]>
Commit | Line | Data |
---|---|---|
b42fe9ca JL |
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 | */ | |
10195b1e | 24 | |
09480072 | 25 | #include <linux/sched/mm.h> |
df0566a6 | 26 | #include <drm/drm_gem.h> |
112ed2d3 | 27 | |
df0566a6 JN |
28 | #include "display/intel_frontbuffer.h" |
29 | ||
30 | #include "gt/intel_engine.h" | |
ccd20945 | 31 | #include "gt/intel_engine_heartbeat.h" |
a1c8a09e | 32 | #include "gt/intel_gt.h" |
ccd20945 | 33 | #include "gt/intel_gt_requests.h" |
b42fe9ca JL |
34 | |
35 | #include "i915_drv.h" | |
103b76ee | 36 | #include "i915_globals.h" |
2850748e | 37 | #include "i915_sw_fence_work.h" |
a09d9a80 | 38 | #include "i915_trace.h" |
df0566a6 | 39 | #include "i915_vma.h" |
b42fe9ca | 40 | |
13f1bfd3 | 41 | static struct i915_global_vma { |
103b76ee | 42 | struct i915_global base; |
13f1bfd3 CW |
43 | struct kmem_cache *slab_vmas; |
44 | } global; | |
45 | ||
46 | struct i915_vma *i915_vma_alloc(void) | |
47 | { | |
48 | return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL); | |
49 | } | |
50 | ||
51 | void i915_vma_free(struct i915_vma *vma) | |
52 | { | |
53 | return kmem_cache_free(global.slab_vmas, vma); | |
54 | } | |
55 | ||
1eca65d9 | 56 | #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM) |
10195b1e CW |
57 | |
58 | #include <linux/stackdepot.h> | |
59 | ||
60 | static void vma_print_allocator(struct i915_vma *vma, const char *reason) | |
61 | { | |
487f3c7f TG |
62 | unsigned long *entries; |
63 | unsigned int nr_entries; | |
10195b1e CW |
64 | char buf[512]; |
65 | ||
66 | if (!vma->node.stack) { | |
67 | DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n", | |
68 | vma->node.start, vma->node.size, reason); | |
69 | return; | |
70 | } | |
71 | ||
487f3c7f TG |
72 | nr_entries = stack_depot_fetch(vma->node.stack, &entries); |
73 | stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0); | |
10195b1e CW |
74 | DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n", |
75 | vma->node.start, vma->node.size, reason, buf); | |
76 | } | |
77 | ||
78 | #else | |
79 | ||
80 | static void vma_print_allocator(struct i915_vma *vma, const char *reason) | |
81 | { | |
82 | } | |
83 | ||
84 | #endif | |
85 | ||
12c255b5 | 86 | static inline struct i915_vma *active_to_vma(struct i915_active *ref) |
b42fe9ca | 87 | { |
12c255b5 CW |
88 | return container_of(ref, typeof(struct i915_vma), active); |
89 | } | |
1ab22356 | 90 | |
12c255b5 CW |
91 | static int __i915_vma_active(struct i915_active *ref) |
92 | { | |
2833ddcc | 93 | return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT; |
12c255b5 CW |
94 | } |
95 | ||
274cbf20 | 96 | __i915_active_call |
12c255b5 CW |
97 | static void __i915_vma_retire(struct i915_active *ref) |
98 | { | |
99 | i915_vma_put(active_to_vma(ref)); | |
b42fe9ca JL |
100 | } |
101 | ||
b42fe9ca | 102 | static struct i915_vma * |
a01cb37a CW |
103 | vma_create(struct drm_i915_gem_object *obj, |
104 | struct i915_address_space *vm, | |
105 | const struct i915_ggtt_view *view) | |
b42fe9ca JL |
106 | { |
107 | struct i915_vma *vma; | |
108 | struct rb_node *rb, **p; | |
b42fe9ca | 109 | |
e1cc3db0 | 110 | /* The aliasing_ppgtt should never be used directly! */ |
71e51ca8 | 111 | GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm); |
e1cc3db0 | 112 | |
13f1bfd3 | 113 | vma = i915_vma_alloc(); |
b42fe9ca JL |
114 | if (vma == NULL) |
115 | return ERR_PTR(-ENOMEM); | |
116 | ||
76f9764c | 117 | kref_init(&vma->ref); |
2850748e CW |
118 | mutex_init(&vma->pages_mutex); |
119 | vma->vm = i915_vm_get(vm); | |
93f2cde2 | 120 | vma->ops = &vm->vma_ops; |
b42fe9ca | 121 | vma->obj = obj; |
ef78f7b1 | 122 | vma->resv = obj->base.resv; |
b42fe9ca | 123 | vma->size = obj->base.size; |
f51455d4 | 124 | vma->display_alignment = I915_GTT_MIN_ALIGNMENT; |
b42fe9ca | 125 | |
b1e3177b | 126 | i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire); |
155ab883 | 127 | |
09480072 CW |
128 | /* Declare ourselves safe for use inside shrinkers */ |
129 | if (IS_ENABLED(CONFIG_LOCKDEP)) { | |
130 | fs_reclaim_acquire(GFP_KERNEL); | |
131 | might_lock(&vma->active.mutex); | |
132 | fs_reclaim_release(GFP_KERNEL); | |
133 | } | |
134 | ||
155ab883 CW |
135 | INIT_LIST_HEAD(&vma->closed_link); |
136 | ||
7c518460 | 137 | if (view && view->type != I915_GGTT_VIEW_NORMAL) { |
b42fe9ca JL |
138 | vma->ggtt_view = *view; |
139 | if (view->type == I915_GGTT_VIEW_PARTIAL) { | |
07e19ea4 | 140 | GEM_BUG_ON(range_overflows_t(u64, |
8bab1193 CW |
141 | view->partial.offset, |
142 | view->partial.size, | |
07e19ea4 | 143 | obj->base.size >> PAGE_SHIFT)); |
8bab1193 | 144 | vma->size = view->partial.size; |
b42fe9ca | 145 | vma->size <<= PAGE_SHIFT; |
7e7367d3 | 146 | GEM_BUG_ON(vma->size > obj->base.size); |
b42fe9ca | 147 | } else if (view->type == I915_GGTT_VIEW_ROTATED) { |
8bab1193 | 148 | vma->size = intel_rotation_info_size(&view->rotated); |
b42fe9ca | 149 | vma->size <<= PAGE_SHIFT; |
1a74fc0b VS |
150 | } else if (view->type == I915_GGTT_VIEW_REMAPPED) { |
151 | vma->size = intel_remapped_info_size(&view->remapped); | |
152 | vma->size <<= PAGE_SHIFT; | |
b42fe9ca JL |
153 | } |
154 | } | |
155 | ||
1fcdaa7e CW |
156 | if (unlikely(vma->size > vm->total)) |
157 | goto err_vma; | |
158 | ||
b00ddb27 CW |
159 | GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE)); |
160 | ||
f524a774 CW |
161 | spin_lock(&obj->vma.lock); |
162 | ||
b42fe9ca | 163 | if (i915_is_ggtt(vm)) { |
1fcdaa7e | 164 | if (unlikely(overflows_type(vma->size, u32))) |
f524a774 | 165 | goto err_unlock; |
1fcdaa7e | 166 | |
91d4e0aa CW |
167 | vma->fence_size = i915_gem_fence_size(vm->i915, vma->size, |
168 | i915_gem_object_get_tiling(obj), | |
169 | i915_gem_object_get_stride(obj)); | |
1fcdaa7e CW |
170 | if (unlikely(vma->fence_size < vma->size || /* overflow */ |
171 | vma->fence_size > vm->total)) | |
f524a774 | 172 | goto err_unlock; |
1fcdaa7e | 173 | |
f51455d4 | 174 | GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT)); |
944397f0 | 175 | |
91d4e0aa CW |
176 | vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size, |
177 | i915_gem_object_get_tiling(obj), | |
178 | i915_gem_object_get_stride(obj)); | |
944397f0 CW |
179 | GEM_BUG_ON(!is_power_of_2(vma->fence_alignment)); |
180 | ||
4dd2fbbf | 181 | __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma)); |
