]> git.karo-electronics.de Git - karo-tx-linux.git/blob - drivers/gpu/drm/i915/i915_vma.c
9e121222c5ebfeb007c45a70f68617dd1e78e531
[karo-tx-linux.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 (GEM_WARN_ON(vma->node.start + vma->node.size < vma->node.start))
180                 return -ENODEV;
181
182         if (GEM_WARN_ON(vma->node.start + vma->node.size > vma->vm->total))
183                 return -ENODEV;
184
185         if (vma_flags == 0 && vma->vm->allocate_va_range) {
186                 trace_i915_va_alloc(vma);
187                 ret = vma->vm->allocate_va_range(vma->vm,
188                                                  vma->node.start,
189                                                  vma->node.size);
190                 if (ret)
191                         return ret;
192         }
193
194         ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
195         if (ret)
196                 return ret;
197
198         vma->flags |= bind_flags;
199         return 0;
200 }
201
202 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
203 {
204         void __iomem *ptr;
205
206         /* Access through the GTT requires the device to be awake. */
207         assert_rpm_wakelock_held(vma->vm->i915);
208
209         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
210         if (WARN_ON(!i915_vma_is_map_and_fenceable(vma)))
211                 return IO_ERR_PTR(-ENODEV);
212
213         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
214         GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
215
216         ptr = vma->iomap;
217         if (ptr == NULL) {
218                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->mappable,
219                                         vma->node.start,
220                                         vma->node.size);
221                 if (ptr == NULL)
222                         return IO_ERR_PTR(-ENOMEM);
223
224                 vma->iomap = ptr;
225         }
226
227         __i915_vma_pin(vma);
228         return ptr;
229 }
230
231 void i915_vma_unpin_and_release(struct i915_vma **p_vma)
232 {
233         struct i915_vma *vma;
234         struct drm_i915_gem_object *obj;
235
236         vma = fetch_and_zero(p_vma);
237         if (!vma)
238                 return;
239
240         obj = vma->obj;
241
242         i915_vma_unpin(vma);
243         i915_vma_close(vma);
244
245         __i915_gem_object_release_unless_active(obj);
246 }
247
248 bool
249 i915_vma_misplaced(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
250 {
251         if (!drm_mm_node_allocated(&vma->node))
252                 return false;
253
254         if (vma->node.size < size)
255                 return true;
256
257         if (alignment && vma->node.start & (alignment - 1))
258                 return true;
259
260         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
261                 return true;
262
263         if (flags & PIN_OFFSET_BIAS &&
264             vma->node.start < (flags & PIN_OFFSET_MASK))
265                 return true;
266
267         if (flags & PIN_OFFSET_FIXED &&
268             vma->node.start != (flags & PIN_OFFSET_MASK))
269                 return true;
270
271         return false;
272 }
273
274 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
275 {
276         struct drm_i915_gem_object *obj = vma->obj;
277         struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
278         bool mappable, fenceable;
279         u32 fence_size, fence_alignment;
280
281         fence_size = i915_gem_get_ggtt_size(dev_priv,
282                                             vma->size,
283                                             i915_gem_object_get_tiling(obj));
284         fence_alignment = i915_gem_get_ggtt_alignment(dev_priv,
285                                                       vma->size,
286                                                       i915_gem_object_get_tiling(obj),
287                                                       true);
288
289         fenceable = (vma->node.size == fence_size &&
290                      (vma->node.start & (fence_alignment - 1)) == 0);
291
292         mappable = (vma->node.start + fence_size <=
293                     dev_priv->ggtt.mappable_end);
294
295         /*
296          * Explicitly disable for rotated VMA since the display does not
297          * need the fence and the VMA is not accessible to other users.
298          */
299         if (mappable && fenceable &&
300             vma->ggtt_view.type != I915_GGTT_VIEW_ROTATED)
301                 vma->flags |= I915_VMA_CAN_FENCE;
302         else
303                 vma->flags &= ~I915_VMA_CAN_FENCE;
304 }
305
306 static bool color_differs(struct drm_mm_node *node, unsigned long color)
307 {
308         return node->allocated && node->color != color;
309 }
310
311 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
312 {
313         struct drm_mm_node *node = &vma->node;
314         struct drm_mm_node *other;
315
316         /*
317          * On some machines we have to be careful when putting differing types
318          * of snoopable memory together to avoid the prefetcher crossing memory
319          * domains and dying. During vm initialisation, we decide whether or not
320          * these constraints apply and set the drm_mm.color_adjust
321          * appropriately.
322          */
323         if (vma->vm->mm.color_adjust == NULL)
324                 return true;
325
326         /* Only valid to be called on an already inserted vma */
327         GEM_BUG_ON(!drm_mm_node_allocated(node));
328         GEM_BUG_ON(list_empty(&node->node_list));
329
330         other = list_prev_entry(node, node_list);
331         if (color_differs(other, cache_level) && !other->hole_follows)
332                 return false;
333
334         other = list_next_entry(node, node_list);
335         if (color_differs(other, cache_level) && !node->hole_follows)
336                 return false;
337
338         return true;
339 }
340
341 /**
342  * i915_vma_insert - finds a slot for the vma in its address space
343  * @vma: the vma
344  * @size: requested size in bytes (can be larger than the VMA)
345  * @alignment: required alignment
346  * @flags: mask of PIN_* flags to use
347  *
348  * First we try to allocate some free space that meets the requirements for
349  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
350  * preferrably the oldest idle entry to make room for the new VMA.
