {
761 size = max_t(typeof(size), size, vma->fence_size);
762 alignment = max_t(typeof(alignment),
763 alignment, vma->fence_alignment);
764 }
765
766 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
767 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
768 GEM_BUG_ON(!is_power_of_2(alignment));
769
770 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
771 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
772
773 end = vma->vm->total;
774 if (flags & PIN_MAPPABLE)
775 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
776 if (flags & PIN_ZONE_4G)
777 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
778 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
779
780 alignment = max(alignment, i915_vm_obj_min_alignment(vma->vm, vma->obj));
781
782 /* If binding the object/GGTT view requires more space than the entire
783 * aperture has, reject it early before evicting everything in a vain
784 * attempt to find space.
785 */
786 if (size > end) {
787 drm_dbg(&to_i915(vma->obj->base.dev)->drm,
788 "Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
789 size, flags & PIN_MAPPABLE ? "mappable" : "total", end);
790 return -ENOSPC;
791 }
792
793 color = 0;
794
795 if (i915_vm_has_cache_coloring(vma->vm))
796 color = vma->obj->cache_level;
797
798 if (flags & PIN_OFFSET_FIXED) {
799 u64 offset = flags & PIN_OFFSET_MASK;
800 if (!IS_ALIGNED(offset, alignment) ||
801 range_overflows(offset, size, end))
802 return -EINVAL;
803
804 ret = i915_gem_gtt_reserve(vma->vm, ww, &vma->node,
805 size, offset, color,
806 flags);
807 if (ret)
808 return ret;
809 } else {
810 /*
811 * We only support huge gtt pages through the 48b PPGTT,
812 * however we also don't want to force any alignment for
813 * objects which need to be tightly packed into the low 32bits.
814 *
815 * Note that we assume that GGTT are limited to 4GiB for the
816 * forseeable future. See also i915_ggtt_offset().
817 */
818 if (upper_32_bits(end - 1) &&
819 vma->page_sizes.sg > I915_GTT_PAGE_SIZE &&
820 !HAS_64K_PAGES(vma->vm->i915)) {
821 /*
822 * We can't mix 64K and 4K PTEs in the same page-table
823 * (2M block), and so to avoid the ugliness and
824 * complexity of coloring we opt for just aligning 64K
825 * objects to 2M.
826 */
827 u64 page_alignment =
828 rounddown_pow_of_two(vma->page_sizes.sg |
829 I915_GTT_PAGE_SIZE_2M);
830
831 /*
832 * Check we don't expand for the limited Global GTT
833 * (mappable aperture is even more precious!). This
834 * also checks that we exclude the aliasing-ppgtt.
835 */
836 GEM_BUG_ON(i915_vma_is_ggtt(vma));
837
838 alignment = max(alignment, page_alignment);
839
840 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
841 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
842 }
843
844 ret = i915_gem_gtt_insert(vma->vm, ww, &vma->node,
845 size, alignment, color,
846 start, end, flags);
847 if (ret)
848 return ret;
849
850 GEM_BUG_ON(vma->node.start < start);
851 GEM_BUG_ON(vma->node.start + vma->node.size > end);
852 }
853 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
854 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
855
856 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
857
858 return 0;
859}
860
861static void
862i915_vma_detach(struct i915_vma *vma)
863{
864 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
865 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
866
867 /*
868 * And finally now the object is completely decoupled from this
869 * vma, we can drop its hold on the backing storage and allow
870 * it to be reaped by the shrinker.
871 */
872 list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
873}
874
875static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
876{
877 unsigned int bound;
878
879 bound = atomic_read(&vma->flags);
880
881 if (flags & PIN_VALIDATE) {
882 flags &= I915_VMA_BIND_MASK;
883
884 return (flags & bound) == flags;
885 }
886
887 /* with the lock mandatory for unbind, we don't race here */
888 flags &= I915_VMA_BIND_MASK;
889 do {
890 if (unlikely(flags & ~bound))
891 return false;
892
893 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
894 return false;
895
896 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
897 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
898
899 return true;
900}
901
902static struct scatterlist *
903rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
904 unsigned int width, unsigned int height,
905 unsigned int src_stride, unsigned int dst_stride,
906 struct sg_table *st, struct scatterlist *sg)
907{
908 unsigned int column, row;
909 unsigned int src_idx;
910
911 for (column = 0; column < width; column++) {
912 unsigned int left;
913
914 src_idx = src_stride * (height - 1) + column + offset;
915 for (row = 0; row < height; row++) {
916 st->nents++;
917 /*
918 * We don't need the pages, but need to initialize
919 * the entries so the sg list can be happily traversed.
920 * The only thing we need are DMA addresses.
921 */
922 sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);
923 sg_dma_address(sg) =
924 i915_gem_object_get_dma_address(obj, src_idx);
925 sg_dma_len(sg) = I915_GTT_PAGE_SIZE;
926 sg = sg_next(sg);
927 src_idx -= src_stride;
928 }
929
930 left = (dst_stride - height) * I915_GTT_PAGE_SIZE;
931
932 if (!left)
933 continue;
934
935 st->nents++;
936
937 /*
938 * The DE ignores the PTEs for the padding tiles, the sg entry
939 * here is just a conenience to indicate how many padding PTEs
940 * to insert at this spot.
