Merge tag 'dmaengine-4.10-rc1' of git://git.infradead.org/users/vkoul/slave-dma
[sfrench/cifs-2.6.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 "i915_vma.h"
26
27 #include "i915_drv.h"
28 #include "intel_ringbuffer.h"
29 #include "intel_frontbuffer.h"
30
31 #include <drm/drm_gem.h>
32
33 static void
34 i915_vma_retire(struct i915_gem_active *active,
35                 struct drm_i915_gem_request *rq)
36 {
37         const unsigned int idx = rq->engine->id;
38         struct i915_vma *vma =
39                 container_of(active, struct i915_vma, last_read[idx]);
40         struct drm_i915_gem_object *obj = vma->obj;
41
42         GEM_BUG_ON(!i915_vma_has_active_engine(vma, idx));
43
44         i915_vma_clear_active(vma, idx);
45         if (i915_vma_is_active(vma))
46                 return;
47
48         list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
49         if (unlikely(i915_vma_is_closed(vma) && !i915_vma_is_pinned(vma)))
50                 WARN_ON(i915_vma_unbind(vma));
51
52         GEM_BUG_ON(!i915_gem_object_is_active(obj));
53         if (--obj->active_count)
54                 return;
55
56         /* Bump our place on the bound list to keep it roughly in LRU order
57          * so that we don't steal from recently used but inactive objects
58          * (unless we are forced to ofc!)
59          */
60         if (obj->bind_count)
61                 list_move_tail(&obj->global_link, &rq->i915->mm.bound_list);
62
63         obj->mm.dirty = true; /* be paranoid  */
64
65         if (i915_gem_object_has_active_reference(obj)) {
66                 i915_gem_object_clear_active_reference(obj);
67                 i915_gem_object_put(obj);
68         }
69 }
70
71 static struct i915_vma *
72 __i915_vma_create(struct drm_i915_gem_object *obj,
73                   struct i915_address_space *vm,
74                   const struct i915_ggtt_view *view)
75 {
76         struct i915_vma *vma;
77         struct rb_node *rb, **p;
78         int i;
79
80         GEM_BUG_ON(vm->closed);
81
82         vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
83         if (vma == NULL)
84                 return ERR_PTR(-ENOMEM);
85
86         INIT_LIST_HEAD(&vma->exec_list);
87         for (i = 0; i < ARRAY_SIZE(vma->last_read); i++)
88                 init_request_active(&vma->last_read[i], i915_vma_retire);
89         init_request_active(&vma->last_fence, NULL);
90         list_add(&vma->vm_link, &vm->unbound_list);
91         vma->vm = vm;
92         vma->obj = obj;
93         vma->size = obj->base.size;
94
95         if (view) {
96                 vma->ggtt_view = *view;
97                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
98                         vma->size = view->params.partial.size;
99                         vma->size <<= PAGE_SHIFT;
100                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
101                         vma->size =
102                                 intel_rotation_info_size(&view->params.rotated);
103                         vma->size <<= PAGE_SHIFT;
104                 }
105         }
106
107         if (i915_is_ggtt(vm)) {
108                 vma->flags |= I915_VMA_GGTT;
109                 list_add(&vma->obj_link, &obj->vma_list);
110         } else {
111                 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
112                 list_add_tail(&vma->obj_link, &obj->vma_list);
113         }
114
115         rb = NULL;
116         p = &obj->vma_tree.rb_node;
117         while (*p) {
118                 struct i915_vma *pos;
119
120                 rb = *p;
121                 pos = rb_entry(rb, struct i915_vma, obj_node);
122                 if (i915_vma_compare(pos, vm, view) < 0)
123                         p = &rb->rb_right;
124                 else
125                         p = &rb->rb_left;
126         }
127         rb_link_node(&vma->obj_node, rb, p);
128         rb_insert_color(&vma->obj_node, &obj->vma_tree);
129
130         return vma;
131 }
132
133 struct i915_vma *
134 i915_vma_create(struct drm_i915_gem_object *obj,
135                 struct i915_address_space *vm,
136                 const struct i915_ggtt_view *view)
137 {
138         lockdep_assert_held(&obj->base.dev->struct_mutex);
139         GEM_BUG_ON(view && !i915_is_ggtt(vm));
140         GEM_BUG_ON(i915_gem_obj_to_vma(obj, vm, view));
141
142         return __i915_vma_create(obj, vm, view);
143 }
144
145 /**
146  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
147  * @vma: VMA to map
148  * @cache_level: mapping cache level
149  * @flags: flags like global or local mapping
150  *
151  * DMA addresses are taken from the scatter-gather table of this object (or of
152  * this VMA in case of non-default GGTT views) and PTE entries set up.