b42fe9ca JL |
182 | } |
183 | ||
184 | rb = NULL; | |
528cbd17 | 185 | p = &obj->vma.tree.rb_node; |
b42fe9ca JL |
186 | while (*p) { |
187 | struct i915_vma *pos; | |
528cbd17 | 188 | long cmp; |
b42fe9ca JL |
189 | |
190 | rb = *p; | |
191 | pos = rb_entry(rb, struct i915_vma, obj_node); | |
528cbd17 CW |
192 | |
193 | /* | |
194 | * If the view already exists in the tree, another thread | |
195 | * already created a matching vma, so return the older instance | |
196 | * and dispose of ours. | |
197 | */ | |
198 | cmp = i915_vma_compare(pos, vm, view); | |
199 | if (cmp == 0) { | |
200 | spin_unlock(&obj->vma.lock); | |
13f1bfd3 | 201 | i915_vma_free(vma); |
528cbd17 CW |
202 | return pos; |
203 | } | |
204 | ||
205 | if (cmp < 0) | |
b42fe9ca JL |
206 | p = &rb->rb_right; |
207 | else | |
208 | p = &rb->rb_left; | |
209 | } | |
210 | rb_link_node(&vma->obj_node, rb, p); | |
528cbd17 CW |
211 | rb_insert_color(&vma->obj_node, &obj->vma.tree); |
212 | ||
213 | if (i915_vma_is_ggtt(vma)) | |
214 | /* | |
215 | * We put the GGTT vma at the start of the vma-list, followed | |
216 | * by the ppGGTT vma. This allows us to break early when | |
217 | * iterating over only the GGTT vma for an object, see | |
218 | * for_each_ggtt_vma() | |
219 | */ | |
220 | list_add(&vma->obj_link, &obj->vma.list); | |
221 | else | |
222 | list_add_tail(&vma->obj_link, &obj->vma.list); | |
223 | ||
224 | spin_unlock(&obj->vma.lock); | |
09d7e46b | 225 | |
b42fe9ca | 226 | return vma; |
1fcdaa7e | 227 | |
f524a774 CW |
228 | err_unlock: |
229 | spin_unlock(&obj->vma.lock); | |
1fcdaa7e | 230 | err_vma: |
13f1bfd3 | 231 | i915_vma_free(vma); |
1fcdaa7e | 232 | return ERR_PTR(-E2BIG); |
b42fe9ca JL |
233 | } |
234 | ||
481a6f7d CW |
235 | static struct i915_vma * |
236 | vma_lookup(struct drm_i915_gem_object *obj, | |
237 | struct i915_address_space *vm, | |
238 | const struct i915_ggtt_view *view) | |
718659a6 CW |
239 | { |
240 | struct rb_node *rb; | |
241 | ||
528cbd17 | 242 | rb = obj->vma.tree.rb_node; |
718659a6 CW |
243 | while (rb) { |
244 | struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node); | |
245 | long cmp; | |
246 | ||
247 | cmp = i915_vma_compare(vma, vm, view); | |
248 | if (cmp == 0) | |
249 | return vma; | |
250 | ||
251 | if (cmp < 0) | |
252 | rb = rb->rb_right; | |
253 | else | |
254 | rb = rb->rb_left; | |
255 | } | |
256 | ||
257 | return NULL; | |
258 | } | |
259 | ||
718659a6 CW |
260 | /** |
261 | * i915_vma_instance - return the singleton instance of the VMA | |
262 | * @obj: parent &struct drm_i915_gem_object to be mapped | |
263 | * @vm: address space in which the mapping is located | |
264 | * @view: additional mapping requirements | |
265 | * | |
266 | * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with | |
267 | * the same @view characteristics. If a match is not found, one is created. | |
268 | * Once created, the VMA is kept until either the object is freed, or the | |
269 | * address space is closed. | |
270 | * | |
718659a6 CW |
271 | * Returns the vma, or an error pointer. |
272 | */ | |
273 | struct i915_vma * | |
274 | i915_vma_instance(struct drm_i915_gem_object *obj, | |
275 | struct i915_address_space *vm, | |
276 | const struct i915_ggtt_view *view) | |
277 | { | |
278 | struct i915_vma *vma; | |
279 | ||
718659a6 | 280 | GEM_BUG_ON(view && !i915_is_ggtt(vm)); |
2850748e | 281 | GEM_BUG_ON(!atomic_read(&vm->open)); |
718659a6 | 282 | |
528cbd17 | 283 | spin_lock(&obj->vma.lock); |
481a6f7d | 284 | vma = vma_lookup(obj, vm, view); |
528cbd17 CW |
285 | spin_unlock(&obj->vma.lock); |
286 | ||
287 | /* vma_create() will resolve the race if another creates the vma */ | |
288 | if (unlikely(!vma)) | |
a01cb37a | 289 | vma = vma_create(obj, vm, view); |
718659a6 | 290 | |
4ea9527c | 291 | GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view)); |
718659a6 CW |
292 | return vma; |
293 | } | |
294 | ||
2850748e CW |
295 | struct i915_vma_work { |
296 | struct dma_fence_work base; | |
297 | struct i915_vma *vma; | |
54d7195f | 298 | struct drm_i915_gem_object *pinned; |
e3793468 | 299 | struct i915_sw_dma_fence_cb cb; |
2850748e CW |
300 | enum i915_cache_level cache_level; |
301 | unsigned int flags; | |
302 | }; | |
303 | ||
304 | static int __vma_bind(struct dma_fence_work *work) | |
305 | { | |
306 | struct i915_vma_work *vw = container_of(work, typeof(*vw), base); | |
307 | struct i915_vma *vma = vw->vma; | |
308 | int err; | |
309 | ||
310 | err = vma->ops->bind_vma(vma, vw->cache_level, vw->flags); | |
311 | if (err) | |
312 | atomic_or(I915_VMA_ERROR, &vma->flags); | |
313 | ||
2850748e CW |
314 | return err; |
315 | } | |
316 | ||
54d7195f CW |
317 | static void __vma_release(struct dma_fence_work *work) |
318 | { | |
319 | struct i915_vma_work *vw = container_of(work, typeof(*vw), base); | |
320 | ||
321 | if (vw->pinned) | |
322 | __i915_gem_object_unpin_pages(vw->pinned); | |
323 | } | |
324 | ||
2850748e CW |
325 | static const struct dma_fence_work_ops bind_ops = { |
326 | .name = "bind", | |
327 | .work = __vma_bind, | |
54d7195f | 328 | .release = __vma_release, |
2850748e CW |
329 | }; |
330 | ||
331 | struct i915_vma_work *i915_vma_work(void) | |
332 | { | |
333 | struct i915_vma_work *vw; | |
334 | ||
335 | vw = kzalloc(sizeof(*vw), GFP_KERNEL); | |
336 | if (!vw) | |
337 | return NULL; | |
338 | ||
339 | dma_fence_work_init(&vw->base, &bind_ops); | |
340 | vw->base.dma.error = -EAGAIN; /* disable the worker by default */ | |
341 | ||
342 | return vw; | |
343 | } | |
344 | ||
e3793468 CW |
345 | int i915_vma_wait_for_bind(struct i915_vma *vma) |
346 | { | |
347 | int err = 0; | |
348 | ||
349 | if (rcu_access_pointer(vma->active.excl.fence)) { | |
350 | struct dma_fence *fence; | |
351 | ||
352 | rcu_read_lock(); | |
353 | fence = dma_fence_get_rcu_safe(&vma->active.excl.fence); | |
354 | rcu_read_unlock(); | |
355 | if (fence) { | |
356 | err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT); | |
357 | dma_fence_put(fence); | |
358 | } | |
359 | } | |
360 | ||
361 | return err; | |
362 | } | |
363 | ||
b42fe9ca JL |
364 | /** |
365 | * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space. | |
366 | * @vma: VMA to map | |
367 | * @cache_level: mapping cache level | |
