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
i915_vma_retire(struct i915_gem_active * active,struct drm_i915_gem_request * rq)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 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
49 list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
50 if (unlikely(i915_vma_is_closed(vma) && !i915_vma_is_pinned(vma)))
51 WARN_ON(i915_vma_unbind(vma));
52
53 GEM_BUG_ON(!i915_gem_object_is_active(obj));
54 if (--obj->active_count)
55 return;
56
57 /* Prune the shared fence arrays iff completely idle (inc. external) */
58 if (reservation_object_trylock(obj->resv)) {
59 if (reservation_object_test_signaled_rcu(obj->resv, true))
60 reservation_object_add_excl_fence(obj->resv, NULL);
61 reservation_object_unlock(obj->resv);
62 }
63
64 /* Bump our place on the bound list to keep it roughly in LRU order
65 * so that we don't steal from recently used but inactive objects
66 * (unless we are forced to ofc!)
67 */
68 lockmgr(&rq->i915->mm.obj_lock, LK_EXCLUSIVE);
69 if (obj->bind_count)
70 list_move_tail(&obj->mm.link, &rq->i915->mm.bound_list);
71 lockmgr(&rq->i915->mm.obj_lock, LK_RELEASE);
72
73 obj->mm.dirty = true; /* be paranoid */
74
75 if (i915_gem_object_has_active_reference(obj)) {
76 i915_gem_object_clear_active_reference(obj);
77 i915_gem_object_put(obj);
78 }
79 }
80
81 static struct i915_vma *
vma_create(struct drm_i915_gem_object * obj,struct i915_address_space * vm,const struct i915_ggtt_view * view)82 vma_create(struct drm_i915_gem_object *obj,
83 struct i915_address_space *vm,
84 const struct i915_ggtt_view *view)
85 {
86 struct i915_vma *vma;
87 struct rb_node *rb, **p;
88 int i;
89
90 /* The aliasing_ppgtt should never be used directly! */
91 GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->base);
92
93 vma = kmem_cache_zalloc(vm->i915->vmas, GFP_KERNEL);
94 if (vma == NULL)
95 return ERR_PTR(-ENOMEM);
96
97 for (i = 0; i < ARRAY_SIZE(vma->last_read); i++)
98 init_request_active(&vma->last_read[i], i915_vma_retire);
99 init_request_active(&vma->last_fence, NULL);
100 vma->vm = vm;
101 vma->obj = obj;
102 vma->resv = obj->resv;
103 vma->size = obj->base.size;
104 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
105
106 if (view && view->type != I915_GGTT_VIEW_NORMAL) {
107 vma->ggtt_view = *view;
108 if (view->type == I915_GGTT_VIEW_PARTIAL) {
109 GEM_BUG_ON(range_overflows_t(u64,
110 view->partial.offset,
111 view->partial.size,
112 obj->base.size >> PAGE_SHIFT));
113 vma->size = view->partial.size;
114 vma->size <<= PAGE_SHIFT;
115 GEM_BUG_ON(vma->size >= obj->base.size);
116 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
117 vma->size = intel_rotation_info_size(&view->rotated);
118 vma->size <<= PAGE_SHIFT;
119 }
120 }
121
122 if (unlikely(vma->size > vm->total))
123 goto err_vma;
124
125 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
126
127 if (i915_is_ggtt(vm)) {
128 if (unlikely(overflows_type(vma->size, u32)))
129 goto err_vma;
130
131 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
132 i915_gem_object_get_tiling(obj),
133 i915_gem_object_get_stride(obj));
134 if (unlikely(vma->fence_size < vma->size || /* overflow */
135 vma->fence_size > vm->total))
136 goto err_vma;
137
138 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
139
140 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
141 i915_gem_object_get_tiling(obj),
142 i915_gem_object_get_stride(obj));
143 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
144
145 vma->flags |= I915_VMA_GGTT;
146 list_add(&vma->obj_link, &obj->vma_list);
147 } else {
148 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
149 list_add_tail(&vma->obj_link, &obj->vma_list);
150 }
151
152 rb = NULL;
153 p = &obj->vma_tree.rb_node;
154 while (*p) {
155 struct i915_vma *pos;
156
157 rb = *p;
158 pos = rb_entry(rb, struct i915_vma, obj_node);
159 if (i915_vma_compare(pos, vm, view) < 0)
160 p = &rb->rb_right;
161 else
162 p = &rb->rb_left;
163 }
164 rb_link_node(&vma->obj_node, rb, p);
165 rb_insert_color(&vma->obj_node, &obj->vma_tree);
166 list_add(&vma->vm_link, &vm->unbound_list);
167
168 return vma;
169
170 err_vma:
171 kmem_cache_free(vm->i915->vmas, vma);
