Merge tag 'docs-5.6-fixes' of git://git.lwn.net/linux
[linux-2.6-microblaze.git] / drivers / gpu / drm / i915 / i915_vma.c
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
41 static struct i915_global_vma {
42         struct i915_global base;
43         struct kmem_cache *slab_vmas;
44 } global;
45
46 struct i915_vma *i915_vma_alloc(void)
47 {
48         return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
49 }
50
51 void 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
60 static 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
80 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
81 {
82 }
83
84 #endif
85
86 static inline struct i915_vma *active_to_vma(struct i915_active *ref)
87 {
88         return container_of(ref, typeof(struct i915_vma), active);
89 }
90
91 static 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
97 static void __i915_vma_retire(struct i915_active *ref)
98 {
99         i915_vma_put(active_to_vma(ref));
100 }
101
102 static struct i915_vma *
103 vma_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 *vma;
108         struct rb_node *rb, **p;
109
110         /* The aliasing_ppgtt should never be used directly! */
111         GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
112
113         vma = i915_vma_alloc();
114         if (vma == NULL)
115                 return ERR_PTR(-ENOMEM);
116
117         kref_init(&vma->ref);
118         mutex_init(&vma->pages_mutex);
119         vma->vm = i915_vm_get(vm);
120         vma->ops = &vm->vma_ops;
121         vma->obj = obj;
122         vma->resv = obj->base.resv;
123         vma->size = obj->base.size;
124         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
125
126         i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire);
127
128         /* Declare ourselves safe for use inside shrinkers */
129         if (IS_ENABLED(CONFIG_LOCKDEP)) {
130                 fs_reclaim_acquire(GFP_KERNEL);
131                 might_lock(&vma->active.mutex);
132                 fs_reclaim_release(GFP_KERNEL);
133         }
134
135         INIT_LIST_HEAD(&vma->closed_link);
136
137         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
138                 vma->ggtt_view = *view;
139                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
140                         GEM_BUG_ON(range_overflows_t(u64,
141                                                      view->partial.offset,
142                                                      view->partial.size,
143                                                      obj->base.size >> PAGE_SHIFT));
144                         vma->size = view->partial.size;
145                         vma->size <<= PAGE_SHIFT;
146                         GEM_BUG_ON(vma->size > obj->base.size);
147                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
148                         vma->size = intel_rotation_info_size(&view->rotated);
149                         vma->size <<= PAGE_SHIFT;
150                 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
151                         vma->size = intel_remapped_info_size(&view->remapped);
152                         vma->size <<= PAGE_SHIFT;
153                 }
154         }
155
156         if (unlikely(vma->size > vm->total))
157                 goto err_vma;
158
159         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
160
161         if (i915_is_ggtt(vm)) {
162                 if (unlikely(overflows_type(vma->size, u32)))
163                         goto err_vma;
164
165                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
166                                                       i915_gem_object_get_tiling(obj),
167                                                       i915_gem_object_get_stride(obj));
168                 if (unlikely(vma->fence_size < vma->size || /* overflow */
169                              vma->fence_size > vm->total))
170                         goto err_vma;
171
172                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
173
174                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
175                                                                 i915_gem_object_get_tiling(obj),
176                                                                 i915_gem_object_get_stride(obj));
177                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
178
179                 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
180         }
181
182         spin_lock(&obj->vma.lock);
183
184         rb = NULL;
185         p = &obj->vma.tree.rb_node;
186         while (*p) {
187                 struct i915_vma *pos;
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                         spin_unlock(&obj->vma.lock);
201                         i915_vma_free(vma);
202                         return pos;
203                 }
204
205                 if (cmp < 0)
206                         p = &rb->rb_right;
207                 else
208                         p = &rb->rb_left;
209         }
210         rb_link_node(&vma->obj_node, rb, p);
211         rb_insert_color(&vma->obj_node, &obj->vma.tree);
212
213         if (i915_vma_is_ggtt(vma))
214                 /*
215                  * We put the GGTT vma at the start of the vma-list, followed
216                  * by the ppGGTT vma. This allows us to break early when
217                  * iterating over only the GGTT vma for an object, see
218                  * for_each_ggtt_vma()
219                  */
220                 list_add(&vma->obj_link, &obj->vma.list);
221         else
222                 list_add_tail(&vma->obj_link, &obj->vma.list);
223
224         spin_unlock(&obj->vma.lock);
225
226         return vma;
227
228 err_vma:
229         i915_vma_free(vma);
230         return ERR_PTR(-E2BIG);
231 }
232
233 static struct i915_vma *
234 vma_lookup(struct drm_i915_gem_object *obj,
235            struct i915_address_space *vm,
236            const struct i915_ggtt_view *view)
237 {
238         struct rb_node *rb;
239
240         rb = obj->vma.tree.rb_node;
241         while (rb) {
242                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
243                 long cmp;
244
245                 cmp = i915_vma_compare(vma, vm, view);
246                 if (cmp == 0)
247                         return vma;
248
249                 if (cmp < 0)
250                         rb = rb->rb_right;
251                 else
252                         rb = rb->rb_left;
253         }
254
255         return NULL;
256 }
257
258 /**
259  * i915_vma_instance - return the singleton instance of the VMA
260  * @obj: parent &struct drm_i915_gem_object to be mapped
261  * @vm: address space in which the mapping is located
262  * @view: additional mapping requirements
263  *
264  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
265  * the same @view characteristics. If a match is not found, one is created.