941 */
942 sg_set_page(sg, NULL, left, 0);
943 sg_dma_address(sg) = 0;
944 sg_dma_len(sg) = left;
945 sg = sg_next(sg);
946 }
947
948 return sg;
949}
950
951static noinline struct sg_table *
952intel_rotate_pages(struct intel_rotation_info *rot_info,
953 struct drm_i915_gem_object *obj)
954{
955 unsigned int size = intel_rotation_info_size(rot_info);
956 struct drm_i915_private *i915 = to_i915(obj->base.dev);
957 struct sg_table *st;
958 struct scatterlist *sg;
959 int ret = -ENOMEM;
960 int i;
961
962 /* Allocate target SG list. */
963 st = kmalloc(sizeof(*st), GFP_KERNEL);
964 if (!st)
965 goto err_st_alloc;
966
967 ret = sg_alloc_table(st, size, GFP_KERNEL);
968 if (ret)
969 goto err_sg_alloc;
970
971 st->nents = 0;
972 sg = st->sgl;
973
974 for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++)
975 sg = rotate_pages(obj, rot_info->plane[i].offset,
976 rot_info->plane[i].width, rot_info->plane[i].height,
977 rot_info->plane[i].src_stride,
978 rot_info->plane[i].dst_stride,
979 st, sg);
980
981 return st;
982
983err_sg_alloc:
984 kfree(st);
985err_st_alloc:
986
987 drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",
988 obj->base.size, rot_info->plane[0].width,
989 rot_info->plane[0].height, size);
990
991 return ERR_PTR(ret);
992}
993
994static struct scatterlist *
995add_padding_pages(unsigned int count,
996 struct sg_table *st, struct scatterlist *sg)
997{
998 st->nents++;
999
1000 /*
1001 * The DE ignores the PTEs for the padding tiles, the sg entry
1002 * here is just a convenience to indicate how many padding PTEs
1003 * to insert at this spot.
1004 */
1005 sg_set_page(sg, NULL, count * I915_GTT_PAGE_SIZE, 0);
1006 sg_dma_address(sg) = 0;
1007 sg_dma_len(sg) = count * I915_GTT_PAGE_SIZE;
1008 sg = sg_next(sg);
1009
1010 return sg;
1011}
1012
1013static struct scatterlist *
1014remap_tiled_color_plane_pages(struct drm_i915_gem_object *obj,
1015 unsigned int offset, unsigned int alignment_pad,
1016 unsigned int width, unsigned int height,
1017 unsigned int src_stride, unsigned int dst_stride,
1018 struct sg_table *st, struct scatterlist *sg,
1019 unsigned int *gtt_offset)
1020{
1021 unsigned int row;
1022
1023 if (!width || !height)
1024 return sg;
1025
1026 if (alignment_pad)
1027 sg = add_padding_pages(alignment_pad, st, sg);
1028
1029 for (row = 0; row < height; row++) {
1030 unsigned int left = width * I915_GTT_PAGE_SIZE;
1031
1032 while (left) {
1033 dma_addr_t addr;
1034 unsigned int length;
1035
1036 /*
1037 * We don't need the pages, but need to initialize
1038 * the entries so the sg list can be happily traversed.
1039 * The only thing we need are DMA addresses.
1040 */
1041
1042 addr = i915_gem_object_get_dma_address_len(obj, offset, &length);
1043
1044 length = min(left, length);
1045
1046 st->nents++;
1047
1048 sg_set_page(sg, NULL, length, 0);
1049 sg_dma_address(sg) = addr;
1050 sg_dma_len(sg) = length;
1051 sg = sg_next(sg);
1052
1053 offset += length / I915_GTT_PAGE_SIZE;
1054 left -= length;
1055 }
1056
1057 offset += src_stride - width;
1058
1059 left = (dst_stride - width) * I915_GTT_PAGE_SIZE;
1060
1061 if (!left)
1062 continue;
1063
1064 sg = add_padding_pages(left >> PAGE_SHIFT, st, sg);
1065 }
1066
1067 *gtt_offset += alignment_pad + dst_stride * height;
1068
1069 return sg;
1070}
1071
1072static struct scatterlist *
1073remap_contiguous_pages(struct drm_i915_gem_object *obj,
1074 unsigned int obj_offset,
1075 unsigned int count,
1076 struct sg_table *st, struct scatterlist *sg)
1077{
1078 struct scatterlist *iter;
1079 unsigned int offset;
1080
1081 iter = i915_gem_object_get_sg_dma(obj, obj_offset, &offset);
1082 GEM_BUG_ON(!iter);
1083
1084 do {
1085 unsigned int len;
1086
1087 len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),
1088 count << PAGE_SHIFT);
1089 sg_set_page(sg, NULL, len, 0);
1090 sg_dma_address(sg) =
1091 sg_dma_address(iter) + (offset << PAGE_SHIFT);
1092 sg_dma_len(sg) = len;
1093
1094 st->nents++;
1095 count -= len >> PAGE_SHIFT;
1096 if (count == 0)
1097 return sg;
1098
1099 sg = __sg_next(sg);
1100 iter = __sg_next(iter);
1101 offset = 0;
1102 } while (1);
1103}
1104
1105static struct scatterlist *
1106remap_linear_color_plane_pages(struct drm_i915_gem_object *obj,
1107 unsigned int obj_offset, unsigned int alignment_pad,
1108 unsigned int size,
1109 struct sg_table *st, struct scatterlist *sg,
1110 unsigned int *gtt_offset)
1111{
1112 if (!size)
1113 return sg;
1114
1115 if (alignment_pad)
1116 sg = add_padding_pages(alignment_pad, st, sg);
1117
1118 sg = remap_contiguous_pages(obj, obj_offset, size, st, sg);
1119 sg = sg_next(sg);
1120
1121 *gtt_offset += alignment_pad + size;
1122
1123 return sg;
1124}
1125
1126static struct scatterlist *
1127remap_color_plane_pages(const struct intel_remapped_info *rem_info,
1128 struct drm_i915_gem_object *obj,
1129 int color_plane,
1130 struct sg_table *st, struct scatterlist *sg,
1131 unsigned int *gtt_offset)
