]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - drivers/gpu/drm/i915/i915_vma.c
drm/i915: Extract reserving space in the GTT to a helper
[mirror_ubuntu-bionic-kernel.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 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
95
96 if (view) {
97 vma->ggtt_view = *view;
98 if (view->type == I915_GGTT_VIEW_PARTIAL) {
99 GEM_BUG_ON(range_overflows_t(u64,
100 view->params.partial.offset,
101 view->params.partial.size,
102 obj->base.size >> PAGE_SHIFT));
103 vma->size = view->params.partial.size;
104 vma->size <<= PAGE_SHIFT;
105 GEM_BUG_ON(vma->size >= obj->base.size);
106 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
107 vma->size =
108 intel_rotation_info_size(&view->params.rotated);
109 vma->size <<= PAGE_SHIFT;
110 }
111 }
112
113 if (i915_is_ggtt(vm)) {
114 GEM_BUG_ON(overflows_type(vma->size, u32));
115 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
116 i915_gem_object_get_tiling(obj),
117 i915_gem_object_get_stride(obj));
118 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
119
120 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
121 i915_gem_object_get_tiling(obj),
122 i915_gem_object_get_stride(obj));
123 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
124
125 vma->flags |= I915_VMA_GGTT;
126 list_add(&vma->obj_link, &obj->vma_list);
127 } else {
128 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
129 list_add_tail(&vma->obj_link, &obj->vma_list);
130 }
131
132 rb = NULL;
133 p = &obj->vma_tree.rb_node;
134 while (*p) {
135 struct i915_vma *pos;
136
137 rb = *p;
138 pos = rb_entry(rb, struct i915_vma, obj_node);
139 if (i915_vma_compare(pos, vm, view) < 0)
140 p = &rb->rb_right;
141 else
142 p = &rb->rb_left;
143 }
144 rb_link_node(&vma->obj_node, rb, p);
145 rb_insert_color(&vma->obj_node, &obj->vma_tree);
146
147 return vma;
148 }
149
150 struct i915_vma *
151 i915_vma_create(struct drm_i915_gem_object *obj,
152 struct i915_address_space *vm,
153 const struct i915_ggtt_view *view)
154 {
155 lockdep_assert_held(&obj->base.dev->struct_mutex);
156 GEM_BUG_ON(view && !i915_is_ggtt(vm));
157 GEM_BUG_ON(i915_gem_obj_to_vma(obj, vm, view));
158
159 return __i915_vma_create(obj, vm, view);
160 }
161
162 /**
163 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
164 * @vma: VMA to map
165 * @cache_level: mapping cache level
166 * @flags: flags like global or local mapping
167 *
168 * DMA addresses are taken from the scatter-gather table of this object (or of
169 * this VMA in case of non-default GGTT views) and PTE entries set up.
170 * Note that DMA addresses are also the only part of the SG table we care about.