266  * Once created, the VMA is kept until either the object is freed, or the
267  * address space is closed.
268  *
269  * Returns the vma, or an error pointer.
270  */
271 struct i915_vma *
272 i915_vma_instance(struct drm_i915_gem_object *obj,
273                   struct i915_address_space *vm,
274                   const struct i915_ggtt_view *view)
275 {
276         struct i915_vma *vma;
277
278         GEM_BUG_ON(view && !i915_is_ggtt(vm));
279         GEM_BUG_ON(!atomic_read(&vm->open));
280
281         spin_lock(&obj->vma.lock);
282         vma = vma_lookup(obj, vm, view);
283         spin_unlock(&obj->vma.lock);
284
285         /* vma_create() will resolve the race if another creates the vma */
286         if (unlikely(!vma))
287                 vma = vma_create(obj, vm, view);
288
289         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
290         return vma;
291 }
292
293 struct i915_vma_work {
294         struct dma_fence_work base;
295         struct i915_vma *vma;
296         struct drm_i915_gem_object *pinned;
297         enum i915_cache_level cache_level;
298         unsigned int flags;
299 };
300
301 static 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, vw->cache_level, vw->flags);
308         if (err)
309                 atomic_or(I915_VMA_ERROR, &vma->flags);
310
311         return err;
312 }
313
314 static 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
322 static const struct dma_fence_work_ops bind_ops = {
323         .name = "bind",
324         .work = __vma_bind,
325         .release = __vma_release,
326 };
327
328 struct 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
342 /**
343  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
344  * @vma: VMA to map
345  * @cache_level: mapping cache level
346  * @flags: flags like global or local mapping
347  * @work: preallocated worker for allocating and binding the PTE
348  *
349  * DMA addresses are taken from the scatter-gather table of this object (or of
350  * this VMA in case of non-default GGTT views) and PTE entries set up.
351  * Note that DMA addresses are also the only part of the SG table we care about.
352  */
353 int i915_vma_bind(struct i915_vma *vma,
354                   enum i915_cache_level cache_level,
355                   u32 flags,
356                   struct i915_vma_work *work)
357 {
358         u32 bind_flags;
359         u32 vma_flags;
360         int ret;
361
362         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
363         GEM_BUG_ON(vma->size > vma->node.size);
364
365         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
366                                               vma->node.size,
367                                               vma->vm->total)))
368                 return -ENODEV;
369
370         if (GEM_DEBUG_WARN_ON(!flags))
371                 return -EINVAL;
372
373         bind_flags = flags;
374         bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
375
376         vma_flags = atomic_read(&vma->flags);
377         vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
378         if (flags & PIN_UPDATE)
379                 bind_flags |= vma_flags;
380         else
381                 bind_flags &= ~vma_flags;
382         if (bind_flags == 0)
383                 return 0;
384
385         GEM_BUG_ON(!vma->pages);
386
387         trace_i915_vma_bind(vma, bind_flags);
388         if (work && (bind_flags & ~vma_flags) & vma->vm->bind_async_flags) {
389                 work->vma = vma;
390                 work->cache_level = cache_level;
391                 work->flags = bind_flags | I915_VMA_ALLOC;
392
393                 /*
394                  * Note we only want to chain up to the migration fence on
395                  * the pages (not the object itself). As we don't track that,
396                  * yet, we have to use the exclusive fence instead.
397                  *
398                  * Also note that we do not want to track the async vma as
399                  * part of the obj->resv->excl_fence as it only affects
400                  * execution and not content or object's backing store lifetime.