368 | * @flags: flags like global or local mapping | |
2850748e | 369 | * @work: preallocated worker for allocating and binding the PTE |
b42fe9ca JL |
370 | * |
371 | * DMA addresses are taken from the scatter-gather table of this object (or of | |
372 | * this VMA in case of non-default GGTT views) and PTE entries set up. | |
373 | * Note that DMA addresses are also the only part of the SG table we care about. | |
374 | */ | |
2850748e CW |
375 | int i915_vma_bind(struct i915_vma *vma, |
376 | enum i915_cache_level cache_level, | |
377 | u32 flags, | |
378 | struct i915_vma_work *work) | |
b42fe9ca JL |
379 | { |
380 | u32 bind_flags; | |
381 | u32 vma_flags; | |
382 | int ret; | |
383 | ||
aa149431 CW |
384 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); |
385 | GEM_BUG_ON(vma->size > vma->node.size); | |
386 | ||
bbb8a9d7 TU |
387 | if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start, |
388 | vma->node.size, | |
389 | vma->vm->total))) | |
aa149431 CW |
390 | return -ENODEV; |
391 | ||
bbb8a9d7 | 392 | if (GEM_DEBUG_WARN_ON(!flags)) |
b42fe9ca JL |
393 | return -EINVAL; |
394 | ||
2850748e CW |
395 | bind_flags = flags; |
396 | bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; | |
b42fe9ca | 397 | |
4dd2fbbf CW |
398 | vma_flags = atomic_read(&vma->flags); |
399 | vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; | |
b42fe9ca JL |
400 | if (flags & PIN_UPDATE) |
401 | bind_flags |= vma_flags; | |
402 | else | |
403 | bind_flags &= ~vma_flags; | |
404 | if (bind_flags == 0) | |
405 | return 0; | |
406 | ||
fa3f46af MA |
407 | GEM_BUG_ON(!vma->pages); |
408 | ||
6146e6da | 409 | trace_i915_vma_bind(vma, bind_flags); |
2850748e | 410 | if (work && (bind_flags & ~vma_flags) & vma->vm->bind_async_flags) { |
e3793468 CW |
411 | struct dma_fence *prev; |
412 | ||
2850748e CW |
413 | work->vma = vma; |
414 | work->cache_level = cache_level; | |
415 | work->flags = bind_flags | I915_VMA_ALLOC; | |
416 | ||
417 | /* | |
418 | * Note we only want to chain up to the migration fence on | |
419 | * the pages (not the object itself). As we don't track that, | |
420 | * yet, we have to use the exclusive fence instead. | |
421 | * | |
422 | * Also note that we do not want to track the async vma as | |
423 | * part of the obj->resv->excl_fence as it only affects | |
424 | * execution and not content or object's backing store lifetime. | |
425 | */ | |
e3793468 | 426 | prev = i915_active_set_exclusive(&vma->active, &work->base.dma); |
30ca04e1 | 427 | if (prev) { |
e3793468 CW |
428 | __i915_sw_fence_await_dma_fence(&work->base.chain, |
429 | prev, | |
430 | &work->cb); | |
30ca04e1 CW |
431 | dma_fence_put(prev); |
432 | } | |
e3793468 | 433 | |
2850748e CW |
434 | work->base.dma.error = 0; /* enable the queue_work() */ |
435 | ||
54d7195f | 436 | if (vma->obj) { |
2850748e | 437 | __i915_gem_object_pin_pages(vma->obj); |
54d7195f CW |
438 | work->pinned = vma->obj; |
439 | } | |
2850748e | 440 | } else { |
2850748e CW |
441 | ret = vma->ops->bind_vma(vma, cache_level, bind_flags); |
442 | if (ret) | |
443 | return ret; | |
444 | } | |
b42fe9ca | 445 | |
4dd2fbbf | 446 | atomic_or(bind_flags, &vma->flags); |
b42fe9ca JL |
447 | return 0; |
448 | } | |
449 | ||
450 | void __iomem *i915_vma_pin_iomap(struct i915_vma *vma) | |
451 | { | |
452 | void __iomem *ptr; | |
b4563f59 | 453 | int err; |
b42fe9ca | 454 | |
2850748e | 455 | if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) { |
b4563f59 CW |
456 | err = -ENODEV; |
457 | goto err; | |
458 | } | |
b42fe9ca JL |
459 | |
460 | GEM_BUG_ON(!i915_vma_is_ggtt(vma)); | |
4dd2fbbf | 461 | GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)); |
b42fe9ca | 462 | |
2850748e | 463 | ptr = READ_ONCE(vma->iomap); |
b42fe9ca | 464 | if (ptr == NULL) { |
73ebd503 | 465 | ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap, |
b42fe9ca JL |
466 | vma->node.start, |
467 | vma->node.size); | |
b4563f59 CW |
468 | if (ptr == NULL) { |
469 | err = -ENOMEM; | |
470 | goto err; | |
471 | } | |
b42fe9ca | 472 | |
2850748e CW |
473 | if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) { |
474 | io_mapping_unmap(ptr); | |
475 | ptr = vma->iomap; | |
476 | } | |
b42fe9ca JL |
477 | } |
478 | ||
479 | __i915_vma_pin(vma); | |
b4563f59 | 480 | |
3bd40735 | 481 | err = i915_vma_pin_fence(vma); |
b4563f59 CW |
482 | if (err) |
483 | goto err_unpin; | |
484 | ||
7125397b | 485 | i915_vma_set_ggtt_write(vma); |
a5972e93 CW |
486 | |
487 | /* NB Access through the GTT requires the device to be awake. */ | |
b42fe9ca | 488 | return ptr; |
b4563f59 CW |
489 | |
490 | err_unpin: | |
491 | __i915_vma_unpin(vma); | |
492 | err: | |
493 | return IO_ERR_PTR(err); | |
494 | } | |
495 | ||
7125397b CW |
496 | void i915_vma_flush_writes(struct i915_vma *vma) |
497 | { | |
2850748e CW |
498 | if (i915_vma_unset_ggtt_write(vma)) |
499 | intel_gt_flush_ggtt_writes(vma->vm->gt); | |
7125397b CW |
500 | } |
501 | ||
b4563f59 CW |
502 | void i915_vma_unpin_iomap(struct i915_vma *vma) |
503 | { | |
b4563f59 CW |
504 | GEM_BUG_ON(vma->iomap == NULL); |
505 | ||
7125397b CW |
506 | i915_vma_flush_writes(vma); |
507 | ||
b4563f59 CW |
508 | i915_vma_unpin_fence(vma); |
509 | i915_vma_unpin(vma); | |
b42fe9ca JL |
510 | } |
511 | ||
6a2f59e4 | 512 | void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags) |
b42fe9ca JL |
513 | { |
514 | struct i915_vma *vma; | |
515 | struct drm_i915_gem_object *obj; | |
516 | ||
517 | vma = fetch_and_zero(p_vma); | |
518 | if (!vma) | |
519 | return; | |
520 | ||
521 | obj = vma->obj; | |
520ea7c5 | 522 | GEM_BUG_ON(!obj); |
b42fe9ca JL |
523 | |
524 | i915_vma_unpin(vma); | |
525 | i915_vma_close(vma); | |
526 | ||
6a2f59e4 CW |
527 | if (flags & I915_VMA_RELEASE_MAP) |
528 | i915_gem_object_unpin_map(obj); | |
529 | ||
c017cf6b | 530 | i915_gem_object_put(obj); |
b42fe9ca JL |
531 | } |
532 | ||
782a3e9e CW |
533 | bool i915_vma_misplaced(const struct i915_vma *vma, |
534 | u64 size, u64 alignment, u64 flags) | |
b42fe9ca JL |
535 | { |
536 | if (!drm_mm_node_allocated(&vma->node)) | |
537 | return false; | |
538 | ||
2850748e CW |
539 | if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma))) |
540 | return true; | |
541 | ||
b42fe9ca JL |
542 | if (vma->node.size < size) |
543 | return true; | |
544 | ||