172 return ERR_PTR(-E2BIG);
173 }
174
175 static struct i915_vma *
vma_lookup(struct drm_i915_gem_object * obj,struct i915_address_space * vm,const struct i915_ggtt_view * view)176 vma_lookup(struct drm_i915_gem_object *obj,
177 struct i915_address_space *vm,
178 const struct i915_ggtt_view *view)
179 {
180 struct rb_node *rb;
181
182 rb = obj->vma_tree.rb_node;
183 while (rb) {
184 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
185 long cmp;
186
187 cmp = i915_vma_compare(vma, vm, view);
188 if (cmp == 0)
189 return vma;
190
191 if (cmp < 0)
192 rb = rb->rb_right;
193 else
194 rb = rb->rb_left;
195 }
196
197 return NULL;
198 }
199
200 /**
201 * i915_vma_instance - return the singleton instance of the VMA
202 * @obj: parent &struct drm_i915_gem_object to be mapped
203 * @vm: address space in which the mapping is located
204 * @view: additional mapping requirements
205 *
206 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
207 * the same @view characteristics. If a match is not found, one is created.
208 * Once created, the VMA is kept until either the object is freed, or the
209 * address space is closed.
210 *
211 * Must be called with struct_mutex held.
212 *
213 * Returns the vma, or an error pointer.
214 */
215 struct i915_vma *
i915_vma_instance(struct drm_i915_gem_object * obj,struct i915_address_space * vm,const struct i915_ggtt_view * view)216 i915_vma_instance(struct drm_i915_gem_object *obj,
217 struct i915_address_space *vm,
218 const struct i915_ggtt_view *view)
219 {
220 struct i915_vma *vma;
221
222 lockdep_assert_held(&obj->base.dev->struct_mutex);
223 GEM_BUG_ON(view && !i915_is_ggtt(vm));
224 GEM_BUG_ON(vm->closed);
225
226 vma = vma_lookup(obj, vm, view);
227 if (!vma)
228 vma = vma_create(obj, vm, view);
229
230 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_is_closed(vma));
231 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
232 GEM_BUG_ON(!IS_ERR(vma) && vma_lookup(obj, vm, view) != vma);
233 return vma;
234 }
235
236 /**
237 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
238 * @vma: VMA to map
239 * @cache_level: mapping cache level
240 * @flags: flags like global or local mapping
241 *
242 * DMA addresses are taken from the scatter-gather table of this object (or of
243 * this VMA in case of non-default GGTT views) and PTE entries set up.
244 * Note that DMA addresses are also the only part of the SG table we care about.
245 */
i915_vma_bind(struct i915_vma * vma,enum i915_cache_level cache_level,u32 flags)246 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
247 u32 flags)
248 {
249 u32 bind_flags;
250 u32 vma_flags;
251 int ret;
252
253 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
254 GEM_BUG_ON(vma->size > vma->node.size);
255
256 if (GEM_WARN_ON(range_overflows(vma->node.start,
257 vma->node.size,
258 vma->vm->total)))
259 return -ENODEV;
260
261 if (GEM_WARN_ON(!flags))
262 return -EINVAL;
263
264 bind_flags = 0;
265 if (flags & PIN_GLOBAL)
266 bind_flags |= I915_VMA_GLOBAL_BIND;
267 if (flags & PIN_USER)
268 bind_flags |= I915_VMA_LOCAL_BIND;
269
270 vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
271 if (flags & PIN_UPDATE)
272 bind_flags |= vma_flags;
273 else
274 bind_flags &= ~vma_flags;
275 if (bind_flags == 0)
276 return 0;
277
278 GEM_BUG_ON(!vma->pages);
279
280 trace_i915_vma_bind(vma, bind_flags);
281 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
282 if (ret)
283 return ret;
284
285 vma->flags |= bind_flags;
286 return 0;
287 }
288
i915_vma_pin_iomap(struct i915_vma * vma)289 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
290 {
291 void __iomem *ptr;
292 int err;
293
294 /* Access through the GTT requires the device to be awake. */
295 assert_rpm_wakelock_held(vma->vm->i915);
296
297 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
298 if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
299 err = -ENODEV;
300 goto err;
301 }
302
303 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
304 GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
305
306 ptr = vma->iomap;
307 if (ptr == NULL) {
308 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->mappable,
309 vma->node.start,
310 vma->node.size);