1132{
1133 unsigned int alignment_pad = 0;
1134
1135 if (rem_info->plane_alignment)
1136 alignment_pad = ALIGN(*gtt_offset, rem_info->plane_alignment) - *gtt_offset;
1137
1138 if (rem_info->plane[color_plane].linear)
1139 sg = remap_linear_color_plane_pages(obj,
1140 rem_info->plane[color_plane].offset,
1141 alignment_pad,
1142 rem_info->plane[color_plane].size,
1143 st, sg,
1144 gtt_offset);
1145
1146 else
1147 sg = remap_tiled_color_plane_pages(obj,
1148 rem_info->plane[color_plane].offset,
1149 alignment_pad,
1150 rem_info->plane[color_plane].width,
1151 rem_info->plane[color_plane].height,
1152 rem_info->plane[color_plane].src_stride,
1153 rem_info->plane[color_plane].dst_stride,
1154 st, sg,
1155 gtt_offset);
1156
1157 return sg;
1158}
1159
1160static noinline struct sg_table *
1161intel_remap_pages(struct intel_remapped_info *rem_info,
1162 struct drm_i915_gem_object *obj)
1163{
1164 unsigned int size = intel_remapped_info_size(rem_info);
1165 struct drm_i915_private *i915 = to_i915(obj->base.dev);
1166 struct sg_table *st;
1167 struct scatterlist *sg;
1168 unsigned int gtt_offset = 0;
1169 int ret = -ENOMEM;
1170 int i;
1171
1172 /* Allocate target SG list. */
1173 st = kmalloc(sizeof(*st), GFP_KERNEL);
1174 if (!st)
1175 goto err_st_alloc;
1176
1177 ret = sg_alloc_table(st, size, GFP_KERNEL);
1178 if (ret)
1179 goto err_sg_alloc;
1180
1181 st->nents = 0;
1182 sg = st->sgl;
1183
1184 for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++)
1185 sg = remap_color_plane_pages(rem_info, obj, i, st, sg, >t_offset);
1186
1187 i915_sg_trim(st);
1188
1189 return st;
1190
1191err_sg_alloc:
1192 kfree(st);
1193err_st_alloc:
1194
1195 drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",
1196 obj->base.size, rem_info->plane[0].width,
1197 rem_info->plane[0].height, size);
1198
1199 return ERR_PTR(ret);
1200}
1201
1202static noinline struct sg_table *
1203intel_partial_pages(const struct i915_gtt_view *view,
1204 struct drm_i915_gem_object *obj)
1205{
1206 struct sg_table *st;
1207 struct scatterlist *sg;
1208 unsigned int count = view->partial.size;
1209 int ret = -ENOMEM;
1210
1211 st = kmalloc(sizeof(*st), GFP_KERNEL);
1212 if (!st)
1213 goto err_st_alloc;
1214
1215 ret = sg_alloc_table(st, count, GFP_KERNEL);
1216 if (ret)
1217 goto err_sg_alloc;
1218
1219 st->nents = 0;
1220
1221 sg = remap_contiguous_pages(obj, view->partial.offset, count, st, st->sgl);
1222
1223 sg_mark_end(sg);
1224 i915_sg_trim(st); /* Drop any unused tail entries. */
1225
1226 return st;
1227
1228err_sg_alloc:
1229 kfree(st);
1230err_st_alloc:
1231 return ERR_PTR(ret);
1232}
1233
1234static int
1235__i915_vma_get_pages(struct i915_vma *vma)
1236{
1237 struct sg_table *pages;
1238
1239 /*
1240 * The vma->pages are only valid within the lifespan of the borrowed
1241 * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so
1242 * must be the vma->pages. A simple rule is that vma->pages must only
1243 * be accessed when the obj->mm.pages are pinned.
1244 */
1245 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));
1246
1247 switch (vma->gtt_view.type) {
1248 default:
1249 GEM_BUG_ON(vma->gtt_view.type);
1250 fallthrough;
1251 case I915_GTT_VIEW_NORMAL:
1252 pages = vma->obj->mm.pages;
1253 break;
1254
1255 case I915_GTT_VIEW_ROTATED:
1256 pages =
1257 intel_rotate_pages(&vma->gtt_view.rotated, vma->obj);
1258 break;
1259
1260 case I915_GTT_VIEW_REMAPPED:
1261 pages =
1262 intel_remap_pages(&vma->gtt_view.remapped, vma->obj);
1263 break;
1264
1265 case I915_GTT_VIEW_PARTIAL:
1266 pages = intel_partial_pages(&vma->gtt_view, vma->obj);
1267 break;
1268 }
1269
1270 if (IS_ERR(pages)) {
1271 drm_err(&vma->vm->i915->drm,
1272 "Failed to get pages for VMA view type %u (%ld)!\n",
1273 vma->gtt_view.type, PTR_ERR(pages));
1274 return PTR_ERR(pages);
1275 }
1276
1277 vma->pages = pages;
1278
1279 return 0;
1280}
1281
1282I915_SELFTEST_EXPORT int i915_vma_get_pages(struct i915_vma *vma)
1283{
1284 int err;
1285
1286 if (atomic_add_unless(&vma->pages_count, 1, 0))
1287 return 0;
1288
1289 err = i915_gem_object_pin_pages(vma->obj);
1290 if (err)
1291 return err;
1292
1293 err = __i915_vma_get_pages(vma);
1294 if (err)
1295 goto err_unpin;
1296
1297 vma->page_sizes = vma->obj->mm.page_sizes;
1298 atomic_inc(&vma->pages_count);
1299
1300 return 0;
1301
1302err_unpin:
1303 __i915_gem_object_unpin_pages(vma->obj);
1304
1305 return err;
1306}
1307
1308void vma_invalidate_tlb(struct i915_address_space *vm, u32 *tlb)
1309{
1310 /*
1311 * Before we release the pages that were bound by this vma, we
1312 * must invalidate all the TLBs that may still have a reference
1313 * back to our physical address. It only needs to be done once,
1314 * so after updating the PTE to point away from the pages, record
1315 * the most recent TLB invalidation seqno, and if we have not yet
1316 * flushed the TLBs upon release, perform a full invalidation.