f51455d4 CW |
545 | GEM_BUG_ON(alignment && !is_power_of_2(alignment)); |
546 | if (alignment && !IS_ALIGNED(vma->node.start, alignment)) | |
b42fe9ca JL |
547 | return true; |
548 | ||
549 | if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma)) | |
550 | return true; | |
551 | ||
552 | if (flags & PIN_OFFSET_BIAS && | |
553 | vma->node.start < (flags & PIN_OFFSET_MASK)) | |
554 | return true; | |
555 | ||
556 | if (flags & PIN_OFFSET_FIXED && | |
557 | vma->node.start != (flags & PIN_OFFSET_MASK)) | |
558 | return true; | |
559 | ||
560 | return false; | |
561 | } | |
562 | ||
563 | void __i915_vma_set_map_and_fenceable(struct i915_vma *vma) | |
564 | { | |
b42fe9ca | 565 | bool mappable, fenceable; |
b42fe9ca | 566 | |
944397f0 CW |
567 | GEM_BUG_ON(!i915_vma_is_ggtt(vma)); |
568 | GEM_BUG_ON(!vma->fence_size); | |
b42fe9ca | 569 | |
944397f0 | 570 | fenceable = (vma->node.size >= vma->fence_size && |
f51455d4 | 571 | IS_ALIGNED(vma->node.start, vma->fence_alignment)); |
944397f0 CW |
572 | |
573 | mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end; | |
574 | ||
575 | if (mappable && fenceable) | |
4dd2fbbf | 576 | set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); |
b42fe9ca | 577 | else |
4dd2fbbf | 578 | clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); |
b42fe9ca JL |
579 | } |
580 | ||
33dd8899 | 581 | bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color) |
7d1d9aea CW |
582 | { |
583 | struct drm_mm_node *node = &vma->node; | |
b42fe9ca JL |
584 | struct drm_mm_node *other; |
585 | ||
586 | /* | |
587 | * On some machines we have to be careful when putting differing types | |
588 | * of snoopable memory together to avoid the prefetcher crossing memory | |
589 | * domains and dying. During vm initialisation, we decide whether or not | |
590 | * these constraints apply and set the drm_mm.color_adjust | |
591 | * appropriately. | |
592 | */ | |
33dd8899 | 593 | if (!i915_vm_has_cache_coloring(vma->vm)) |
b42fe9ca JL |
594 | return true; |
595 | ||
7d1d9aea CW |
596 | /* Only valid to be called on an already inserted vma */ |
597 | GEM_BUG_ON(!drm_mm_node_allocated(node)); | |
598 | GEM_BUG_ON(list_empty(&node->node_list)); | |
b42fe9ca | 599 | |
7d1d9aea | 600 | other = list_prev_entry(node, node_list); |
33dd8899 | 601 | if (i915_node_color_differs(other, color) && |
1e0a96e5 | 602 | !drm_mm_hole_follows(other)) |
b42fe9ca JL |
603 | return false; |
604 | ||
7d1d9aea | 605 | other = list_next_entry(node, node_list); |
33dd8899 | 606 | if (i915_node_color_differs(other, color) && |
1e0a96e5 | 607 | !drm_mm_hole_follows(node)) |
b42fe9ca JL |
608 | return false; |
609 | ||
610 | return true; | |
611 | } | |
612 | ||
83d317ad CW |
613 | static void assert_bind_count(const struct drm_i915_gem_object *obj) |
614 | { | |
615 | /* | |
616 | * Combine the assertion that the object is bound and that we have | |
617 | * pinned its pages. But we should never have bound the object | |
618 | * more than we have pinned its pages. (For complete accuracy, we | |
619 | * assume that no else is pinning the pages, but as a rough assertion | |
620 | * that we will not run into problems later, this will do!) | |
621 | */ | |
ecab9be1 | 622 | GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < atomic_read(&obj->bind_count)); |
83d317ad CW |
623 | } |
624 | ||
b42fe9ca JL |
625 | /** |
626 | * i915_vma_insert - finds a slot for the vma in its address space | |
627 | * @vma: the vma | |
628 | * @size: requested size in bytes (can be larger than the VMA) | |
629 | * @alignment: required alignment | |
630 | * @flags: mask of PIN_* flags to use | |
631 | * | |
632 | * First we try to allocate some free space that meets the requirements for | |
633 | * the VMA. Failiing that, if the flags permit, it will evict an old VMA, | |
634 | * preferrably the oldest idle entry to make room for the new VMA. | |
635 | * | |
636 | * Returns: | |
637 | * 0 on success, negative error code otherwise. | |
638 | */ | |
639 | static int | |
640 | i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags) | |
641 | { | |
33dd8899 | 642 | unsigned long color; |
b42fe9ca JL |
643 | u64 start, end; |
644 | int ret; | |
645 | ||
4dd2fbbf | 646 | GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); |
b42fe9ca JL |
647 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); |
648 | ||
649 | size = max(size, vma->size); | |
944397f0 CW |
650 | alignment = max(alignment, vma->display_alignment); |
651 | if (flags & PIN_MAPPABLE) { | |
652 | size = max_t(typeof(size), size, vma->fence_size); | |
653 | alignment = max_t(typeof(alignment), | |
654 | alignment, vma->fence_alignment); | |
655 | } | |
b42fe9ca | 656 | |
f51455d4 CW |
657 | GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); |
658 | GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT)); | |
659 | GEM_BUG_ON(!is_power_of_2(alignment)); | |
660 | ||
b42fe9ca | 661 | start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0; |
f51455d4 | 662 | GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); |
b42fe9ca JL |
663 | |
664 | end = vma->vm->total; | |
665 | if (flags & PIN_MAPPABLE) | |
2850748e | 666 | end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end); |
b42fe9ca | 667 | if (flags & PIN_ZONE_4G) |
f51455d4 CW |
668 | end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE); |
669 | GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); | |
b42fe9ca JL |
670 | |
671 | /* If binding the object/GGTT view requires more space than the entire | |
672 | * aperture has, reject it early before evicting everything in a vain | |
673 | * attempt to find space. | |
674 | */ | |
675 | if (size > end) { | |
520ea7c5 CW |
676 | DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n", |
677 | size, flags & PIN_MAPPABLE ? "mappable" : "total", | |
b42fe9ca | 678 | end); |
2889caa9 | 679 | return -ENOSPC; |
b42fe9ca JL |
680 | } |
681 | ||
33dd8899 | 682 | color = 0; |
2850748e CW |
683 | if (vma->obj && i915_vm_has_cache_coloring(vma->vm)) |
684 | color = vma->obj->cache_level; | |
fa3f46af | 685 | |
b42fe9ca JL |
686 | if (flags & PIN_OFFSET_FIXED) { |
687 | u64 offset = flags & PIN_OFFSET_MASK; | |
f51455d4 | 688 | if (!IS_ALIGNED(offset, alignment) || |
2850748e CW |
689 | range_overflows(offset, size, end)) |