311 if (ptr == NULL) {
312 err = -ENOMEM;
313 goto err;
314 }
315
316 vma->iomap = ptr;
317 }
318
319 __i915_vma_pin(vma);
320
321 err = i915_vma_pin_fence(vma);
322 if (err)
323 goto err_unpin;
324
325 return ptr;
326
327 err_unpin:
328 __i915_vma_unpin(vma);
329 err:
330 return IO_ERR_PTR(err);
331 }
332
i915_vma_unpin_iomap(struct i915_vma * vma)333 void i915_vma_unpin_iomap(struct i915_vma *vma)
334 {
335 lockdep_assert_held(&vma->obj->base.dev->struct_mutex);
336
337 GEM_BUG_ON(vma->iomap == NULL);
338
339 i915_vma_unpin_fence(vma);
340 i915_vma_unpin(vma);
341 }
342
i915_vma_unpin_and_release(struct i915_vma ** p_vma)343 void i915_vma_unpin_and_release(struct i915_vma **p_vma)
344 {
345 struct i915_vma *vma;
346 struct drm_i915_gem_object *obj;
347
348 vma = fetch_and_zero(p_vma);
349 if (!vma)
350 return;
351
352 obj = vma->obj;
353
354 i915_vma_unpin(vma);
355 i915_vma_close(vma);
356
357 __i915_gem_object_release_unless_active(obj);
358 }
359
i915_vma_misplaced(const struct i915_vma * vma,u64 size,u64 alignment,u64 flags)360 bool i915_vma_misplaced(const struct i915_vma *vma,
361 u64 size, u64 alignment, u64 flags)
362 {
363 if (!drm_mm_node_allocated(&vma->node))
364 return false;
365
366 if (vma->node.size < size)
367 return true;
368
369 GEM_BUG_ON(alignment && !is_power_of_2(alignment));
370 if (alignment && !IS_ALIGNED(vma->node.start, alignment))
371 return true;
372
373 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
374 return true;
375
376 if (flags & PIN_OFFSET_BIAS &&
377 vma->node.start < (flags & PIN_OFFSET_MASK))
378 return true;
379
380 if (flags & PIN_OFFSET_FIXED &&
381 vma->node.start != (flags & PIN_OFFSET_MASK))
382 return true;
383
384 return false;
385 }
386
__i915_vma_set_map_and_fenceable(struct i915_vma * vma)387 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
388 {
389 bool mappable, fenceable;
390
391 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
392 GEM_BUG_ON(!vma->fence_size);
393
394 /*
395 * Explicitly disable for rotated VMA since the display does not
396 * need the fence and the VMA is not accessible to other users.
397 */
398 if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
399 return;
400
401 fenceable = (vma->node.size >= vma->fence_size &&
402 IS_ALIGNED(vma->node.start, vma->fence_alignment));
403
404 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
405
406 if (mappable && fenceable)
407 vma->flags |= I915_VMA_CAN_FENCE;
408 else
409 vma->flags &= ~I915_VMA_CAN_FENCE;
410 }
411
color_differs(struct drm_mm_node * node,unsigned long color)412 static bool color_differs(struct drm_mm_node *node, unsigned long color)
413 {
414 return node->allocated && node->color != color;
415 }
416
i915_gem_valid_gtt_space(struct i915_vma * vma,unsigned long cache_level)417 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
418 {
419 struct drm_mm_node *node = &vma->node;
420 struct drm_mm_node *other;
421
422 /*
423 * On some machines we have to be careful when putting differing types
424 * of snoopable memory together to avoid the prefetcher crossing memory
425 * domains and dying. During vm initialisation, we decide whether or not
426 * these constraints apply and set the drm_mm.color_adjust
427 * appropriately.
428 */
429 if (vma->vm->mm.color_adjust == NULL)
430 return true;
431
432 /* Only valid to be called on an already inserted vma */
433 GEM_BUG_ON(!drm_mm_node_allocated(node));
434 GEM_BUG_ON(list_empty(&node->node_list));
435
436 other = list_prev_entry(node, node_list);
437 if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
438 return false;
439
440 other = list_next_entry(node, node_list);
441 if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
442 return false;
443
444 return true;
445 }
446
447 /**
448 * i915_vma_insert - finds a slot for the vma in its address space
449 * @vma: the vma
450 * @size: requested size in bytes (can be larger than the VMA)
451 * @alignment: required alignment
452 * @flags: mask of PIN_* flags to use
453 *
454 * First we try to allocate some free space that meets the requirements for
455 * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
456 * preferrably the oldest idle entry to make room for the new VMA.