1317 */
1318 WRITE_ONCE(*tlb, intel_gt_next_invalidate_tlb_full(vm->gt));
1319}
1320
1321static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
1322{
1323 /* We allocate under vma_get_pages, so beware the shrinker */
1324 GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
1325
1326 if (atomic_sub_return(count, &vma->pages_count) == 0) {
1327 if (vma->pages != vma->obj->mm.pages) {
1328 sg_free_table(vma->pages);
1329 kfree(vma->pages);
1330 }
1331 vma->pages = NULL;
1332
1333 i915_gem_object_unpin_pages(vma->obj);
1334 }
1335}
1336
1337I915_SELFTEST_EXPORT void i915_vma_put_pages(struct i915_vma *vma)
1338{
1339 if (atomic_add_unless(&vma->pages_count, -1, 1))
1340 return;
1341
1342 __vma_put_pages(vma, 1);
1343}
1344
1345static void vma_unbind_pages(struct i915_vma *vma)
1346{
1347 unsigned int count;
1348
1349 lockdep_assert_held(&vma->vm->mutex);
1350
1351 /* The upper portion of pages_count is the number of bindings */
1352 count = atomic_read(&vma->pages_count);
1353 count >>= I915_VMA_PAGES_BIAS;
1354 GEM_BUG_ON(!count);
1355
1356 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
1357}
1358
1359int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1360 u64 size, u64 alignment, u64 flags)
1361{
1362 struct i915_vma_work *work = NULL;
1363 struct dma_fence *moving = NULL;
1364 struct i915_vma_resource *vma_res = NULL;
1365 intel_wakeref_t wakeref = 0;
1366 unsigned int bound;
1367 int err;
1368
1369 assert_vma_held(vma);
1370 GEM_BUG_ON(!ww);
1371
1372 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
1373 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
1374
1375 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
1376
1377 /* First try and grab the pin without rebinding the vma */
1378 if (try_qad_pin(vma, flags))
1379 return 0;
1380
1381 err = i915_vma_get_pages(vma);
1382 if (err)
1383 return err;
1384
1385 if (flags & PIN_GLOBAL)
1386 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
1387
1388 if (flags & vma->vm->bind_async_flags) {
1389 /* lock VM */
1390 err = i915_vm_lock_objects(vma->vm, ww);
1391 if (err)
1392 goto err_rpm;
1393
1394 work = i915_vma_work();
1395 if (!work) {
1396 err = -ENOMEM;
1397 goto err_rpm;
1398 }
1399
1400 work->vm = vma->vm;
1401
1402 err = i915_gem_object_get_moving_fence(vma->obj, &moving);
1403 if (err)
1404 goto err_rpm;
1405
1406 dma_fence_work_chain(&work->base, moving);
1407
1408 /* Allocate enough page directories to used PTE */
1409 if (vma->vm->allocate_va_range) {
1410 err = i915_vm_alloc_pt_stash(vma->vm,
1411 &work->stash,
1412 vma->size);
1413 if (err)
1414 goto err_fence;
1415
1416 err = i915_vm_map_pt_stash(vma->vm, &work->stash);
1417 if (err)
1418 goto err_fence;
1419 }
1420 }
1421
1422 vma_res = i915_vma_resource_alloc();
1423 if (IS_ERR(vma_res)) {
1424 err = PTR_ERR(vma_res);
1425 goto err_fence;
1426 }
1427
1428 /*
1429 * Differentiate between user/kernel vma inside the aliasing-ppgtt.
1430 *
1431 * We conflate the Global GTT with the user's vma when using the
1432 * aliasing-ppgtt, but it is still vitally important to try and
1433 * keep the use cases distinct. For example, userptr objects are
1434 * not allowed inside the Global GTT as that will cause lock
1435 * inversions when we have to evict them the mmu_notifier callbacks -
1436 * but they are allowed to be part of the user ppGTT which can never
1437 * be mapped. As such we try to give the distinct users of the same
1438 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
1439 * and i915_ppgtt separate].
1440 *
1441 * NB this may cause us to mask real lock inversions -- while the
1442 * code is safe today, lockdep may not be able to spot future
1443 * transgressions.