690 | return -EINVAL; | |
b42fe9ca | 691 | |
625d988a | 692 | ret = i915_gem_gtt_reserve(vma->vm, &vma->node, |
33dd8899 | 693 | size, offset, color, |
625d988a CW |
694 | flags); |
695 | if (ret) | |
2850748e | 696 | return ret; |
b42fe9ca | 697 | } else { |
7464284b MA |
698 | /* |
699 | * We only support huge gtt pages through the 48b PPGTT, | |
700 | * however we also don't want to force any alignment for | |
701 | * objects which need to be tightly packed into the low 32bits. | |
702 | * | |
703 | * Note that we assume that GGTT are limited to 4GiB for the | |
704 | * forseeable future. See also i915_ggtt_offset(). | |
705 | */ | |
706 | if (upper_32_bits(end - 1) && | |
707 | vma->page_sizes.sg > I915_GTT_PAGE_SIZE) { | |
855822be MA |
708 | /* |
709 | * We can't mix 64K and 4K PTEs in the same page-table | |
710 | * (2M block), and so to avoid the ugliness and | |
711 | * complexity of coloring we opt for just aligning 64K | |
712 | * objects to 2M. | |
713 | */ | |
7464284b | 714 | u64 page_alignment = |
855822be MA |
715 | rounddown_pow_of_two(vma->page_sizes.sg | |
716 | I915_GTT_PAGE_SIZE_2M); | |
7464284b | 717 | |
bef27bdb CW |
718 | /* |
719 | * Check we don't expand for the limited Global GTT | |
720 | * (mappable aperture is even more precious!). This | |
721 | * also checks that we exclude the aliasing-ppgtt. | |
722 | */ | |
723 | GEM_BUG_ON(i915_vma_is_ggtt(vma)); | |
724 | ||
7464284b | 725 | alignment = max(alignment, page_alignment); |
855822be MA |
726 | |
727 | if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) | |
728 | size = round_up(size, I915_GTT_PAGE_SIZE_2M); | |
7464284b MA |
729 | } |
730 | ||
e007b19d | 731 | ret = i915_gem_gtt_insert(vma->vm, &vma->node, |
33dd8899 | 732 | size, alignment, color, |
e007b19d CW |
733 | start, end, flags); |
734 | if (ret) | |
2850748e | 735 | return ret; |
b42fe9ca JL |
736 | |
737 | GEM_BUG_ON(vma->node.start < start); | |
738 | GEM_BUG_ON(vma->node.start + vma->node.size > end); | |
739 | } | |
44a0ec0d | 740 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); |
33dd8899 | 741 | GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color)); |
b42fe9ca | 742 | |
520ea7c5 | 743 | if (vma->obj) { |
dde01d94 CW |
744 | struct drm_i915_gem_object *obj = vma->obj; |
745 | ||
746 | atomic_inc(&obj->bind_count); | |
747 | assert_bind_count(obj); | |
520ea7c5 | 748 | } |
dde01d94 | 749 | list_add_tail(&vma->vm_link, &vma->vm->bound_list); |
b42fe9ca JL |
750 | |
751 | return 0; | |
b42fe9ca JL |
752 | } |
753 | ||
31c7effa | 754 | static void |
dde01d94 | 755 | i915_vma_detach(struct i915_vma *vma) |
31c7effa | 756 | { |
31c7effa | 757 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); |
4dd2fbbf | 758 | GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); |
31c7effa | 759 | |
520ea7c5 | 760 | /* |
dde01d94 CW |
761 | * And finally now the object is completely decoupled from this |
762 | * vma, we can drop its hold on the backing storage and allow | |
763 | * it to be reaped by the shrinker. | |
31c7effa | 764 | */ |
dde01d94 | 765 | list_del(&vma->vm_link); |
520ea7c5 CW |
766 | if (vma->obj) { |
767 | struct drm_i915_gem_object *obj = vma->obj; | |
d82b4b26 | 768 | |
520ea7c5 | 769 | assert_bind_count(obj); |
dde01d94 | 770 | atomic_dec(&obj->bind_count); |
520ea7c5 | 771 | } |
31c7effa CW |
772 | } |
773 | ||
2850748e | 774 | static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) |
b42fe9ca | 775 | { |
2850748e CW |
776 | unsigned int bound; |
777 | bool pinned = true; | |
b42fe9ca | 778 | |
2850748e CW |
779 | bound = atomic_read(&vma->flags); |
780 | do { | |
781 | if (unlikely(flags & ~bound)) | |
782 | return false; | |
b42fe9ca | 783 | |
2850748e CW |
784 | if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) |
785 | return false; | |
786 | ||
787 | if (!(bound & I915_VMA_PIN_MASK)) | |
788 | goto unpinned; | |
789 | ||
790 | GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0); | |
791 | } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1)); | |
792 | ||
793 | return true; | |
794 | ||
795 | unpinned: | |
796 | /* | |
797 | * If pin_count==0, but we are bound, check under the lock to avoid | |
798 | * racing with a concurrent i915_vma_unbind(). | |
799 | */ | |
800 | mutex_lock(&vma->vm->mutex); | |
801 | do { | |
802 | if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) { | |
803 | pinned = false; | |
804 | break; | |
805 | } | |
806 | ||
807 | if (unlikely(flags & ~bound)) { | |
808 | pinned = false; | |
809 | break; | |
810 | } | |
811 | } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1)); | |
812 | mutex_unlock(&vma->vm->mutex); | |
813 | ||
814 | return pinned; | |
815 | } | |
816 | ||
817 | static int vma_get_pages(struct i915_vma *vma) | |
818 | { | |
819 | int err = 0; | |
820 | ||
821 | if (atomic_add_unless(&vma->pages_count, 1, 0)) | |
822 | return 0; | |
823 | ||
824 | /* Allocations ahoy! */ | |
825 | if (mutex_lock_interruptible(&vma->pages_mutex)) | |
826 | return -EINTR; | |
827 | ||
828 | if (!atomic_read(&vma->pages_count)) { | |
829 | if (vma->obj) { | |
830 | err = i915_gem_object_pin_pages(vma->obj); | |
831 | if (err) | |
832 | goto unlock; | |
833 | } | |
834 | ||
835 | err = vma->ops->set_pages(vma); | |
56184a20 CW |
836 | if (err) { |
837 | if (vma->obj) | |
838 | i915_gem_object_unpin_pages(vma->obj); | |
2850748e | 839 | goto unlock; |
56184a20 | 840 | } |
b42fe9ca | 841 | } |
2850748e | 842 | atomic_inc(&vma->pages_count); |
b42fe9ca | 843 | |
2850748e CW |
844 | unlock: |
845 | mutex_unlock(&vma->pages_mutex); | |
846 | ||
847 | return err; | |
848 | } | |
849 | ||
850 | static void __vma_put_pages(struct i915_vma *vma, unsigned int count) | |
851 | { | |
852 | /* We allocate under vma_get_pages, so beware the shrinker */ | |
853 | mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING); | |
854 | GEM_BUG_ON(atomic_read(&vma->pages_count) < count); | |
855 | if (atomic_sub_return(count, &vma->pages_count) == 0) { | |
856 | vma->ops->clear_pages(vma); | |
857 | GEM_BUG_ON(vma->pages); | |
858 | if (vma->obj) | |
859 | i915_gem_object_unpin_pages(vma->obj); | |
b42fe9ca | 860 | } |
2850748e CW |
861 | mutex_unlock(&vma->pages_mutex); |
862 | } | |
b42fe9ca | 863 | |
2850748e CW |
864 | static void vma_put_pages(struct i915_vma *vma) |