153  * Note that DMA addresses are also the only part of the SG table we care about.
154  */
155 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
156                   u32 flags)
157 {
158         u32 bind_flags;
159         u32 vma_flags;
160         int ret;
161
162         if (WARN_ON(flags == 0))
163                 return -EINVAL;
164
165         bind_flags = 0;
166         if (flags & PIN_GLOBAL)
167                 bind_flags |= I915_VMA_GLOBAL_BIND;
168         if (flags & PIN_USER)
169                 bind_flags |= I915_VMA_LOCAL_BIND;
170
171         vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
172         if (flags & PIN_UPDATE)
173                 bind_flags |= vma_flags;
174         else
175                 bind_flags &= ~vma_flags;
176         if (bind_flags == 0)
177                 return 0;
178
179         if (vma_flags == 0 && vma->vm->allocate_va_range) {
180                 trace_i915_va_alloc(vma);
181                 ret = vma->vm->allocate_va_range(vma->vm,
182                                                  vma->node.start,
183                                                  vma->node.size);
184                 if (ret)
185                         return ret;
186         }
187
188         ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
189         if (ret)
190                 return ret;
191
192         vma->flags |= bind_flags;
193         return 0;
194 }
195
196 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
197 {
198         void __iomem *ptr;
199
200         /* Access through the GTT requires the device to be awake. */
201         assert_rpm_wakelock_held(to_i915(vma->vm->dev));
202
203         lockdep_assert_held(&vma->vm->dev->struct_mutex);
204         if (WARN_ON(!i915_vma_is_map_and_fenceable(vma)))
205                 return IO_ERR_PTR(-ENODEV);
206
207         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
208         GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
209
210         ptr = vma->iomap;
211         if (ptr == NULL) {
212                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->mappable,
213                                         vma->node.start,
214                                         vma->node.size);
215                 if (ptr == NULL)
216                         return IO_ERR_PTR(-ENOMEM);
217
218                 vma->iomap = ptr;
219         }
220
221         __i915_vma_pin(vma);
222         return ptr;
223 }
224
225 void i915_vma_unpin_and_release(struct i915_vma **p_vma)
226 {
227         struct i915_vma *vma;
228         struct drm_i915_gem_object *obj;
229
230         vma = fetch_and_zero(p_vma);
231         if (!vma)
232                 return;
233
234         obj = vma->obj;
235
236         i915_vma_unpin(vma);
237         i915_vma_close(vma);
238
239         __i915_gem_object_release_unless_active(obj);
240 }
241
242 bool
243 i915_vma_misplaced(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
244 {
245         if (!drm_mm_node_allocated(&vma->node))
246                 return false;
247
248         if (vma->node.size < size)
249                 return true;
250
251         if (alignment && vma->node.start & (alignment - 1))
252                 return true;
253
254         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
255                 return true;
256
257         if (flags & PIN_OFFSET_BIAS &&
258             vma->node.start < (flags & PIN_OFFSET_MASK))
259                 return true;
260
261         if (flags & PIN_OFFSET_FIXED &&
262             vma->node.start != (flags & PIN_OFFSET_MASK))
263                 return true;
264
265         return false;
266 }
267
268 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
269 {
270         struct drm_i915_gem_object *obj = vma->obj;
271         struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
272         bool mappable, fenceable;
273         u32 fence_size, fence_alignment;
274
275         fence_size = i915_gem_get_ggtt_size(dev_priv,
276                                             vma->size,
277                                             i915_gem_object_get_tiling(obj));
278         fence_alignment = i915_gem_get_ggtt_alignment(dev_priv,
279                                                       vma->size,
280                                                       i915_gem_object_get_tiling(obj),
281                                                       true);
282
283         fenceable = (vma->node.size == fence_size &&
284                      (vma->node.start & (fence_alignment - 1)) == 0);
285
286         mappable = (vma->node.start + fence_size <=
287                     dev_priv->ggtt.mappable_end);
288
289         /*
290          * Explicitly disable for rotated VMA since the display does not
291          * need the fence and the VMA is not accessible to other users.