401                  */
402                 GEM_BUG_ON(i915_active_has_exclusive(&vma->active));
403                 i915_active_set_exclusive(&vma->active, &work->base.dma);
404                 work->base.dma.error = 0; /* enable the queue_work() */
405
406                 if (vma->obj) {
407                         __i915_gem_object_pin_pages(vma->obj);
408                         work->pinned = vma->obj;
409                 }
410         } else {
411                 GEM_BUG_ON((bind_flags & ~vma_flags) & vma->vm->bind_async_flags);
412                 ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
413                 if (ret)
414                         return ret;
415         }
416
417         atomic_or(bind_flags, &vma->flags);
418         return 0;
419 }
420
421 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
422 {
423         void __iomem *ptr;
424         int err;
425
426         if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
427                 err = -ENODEV;
428                 goto err;
429         }
430
431         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
432         GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
433
434         ptr = READ_ONCE(vma->iomap);
435         if (ptr == NULL) {
436                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
437                                         vma->node.start,
438                                         vma->node.size);
439                 if (ptr == NULL) {
440                         err = -ENOMEM;
441                         goto err;
442                 }
443
444                 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
445                         io_mapping_unmap(ptr);
446                         ptr = vma->iomap;
447                 }
448         }
449
450         __i915_vma_pin(vma);
451
452         err = i915_vma_pin_fence(vma);
453         if (err)
454                 goto err_unpin;
455
456         i915_vma_set_ggtt_write(vma);
457
458         /* NB Access through the GTT requires the device to be awake. */
459         return ptr;
460
461 err_unpin:
462         __i915_vma_unpin(vma);
463 err:
464         return IO_ERR_PTR(err);
465 }
466
467 void i915_vma_flush_writes(struct i915_vma *vma)
468 {
469         if (i915_vma_unset_ggtt_write(vma))
470                 intel_gt_flush_ggtt_writes(vma->vm->gt);
471 }
472
473 void i915_vma_unpin_iomap(struct i915_vma *vma)
474 {
475         GEM_BUG_ON(vma->iomap == NULL);
476
477         i915_vma_flush_writes(vma);
478
479         i915_vma_unpin_fence(vma);
480         i915_vma_unpin(vma);
481 }
482
483 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
484 {
485         struct i915_vma *vma;
486         struct drm_i915_gem_object *obj;
487
488         vma = fetch_and_zero(p_vma);
489         if (!vma)
490                 return;
491
492         obj = vma->obj;
493         GEM_BUG_ON(!obj);
494
495         i915_vma_unpin(vma);
496         i915_vma_close(vma);
497
498         if (flags & I915_VMA_RELEASE_MAP)
499                 i915_gem_object_unpin_map(obj);
500
501         i915_gem_object_put(obj);
502 }
503
504 bool i915_vma_misplaced(const struct i915_vma *vma,
505                         u64 size, u64 alignment, u64 flags)
506 {
507         if (!drm_mm_node_allocated(&vma->node))
508                 return false;
509
510         if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
511                 return true;
512
513         if (vma->node.size < size)
514                 return true;
515
516         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
517         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
518                 return true;
519
520         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
521                 return true;
522
523         if (flags & PIN_OFFSET_BIAS &&
524             vma->node.start < (flags & PIN_OFFSET_MASK))
525                 return true;
526
527         if (flags & PIN_OFFSET_FIXED &&
528             vma->node.start != (flags & PIN_OFFSET_MASK))
529                 return true;
530
531         return false;
532 }
533
534 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
535 {
536         bool mappable, fenceable;
537
538         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
539         GEM_BUG_ON(!vma->fence_size);
540
541         fenceable = (vma->node.size >= vma->fence_size &&
542                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
543
544         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
545
546         if (mappable && fenceable)
547                 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
548         else
549                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
550 }
551
552 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
553 {
554         struct drm_mm_node *node = &vma->node;
555         struct drm_mm_node *other;
556
557         /*
558          * On some machines we have to be careful when putting differing types
559          * of snoopable memory together to avoid the prefetcher crossing memory
560          * domains and dying. During vm initialisation, we decide whether or not
561          * these constraints apply and set the drm_mm.color_adjust
562          * appropriately.