865 | { | |
866 | if (atomic_add_unless(&vma->pages_count, -1, 1)) | |
867 | return; | |
868 | ||
869 | __vma_put_pages(vma, 1); | |
870 | } | |
871 | ||
872 | static void vma_unbind_pages(struct i915_vma *vma) | |
873 | { | |
874 | unsigned int count; | |
875 | ||
876 | lockdep_assert_held(&vma->vm->mutex); | |
877 | ||
878 | /* The upper portion of pages_count is the number of bindings */ | |
879 | count = atomic_read(&vma->pages_count); | |
880 | count >>= I915_VMA_PAGES_BIAS; | |
881 | GEM_BUG_ON(!count); | |
882 | ||
883 | __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS); | |
884 | } | |
885 | ||
886 | int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags) | |
887 | { | |
888 | struct i915_vma_work *work = NULL; | |
c0e60347 | 889 | intel_wakeref_t wakeref = 0; |
2850748e CW |
890 | unsigned int bound; |
891 | int err; | |
892 | ||
893 | BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND); | |
894 | BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND); | |
895 | ||
896 | GEM_BUG_ON(flags & PIN_UPDATE); | |
897 | GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL))); | |
898 | ||
899 | /* First try and grab the pin without rebinding the vma */ | |
900 | if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK)) | |
901 | return 0; | |
902 | ||
903 | err = vma_get_pages(vma); | |
904 | if (err) | |
905 | return err; | |
906 | ||
907 | if (flags & vma->vm->bind_async_flags) { | |
908 | work = i915_vma_work(); | |
909 | if (!work) { | |
910 | err = -ENOMEM; | |
911 | goto err_pages; | |
912 | } | |
913 | } | |
914 | ||
c0e60347 CW |
915 | if (flags & PIN_GLOBAL) |
916 | wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm); | |
917 | ||
2850748e CW |
918 | /* No more allocations allowed once we hold vm->mutex */ |
919 | err = mutex_lock_interruptible(&vma->vm->mutex); | |
920 | if (err) | |
921 | goto err_fence; | |
922 | ||
00de702c CW |
923 | if (unlikely(i915_vma_is_closed(vma))) { |
924 | err = -ENOENT; | |
925 | goto err_unlock; | |
926 | } | |
927 | ||
2850748e CW |
928 | bound = atomic_read(&vma->flags); |
929 | if (unlikely(bound & I915_VMA_ERROR)) { | |
930 | err = -ENOMEM; | |
931 | goto err_unlock; | |
932 | } | |
933 | ||
934 | if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) { | |
935 | err = -EAGAIN; /* pins are meant to be fairly temporary */ | |
936 | goto err_unlock; | |
937 | } | |
938 | ||
939 | if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) { | |
940 | __i915_vma_pin(vma); | |
941 | goto err_unlock; | |
942 | } | |
943 | ||
944 | err = i915_active_acquire(&vma->active); | |
945 | if (err) | |
946 | goto err_unlock; | |
947 | ||
948 | if (!(bound & I915_VMA_BIND_MASK)) { | |
949 | err = i915_vma_insert(vma, size, alignment, flags); | |
950 | if (err) | |
951 | goto err_active; | |
952 | ||
953 | if (i915_is_ggtt(vma->vm)) | |
954 | __i915_vma_set_map_and_fenceable(vma); | |
955 | } | |
b42fe9ca | 956 | |
2850748e CW |
957 | GEM_BUG_ON(!vma->pages); |
958 | err = i915_vma_bind(vma, | |
959 | vma->obj ? vma->obj->cache_level : 0, | |
960 | flags, work); | |
961 | if (err) | |
962 | goto err_remove; | |
d36caeea | 963 | |
2850748e CW |
964 | /* There should only be at most 2 active bindings (user, global) */ |
965 | GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound); | |
966 | atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count); | |
967 | list_move_tail(&vma->vm_link, &vma->vm->bound_list); | |
b42fe9ca | 968 | |
2850748e CW |
969 | __i915_vma_pin(vma); |
970 | GEM_BUG_ON(!i915_vma_is_pinned(vma)); | |
971 | GEM_BUG_ON(!i915_vma_is_bound(vma, flags)); | |
b42fe9ca | 972 | GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags)); |
b42fe9ca | 973 | |
31c7effa | 974 | err_remove: |
dde01d94 CW |
975 | if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) { |
976 | i915_vma_detach(vma); | |
977 | drm_mm_remove_node(&vma->node); | |
978 | } | |
2850748e CW |
979 | err_active: |
980 | i915_active_release(&vma->active); | |
981 | err_unlock: | |
982 | mutex_unlock(&vma->vm->mutex); | |
983 | err_fence: | |
984 | if (work) | |
985 | dma_fence_work_commit(&work->base); | |
c0e60347 CW |
986 | if (wakeref) |
987 | intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref); | |
2850748e CW |
988 | err_pages: |
989 | vma_put_pages(vma); | |
990 | return err; | |
b42fe9ca JL |
991 | } |
992 | ||
ccd20945 CW |
993 | static void flush_idle_contexts(struct intel_gt *gt) |
994 | { | |
995 | struct intel_engine_cs *engine; | |
996 | enum intel_engine_id id; | |
997 | ||
998 | for_each_engine(engine, gt, id) | |
999 | intel_engine_flush_barriers(engine); | |
1000 | ||
1001 | intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT); | |
1002 | } | |
1003 | ||
1004 | int i915_ggtt_pin(struct i915_vma *vma, u32 align, unsigned int flags) | |
1005 | { | |
1006 | struct i915_address_space *vm = vma->vm; | |
1007 | int err; | |
1008 | ||
1009 | GEM_BUG_ON(!i915_vma_is_ggtt(vma)); | |
1010 | ||
1011 | do { | |
1012 | err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL); | |
e3793468 CW |
1013 | if (err != -ENOSPC) { |
1014 | if (!err) { | |
1015 | err = i915_vma_wait_for_bind(vma); | |
1016 | if (err) | |
1017 | i915_vma_unpin(vma); | |
1018 | } | |
ccd20945 | 1019 | return err; |
e3793468 | 1020 | } |
ccd20945 CW |
1021 | |
1022 | /* Unlike i915_vma_pin, we don't take no for an answer! */ | |
1023 | flush_idle_contexts(vm->gt); | |
1024 | if (mutex_lock_interruptible(&vm->mutex) == 0) { | |
1025 | i915_gem_evict_vm(vm); | |
1026 | mutex_unlock(&vm->mutex); | |
1027 | } | |
1028 | } while (1); | |
1029 | } | |
1030 | ||
3365e226 CW |
1031 | void i915_vma_close(struct i915_vma *vma) |
1032 | { | |
71e51ca8 | 1033 | struct intel_gt *gt = vma->vm->gt; |
155ab883 | 1034 | unsigned long flags; |
3365e226 CW |
1035 | |
1036 | GEM_BUG_ON(i915_vma_is_closed(vma)); | |
3365e226 CW |
1037 | |
1038 | /* | |
1039 | * We defer actually closing, unbinding and destroying the VMA until | |
1040 | * the next idle point, or if the object is freed in the meantime. By | |
1041 | * postponing the unbind, we allow for it to be resurrected by the | |
1042 | * client, avoiding the work required to rebind the VMA. This is | |
1043 | * advantageous for DRI, where the client/server pass objects | |