351  *
352  * Returns:
353  * 0 on success, negative error code otherwise.
354  */
355 static int
356 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
357 {
358         struct drm_i915_private *dev_priv = vma->vm->i915;
359         struct drm_i915_gem_object *obj = vma->obj;
360         u64 start, end;
361         int ret;
362
363         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
364         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
365
366         size = max(size, vma->size);
367         if (flags & PIN_MAPPABLE)
368                 size = i915_gem_get_ggtt_size(dev_priv, size,
369                                               i915_gem_object_get_tiling(obj));
370
371         alignment = max(max(alignment, vma->display_alignment),
372                         i915_gem_get_ggtt_alignment(dev_priv, size,
373                                                     i915_gem_object_get_tiling(obj),
374                                                     flags & PIN_MAPPABLE));
375
376         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
377
378         end = vma->vm->total;
379         if (flags & PIN_MAPPABLE)
380                 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
381         if (flags & PIN_ZONE_4G)
382                 end = min_t(u64, end, (1ULL << 32) - PAGE_SIZE);
383
384         /* If binding the object/GGTT view requires more space than the entire
385          * aperture has, reject it early before evicting everything in a vain
386          * attempt to find space.
387          */
388         if (size > end) {
389                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu [object=%zd] > %s aperture=%llu\n",
390                           size, obj->base.size,
391                           flags & PIN_MAPPABLE ? "mappable" : "total",
392                           end);
393                 return -E2BIG;
394         }
395
396         ret = i915_gem_object_pin_pages(obj);
397         if (ret)
398                 return ret;
399
400         if (flags & PIN_OFFSET_FIXED) {
401                 u64 offset = flags & PIN_OFFSET_MASK;
402                 if (offset & (alignment - 1) || offset > end - size) {
403                         ret = -EINVAL;
404                         goto err_unpin;
405                 }
406
407                 vma->node.start = offset;
408                 vma->node.size = size;
409                 vma->node.color = obj->cache_level;
410                 ret = drm_mm_reserve_node(&vma->vm->mm, &vma->node);
411                 if (ret) {
412                         ret = i915_gem_evict_for_vma(vma, flags);
413                         if (ret == 0)
414                                 ret = drm_mm_reserve_node(&vma->vm->mm, &vma->node);
415                         if (ret)
416                                 goto err_unpin;
417                 }
418         } else {
419                 u32 search_flag, alloc_flag;
420
421                 if (flags & PIN_HIGH) {
422                         search_flag = DRM_MM_SEARCH_BELOW;
423                         alloc_flag = DRM_MM_CREATE_TOP;
424                 } else {
425                         search_flag = DRM_MM_SEARCH_DEFAULT;
426                         alloc_flag = DRM_MM_CREATE_DEFAULT;
427                 }
428
429                 /* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks,
430                  * so we know that we always have a minimum alignment of 4096.
431                  * The drm_mm range manager is optimised to return results
432                  * with zero alignment, so where possible use the optimal
433                  * path.
434                  */
435                 if (alignment <= 4096)
436                         alignment = 0;
437
438 search_free:
439                 ret = drm_mm_insert_node_in_range_generic(&vma->vm->mm,
440                                                           &vma->node,
441                                                           size, alignment,
442                                                           obj->cache_level,
443                                                           start, end,
444                                                           search_flag,
445                                                           alloc_flag);
446                 if (ret) {
447                         ret = i915_gem_evict_something(vma->vm, size, alignment,
448                                                        obj->cache_level,
449                                                        start, end,
450                                                        flags);
451                         if (ret == 0)
452                                 goto search_free;
453
454                         goto err_unpin;
455                 }
456
457                 GEM_BUG_ON(vma->node.start < start);
458                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
459         }
460         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level));
461
462         list_move_tail(&obj->global_link, &dev_priv->mm.bound_list);
463         list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
464         obj->bind_count++;
465         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
466
467         return 0;
468
469 err_unpin:
470         i915_gem_object_unpin_pages(obj);
471         return ret;
472 }
473
474 int __i915_vma_do_pin(struct i915_vma *vma,
475                       u64 size, u64 alignment, u64 flags)
476 {
477         unsigned int bound = vma->flags;
478         int ret;
479
480         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
481         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
482         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