1444 */
1445 err = mutex_lock_interruptible_nested(&vma->vm->mutex,
1446 !(flags & PIN_GLOBAL));
1447 if (err)
1448 goto err_vma_res;
1449
1450 /* No more allocations allowed now we hold vm->mutex */
1451
1452 if (unlikely(i915_vma_is_closed(vma))) {
1453 err = -ENOENT;
1454 goto err_unlock;
1455 }
1456
1457 bound = atomic_read(&vma->flags);
1458 if (unlikely(bound & I915_VMA_ERROR)) {
1459 err = -ENOMEM;
1460 goto err_unlock;
1461 }
1462
1463 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
1464 err = -EAGAIN; /* pins are meant to be fairly temporary */
1465 goto err_unlock;
1466 }
1467
1468 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
1469 if (!(flags & PIN_VALIDATE))
1470 __i915_vma_pin(vma);
1471 goto err_unlock;
1472 }
1473
1474 err = i915_active_acquire(&vma->active);
1475 if (err)
1476 goto err_unlock;
1477
1478 if (!(bound & I915_VMA_BIND_MASK)) {
1479 err = i915_vma_insert(vma, ww, size, alignment, flags);
1480 if (err)
1481 goto err_active;
1482
1483 if (i915_is_ggtt(vma->vm))
1484 __i915_vma_set_map_and_fenceable(vma);
1485 }
1486
1487 GEM_BUG_ON(!vma->pages);
1488 err = i915_vma_bind(vma,
1489 vma->obj->cache_level,
1490 flags, work, vma_res);
1491 vma_res = NULL;
1492 if (err)
1493 goto err_remove;
1494
1495 /* There should only be at most 2 active bindings (user, global) */
1496 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
1497 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
1498 list_move_tail(&vma->vm_link, &vma->vm->bound_list);
1499
1500 if (!(flags & PIN_VALIDATE)) {
1501 __i915_vma_pin(vma);
1502 GEM_BUG_ON(!i915_vma_is_pinned(vma));
1503 }
1504 GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
1505 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
1506
1507err_remove:
1508 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
1509 i915_vma_detach(vma);
1510 drm_mm_remove_node(&vma->node);
1511 }
1512err_active:
1513 i915_active_release(&vma->active);
1514err_unlock:
1515 mutex_unlock(&vma->vm->mutex);
1516err_vma_res:
1517 i915_vma_resource_free(vma_res);
1518err_fence:
1519 if (work)
1520 dma_fence_work_commit_imm(&work->base);
1521err_rpm:
1522 if (wakeref)
1523 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
1524
1525 if (moving)
1526 dma_fence_put(moving);
1527
1528 i915_vma_put_pages(vma);
1529 return err;
1530}
1531
1532static void flush_idle_contexts(struct intel_gt *gt)
1533{
1534 struct intel_engine_cs *engine;
1535 enum intel_engine_id id;
1536
1537 for_each_engine(engine, gt, id)
1538 intel_engine_flush_barriers(engine);
1539
1540 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1541}
1542
1543static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1544 u32 align, unsigned int flags)
1545{
1546 struct i915_address_space *vm = vma->vm;
1547 int err;
1548
1549 do {
1550 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);
1551
1552 if (err != -ENOSPC) {
1553 if (!err) {
1554 err = i915_vma_wait_for_bind(vma);
1555 if (err)
1556 i915_vma_unpin(vma);
1557 }
1558 return err;
1559 }
1560
1561 /* Unlike i915_vma_pin, we don't take no for an answer! */
1562 flush_idle_contexts(vm->gt);
1563 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1564 /*
1565 * We pass NULL ww here, as we don't want to unbind
1566 * locked objects when called from execbuf when pinning
1567 * is removed. This would probably regress badly.
1568 */
1569 i915_gem_evict_vm(vm, NULL, NULL);
1570 mutex_unlock(&vm->mutex);
1571 }
1572 } while (1);
1573}
1574
1575int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1576 u32 align, unsigned int flags)
1577{
1578 struct i915_gem_ww_ctx _ww;
1579 int err;
1580
1581 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1582
1583 if (ww)
1584 return __i915_ggtt_pin(vma, ww, align, flags);
1585
1586 lockdep_assert_not_held(&vma->obj->base.resv->lock.base);
1587
1588 for_i915_gem_ww(&_ww, err, true) {
1589 err = i915_gem_object_lock(vma->obj, &_ww);
1590 if (!err)
1591 err = __i915_ggtt_pin(vma, &_ww, align, flags);
1592 }
1593
1594 return err;
1595}
1596
1597static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1598{
1599 /*
1600 * We defer actually closing, unbinding and destroying the VMA until
1601 * the next idle point, or if the object is freed in the meantime. By
1602 * postponing the unbind, we allow for it to be resurrected by the
1603 * client, avoiding the work required to rebind the VMA. This is
1604 * advantageous for DRI, where the client/server pass objects
1605 * between themselves, temporarily opening a local VMA to the
1606 * object, and then closing it again. The same object is then reused
1607 * on the next frame (or two, depending on the depth of the swap queue)
1608 * causing us to rebind the VMA once more. This ends up being a lot
1609 * of wasted work for the steady state.
1610 */
1611 GEM_BUG_ON(i915_vma_is_closed(vma));
1612 list_add(&vma->closed_link, >->closed_vma);
1613}
1614
1615void i915_vma_close(struct i915_vma *vma)
1616{
1617 struct intel_gt *gt = vma->vm->gt;
1618 unsigned long flags;
1619
1620 if (i915_vma_is_ggtt(vma))
1621 return;
1622
1623 GEM_BUG_ON(!atomic_read(&vma->open_count));
1624 if (atomic_dec_and_lock_irqsave(&vma->open_count,
1625 >->closed_lock,
1626 flags)) {
1627 __vma_close(vma, gt);
1628 spin_unlock_irqrestore(>->closed_lock, flags);
1629 }
1630}
1631
1632static void __i915_vma_remove_closed(struct i915_vma *vma)
1633{
1634 list_del_init(&vma->closed_link);
1635}
1636
1637void i915_vma_reopen(struct i915_vma *vma)
1638{
1639 struct intel_gt *gt = vma->vm->gt;
1640
1641 spin_lock_irq(>->closed_lock);
1642 if (i915_vma_is_closed(vma))
1643 __i915_vma_remove_closed(vma);
1644 spin_unlock_irq(>->closed_lock);
1645}
1646
1647static void force_unbind(struct i915_vma *vma)
1648{
1649 if (!drm_mm_node_allocated(&vma->node))
1650 return;
1651
1652 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1653 WARN_ON(__i915_vma_unbind(vma));
1654 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1655}
1656
1657static void release_references(struct i915_vma *vma, struct intel_gt *gt,
1658 bool vm_ddestroy)
1659{
1660 struct drm_i915_gem_object *obj = vma->obj;
1661
1662 GEM_BUG_ON(i915_vma_is_active(vma));
1663
1664 spin_lock(&obj->vma.lock);
1665 list_del(&vma->obj_link);
1666 if (!RB_EMPTY_NODE(&vma->obj_node))
1667 rb_erase(&vma->obj_node, &obj->vma.tree);
1668
1669 spin_unlock(&obj->vma.lock);
1670
1671 spin_lock_irq(>->closed_lock);
1672 __i915_vma_remove_closed(vma);
1673 spin_unlock_irq(>->closed_lock);
1674
1675 if (vm_ddestroy)
1676 i915_vm_resv_put(vma->vm);
1677
1678 i915_active_fini(&vma->active);
1679 GEM_WARN_ON(vma->resource);
1680 i915_vma_free(vma);
1681}
1682
1683/**
1684 * i915_vma_destroy_locked - Remove all weak reference to the vma and put
1685 * the initial reference.