457 *
458 * Returns:
459 * 0 on success, negative error code otherwise.
460 */
461 static int
i915_vma_insert(struct i915_vma * vma,u64 size,u64 alignment,u64 flags)462 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
463 {
464 struct drm_i915_private *dev_priv = vma->vm->i915;
465 struct drm_i915_gem_object *obj = vma->obj;
466 u64 start, end;
467 int ret;
468
469 GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
470 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
471
472 size = max(size, vma->size);
473 alignment = max(alignment, vma->display_alignment);
474 if (flags & PIN_MAPPABLE) {
475 size = max_t(typeof(size), size, vma->fence_size);
476 alignment = max_t(typeof(alignment),
477 alignment, vma->fence_alignment);
478 }
479
480 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
481 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
482 GEM_BUG_ON(!is_power_of_2(alignment));
483
484 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
485 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
486
487 end = vma->vm->total;
488 if (flags & PIN_MAPPABLE)
489 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
490 if (flags & PIN_ZONE_4G)
491 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
492 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
493
494 /* If binding the object/GGTT view requires more space than the entire
495 * aperture has, reject it early before evicting everything in a vain
496 * attempt to find space.
497 */
498 if (size > end) {
499 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu [object=%zd] > %s aperture=%llu\n",
500 size, obj->base.size,
501 flags & PIN_MAPPABLE ? "mappable" : "total",
502 end);
503 return -ENOSPC;
504 }
505
506 ret = i915_gem_object_pin_pages(obj);
507 if (ret)
508 return ret;
509
510 GEM_BUG_ON(vma->pages);
511
512 ret = vma->vm->set_pages(vma);
513 if (ret)
514 goto err_unpin;
515
516 if (flags & PIN_OFFSET_FIXED) {
517 u64 offset = flags & PIN_OFFSET_MASK;
518 if (!IS_ALIGNED(offset, alignment) ||
519 range_overflows(offset, size, end)) {
520 ret = -EINVAL;
521 goto err_clear;
522 }
523
524 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
525 size, offset, obj->cache_level,
526 flags);
527 if (ret)
528 goto err_clear;
529 } else {
530 /*
531 * We only support huge gtt pages through the 48b PPGTT,
532 * however we also don't want to force any alignment for
533 * objects which need to be tightly packed into the low 32bits.
534 *
535 * Note that we assume that GGTT are limited to 4GiB for the
536 * forseeable future. See also i915_ggtt_offset().
537 */
538 if (upper_32_bits(end - 1) &&
539 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
540 /*
541 * We can't mix 64K and 4K PTEs in the same page-table
542 * (2M block), and so to avoid the ugliness and
543 * complexity of coloring we opt for just aligning 64K
544 * objects to 2M.
545 */
546 u64 page_alignment =
547 rounddown_pow_of_two(vma->page_sizes.sg |
548 I915_GTT_PAGE_SIZE_2M);
549
550 /*
551 * Check we don't expand for the limited Global GTT
552 * (mappable aperture is even more precious!). This
553 * also checks that we exclude the aliasing-ppgtt.
554 */
555 GEM_BUG_ON(i915_vma_is_ggtt(vma));
556
557 alignment = max(alignment, page_alignment);
558
559 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
560 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
561 }
562
563 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
564 size, alignment, obj->cache_level,
565 start, end, flags);
566 if (ret)
567 goto err_clear;
568
569 GEM_BUG_ON(vma->node.start < start);
570 GEM_BUG_ON(vma->node.start + vma->node.size > end);
571 }
572 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
573 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level));
574
575 list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
576
577 lockmgr(&dev_priv->mm.obj_lock, LK_EXCLUSIVE);
578 list_move_tail(&obj->mm.link, &dev_priv->mm.bound_list);
579 obj->bind_count++;
580 lockmgr(&dev_priv->mm.obj_lock, LK_RELEASE);
581
582 GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
583
584 return 0;
585
586 err_clear:
587 vma->vm->clear_pages(vma);
588 err_unpin:
589 i915_gem_object_unpin_pages(obj);
590 return ret;
591 }
592
593 static void
i915_vma_remove(struct i915_vma * vma)594 i915_vma_remove(struct i915_vma *vma)
595 {
596 struct drm_i915_private *i915 = vma->vm->i915;
597 struct drm_i915_gem_object *obj = vma->obj;
598
599 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
600 GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
601
602 vma->vm->clear_pages(vma);
603
604 drm_mm_remove_node(&vma->node);
605 list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
606
607 /* Since the unbound list is global, only move to that list if
608 * no more VMAs exist.