483
484         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
485                 ret = -EBUSY;
486                 goto err;
487         }
488
489         if ((bound & I915_VMA_BIND_MASK) == 0) {
490                 ret = i915_vma_insert(vma, size, alignment, flags);
491                 if (ret)
492                         goto err;
493         }
494
495         ret = i915_vma_bind(vma, vma->obj->cache_level, flags);
496         if (ret)
497                 goto err;
498
499         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
500                 __i915_vma_set_map_and_fenceable(vma);
501
502         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
503         return 0;
504
505 err:
506         __i915_vma_unpin(vma);
507         return ret;
508 }
509
510 void i915_vma_destroy(struct i915_vma *vma)
511 {
512         GEM_BUG_ON(vma->node.allocated);
513         GEM_BUG_ON(i915_vma_is_active(vma));
514         GEM_BUG_ON(!i915_vma_is_closed(vma));
515         GEM_BUG_ON(vma->fence);
516
517         list_del(&vma->vm_link);
518         if (!i915_vma_is_ggtt(vma))
519                 i915_ppgtt_put(i915_vm_to_ppgtt(vma->vm));
520
521         kmem_cache_free(to_i915(vma->obj->base.dev)->vmas, vma);
522 }
523
524 void i915_vma_close(struct i915_vma *vma)
525 {
526         GEM_BUG_ON(i915_vma_is_closed(vma));
527         vma->flags |= I915_VMA_CLOSED;
528
529         list_del(&vma->obj_link);
530         rb_erase(&vma->obj_node, &vma->obj->vma_tree);
531
532         if (!i915_vma_is_active(vma) && !i915_vma_is_pinned(vma))
533                 WARN_ON(i915_vma_unbind(vma));
534 }
535
536 static void __i915_vma_iounmap(struct i915_vma *vma)
537 {
538         GEM_BUG_ON(i915_vma_is_pinned(vma));
539
540         if (vma->iomap == NULL)
541                 return;
542
543         io_mapping_unmap(vma->iomap);
544         vma->iomap = NULL;
545 }
546
547 int i915_vma_unbind(struct i915_vma *vma)
548 {
549         struct drm_i915_gem_object *obj = vma->obj;
550         unsigned long active;
551         int ret;
552
553         lockdep_assert_held(&obj->base.dev->struct_mutex);
554
555         /* First wait upon any activity as retiring the request may
556          * have side-effects such as unpinning or even unbinding this vma.
557          */
558         active = i915_vma_get_active(vma);
559         if (active) {
560                 int idx;
561
562                 /* When a closed VMA is retired, it is unbound - eek.
563                  * In order to prevent it from being recursively closed,
564                  * take a pin on the vma so that the second unbind is
565                  * aborted.
566                  *
567                  * Even more scary is that the retire callback may free
568                  * the object (last active vma). To prevent the explosion
569                  * we defer the actual object free to a worker that can
570                  * only proceed once it acquires the struct_mutex (which
571                  * we currently hold, therefore it cannot free this object
572                  * before we are finished).
573                  */
574                 __i915_vma_pin(vma);
575
576                 for_each_active(active, idx) {
577                         ret = i915_gem_active_retire(&vma->last_read[idx],
578                                                      &vma->vm->i915->drm.struct_mutex);
579                         if (ret)
580                                 break;
581                 }
582
583                 __i915_vma_unpin(vma);
584                 if (ret)
585                         return ret;
586
587                 GEM_BUG_ON(i915_vma_is_active(vma));
588         }
589
590         if (i915_vma_is_pinned(vma))
591                 return -EBUSY;
592
593         if (!drm_mm_node_allocated(&vma->node))
594                 goto destroy;
595
596         GEM_BUG_ON(obj->bind_count == 0);
597         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
598
599         if (i915_vma_is_map_and_fenceable(vma)) {
600                 /* release the fence reg _after_ flushing */
601                 ret = i915_vma_put_fence(vma);
602                 if (ret)
603                         return ret;
604
605                 /* Force a pagefault for domain tracking on next user access */
606                 i915_gem_release_mmap(obj);
607
608                 __i915_vma_iounmap(vma);
609                 vma->flags &= ~I915_VMA_CAN_FENCE;
610         }
611
612         if (likely(!vma->vm->closed)) {
613                 trace_i915_vma_unbind(vma);
614                 vma->vm->unbind_vma(vma);
615         }
616         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
617
618         drm_mm_remove_node(&vma->node);
619         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
620
621         if (vma->pages != obj->mm.pages) {
622                 GEM_BUG_ON(!vma->pages);
623                 sg_free_table(vma->pages);
624                 kfree(vma->pages);
625         }
626         vma->pages = NULL;
627
628         /* Since the unbound list is global, only move to that list if
629          * no more VMAs exist. */
630         if (--obj->bind_count == 0)
631                 list_move_tail(&obj->global_link,
632                                &to_i915(obj->base.dev)->mm.unbound_list);
633
634         /* And finally now the object is completely decoupled from this vma,
635          * we can drop its hold on the backing storage and allow it to be
636          * reaped by the shrinker.
637          */
638         i915_gem_object_unpin_pages(obj);
639
640 destroy:
641         if (unlikely(i915_vma_is_closed(vma)))
642                 i915_vma_destroy(vma);
643
644         return 0;
645 }
646