1686 *
1687 * This function should be called when it's decided the vma isn't needed
1688 * anymore. The caller must assure that it doesn't race with another lookup
1689 * plus destroy, typically by taking an appropriate reference.
1690 *
1691 * Current callsites are
1692 * - __i915_gem_object_pages_fini()
1693 * - __i915_vm_close() - Blocks the above function by taking a reference on
1694 * the object.
1695 * - __i915_vma_parked() - Blocks the above functions by taking a reference
1696 * on the vm and a reference on the object. Also takes the object lock so
1697 * destruction from __i915_vma_parked() can be blocked by holding the
1698 * object lock. Since the object lock is only allowed from within i915 with
1699 * an object refcount, holding the object lock also implicitly blocks the
1700 * vma freeing from __i915_gem_object_pages_fini().
1701 *
1702 * Because of locks taken during destruction, a vma is also guaranteed to
1703 * stay alive while the following locks are held if it was looked up while
1704 * holding one of the locks:
1705 * - vm->mutex
1706 * - obj->vma.lock
1707 * - gt->closed_lock
1708 */
1709void i915_vma_destroy_locked(struct i915_vma *vma)
1710{
1711 lockdep_assert_held(&vma->vm->mutex);
1712
1713 force_unbind(vma);
1714 list_del_init(&vma->vm_link);
1715 release_references(vma, vma->vm->gt, false);
1716}
1717
1718void i915_vma_destroy(struct i915_vma *vma)
1719{
1720 struct intel_gt *gt;
1721 bool vm_ddestroy;
1722
1723 mutex_lock(&vma->vm->mutex);
1724 force_unbind(vma);
1725 list_del_init(&vma->vm_link);
1726 vm_ddestroy = vma->vm_ddestroy;
1727 vma->vm_ddestroy = false;
1728
1729 /* vma->vm may be freed when releasing vma->vm->mutex. */
1730 gt = vma->vm->gt;
1731 mutex_unlock(&vma->vm->mutex);
1732 release_references(vma, gt, vm_ddestroy);
1733}
1734
1735void i915_vma_parked(struct intel_gt *gt)
1736{
1737 struct i915_vma *vma, *next;
1738 LIST_HEAD(closed);
1739
1740 spin_lock_irq(>->closed_lock);
1741 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
1742 struct drm_i915_gem_object *obj = vma->obj;
1743 struct i915_address_space *vm = vma->vm;
1744
1745 /* XXX All to avoid keeping a reference on i915_vma itself */
1746
1747 if (!kref_get_unless_zero(&obj->base.refcount))
1748 continue;
1749
1750 if (!i915_vm_tryget(vm)) {
1751 i915_gem_object_put(obj);
1752 continue;
1753 }
1754
1755 list_move(&vma->closed_link, &closed);
1756 }
1757 spin_unlock_irq(>->closed_lock);
1758
1759 /* As the GT is held idle, no vma can be reopened as we destroy them */
1760 list_for_each_entry_safe(vma, next, &closed, closed_link) {
1761 struct drm_i915_gem_object *obj = vma->obj;
1762 struct i915_address_space *vm = vma->vm;
1763
1764 if (i915_gem_object_trylock(obj, NULL)) {
1765 INIT_LIST_HEAD(&vma->closed_link);
1766 i915_vma_destroy(vma);
1767 i915_gem_object_unlock(obj);
1768 } else {
1769 /* back you go.. */
1770 spin_lock_irq(>->closed_lock);
1771 list_add(&vma->closed_link, >->closed_vma);
1772 spin_unlock_irq(>->closed_lock);
1773 }
1774
1775 i915_gem_object_put(obj);
1776 i915_vm_put(vm);
1777 }
1778}
1779
1780static void __i915_vma_iounmap(struct i915_vma *vma)
1781{
1782 GEM_BUG_ON(i915_vma_is_pinned(vma));
1783
1784 if (vma->iomap == NULL)
1785 return;
1786
1787 if (page_unmask_bits(vma->iomap))
1788 __i915_gem_object_release_map(vma->obj);
1789 else
1790 io_mapping_unmap(vma->iomap);
1791 vma->iomap = NULL;
1792}
1793
1794void i915_vma_revoke_mmap(struct i915_vma *vma)
1795{
1796 struct drm_vma_offset_node *node;
1797 u64 vma_offset;
1798
1799 if (!i915_vma_has_userfault(vma))
1800 return;
1801
1802 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1803 GEM_BUG_ON(!vma->obj->userfault_count);
1804
1805 node = &vma->mmo->vma_node;
1806 vma_offset = vma->gtt_view.partial.offset << PAGE_SHIFT;
1807 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1808 drm_vma_node_offset_addr(node) + vma_offset,
1809 vma->size,
1810 1);
1811
1812 i915_vma_unset_userfault(vma);
1813 if (!--vma->obj->userfault_count)
1814 list_del(&vma->obj->userfault_link);
1815}
1816
1817static int
1818__i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma)
1819{
1820 return __i915_request_await_exclusive(rq, &vma->active);
1821}
1822
1823static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1824{
1825 int err;
1826
1827 /* Wait for the vma to be bound before we start! */
1828 err = __i915_request_await_bind(rq, vma);
1829 if (err)
1830 return err;
1831
1832 return i915_active_add_request(&vma->active, rq);
1833}
1834
1835int _i915_vma_move_to_active(struct i915_vma *vma,
1836 struct i915_request *rq,
1837 struct dma_fence *fence,
1838 unsigned int flags)
1839{
1840 struct drm_i915_gem_object *obj = vma->obj;
1841 int err;
1842
1843 assert_object_held(obj);
1844
1845 GEM_BUG_ON(!vma->pages);
1846
1847 if (!(flags & __EXEC_OBJECT_NO_REQUEST_AWAIT)) {
1848 err = i915_request_await_object(rq, vma->obj, flags & EXEC_OBJECT_WRITE);
1849 if (unlikely(err))
1850 return err;
1851 }
1852 err = __i915_vma_move_to_active(vma, rq);
1853 if (unlikely(err))
1854 return err;
1855
1856 /*
1857 * Reserve fences slot early to prevent an allocation after preparing
1858 * the workload and associating fences with dma_resv.