563          */
564         if (!i915_vm_has_cache_coloring(vma->vm))
565                 return true;
566
567         /* Only valid to be called on an already inserted vma */
568         GEM_BUG_ON(!drm_mm_node_allocated(node));
569         GEM_BUG_ON(list_empty(&node->node_list));
570
571         other = list_prev_entry(node, node_list);
572         if (i915_node_color_differs(other, color) &&
573             !drm_mm_hole_follows(other))
574                 return false;
575
576         other = list_next_entry(node, node_list);
577         if (i915_node_color_differs(other, color) &&
578             !drm_mm_hole_follows(node))
579                 return false;
580
581         return true;
582 }
583
584 static void assert_bind_count(const struct drm_i915_gem_object *obj)
585 {
586         /*
587          * Combine the assertion that the object is bound and that we have
588          * pinned its pages. But we should never have bound the object
589          * more than we have pinned its pages. (For complete accuracy, we
590          * assume that no else is pinning the pages, but as a rough assertion
591          * that we will not run into problems later, this will do!)
592          */
593         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < atomic_read(&obj->bind_count));
594 }
595
596 /**
597  * i915_vma_insert - finds a slot for the vma in its address space
598  * @vma: the vma
599  * @size: requested size in bytes (can be larger than the VMA)
600  * @alignment: required alignment
601  * @flags: mask of PIN_* flags to use
602  *
603  * First we try to allocate some free space that meets the requirements for
604  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
605  * preferrably the oldest idle entry to make room for the new VMA.
606  *
607  * Returns:
608  * 0 on success, negative error code otherwise.
609  */
610 static int
611 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
612 {
613         unsigned long color;
614         u64 start, end;
615         int ret;
616
617         GEM_BUG_ON(i915_vma_is_closed(vma));
618         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
619         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
620
621         size = max(size, vma->size);
622         alignment = max(alignment, vma->display_alignment);
623         if (flags & PIN_MAPPABLE) {
624                 size = max_t(typeof(size), size, vma->fence_size);
625                 alignment = max_t(typeof(alignment),
626                                   alignment, vma->fence_alignment);
627         }
628
629         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
630         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
631         GEM_BUG_ON(!is_power_of_2(alignment));
632
633         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
634         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
635
636         end = vma->vm->total;
637         if (flags & PIN_MAPPABLE)
638                 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
639         if (flags & PIN_ZONE_4G)
640                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
641         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
642
643         /* If binding the object/GGTT view requires more space than the entire
644          * aperture has, reject it early before evicting everything in a vain
645          * attempt to find space.
646          */
647         if (size > end) {
648                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
649                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
650                           end);
651                 return -ENOSPC;
652         }
653
654         color = 0;
655         if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
656                 color = vma->obj->cache_level;
657
658         if (flags & PIN_OFFSET_FIXED) {
659                 u64 offset = flags & PIN_OFFSET_MASK;
660                 if (!IS_ALIGNED(offset, alignment) ||
661                     range_overflows(offset, size, end))
662                         return -EINVAL;
663
664                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
665                                            size, offset, color,
666                                            flags);
667                 if (ret)
668                         return ret;
669         } else {
670                 /*
671                  * We only support huge gtt pages through the 48b PPGTT,
672                  * however we also don't want to force any alignment for
673                  * objects which need to be tightly packed into the low 32bits.
674                  *
675                  * Note that we assume that GGTT are limited to 4GiB for the
676                  * forseeable future. See also i915_ggtt_offset().
677                  */
678                 if (upper_32_bits(end - 1) &&
679                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
680                         /*
681                          * We can't mix 64K and 4K PTEs in the same page-table
682                          * (2M block), and so to avoid the ugliness and
683                          * complexity of coloring we opt for just aligning 64K
684                          * objects to 2M.
685                          */
686                         u64 page_alignment =
687                                 rounddown_pow_of_two(vma->page_sizes.sg |
688                                                      I915_GTT_PAGE_SIZE_2M);
689
690                         /*
691                          * Check we don't expand for the limited Global GTT
692                          * (mappable aperture is even more precious!). This
693                          * also checks that we exclude the aliasing-ppgtt.
694                          */
695                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
696
697                         alignment = max(alignment, page_alignment);
698
699                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
700                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
701                 }
702
703                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
704                                           size, alignment, color,
705                                           start, end, flags);
706                 if (ret)
707                         return ret;
708
709                 GEM_BUG_ON(vma->node.start < start);
710                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
711         }
712         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
713         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
714
715         if (vma->obj) {
716                 struct drm_i915_gem_object *obj = vma->obj;
717
718                 atomic_inc(&obj->bind_count);
719                 assert_bind_count(obj);
720         }
721         list_add_tail(&vma->vm_link, &vma->vm->bound_list);
722
723         return 0;
724 }
725
726 static void
727 i915_vma_detach(struct i915_vma *vma)
728 {
729         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
730         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
731
732         /*
733          * And finally now the object is completely decoupled from this
734          * vma, we can drop its hold on the backing storage and allow
735          * it to be reaped by the shrinker.