609 */
610 lockmgr(&i915->mm.obj_lock, LK_EXCLUSIVE);
611 if (--obj->bind_count == 0)
612 list_move_tail(&obj->mm.link, &i915->mm.unbound_list);
613 lockmgr(&i915->mm.obj_lock, LK_RELEASE);
614
615 /* And finally now the object is completely decoupled from this vma,
616 * we can drop its hold on the backing storage and allow it to be
617 * reaped by the shrinker.
618 */
619 i915_gem_object_unpin_pages(obj);
620 GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
621 }
622
__i915_vma_do_pin(struct i915_vma * vma,u64 size,u64 alignment,u64 flags)623 int __i915_vma_do_pin(struct i915_vma *vma,
624 u64 size, u64 alignment, u64 flags)
625 {
626 const unsigned int bound = vma->flags;
627 int ret;
628
629 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
630 GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
631 GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
632
633 if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
634 ret = -EBUSY;
635 goto err_unpin;
636 }
637
638 if ((bound & I915_VMA_BIND_MASK) == 0) {
639 ret = i915_vma_insert(vma, size, alignment, flags);
640 if (ret)
641 goto err_unpin;
642 }
643
644 ret = i915_vma_bind(vma, vma->obj->cache_level, flags);
645 if (ret)
646 goto err_remove;
647
648 if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
649 __i915_vma_set_map_and_fenceable(vma);
650
651 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
652 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
653 return 0;
654
655 err_remove:
656 if ((bound & I915_VMA_BIND_MASK) == 0) {
657 i915_vma_remove(vma);
658 GEM_BUG_ON(vma->pages);
659 }
660 err_unpin:
661 __i915_vma_unpin(vma);
662 return ret;
663 }
664
i915_vma_destroy(struct i915_vma * vma)665 static void i915_vma_destroy(struct i915_vma *vma)
666 {
667 int i;
668
669 GEM_BUG_ON(vma->node.allocated);
670 GEM_BUG_ON(i915_vma_is_active(vma));
671 GEM_BUG_ON(!i915_vma_is_closed(vma));
672 GEM_BUG_ON(vma->fence);
673
674 for (i = 0; i < ARRAY_SIZE(vma->last_read); i++)
675 GEM_BUG_ON(i915_gem_active_isset(&vma->last_read[i]));
676 GEM_BUG_ON(i915_gem_active_isset(&vma->last_fence));
677
678 list_del(&vma->vm_link);
679 if (!i915_vma_is_ggtt(vma))
680 i915_ppgtt_put(i915_vm_to_ppgtt(vma->vm));
681
682 kmem_cache_free(to_i915(vma->obj->base.dev)->vmas, vma);
683 }
684
i915_vma_close(struct i915_vma * vma)685 void i915_vma_close(struct i915_vma *vma)
686 {
687 GEM_BUG_ON(i915_vma_is_closed(vma));
688 vma->flags |= I915_VMA_CLOSED;
689
690 list_del(&vma->obj_link);
691 rb_erase(&vma->obj_node, &vma->obj->vma_tree);
692
693 if (!i915_vma_is_active(vma) && !i915_vma_is_pinned(vma))
694 WARN_ON(i915_vma_unbind(vma));
695 }
696
__i915_vma_iounmap(struct i915_vma * vma)697 static void __i915_vma_iounmap(struct i915_vma *vma)
698 {
699 GEM_BUG_ON(i915_vma_is_pinned(vma));
700
701 if (vma->iomap == NULL)
702 return;
703
704 io_mapping_unmap(vma->iomap);
705 vma->iomap = NULL;
706 }
707
708 #ifdef __DragonFly__
709 /*
710 * XXX: this is a bit of a hammer approach
711 * On DragonFly, we invalidate the entire set of gem object pages
712 * and not just the ones belonging to a particular VMA
713 */
drm_vma_node_unmap(struct drm_vma_offset_node * node,struct address_space * file_mapping)714 static inline void drm_vma_node_unmap(struct drm_vma_offset_node *node,
715 struct address_space *file_mapping)
716 {
717 struct drm_i915_gem_object *obj = container_of(
718 node,struct drm_i915_gem_object, base.vma_node);
719 vm_object_t devobj;
720 vm_page_t m;
721 int i, page_count;
722
723 devobj = cdev_pager_lookup(obj);
724 if (devobj != NULL) {
725 page_count = OFF_TO_IDX(obj->base.size);
726