1859 */
1860 if (fence && !(flags & __EXEC_OBJECT_NO_RESERVE)) {
1861 struct dma_fence *curr;
1862 int idx;
1863
1864 dma_fence_array_for_each(curr, idx, fence)
1865 ;
1866 err = dma_resv_reserve_fences(vma->obj->base.resv, idx);
1867 if (unlikely(err))
1868 return err;
1869 }
1870
1871 if (flags & EXEC_OBJECT_WRITE) {
1872 struct intel_frontbuffer *front;
1873
1874 front = __intel_frontbuffer_get(obj);
1875 if (unlikely(front)) {
1876 if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1877 i915_active_add_request(&front->write, rq);
1878 intel_frontbuffer_put(front);
1879 }
1880 }
1881
1882 if (fence) {
1883 struct dma_fence *curr;
1884 enum dma_resv_usage usage;
1885 int idx;
1886
1887 if (flags & EXEC_OBJECT_WRITE) {
1888 usage = DMA_RESV_USAGE_WRITE;
1889 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1890 obj->read_domains = 0;
1891 } else {
1892 usage = DMA_RESV_USAGE_READ;
1893 obj->write_domain = 0;
1894 }
1895
1896 dma_fence_array_for_each(curr, idx, fence)
1897 dma_resv_add_fence(vma->obj->base.resv, curr, usage);
1898 }
1899
1900 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1901 i915_active_add_request(&vma->fence->active, rq);
1902
1903 obj->read_domains |= I915_GEM_GPU_DOMAINS;
1904 obj->mm.dirty = true;
1905
1906 GEM_BUG_ON(!i915_vma_is_active(vma));
1907 return 0;
1908}
1909
1910struct dma_fence *__i915_vma_evict(struct i915_vma *vma, bool async)
1911{
1912 struct i915_vma_resource *vma_res = vma->resource;
1913 struct dma_fence *unbind_fence;
1914
1915 GEM_BUG_ON(i915_vma_is_pinned(vma));
1916 assert_vma_held_evict(vma);
1917
1918 if (i915_vma_is_map_and_fenceable(vma)) {
1919 /* Force a pagefault for domain tracking on next user access */
1920 i915_vma_revoke_mmap(vma);
1921
1922 /*
1923 * Check that we have flushed all writes through the GGTT
1924 * before the unbind, other due to non-strict nature of those
1925 * indirect writes they may end up referencing the GGTT PTE
1926 * after the unbind.
1927 *
1928 * Note that we may be concurrently poking at the GGTT_WRITE
1929 * bit from set-domain, as we mark all GGTT vma associated
1930 * with an object. We know this is for another vma, as we
1931 * are currently unbinding this one -- so if this vma will be
1932 * reused, it will be refaulted and have its dirty bit set
1933 * before the next write.
1934 */
1935 i915_vma_flush_writes(vma);
1936
1937 /* release the fence reg _after_ flushing */
1938 i915_vma_revoke_fence(vma);
1939
1940 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1941 }
1942
1943 __i915_vma_iounmap(vma);
1944
1945 GEM_BUG_ON(vma->fence);
1946 GEM_BUG_ON(i915_vma_has_userfault(vma));
1947
1948 /* Object backend must be async capable. */
1949 GEM_WARN_ON(async && !vma->resource->bi.pages_rsgt);
1950
1951 /* If vm is not open, unbind is a nop. */
1952 vma_res->needs_wakeref = i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND) &&
1953 kref_read(&vma->vm->ref);
1954 vma_res->skip_pte_rewrite = !kref_read(&vma->vm->ref) ||
1955 vma->vm->skip_pte_rewrite;
1956 trace_i915_vma_unbind(vma);
1957
1958 if (async)
1959 unbind_fence = i915_vma_resource_unbind(vma_res,
1960 &vma->obj->mm.tlb);
1961 else
1962 unbind_fence = i915_vma_resource_unbind(vma_res, NULL);
1963
1964 vma->resource = NULL;
1965
1966 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1967 &vma->flags);
1968
1969 i915_vma_detach(vma);
1970
1971 if (!async) {
1972 if (unbind_fence) {
1973 dma_fence_wait(unbind_fence, false);
1974 dma_fence_put(unbind_fence);
1975 unbind_fence = NULL;
1976 }
1977 vma_invalidate_tlb(vma->vm, &vma->obj->mm.tlb);
1978 }
1979
1980 /*
1981 * Binding itself may not have completed until the unbind fence signals,
1982 * so don't drop the pages until that happens, unless the resource is
1983 * async_capable.
1984 */
1985
1986 vma_unbind_pages(vma);
1987 return unbind_fence;
1988}
1989
1990int __i915_vma_unbind(struct i915_vma *vma)
1991{
1992 int ret;
1993
1994 lockdep_assert_held(&vma->vm->mutex);
1995 assert_vma_held_evict(vma);
1996
1997 if (!drm_mm_node_allocated(&vma->node))
1998 return 0;
1999
2000 if (i915_vma_is_pinned(vma)) {
2001 vma_print_allocator(vma, "is pinned");
2002 return -EAGAIN;
2003 }
2004
2005 /*
2006 * After confirming that no one else is pinning this vma, wait for
2007 * any laggards who may have crept in during the wait (through
2008 * a residual pin skipping the vm->mutex) to complete.