736          */
737         list_del(&vma->vm_link);
738         if (vma->obj) {
739                 struct drm_i915_gem_object *obj = vma->obj;
740
741                 assert_bind_count(obj);
742                 atomic_dec(&obj->bind_count);
743         }
744 }
745
746 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
747 {
748         unsigned int bound;
749         bool pinned = true;
750
751         bound = atomic_read(&vma->flags);
752         do {
753                 if (unlikely(flags & ~bound))
754                         return false;
755
756                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
757                         return false;
758
759                 if (!(bound & I915_VMA_PIN_MASK))
760                         goto unpinned;
761
762                 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
763         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
764
765         return true;
766
767 unpinned:
768         /*
769          * If pin_count==0, but we are bound, check under the lock to avoid
770          * racing with a concurrent i915_vma_unbind().
771          */
772         mutex_lock(&vma->vm->mutex);
773         do {
774                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
775                         pinned = false;
776                         break;
777                 }
778
779                 if (unlikely(flags & ~bound)) {
780                         pinned = false;
781                         break;
782                 }
783         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
784         mutex_unlock(&vma->vm->mutex);
785
786         return pinned;
787 }
788
789 static int vma_get_pages(struct i915_vma *vma)
790 {
791         int err = 0;
792
793         if (atomic_add_unless(&vma->pages_count, 1, 0))
794                 return 0;
795
796         /* Allocations ahoy! */
797         if (mutex_lock_interruptible(&vma->pages_mutex))
798                 return -EINTR;
799
800         if (!atomic_read(&vma->pages_count)) {
801                 if (vma->obj) {
802                         err = i915_gem_object_pin_pages(vma->obj);
803                         if (err)
804                                 goto unlock;
805                 }
806
807                 err = vma->ops->set_pages(vma);
808                 if (err) {
809                         if (vma->obj)
810                                 i915_gem_object_unpin_pages(vma->obj);
811                         goto unlock;
812                 }
813         }
814         atomic_inc(&vma->pages_count);
815
816 unlock:
817         mutex_unlock(&vma->pages_mutex);
818
819         return err;
820 }
821
822 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
823 {
824         /* We allocate under vma_get_pages, so beware the shrinker */
825         mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
826         GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
827         if (atomic_sub_return(count, &vma->pages_count) == 0) {
828                 vma->ops->clear_pages(vma);
829                 GEM_BUG_ON(vma->pages);
830                 if (vma->obj)
831                         i915_gem_object_unpin_pages(vma->obj);
832         }
833         mutex_unlock(&vma->pages_mutex);
834 }
835
836 static void vma_put_pages(struct i915_vma *vma)
837 {
838         if (atomic_add_unless(&vma->pages_count, -1, 1))
839                 return;
840
841         __vma_put_pages(vma, 1);
842 }
843
844 static void vma_unbind_pages(struct i915_vma *vma)
845 {
846         unsigned int count;
847
848         lockdep_assert_held(&vma->vm->mutex);
849
850         /* The upper portion of pages_count is the number of bindings */
851         count = atomic_read(&vma->pages_count);
852         count >>= I915_VMA_PAGES_BIAS;
853         GEM_BUG_ON(!count);
854
855         __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
856 }
857
858 int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
859 {
860         struct i915_vma_work *work = NULL;
861         intel_wakeref_t wakeref = 0;
862         unsigned int bound;
863         int err;
864
865         BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
866         BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
867
868         GEM_BUG_ON(flags & PIN_UPDATE);
869         GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
870
871         /* First try and grab the pin without rebinding the vma */
872         if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
873                 return 0;
874
875         err = vma_get_pages(vma);
876         if (err)
877                 return err;
878
879         if (flags & vma->vm->bind_async_flags) {
880                 work = i915_vma_work();
881                 if (!work) {
882                         err = -ENOMEM;