1044 | * between themselves, temporarily opening a local VMA to the | |
1045 | * object, and then closing it again. The same object is then reused | |
1046 | * on the next frame (or two, depending on the depth of the swap queue) | |
1047 | * causing us to rebind the VMA once more. This ends up being a lot | |
1048 | * of wasted work for the steady state. | |
1049 | */ | |
71e51ca8 CW |
1050 | spin_lock_irqsave(>->closed_lock, flags); |
1051 | list_add(&vma->closed_link, >->closed_vma); | |
1052 | spin_unlock_irqrestore(>->closed_lock, flags); | |
3365e226 CW |
1053 | } |
1054 | ||
155ab883 | 1055 | static void __i915_vma_remove_closed(struct i915_vma *vma) |
3365e226 | 1056 | { |
71e51ca8 | 1057 | struct intel_gt *gt = vma->vm->gt; |
3365e226 | 1058 | |
71e51ca8 | 1059 | spin_lock_irq(>->closed_lock); |
155ab883 | 1060 | list_del_init(&vma->closed_link); |
71e51ca8 | 1061 | spin_unlock_irq(>->closed_lock); |
155ab883 CW |
1062 | } |
1063 | ||
1064 | void i915_vma_reopen(struct i915_vma *vma) | |
1065 | { | |
2850748e CW |
1066 | if (i915_vma_is_closed(vma)) |
1067 | __i915_vma_remove_closed(vma); | |
3365e226 CW |
1068 | } |
1069 | ||
76f9764c | 1070 | void i915_vma_release(struct kref *ref) |
b42fe9ca | 1071 | { |
76f9764c CW |
1072 | struct i915_vma *vma = container_of(ref, typeof(*vma), ref); |
1073 | ||
2850748e CW |
1074 | if (drm_mm_node_allocated(&vma->node)) { |
1075 | mutex_lock(&vma->vm->mutex); | |
1076 | atomic_and(~I915_VMA_PIN_MASK, &vma->flags); | |
1077 | WARN_ON(__i915_vma_unbind(vma)); | |
1078 | mutex_unlock(&vma->vm->mutex); | |
1079 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); | |
1080 | } | |
1081 | GEM_BUG_ON(i915_vma_is_active(vma)); | |
b42fe9ca | 1082 | |
528cbd17 CW |
1083 | if (vma->obj) { |
1084 | struct drm_i915_gem_object *obj = vma->obj; | |
1085 | ||
1086 | spin_lock(&obj->vma.lock); | |
1087 | list_del(&vma->obj_link); | |
2850748e | 1088 | rb_erase(&vma->obj_node, &obj->vma.tree); |
528cbd17 CW |
1089 | spin_unlock(&obj->vma.lock); |
1090 | } | |
010e3e68 | 1091 | |
155ab883 | 1092 | __i915_vma_remove_closed(vma); |
2850748e | 1093 | i915_vm_put(vma->vm); |
3365e226 | 1094 | |
2850748e CW |
1095 | i915_active_fini(&vma->active); |
1096 | i915_vma_free(vma); | |
3365e226 CW |
1097 | } |
1098 | ||
71e51ca8 | 1099 | void i915_vma_parked(struct intel_gt *gt) |
3365e226 CW |
1100 | { |
1101 | struct i915_vma *vma, *next; | |
b0647a5e | 1102 | LIST_HEAD(closed); |
b42fe9ca | 1103 | |
71e51ca8 CW |
1104 | spin_lock_irq(>->closed_lock); |
1105 | list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) { | |
2850748e CW |
1106 | struct drm_i915_gem_object *obj = vma->obj; |
1107 | struct i915_address_space *vm = vma->vm; | |
1108 | ||
1109 | /* XXX All to avoid keeping a reference on i915_vma itself */ | |
1110 | ||
1111 | if (!kref_get_unless_zero(&obj->base.refcount)) | |
1112 | continue; | |
1113 | ||
b0647a5e | 1114 | if (!i915_vm_tryopen(vm)) { |
2850748e | 1115 | i915_gem_object_put(obj); |
b0647a5e | 1116 | continue; |
2850748e CW |
1117 | } |
1118 | ||
b0647a5e CW |
1119 | list_move(&vma->closed_link, &closed); |
1120 | } | |
1121 | spin_unlock_irq(>->closed_lock); | |
155ab883 | 1122 | |
b0647a5e CW |
1123 | /* As the GT is held idle, no vma can be reopened as we destroy them */ |
1124 | list_for_each_entry_safe(vma, next, &closed, closed_link) { | |
1125 | struct drm_i915_gem_object *obj = vma->obj; | |
1126 | struct i915_address_space *vm = vma->vm; | |
3365e226 | 1127 | |
b0647a5e CW |
1128 | INIT_LIST_HEAD(&vma->closed_link); |
1129 | __i915_vma_put(vma); | |
2850748e | 1130 | |
b0647a5e CW |
1131 | i915_gem_object_put(obj); |
1132 | i915_vm_close(vm); | |
155ab883 | 1133 | } |
b42fe9ca JL |
1134 | } |
1135 | ||
1136 | static void __i915_vma_iounmap(struct i915_vma *vma) | |
1137 | { | |
1138 | GEM_BUG_ON(i915_vma_is_pinned(vma)); | |
1139 | ||
1140 | if (vma->iomap == NULL) | |
1141 | return; | |
1142 | ||
1143 | io_mapping_unmap(vma->iomap); | |
1144 | vma->iomap = NULL; | |
1145 | } | |
1146 | ||
a65adaf8 CW |
1147 | void i915_vma_revoke_mmap(struct i915_vma *vma) |
1148 | { | |
cc662126 | 1149 | struct drm_vma_offset_node *node; |
a65adaf8 CW |
1150 | u64 vma_offset; |
1151 | ||
a65adaf8 CW |
1152 | if (!i915_vma_has_userfault(vma)) |
1153 | return; | |
1154 | ||
1155 | GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma)); | |
1156 | GEM_BUG_ON(!vma->obj->userfault_count); | |
1157 | ||
cc662126 | 1158 | node = &vma->mmo->vma_node; |
a65adaf8 CW |
1159 | vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT; |
1160 | unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping, | |
1161 | drm_vma_node_offset_addr(node) + vma_offset, | |
1162 | vma->size, | |
1163 | 1); | |
1164 | ||
1165 | i915_vma_unset_userfault(vma); | |
1166 | if (!--vma->obj->userfault_count) | |
1167 | list_del(&vma->obj->userfault_link); | |
1168 | } | |
1169 | ||
2850748e CW |
1170 | int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq) |
1171 | { | |
1172 | int err; | |
1173 | ||
1174 | GEM_BUG_ON(!i915_vma_is_pinned(vma)); | |
1175 | ||
1176 | /* Wait for the vma to be bound before we start! */ | |
29e6ecf3 | 1177 | err = i915_request_await_active(rq, &vma->active, 0); |
2850748e CW |
1178 | if (err) |
1179 | return err; | |
1180 | ||
1181 | return i915_active_add_request(&vma->active, rq); | |
1182 | } | |
1183 | ||
e6bb1d7f CW |
1184 | int i915_vma_move_to_active(struct i915_vma *vma, |
1185 | struct i915_request *rq, | |
1186 | unsigned int flags) | |
1187 | { | |
1188 | struct drm_i915_gem_object *obj = vma->obj; | |
a93615f9 | 1189 | int err; |
e6bb1d7f | 1190 | |
6951e589 | 1191 | assert_object_held(obj); |
e6bb1d7f | 1192 | |
2850748e | 1193 | err = __i915_vma_move_to_active(vma, rq); |
a93615f9 CW |
1194 | if (unlikely(err)) |
1195 | return err; | |
e6bb1d7f | 1196 | |
e6bb1d7f | 1197 | if (flags & EXEC_OBJECT_WRITE) { |
da42104f CW |
1198 | struct intel_frontbuffer *front; |
1199 | ||
1200 | front = __intel_frontbuffer_get(obj); | |
1201 | if (unlikely(front)) { | |
1202 | if (intel_frontbuffer_invalidate(front, ORIGIN_CS)) | |
1203 | i915_active_add_request(&front->write, rq); | |
1204 | intel_frontbuffer_put(front); | |
1205 | } | |
e6bb1d7f | 1206 | |
829e8def | 1207 | dma_resv_add_excl_fence(vma->resv, &rq->fence); |