727 VM_OBJECT_LOCK(devobj);
728 vm_object_page_remove(devobj, 0, 0, false);
729 #if 1
730 for (i = 0; i < page_count; i++) {
731 m = vm_page_lookup_busy_wait(devobj, i, TRUE, "915unm");
732 if (m == NULL)
733 continue;
734 cdev_pager_free_page(devobj, m);
735 }
736 #endif
737 VM_OBJECT_UNLOCK(devobj);
738 vm_object_deallocate(devobj);
739 }
740 }
741 #endif
742
i915_vma_revoke_mmap(struct i915_vma * vma)743 void i915_vma_revoke_mmap(struct i915_vma *vma)
744 {
745 struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
746 u64 vma_offset;
747
748 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
749
750 if (!i915_vma_has_userfault(vma))
751 return;
752
753 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
754 GEM_BUG_ON(!vma->obj->userfault_count);
755
756 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
757 #ifdef __DragonFly__
758 drm_vma_node_unmap(node, NULL);
759 #else
760 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
761 drm_vma_node_offset_addr(node) + vma_offset,
762 vma->size,
763 1);
764 #endif
765
766 i915_vma_unset_userfault(vma);
767 if (!--vma->obj->userfault_count)
768 list_del(&vma->obj->userfault_link);
769 }
770
i915_vma_unbind(struct i915_vma * vma)771 int i915_vma_unbind(struct i915_vma *vma)
772 {
773 struct drm_i915_gem_object *obj = vma->obj;
774 unsigned long active;
775 int ret;
776
777 lockdep_assert_held(&obj->base.dev->struct_mutex);
778
779 /* First wait upon any activity as retiring the request may
780 * have side-effects such as unpinning or even unbinding this vma.
781 */
782 active = i915_vma_get_active(vma);
783 if (active) {
784 int idx;
785
786 /* When a closed VMA is retired, it is unbound - eek.
787 * In order to prevent it from being recursively closed,
788 * take a pin on the vma so that the second unbind is
789 * aborted.
790 *
791 * Even more scary is that the retire callback may free
792 * the object (last active vma). To prevent the explosion
793 * we defer the actual object free to a worker that can
794 * only proceed once it acquires the struct_mutex (which
795 * we currently hold, therefore it cannot free this object
796 * before we are finished).
797 */
798 __i915_vma_pin(vma);
799
800 for_each_active(active, idx) {
801 ret = i915_gem_active_retire(&vma->last_read[idx],
802 &vma->vm->i915->drm.struct_mutex);
803 if (ret)
804 break;
805 }
806
807 if (!ret) {
808 ret = i915_gem_active_retire(&vma->last_fence,
809 &vma->vm->i915->drm.struct_mutex);
810 }
811
812 __i915_vma_unpin(vma);
813 if (ret)
814 return ret;
815 }
816 GEM_BUG_ON(i915_vma_is_active(vma));
817
818 if (i915_vma_is_pinned(vma))
819 return -EBUSY;
820
821 if (!drm_mm_node_allocated(&vma->node))
822 goto destroy;
823
824 GEM_BUG_ON(obj->bind_count == 0);
825 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
826
827 if (i915_vma_is_map_and_fenceable(vma)) {
828 /* release the fence reg _after_ flushing */
829 ret = i915_vma_put_fence(vma);
830 if (ret)
831 return ret;
832
833 /* Force a pagefault for domain tracking on next user access */
834 i915_vma_revoke_mmap(vma);
835
836 __i915_vma_iounmap(vma);
837 vma->flags &= ~I915_VMA_CAN_FENCE;
838 }
839 GEM_BUG_ON(vma->fence);
840 GEM_BUG_ON(i915_vma_has_userfault(vma));
841
842 if (likely(!vma->vm->closed)) {
843 trace_i915_vma_unbind(vma);
844 vma->vm->unbind_vma(vma);
845 }
846 vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
847
848 i915_vma_remove(vma);
849
850 destroy:
851 if (unlikely(i915_vma_is_closed(vma)))
852 i915_vma_destroy(vma);
853
854 return 0;
855 }
856
857 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
858 #include "selftests/i915_vma.c"
859 #endif
860