883                         goto err_pages;
884                 }
885         }
886
887         if (flags & PIN_GLOBAL)
888                 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
889
890         /* No more allocations allowed once we hold vm->mutex */
891         err = mutex_lock_interruptible(&vma->vm->mutex);
892         if (err)
893                 goto err_fence;
894
895         bound = atomic_read(&vma->flags);
896         if (unlikely(bound & I915_VMA_ERROR)) {
897                 err = -ENOMEM;
898                 goto err_unlock;
899         }
900
901         if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
902                 err = -EAGAIN; /* pins are meant to be fairly temporary */
903                 goto err_unlock;
904         }
905
906         if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
907                 __i915_vma_pin(vma);
908                 goto err_unlock;
909         }
910
911         err = i915_active_acquire(&vma->active);
912         if (err)
913                 goto err_unlock;
914
915         if (!(bound & I915_VMA_BIND_MASK)) {
916                 err = i915_vma_insert(vma, size, alignment, flags);
917                 if (err)
918                         goto err_active;
919
920                 if (i915_is_ggtt(vma->vm))
921                         __i915_vma_set_map_and_fenceable(vma);
922         }
923
924         GEM_BUG_ON(!vma->pages);
925         err = i915_vma_bind(vma,
926                             vma->obj ? vma->obj->cache_level : 0,
927                             flags, work);
928         if (err)
929                 goto err_remove;
930
931         /* There should only be at most 2 active bindings (user, global) */
932         GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
933         atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
934         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
935
936         __i915_vma_pin(vma);
937         GEM_BUG_ON(!i915_vma_is_pinned(vma));
938         GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
939         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
940
941 err_remove:
942         if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
943                 i915_vma_detach(vma);
944                 drm_mm_remove_node(&vma->node);
945         }
946 err_active:
947         i915_active_release(&vma->active);
948 err_unlock:
949         mutex_unlock(&vma->vm->mutex);
950 err_fence:
951         if (work)
952                 dma_fence_work_commit(&work->base);
953         if (wakeref)
954                 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
955 err_pages:
956         vma_put_pages(vma);
957         return err;
958 }
959
960 static void flush_idle_contexts(struct intel_gt *gt)
961 {
962         struct intel_engine_cs *engine;
963         enum intel_engine_id id;
964
965         for_each_engine(engine, gt, id)
966                 intel_engine_flush_barriers(engine);
967
968         intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
969 }
970
971 int i915_ggtt_pin(struct i915_vma *vma, u32 align, unsigned int flags)
972 {
973         struct i915_address_space *vm = vma->vm;
974         int err;
975
976         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
977
978         do {
979                 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
980                 if (err != -ENOSPC)
981                         return err;
982
983                 /* Unlike i915_vma_pin, we don't take no for an answer! */
984                 flush_idle_contexts(vm->gt);
985                 if (mutex_lock_interruptible(&vm->mutex) == 0) {
986                         i915_gem_evict_vm(vm);
987                         mutex_unlock(&vm->mutex);
988                 }
989         } while (1);
990 }
991
992 void i915_vma_close(struct i915_vma *vma)
993 {
994         struct intel_gt *gt = vma->vm->gt;
995         unsigned long flags;
996
997         GEM_BUG_ON(i915_vma_is_closed(vma));
998
999         /*
1000          * We defer actually closing, unbinding and destroying the VMA until
1001          * the next idle point, or if the object is freed in the meantime. By
1002          * postponing the unbind, we allow for it to be resurrected by the
1003          * client, avoiding the work required to rebind the VMA. This is
1004          * advantageous for DRI, where the client/server pass objects
1005          * between themselves, temporarily opening a local VMA to the
1006          * object, and then closing it again. The same object is then reused
1007          * on the next frame (or two, depending on the depth of the swap queue)
1008          * causing us to rebind the VMA once more. This ends up being a lot
1009          * of wasted work for the steady state.