292          */
293         if (mappable && fenceable &&
294             vma->ggtt_view.type != I915_GGTT_VIEW_ROTATED)
295                 vma->flags |= I915_VMA_CAN_FENCE;
296         else
297                 vma->flags &= ~I915_VMA_CAN_FENCE;
298 }
299
300 bool i915_gem_valid_gtt_space(struct i915_vma *vma,
301                               unsigned long cache_level)
302 {
303         struct drm_mm_node *gtt_space = &vma->node;
304         struct drm_mm_node *other;
305
306         /*
307          * On some machines we have to be careful when putting differing types
308          * of snoopable memory together to avoid the prefetcher crossing memory
309          * domains and dying. During vm initialisation, we decide whether or not
310          * these constraints apply and set the drm_mm.color_adjust
311          * appropriately.
312          */
313         if (vma->vm->mm.color_adjust == NULL)
314                 return true;
315
316         if (!drm_mm_node_allocated(gtt_space))
317                 return true;
318
319         if (list_empty(&gtt_space->node_list))
320                 return true;
321
322         other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
323         if (other->allocated && !other->hole_follows && other->color != cache_level)
324                 return false;
325
326         other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
327         if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
328                 return false;
329
330         return true;
331 }
332
333 /**
334  * i915_vma_insert - finds a slot for the vma in its address space
335  * @vma: the vma
336  * @size: requested size in bytes (can be larger than the VMA)
337  * @alignment: required alignment
338  * @flags: mask of PIN_* flags to use
339  *
340  * First we try to allocate some free space that meets the requirements for
341  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
342  * preferrably the oldest idle entry to make room for the new VMA.
343  *
344  * Returns:
345  * 0 on success, negative error code otherwise.
346  */
347 static int
348 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
349 {
350         struct drm_i915_private *dev_priv = to_i915(vma->vm->dev);
351         struct drm_i915_gem_object *obj = vma->obj;
352         u64 start, end;
353         int ret;
354
355         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
356         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
357
358         size = max(size, vma->size);
359         if (flags & PIN_MAPPABLE)
360                 size = i915_gem_get_ggtt_size(dev_priv, size,
361                                               i915_gem_object_get_tiling(obj));
362
363         alignment = max(max(alignment, vma->display_alignment),
364                         i915_gem_get_ggtt_alignment(dev_priv, size,
365                                                     i915_gem_object_get_tiling(obj),
366                                                     flags & PIN_MAPPABLE));
367
368         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
369
370         end = vma->vm->total;
371         if (flags & PIN_MAPPABLE)
372                 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
373         if (flags & PIN_ZONE_4G)
374                 end = min_t(u64, end, (1ULL << 32) - PAGE_SIZE);
375
376         /* If binding the object/GGTT view requires more space than the entire
377          * aperture has, reject it early before evicting everything in a vain
378          * attempt to find space.