171 */
172 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
173 u32 flags)
174 {
175 u32 bind_flags;
176 u32 vma_flags;
177 int ret;
178
179 if (WARN_ON(flags == 0))
180 return -EINVAL;
181
182 bind_flags = 0;
183 if (flags & PIN_GLOBAL)
184 bind_flags |= I915_VMA_GLOBAL_BIND;
185 if (flags & PIN_USER)
186 bind_flags |= I915_VMA_LOCAL_BIND;
187
188 vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
189 if (flags & PIN_UPDATE)
190 bind_flags |= vma_flags;
191 else
192 bind_flags &= ~vma_flags;
193 if (bind_flags == 0)
194 return 0;
195
196 if (GEM_WARN_ON(range_overflows(vma->node.start,
197 vma->node.size,
198 vma->vm->total)))
199 return -ENODEV;
200
201 if (vma_flags == 0 && vma->vm->allocate_va_range) {
202 trace_i915_va_alloc(vma);
203 ret = vma->vm->allocate_va_range(vma->vm,
204 vma->node.start,
205 vma->node.size);
206 if (ret)
207 return ret;
208 }
209
210 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
211 if (ret)
212 return ret;
213
214 vma->flags |= bind_flags;
215 return 0;
216 }
217
218 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
219 {
220 void __iomem *ptr;
221
222 /* Access through the GTT requires the device to be awake. */
223 assert_rpm_wakelock_held(vma->vm->i915);
224
225 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
226 if (WARN_ON(!i915_vma_is_map_and_fenceable(vma)))
227 return IO_ERR_PTR(-ENODEV);
228
229 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
230 GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
231
232 ptr = vma->iomap;
233 if (ptr == NULL) {
234 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->mappable,
235 vma->node.start,
236 vma->node.size);
237 if (ptr == NULL)
238 return IO_ERR_PTR(-ENOMEM);
239
240 vma->iomap = ptr;
241 }
242
243 __i915_vma_pin(vma);
244 return ptr;
245 }
246
247 void i915_vma_unpin_and_release(struct i915_vma **p_vma)
248 {
249 struct i915_vma *vma;
250 struct drm_i915_gem_object *obj;
251
252 vma = fetch_and_zero(p_vma);
253 if (!vma)
254 return;
255
256 obj = vma->obj;
257
258 i915_vma_unpin(vma);
259 i915_vma_close(vma);
260
261 __i915_gem_object_release_unless_active(obj);
262 }
263
264 bool
265 i915_vma_misplaced(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
266 {
267 if (!drm_mm_node_allocated(&vma->node))
268 return false;
269
270 if (vma->node.size < size)
271 return true;
272
273 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
274 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
275 return true;
276
277 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
278 return true;
279
280 if (flags & PIN_OFFSET_BIAS &&
281 vma->node.start < (flags & PIN_OFFSET_MASK))
282 return true;
283
284 if (flags & PIN_OFFSET_FIXED &&
285 vma->node.start != (flags & PIN_OFFSET_MASK))
286 return true;
287
288 return false;
289 }
290
291 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
292 {
293 bool mappable, fenceable;
294
295 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
296 GEM_BUG_ON(!vma->fence_size);
297
298 /*
299 * Explicitly disable for rotated VMA since the display does not
300 * need the fence and the VMA is not accessible to other users.
301 */
302 if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
303 return;
304
305 fenceable = (vma->node.size >= vma->fence_size &&
306 IS_ALIGNED(vma->node.start, vma->fence_alignment));
307
308 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
309
310 if (mappable && fenceable)
311 vma->flags |= I915_VMA_CAN_FENCE;
312 else
313 vma->flags &= ~I915_VMA_CAN_FENCE;
314 }
315
316 static bool color_differs(struct drm_mm_node *node, unsigned long color)
317 {
318 return node->allocated && node->color != color;
319 }
320
321 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
322 {
323 struct drm_mm_node *node = &vma->node;
324 struct drm_mm_node *other;
325
326 /*
327 * On some machines we have to be careful when putting differing types
328 * of snoopable memory together to avoid the prefetcher crossing memory
329 * domains and dying. During vm initialisation, we decide whether or not
330 * these constraints apply and set the drm_mm.color_adjust
331 * appropriately.
332 */
333 if (vma->vm->mm.color_adjust == NULL)
334 return true;
335
336 /* Only valid to be called on an already inserted vma */
337 GEM_BUG_ON(!drm_mm_node_allocated(node));
338 GEM_BUG_ON(list_empty(&node->node_list));
339
340 other = list_prev_entry(node, node_list);
341 if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
342 return false;
343
344 other = list_next_entry(node, node_list);
345 if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
346 return false;
347
348 return true;
349 }
350
351 /**
352 * i915_vma_insert - finds a slot for the vma in its address space
353 * @vma: the vma
354 * @size: requested size in bytes (can be larger than the VMA)
355 * @alignment: required alignment
356 * @flags: mask of PIN_* flags to use
357 *
358 * First we try to allocate some free space that meets the requirements for
359 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
360 * preferrably the oldest idle entry to make room for the new VMA.