1010          */
1011         spin_lock_irqsave(&gt->closed_lock, flags);
1012         list_add(&vma->closed_link, &gt->closed_vma);
1013         spin_unlock_irqrestore(&gt->closed_lock, flags);
1014 }
1015
1016 static void __i915_vma_remove_closed(struct i915_vma *vma)
1017 {
1018         struct intel_gt *gt = vma->vm->gt;
1019
1020         spin_lock_irq(&gt->closed_lock);
1021         list_del_init(&vma->closed_link);
1022         spin_unlock_irq(&gt->closed_lock);
1023 }
1024
1025 void i915_vma_reopen(struct i915_vma *vma)
1026 {
1027         if (i915_vma_is_closed(vma))
1028                 __i915_vma_remove_closed(vma);
1029 }
1030
1031 void i915_vma_release(struct kref *ref)
1032 {
1033         struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1034
1035         if (drm_mm_node_allocated(&vma->node)) {
1036                 mutex_lock(&vma->vm->mutex);
1037                 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1038                 WARN_ON(__i915_vma_unbind(vma));
1039                 mutex_unlock(&vma->vm->mutex);
1040                 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1041         }
1042         GEM_BUG_ON(i915_vma_is_active(vma));
1043
1044         if (vma->obj) {
1045                 struct drm_i915_gem_object *obj = vma->obj;
1046
1047                 spin_lock(&obj->vma.lock);
1048                 list_del(&vma->obj_link);
1049                 rb_erase(&vma->obj_node, &obj->vma.tree);
1050                 spin_unlock(&obj->vma.lock);
1051         }
1052
1053         __i915_vma_remove_closed(vma);
1054         i915_vm_put(vma->vm);
1055
1056         i915_active_fini(&vma->active);
1057         i915_vma_free(vma);
1058 }
1059
1060 void i915_vma_parked(struct intel_gt *gt)
1061 {
1062         struct i915_vma *vma, *next;
1063
1064         spin_lock_irq(&gt->closed_lock);
1065         list_for_each_entry_safe(vma, next, &gt->closed_vma, closed_link) {
1066                 struct drm_i915_gem_object *obj = vma->obj;
1067                 struct i915_address_space *vm = vma->vm;
1068
1069                 /* XXX All to avoid keeping a reference on i915_vma itself */
1070
1071                 if (!kref_get_unless_zero(&obj->base.refcount))
1072                         continue;
1073
1074                 if (i915_vm_tryopen(vm)) {
1075                         list_del_init(&vma->closed_link);
1076                 } else {
1077                         i915_gem_object_put(obj);
1078                         obj = NULL;
1079                 }
1080
1081                 spin_unlock_irq(&gt->closed_lock);
1082
1083                 if (obj) {
1084                         __i915_vma_put(vma);
1085                         i915_gem_object_put(obj);
1086                 }
1087
1088                 i915_vm_close(vm);
1089
1090                 /* Restart after dropping lock */
1091                 spin_lock_irq(&gt->closed_lock);
1092                 next = list_first_entry(&gt->closed_vma,
1093                                         typeof(*next), closed_link);
1094         }
1095         spin_unlock_irq(&gt->closed_lock);
1096 }
1097
1098 static void __i915_vma_iounmap(struct i915_vma *vma)
1099 {
1100         GEM_BUG_ON(i915_vma_is_pinned(vma));
1101
1102         if (vma->iomap == NULL)
1103                 return;
1104
1105         io_mapping_unmap(vma->iomap);
1106         vma->iomap = NULL;
1107 }
1108
1109 void i915_vma_revoke_mmap(struct i915_vma *vma)
1110 {
1111         struct drm_vma_offset_node *node;
1112         u64 vma_offset;
1113
1114         if (!i915_vma_has_userfault(vma))
1115                 return;
1116
1117         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1118         GEM_BUG_ON(!vma->obj->userfault_count);
1119
1120         node = &vma->mmo->vma_node;
1121         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1122         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1123                             drm_vma_node_offset_addr(node) + vma_offset,
1124                             vma->size,
1125                             1);
1126
1127         i915_vma_unset_userfault(vma);
1128         if (!--vma->obj->userfault_count)
1129                 list_del(&vma->obj->userfault_link);
1130 }
1131
1132 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1133 {
1134         int err;
1135
1136         GEM_BUG_ON(!i915_vma_is_pinned(vma));
1137
1138         /* Wait for the vma to be bound before we start! */
1139         err = i915_request_await_active(rq, &vma->active);
1140         if (err)
1141                 return err;
1142
1143         return i915_active_add_request(&vma->active, rq);
1144 }
1145
1146 int i915_vma_move_to_active(struct i915_vma *vma,
1147                             struct i915_request *rq,
1148                             unsigned int flags)
1149 {
1150         struct drm_i915_gem_object *obj = vma->obj;
1151         int err;
1152
1153         assert_object_held(obj);
1154
1155         err = __i915_vma_move_to_active(vma, rq);
1156         if (unlikely(err))
1157                 return err;
1158
1159         if (flags & EXEC_OBJECT_WRITE) {
1160                 struct intel_frontbuffer *front;
1161
1162                 front = __intel_frontbuffer_get(obj);
1163                 if (unlikely(front)) {
1164                         if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1165                                 i915_active_add_request(&front->write, rq);
1166                         intel_frontbuffer_put(front);
1167                 }
1168
1169                 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1170                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1171                 obj->read_domains = 0;
1172         } else {
1173                 err = dma_resv_reserve_shared(vma->resv, 1);
1174                 if (unlikely(err))
1175                         return err;
1176
1177                 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1178                 obj->write_domain = 0;
1179         }
1180         obj->read_domains |= I915_GEM_GPU_DOMAINS;
1181         obj->mm.dirty = true;
1182
1183         GEM_BUG_ON(!i915_vma_is_active(vma));
1184         return 0;
1185 }
1186
1187 int __i915_vma_unbind(struct i915_vma *vma)
1188 {
1189         int ret;
1190
1191         lockdep_assert_held(&vma->vm->mutex);
1192
1193         /*
1194          * First wait upon any activity as retiring the request may
1195          * have side-effects such as unpinning or even unbinding this vma.