379          */
380         if (size > end) {
381                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu [object=%zd] > %s aperture=%llu\n",
382                           size, obj->base.size,
383                           flags & PIN_MAPPABLE ? "mappable" : "total",
384                           end);
385                 return -E2BIG;
386         }
387
388         ret = i915_gem_object_pin_pages(obj);
389         if (ret)
390                 return ret;
391
392         if (flags & PIN_OFFSET_FIXED) {
393                 u64 offset = flags & PIN_OFFSET_MASK;
394                 if (offset & (alignment - 1) || offset > end - size) {
395                         ret = -EINVAL;
396                         goto err_unpin;
397                 }
398
399                 vma->node.start = offset;
400                 vma->node.size = size;
401                 vma->node.color = obj->cache_level;
402                 ret = drm_mm_reserve_node(&vma->vm->mm, &vma->node);
403                 if (ret) {
404                         ret = i915_gem_evict_for_vma(vma);
405                         if (ret == 0)
406                                 ret = drm_mm_reserve_node(&vma->vm->mm, &vma->node);
407                         if (ret)
408                                 goto err_unpin;
409                 }
410         } else {
411                 u32 search_flag, alloc_flag;
412
413                 if (flags & PIN_HIGH) {
414                         search_flag = DRM_MM_SEARCH_BELOW;
415                         alloc_flag = DRM_MM_CREATE_TOP;
416                 } else {
417                         search_flag = DRM_MM_SEARCH_DEFAULT;
418                         alloc_flag = DRM_MM_CREATE_DEFAULT;
419                 }
420
421                 /* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks,
422                  * so we know that we always have a minimum alignment of 4096.
423                  * The drm_mm range manager is optimised to return results
424                  * with zero alignment, so where possible use the optimal
425                  * path.
426                  */
427                 if (alignment <= 4096)
428                         alignment = 0;
429
430 search_free:
431                 ret = drm_mm_insert_node_in_range_generic(&vma->vm->mm,
432                                                           &vma->node,
433                                                           size, alignment,
434                                                           obj->cache_level,
435                                                           start, end,
436                                                           search_flag,
437                                                           alloc_flag);
438                 if (ret) {
439                         ret = i915_gem_evict_something(vma->vm, size, alignment,
440                                                        obj->cache_level,
441                                                        start, end,
442                                                        flags);
443                         if (ret == 0)
444                                 goto search_free;
445
446                         goto err_unpin;
447                 }
448
449                 GEM_BUG_ON(vma->node.start < start);
450                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
451         }
452         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level));
453
454         list_move_tail(&obj->global_link, &dev_priv->mm.bound_list);
455         list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
456         obj->bind_count++;
457         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
458
459         return 0;
460
461 err_unpin:
462         i915_gem_object_unpin_pages(obj);
463         return ret;
464 }
465
466 int __i915_vma_do_pin(struct i915_vma *vma,
467                       u64 size, u64 alignment, u64 flags)
468 {
469         unsigned int bound = vma->flags;
470         int ret;
471
472         lockdep_assert_held(&vma->vm->dev->struct_mutex);
473         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
474         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
475
476         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
477                 ret = -EBUSY;
478                 goto err;
479         }
480
481         if ((bound & I915_VMA_BIND_MASK) == 0) {
482                 ret = i915_vma_insert(vma, size, alignment, flags);
483                 if (ret)
484                         goto err;
485         }
486
487         ret = i915_vma_bind(vma, vma->obj->cache_level, flags);
488         if (ret)
489                 goto err;
490
491         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
492                 __i915_vma_set_map_and_fenceable(vma);
493
494         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
495         return 0;
496
497 err:
498         __i915_vma_unpin(vma);
499         return ret;
500 }
501
502 void i915_vma_destroy(struct i915_vma *vma)
503 {
504         GEM_BUG_ON(vma->node.allocated);
505         GEM_BUG_ON(i915_vma_is_active(vma));
506         GEM_BUG_ON(!i915_vma_is_closed(vma));
507         GEM_BUG_ON(vma->fence);
508
509         list_del(&vma->vm_link);
510         if (!i915_vma_is_ggtt(vma))
511                 i915_ppgtt_put(i915_vm_to_ppgtt(vma->vm));
512
513         kmem_cache_free(to_i915(vma->obj->base.dev)->vmas, vma);
514 }
515
516 void i915_vma_close(struct i915_vma *vma)
517 {
518         GEM_BUG_ON(i915_vma_is_closed(vma));
519         vma->flags |= I915_VMA_CLOSED;
520
521         list_del(&vma->obj_link);
522         rb_erase(&vma->obj_node, &vma->obj->vma_tree);
523
524         if (!i915_vma_is_active(vma) && !i915_vma_is_pinned(vma))
525                 WARN_ON(i915_vma_unbind(vma));
526 }
527
528 static void __i915_vma_iounmap(struct i915_vma *vma)
529 {
530         GEM_BUG_ON(i915_vma_is_pinned(vma));
531
532         if (vma->iomap == NULL)
533                 return;
534
535         io_mapping_unmap(vma->iomap);
536         vma->iomap = NULL;
537 }
538
539 int i915_vma_unbind(struct i915_vma *vma)
540 {
541         struct drm_i915_gem_object *obj = vma->obj;
542         unsigned long active;
543         int ret;
544
545         lockdep_assert_held(&obj->base.dev->struct_mutex);
546
547         /* First wait upon any activity as retiring the request may
548          * have side-effects such as unpinning or even unbinding this vma.