361 *
362 * Returns:
363 * 0 on success, negative error code otherwise.
364 */
365 static int
366 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
367 {
368 struct drm_i915_private *dev_priv = vma->vm->i915;
369 struct drm_i915_gem_object *obj = vma->obj;
370 u64 start, end;
371 int ret;
372
373 GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
374 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
375
376 size = max(size, vma->size);
377 alignment = max(alignment, vma->display_alignment);
378 if (flags & PIN_MAPPABLE) {
379 size = max_t(typeof(size), size, vma->fence_size);
380 alignment = max_t(typeof(alignment),
381 alignment, vma->fence_alignment);
382 }
383
384 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
385 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
386 GEM_BUG_ON(!is_power_of_2(alignment));
387
388 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
389 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
390
391 end = vma->vm->total;
392 if (flags & PIN_MAPPABLE)
393 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
394 if (flags & PIN_ZONE_4G)
395 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
396 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
397
398 /* If binding the object/GGTT view requires more space than the entire
399 * aperture has, reject it early before evicting everything in a vain
400 * attempt to find space.
401 */
402 if (size > end) {
403 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu [object=%zd] > %s aperture=%llu\n",
404 size, obj->base.size,
405 flags & PIN_MAPPABLE ? "mappable" : "total",
406 end);
407 return -E2BIG;
408 }
409
410 ret = i915_gem_object_pin_pages(obj);
411 if (ret)
412 return ret;
413
414 if (flags & PIN_OFFSET_FIXED) {
415 u64 offset = flags & PIN_OFFSET_MASK;
416 if (!IS_ALIGNED(offset, alignment) ||
417 range_overflows(offset, size, end)) {
418 ret = -EINVAL;
419 goto err_unpin;
420 }
421
422 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
423 size, offset, obj->cache_level,
424 flags);
425 if (ret)
426 goto err_unpin;
427 } else {
428 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
429 size, alignment, obj->cache_level,
430 start, end, flags);
431 if (ret)
432 goto err_unpin;
433
434 GEM_BUG_ON(vma->node.start < start);
435 GEM_BUG_ON(vma->node.start + vma->node.size > end);
436 }
437 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level));
438
439 list_move_tail(&obj->global_link, &dev_priv->mm.bound_list);
440 list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
441 obj->bind_count++;
442 GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
443
444 return 0;
445
446 err_unpin:
447 i915_gem_object_unpin_pages(obj);
448 return ret;
449 }
450
451 int __i915_vma_do_pin(struct i915_vma *vma,
452 u64 size, u64 alignment, u64 flags)
453 {
454 unsigned int bound = vma->flags;
455 int ret;
456
457 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
458 GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
459 GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
460
461 if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
462 ret = -EBUSY;
463 goto err;
464 }
465
466 if ((bound & I915_VMA_BIND_MASK) == 0) {
467 ret = i915_vma_insert(vma, size, alignment, flags);
468 if (ret)
469 goto err;
470 }
471
472 ret = i915_vma_bind(vma, vma->obj->cache_level, flags);
473 if (ret)
474 goto err;
475
476 if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
477 __i915_vma_set_map_and_fenceable(vma);
478
479 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
480 return 0;
481
482 err:
483 __i915_vma_unpin(vma);
484 return ret;
485 }
486
487 void i915_vma_destroy(struct i915_vma *vma)
488 {
489 GEM_BUG_ON(vma->node.allocated);
490 GEM_BUG_ON(i915_vma_is_active(vma));
491 GEM_BUG_ON(!i915_vma_is_closed(vma));
492 GEM_BUG_ON(vma->fence);
493
494 list_del(&vma->vm_link);
495 if (!i915_vma_is_ggtt(vma))
496 i915_ppgtt_put(i915_vm_to_ppgtt(vma->vm));
497
498 kmem_cache_free(to_i915(vma->obj->base.dev)->vmas, vma);
499 }
500
501 void i915_vma_close(struct i915_vma *vma)
502 {
503 GEM_BUG_ON(i915_vma_is_closed(vma));
504 vma->flags |= I915_VMA_CLOSED;
505
506 list_del(&vma->obj_link);
507 rb_erase(&vma->obj_node, &vma->obj->vma_tree);
508
509 if (!i915_vma_is_active(vma) && !i915_vma_is_pinned(vma))
510 WARN_ON(i915_vma_unbind(vma));
511 }
512
513 static void __i915_vma_iounmap(struct i915_vma *vma)
514 {
515 GEM_BUG_ON(i915_vma_is_pinned(vma));
516
517 if (vma->iomap == NULL)
518 return;
519
520 io_mapping_unmap(vma->iomap);
521 vma->iomap = NULL;
522 }
523
524 int i915_vma_unbind(struct i915_vma *vma)
525 {
526 struct drm_i915_gem_object *obj = vma->obj;
527 unsigned long active;
528 int ret;
529
530 lockdep_assert_held(&obj->base.dev->struct_mutex);
531
532 /* First wait upon any activity as retiring the request may
533 * have side-effects such as unpinning or even unbinding this vma.