cd2a4eaf | 1208 | obj->write_domain = I915_GEM_DOMAIN_RENDER; |
e6bb1d7f | 1209 | obj->read_domains = 0; |
cd2a4eaf | 1210 | } else { |
829e8def | 1211 | err = dma_resv_reserve_shared(vma->resv, 1); |
cd2a4eaf CW |
1212 | if (unlikely(err)) |
1213 | return err; | |
1214 | ||
829e8def | 1215 | dma_resv_add_shared_fence(vma->resv, &rq->fence); |
cd2a4eaf | 1216 | obj->write_domain = 0; |
e6bb1d7f CW |
1217 | } |
1218 | obj->read_domains |= I915_GEM_GPU_DOMAINS; | |
a93615f9 | 1219 | obj->mm.dirty = true; |
e6bb1d7f | 1220 | |
a93615f9 | 1221 | GEM_BUG_ON(!i915_vma_is_active(vma)); |
e6bb1d7f CW |
1222 | return 0; |
1223 | } | |
1224 | ||
2850748e | 1225 | int __i915_vma_unbind(struct i915_vma *vma) |
b42fe9ca | 1226 | { |
b42fe9ca JL |
1227 | int ret; |
1228 | ||
2850748e | 1229 | lockdep_assert_held(&vma->vm->mutex); |
b42fe9ca | 1230 | |
10195b1e CW |
1231 | if (i915_vma_is_pinned(vma)) { |
1232 | vma_print_allocator(vma, "is pinned"); | |
d3e48352 | 1233 | return -EAGAIN; |
10195b1e | 1234 | } |
b42fe9ca | 1235 | |
60e94557 CW |
1236 | /* |
1237 | * After confirming that no one else is pinning this vma, wait for | |
1238 | * any laggards who may have crept in during the wait (through | |
1239 | * a residual pin skipping the vm->mutex) to complete. | |
1240 | */ | |
1241 | ret = i915_vma_sync(vma); | |
1242 | if (ret) | |
1243 | return ret; | |
1244 | ||
b42fe9ca | 1245 | if (!drm_mm_node_allocated(&vma->node)) |
3365e226 | 1246 | return 0; |
b42fe9ca | 1247 | |
60e94557 CW |
1248 | GEM_BUG_ON(i915_vma_is_pinned(vma)); |
1249 | GEM_BUG_ON(i915_vma_is_active(vma)); | |
1250 | ||
b42fe9ca | 1251 | if (i915_vma_is_map_and_fenceable(vma)) { |
7125397b CW |
1252 | /* |
1253 | * Check that we have flushed all writes through the GGTT | |
1254 | * before the unbind, other due to non-strict nature of those | |
1255 | * indirect writes they may end up referencing the GGTT PTE | |
1256 | * after the unbind. | |
5424f5d7 CW |
1257 | * |
1258 | * Note that we may be concurrently poking at the GGTT_WRITE | |
1259 | * bit from set-domain, as we mark all GGTT vma associated | |
1260 | * with an object. We know this is for another vma, as we | |
1261 | * are currently unbinding this one -- so if this vma will be | |
1262 | * reused, it will be refaulted and have its dirty bit set | |
1263 | * before the next write. | |
7125397b CW |
1264 | */ |
1265 | i915_vma_flush_writes(vma); | |
7125397b | 1266 | |
b42fe9ca | 1267 | /* release the fence reg _after_ flushing */ |
1f7fd484 | 1268 | ret = i915_vma_revoke_fence(vma); |
b42fe9ca JL |
1269 | if (ret) |
1270 | return ret; | |
1271 | ||
1272 | /* Force a pagefault for domain tracking on next user access */ | |
a65adaf8 | 1273 | i915_vma_revoke_mmap(vma); |
b42fe9ca JL |
1274 | |
1275 | __i915_vma_iounmap(vma); | |
4dd2fbbf | 1276 | clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); |
b42fe9ca | 1277 | } |
a65adaf8 CW |
1278 | GEM_BUG_ON(vma->fence); |
1279 | GEM_BUG_ON(i915_vma_has_userfault(vma)); | |
b42fe9ca | 1280 | |
2850748e | 1281 | if (likely(atomic_read(&vma->vm->open))) { |
b42fe9ca | 1282 | trace_i915_vma_unbind(vma); |
93f2cde2 | 1283 | vma->ops->unbind_vma(vma); |
b42fe9ca | 1284 | } |
5424f5d7 CW |
1285 | atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE), |
1286 | &vma->flags); | |
b42fe9ca | 1287 | |
dde01d94 | 1288 | i915_vma_detach(vma); |
2850748e | 1289 | vma_unbind_pages(vma); |
b42fe9ca | 1290 | |
76f9764c | 1291 | drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ |
b42fe9ca JL |
1292 | return 0; |
1293 | } | |
1294 | ||
2850748e CW |
1295 | int i915_vma_unbind(struct i915_vma *vma) |
1296 | { | |
1297 | struct i915_address_space *vm = vma->vm; | |
c0e60347 | 1298 | intel_wakeref_t wakeref = 0; |
2850748e CW |
1299 | int err; |
1300 | ||
e6ba7648 CW |
1301 | if (!drm_mm_node_allocated(&vma->node)) |
1302 | return 0; | |
1303 | ||
d62f416f CW |
1304 | /* Optimistic wait before taking the mutex */ |
1305 | err = i915_vma_sync(vma); | |
1306 | if (err) | |
1307 | goto out_rpm; | |
1308 | ||
af23facc CW |
1309 | if (i915_vma_is_pinned(vma)) { |
1310 | vma_print_allocator(vma, "is pinned"); | |
1311 | return -EAGAIN; | |
1312 | } | |
1313 | ||
1314 | if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) | |
1315 | /* XXX not always required: nop_clear_range */ | |
1316 | wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); | |
1317 | ||
2850748e CW |
1318 | err = mutex_lock_interruptible(&vm->mutex); |
1319 | if (err) | |
d62f416f | 1320 | goto out_rpm; |
2850748e CW |
1321 | |
1322 | err = __i915_vma_unbind(vma); | |
1323 | mutex_unlock(&vm->mutex); | |
1324 | ||
d62f416f | 1325 | out_rpm: |
c0e60347 CW |
1326 | if (wakeref) |
1327 | intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); | |
2850748e CW |
1328 | return err; |
1329 | } | |
1330 | ||
1aff1903 CW |
1331 | struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma) |
1332 | { | |
1333 | i915_gem_object_make_unshrinkable(vma->obj); | |
1334 | return vma; | |
1335 | } | |
1336 | ||
1337 | void i915_vma_make_shrinkable(struct i915_vma *vma) | |
1338 | { | |
1339 | i915_gem_object_make_shrinkable(vma->obj); | |
1340 | } | |
1341 | ||
1342 | void i915_vma_make_purgeable(struct i915_vma *vma) | |
1343 | { | |
1344 | i915_gem_object_make_purgeable(vma->obj); | |
1345 | } | |
1346 | ||
e3c7a1c5 CW |
1347 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) |
1348 | #include "selftests/i915_vma.c" | |
1349 | #endif | |
13f1bfd3 | 1350 | |
103b76ee | 1351 | static void i915_global_vma_shrink(void) |
13f1bfd3 | 1352 | { |
103b76ee | 1353 | kmem_cache_shrink(global.slab_vmas); |
13f1bfd3 CW |
1354 | } |
1355 | ||
103b76ee | 1356 | static void i915_global_vma_exit(void) |
13f1bfd3 | 1357 | { |
103b76ee | 1358 | kmem_cache_destroy(global.slab_vmas); |
13f1bfd3 CW |
1359 | } |
1360 | ||
103b76ee CW |
1361 | static struct i915_global_vma global = { { |
1362 | .shrink = i915_global_vma_shrink, | |
1363 | .exit = i915_global_vma_exit, | |
1364 | } }; | |
1365 | ||
1366 | int __init i915_global_vma_init(void) | |
13f1bfd3 | 1367 | { |
103b76ee CW |
1368 | global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN); |
1369 | if (!global.slab_vmas) | |
1370 | return -ENOMEM; | |
1371 | ||
1372 | i915_global_register(&global.base); | |
1373 | return 0; | |
13f1bfd3 | 1374 | } |