549          */
550         active = i915_vma_get_active(vma);
551         if (active) {
552                 int idx;
553
554                 /* When a closed VMA is retired, it is unbound - eek.
555                  * In order to prevent it from being recursively closed,
556                  * take a pin on the vma so that the second unbind is
557                  * aborted.
558                  *
559                  * Even more scary is that the retire callback may free
560                  * the object (last active vma). To prevent the explosion
561                  * we defer the actual object free to a worker that can
562                  * only proceed once it acquires the struct_mutex (which
563                  * we currently hold, therefore it cannot free this object
564                  * before we are finished).
565                  */
566                 __i915_vma_pin(vma);
567
568                 for_each_active(active, idx) {
569                         ret = i915_gem_active_retire(&vma->last_read[idx],
570                                                    &vma->vm->dev->struct_mutex);
571                         if (ret)
572                                 break;
573                 }
574
575                 __i915_vma_unpin(vma);
576                 if (ret)
577                         return ret;
578
579                 GEM_BUG_ON(i915_vma_is_active(vma));
580         }
581
582         if (i915_vma_is_pinned(vma))
583                 return -EBUSY;
584
585         if (!drm_mm_node_allocated(&vma->node))
586                 goto destroy;
587
588         GEM_BUG_ON(obj->bind_count == 0);
589         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
590
591         if (i915_vma_is_map_and_fenceable(vma)) {
592                 /* release the fence reg _after_ flushing */
593                 ret = i915_vma_put_fence(vma);
594                 if (ret)
595                         return ret;
596
597                 /* Force a pagefault for domain tracking on next user access */
598                 i915_gem_release_mmap(obj);
599
600                 __i915_vma_iounmap(vma);
601                 vma->flags &= ~I915_VMA_CAN_FENCE;
602         }
603
604         if (likely(!vma->vm->closed)) {
605                 trace_i915_vma_unbind(vma);
606                 vma->vm->unbind_vma(vma);
607         }
608         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
609
610         drm_mm_remove_node(&vma->node);
611         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
612
613         if (vma->pages != obj->mm.pages) {
614                 GEM_BUG_ON(!vma->pages);
615                 sg_free_table(vma->pages);
616                 kfree(vma->pages);
617         }
618         vma->pages = NULL;
619
620         /* Since the unbound list is global, only move to that list if
621          * no more VMAs exist. */
622         if (--obj->bind_count == 0)
623                 list_move_tail(&obj->global_link,
624                                &to_i915(obj->base.dev)->mm.unbound_list);
625
626         /* And finally now the object is completely decoupled from this vma,
627          * we can drop its hold on the backing storage and allow it to be
628          * reaped by the shrinker.
629          */
630         i915_gem_object_unpin_pages(obj);
631
632 destroy:
633         if (unlikely(i915_vma_is_closed(vma)))
634                 i915_vma_destroy(vma);
635
636         return 0;
637 }
638