534 */
535 active = i915_vma_get_active(vma);
536 if (active) {
537 int idx;
538
539 /* When a closed VMA is retired, it is unbound - eek.
540 * In order to prevent it from being recursively closed,
541 * take a pin on the vma so that the second unbind is
542 * aborted.
543 *
544 * Even more scary is that the retire callback may free
545 * the object (last active vma). To prevent the explosion
546 * we defer the actual object free to a worker that can
547 * only proceed once it acquires the struct_mutex (which
548 * we currently hold, therefore it cannot free this object
549 * before we are finished).
550 */
551 __i915_vma_pin(vma);
552
553 for_each_active(active, idx) {
554 ret = i915_gem_active_retire(&vma->last_read[idx],
555 &vma->vm->i915->drm.struct_mutex);
556 if (ret)
557 break;
558 }
559
560 __i915_vma_unpin(vma);
561 if (ret)
562 return ret;
563
564 GEM_BUG_ON(i915_vma_is_active(vma));
565 }
566
567 if (i915_vma_is_pinned(vma))
568 return -EBUSY;
569
570 if (!drm_mm_node_allocated(&vma->node))
571 goto destroy;
572
573 GEM_BUG_ON(obj->bind_count == 0);
574 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
575
576 if (i915_vma_is_map_and_fenceable(vma)) {
577 /* release the fence reg _after_ flushing */
578 ret = i915_vma_put_fence(vma);
579 if (ret)
580 return ret;
581
582 /* Force a pagefault for domain tracking on next user access */
583 i915_gem_release_mmap(obj);
584
585 __i915_vma_iounmap(vma);
586 vma->flags &= ~I915_VMA_CAN_FENCE;
587 }
588
589 if (likely(!vma->vm->closed)) {
590 trace_i915_vma_unbind(vma);
591 vma->vm->unbind_vma(vma);
592 }
593 vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
594
595 drm_mm_remove_node(&vma->node);
596 list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
597
598 if (vma->pages != obj->mm.pages) {
599 GEM_BUG_ON(!vma->pages);
600 sg_free_table(vma->pages);
601 kfree(vma->pages);
602 }
603 vma->pages = NULL;
604
605 /* Since the unbound list is global, only move to that list if
606 * no more VMAs exist. */
607 if (--obj->bind_count == 0)
608 list_move_tail(&obj->global_link,
609 &to_i915(obj->base.dev)->mm.unbound_list);
610
611 /* And finally now the object is completely decoupled from this vma,
612 * we can drop its hold on the backing storage and allow it to be
613 * reaped by the shrinker.
614 */
615 i915_gem_object_unpin_pages(obj);
616 GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
617
618 destroy:
619 if (unlikely(i915_vma_is_closed(vma)))
620 i915_vma_destroy(vma);
621
622 return 0;
623 }
624