2009 */
2010 ret = i915_vma_sync(vma);
2011 if (ret)
2012 return ret;
2013
2014 GEM_BUG_ON(i915_vma_is_active(vma));
2015 __i915_vma_evict(vma, false);
2016
2017 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
2018 return 0;
2019}
2020
2021static struct dma_fence *__i915_vma_unbind_async(struct i915_vma *vma)
2022{
2023 struct dma_fence *fence;
2024
2025 lockdep_assert_held(&vma->vm->mutex);
2026
2027 if (!drm_mm_node_allocated(&vma->node))
2028 return NULL;
2029
2030 if (i915_vma_is_pinned(vma) ||
2031 &vma->obj->mm.rsgt->table != vma->resource->bi.pages)
2032 return ERR_PTR(-EAGAIN);
2033
2034 /*
2035 * We probably need to replace this with awaiting the fences of the
2036 * object's dma_resv when the vma active goes away. When doing that
2037 * we need to be careful to not add the vma_resource unbind fence
2038 * immediately to the object's dma_resv, because then unbinding
2039 * the next vma from the object, in case there are many, will
2040 * actually await the unbinding of the previous vmas, which is
2041 * undesirable.
2042 */
2043 if (i915_sw_fence_await_active(&vma->resource->chain, &vma->active,
2044 I915_ACTIVE_AWAIT_EXCL |
2045 I915_ACTIVE_AWAIT_ACTIVE) < 0) {
2046 return ERR_PTR(-EBUSY);
2047 }
2048
2049 fence = __i915_vma_evict(vma, true);
2050
2051 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
2052
2053 return fence;
2054}
2055
2056int i915_vma_unbind(struct i915_vma *vma)
2057{
2058 struct i915_address_space *vm = vma->vm;
2059 intel_wakeref_t wakeref = 0;
2060 int err;
2061
2062 assert_object_held_shared(vma->obj);
2063
2064 /* Optimistic wait before taking the mutex */
2065 err = i915_vma_sync(vma);
2066 if (err)
2067 return err;
2068
2069 if (!drm_mm_node_allocated(&vma->node))
2070 return 0;
2071
2072 if (i915_vma_is_pinned(vma)) {
2073 vma_print_allocator(vma, "is pinned");
2074 return -EAGAIN;
2075 }
2076
2077 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
2078 /* XXX not always required: nop_clear_range */
2079 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
2080
2081 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
2082 if (err)
2083 goto out_rpm;
2084
2085 err = __i915_vma_unbind(vma);
2086 mutex_unlock(&vm->mutex);
2087
2088out_rpm:
2089 if (wakeref)
2090 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
2091 return err;
2092}
2093
2094int i915_vma_unbind_async(struct i915_vma *vma, bool trylock_vm)
2095{
2096 struct drm_i915_gem_object *obj = vma->obj;
2097 struct i915_address_space *vm = vma->vm;
2098 intel_wakeref_t wakeref = 0;
2099 struct dma_fence *fence;
2100 int err;
2101
2102 /*
2103 * We need the dma-resv lock since we add the
2104 * unbind fence to the dma-resv object.
2105 */
2106 assert_object_held(obj);
2107
2108 if (!drm_mm_node_allocated(&vma->node))
2109 return 0;
2110
2111 if (i915_vma_is_pinned(vma)) {
2112 vma_print_allocator(vma, "is pinned");
2113 return -EAGAIN;
2114 }
2115
2116 if (!obj->mm.rsgt)
2117 return -EBUSY;
2118
2119 err = dma_resv_reserve_fences(obj->base.resv, 2);
2120 if (err)
2121 return -EBUSY;
2122
2123 /*
2124 * It would be great if we could grab this wakeref from the
2125 * async unbind work if needed, but we can't because it uses
2126 * kmalloc and it's in the dma-fence signalling critical path.
2127 */
2128 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
2129 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
2130
2131 if (trylock_vm && !mutex_trylock(&vm->mutex)) {
2132 err = -EBUSY;
2133 goto out_rpm;
2134 } else if (!trylock_vm) {
2135 err = mutex_lock_interruptible_nested(&vm->mutex, !wakeref);
2136 if (err)
2137 goto out_rpm;
2138 }
2139
2140 fence = __i915_vma_unbind_async(vma);
2141 mutex_unlock(&vm->mutex);
2142 if (IS_ERR_OR_NULL(fence)) {
2143 err = PTR_ERR_OR_ZERO(fence);
2144 goto out_rpm;
2145 }
2146
2147 dma_resv_add_fence(obj->base.resv, fence, DMA_RESV_USAGE_READ);
2148 dma_fence_put(fence);
2149
2150out_rpm:
2151 if (wakeref)
2152 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
2153 return err;
2154}
2155
2156int i915_vma_unbind_unlocked(struct i915_vma *vma)
2157{
2158 int err;
2159
2160 i915_gem_object_lock(vma->obj, NULL);
2161 err = i915_vma_unbind(vma);
2162 i915_gem_object_unlock(vma->obj);
2163
2164 return err;
2165}
2166
2167struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
2168{
2169 i915_gem_object_make_unshrinkable(vma->obj);
2170 return vma;
2171}
2172
2173void i915_vma_make_shrinkable(struct i915_vma *vma)
2174{
2175 i915_gem_object_make_shrinkable(vma->obj);
2176}
2177
2178void i915_vma_make_purgeable(struct i915_vma *vma)
2179{
2180 i915_gem_object_make_purgeable(vma->obj);
2181}
2182
2183#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2184#include "selftests/i915_vma.c"
2185#endif
2186
2187void i915_vma_module_exit(void)
2188{
2189 kmem_cache_destroy(slab_vmas);
2190}
2191
2192int __init i915_vma_module_init(void)
2193{
2194 slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
2195 if (!slab_vmas)
2196 return -ENOMEM;
2197
2198 return 0;
2199}