1196          *
1197          * XXX Actually waiting under the vm->mutex is a hinderance and
1198          * should be pipelined wherever possible. In cases where that is
1199          * unavoidable, we should lift the wait to before the mutex.
1200          */
1201         ret = i915_vma_sync(vma);
1202         if (ret)
1203                 return ret;
1204
1205         if (i915_vma_is_pinned(vma)) {
1206                 vma_print_allocator(vma, "is pinned");
1207                 return -EAGAIN;
1208         }
1209
1210         /*
1211          * After confirming that no one else is pinning this vma, wait for
1212          * any laggards who may have crept in during the wait (through
1213          * a residual pin skipping the vm->mutex) to complete.
1214          */
1215         ret = i915_vma_sync(vma);
1216         if (ret)
1217                 return ret;
1218
1219         if (!drm_mm_node_allocated(&vma->node))
1220                 return 0;
1221
1222         GEM_BUG_ON(i915_vma_is_pinned(vma));
1223         GEM_BUG_ON(i915_vma_is_active(vma));
1224
1225         if (i915_vma_is_map_and_fenceable(vma)) {
1226                 /*
1227                  * Check that we have flushed all writes through the GGTT
1228                  * before the unbind, other due to non-strict nature of those
1229                  * indirect writes they may end up referencing the GGTT PTE
1230                  * after the unbind.
1231                  */
1232                 i915_vma_flush_writes(vma);
1233                 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
1234
1235                 /* release the fence reg _after_ flushing */
1236                 ret = i915_vma_revoke_fence(vma);
1237                 if (ret)
1238                         return ret;
1239
1240                 /* Force a pagefault for domain tracking on next user access */
1241                 i915_vma_revoke_mmap(vma);
1242
1243                 __i915_vma_iounmap(vma);
1244                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1245         }
1246         GEM_BUG_ON(vma->fence);
1247         GEM_BUG_ON(i915_vma_has_userfault(vma));
1248
1249         if (likely(atomic_read(&vma->vm->open))) {
1250                 trace_i915_vma_unbind(vma);
1251                 vma->ops->unbind_vma(vma);
1252         }
1253         atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR), &vma->flags);
1254
1255         i915_vma_detach(vma);
1256         vma_unbind_pages(vma);
1257
1258         drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1259         return 0;
1260 }
1261
1262 int i915_vma_unbind(struct i915_vma *vma)
1263 {
1264         struct i915_address_space *vm = vma->vm;
1265         intel_wakeref_t wakeref = 0;
1266         int err;
1267
1268         if (!drm_mm_node_allocated(&vma->node))
1269                 return 0;
1270
1271         if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1272                 /* XXX not always required: nop_clear_range */
1273                 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1274
1275         err = mutex_lock_interruptible(&vm->mutex);
1276         if (err)
1277                 return err;
1278
1279         err = __i915_vma_unbind(vma);
1280         mutex_unlock(&vm->mutex);
1281
1282         if (wakeref)
1283                 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1284
1285         return err;
1286 }
1287
1288 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1289 {
1290         i915_gem_object_make_unshrinkable(vma->obj);
1291         return vma;
1292 }
1293
1294 void i915_vma_make_shrinkable(struct i915_vma *vma)
1295 {
1296         i915_gem_object_make_shrinkable(vma->obj);
1297 }
1298
1299 void i915_vma_make_purgeable(struct i915_vma *vma)
1300 {
1301         i915_gem_object_make_purgeable(vma->obj);
1302 }
1303
1304 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1305 #include "selftests/i915_vma.c"
1306 #endif
1307
1308 static void i915_global_vma_shrink(void)
1309 {
1310         kmem_cache_shrink(global.slab_vmas);
1311 }
1312
1313 static void i915_global_vma_exit(void)
1314 {
1315         kmem_cache_destroy(global.slab_vmas);
1316 }
1317
1318 static struct i915_global_vma global = { {
1319         .shrink = i915_global_vma_shrink,
1320         .exit = i915_global_vma_exit,
1321 } };
1322
1323 int __init i915_global_vma_init(void)
1324 {
1325         global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1326         if (!global.slab_vmas)
1327                 return -ENOMEM;
1328
1329         i915_global_register(&global.base);
1330         return 0;
1331 }