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 <linux/sched/mm.h> |
26 | #include <linux/dma-fence-array.h> |
27 | #include <drm/drm_gem.h> |
28 | |
29 | #include "display/intel_display.h" |
30 | #include "display/intel_frontbuffer.h" |
31 | #include "gem/i915_gem_lmem.h" |
32 | #include "gem/i915_gem_object_frontbuffer.h" |
33 | #include "gem/i915_gem_tiling.h" |
34 | #include "gt/intel_engine.h" |
35 | #include "gt/intel_engine_heartbeat.h" |
36 | #include "gt/intel_gt.h" |
37 | #include "gt/intel_gt_pm.h" |
38 | #include "gt/intel_gt_requests.h" |
39 | #include "gt/intel_tlb.h" |
40 | |
41 | #include "i915_drv.h" |
42 | #include "i915_gem_evict.h" |
43 | #include "i915_sw_fence_work.h" |
44 | #include "i915_trace.h" |
45 | #include "i915_vma.h" |
46 | #include "i915_vma_resource.h" |
47 | |
48 | static inline void assert_vma_held_evict(const struct i915_vma *vma) |
49 | { |
50 | /* |
51 | * We may be forced to unbind when the vm is dead, to clean it up. |
52 | * This is the only exception to the requirement of the object lock |
53 | * being held. |
54 | */ |
55 | if (kref_read(kref: &vma->vm->ref)) |
56 | assert_object_held_shared(obj: vma->obj); |
57 | } |
58 | |
59 | static struct kmem_cache *slab_vmas; |
60 | |
61 | static struct i915_vma *i915_vma_alloc(void) |
62 | { |
63 | return kmem_cache_zalloc(k: slab_vmas, GFP_KERNEL); |
64 | } |
65 | |
66 | static void i915_vma_free(struct i915_vma *vma) |
67 | { |
68 | return kmem_cache_free(s: slab_vmas, objp: vma); |
69 | } |
70 | |
71 | #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM) |
72 | |
73 | #include <linux/stackdepot.h> |
74 | |
75 | static void vma_print_allocator(struct i915_vma *vma, const char *reason) |
76 | { |
77 | char buf[512]; |
78 | |
79 | if (!vma->node.stack) { |
80 | drm_dbg(vma->obj->base.dev, |
81 | "vma.node [%08llx + %08llx] %s: unknown owner\n" , |
82 | vma->node.start, vma->node.size, reason); |
83 | return; |
84 | } |
85 | |
86 | stack_depot_snprint(vma->node.stack, buf, sizeof(buf), 0); |
87 | drm_dbg(vma->obj->base.dev, |
88 | "vma.node [%08llx + %08llx] %s: inserted at %s\n" , |
89 | vma->node.start, vma->node.size, reason, buf); |
90 | } |
91 | |
92 | #else |
93 | |
94 | static void vma_print_allocator(struct i915_vma *vma, const char *reason) |
95 | { |
96 | } |
97 | |
98 | #endif |
99 | |
100 | static inline struct i915_vma *active_to_vma(struct i915_active *ref) |
101 | { |
102 | return container_of(ref, typeof(struct i915_vma), active); |
103 | } |
104 | |
105 | static int __i915_vma_active(struct i915_active *ref) |
106 | { |
107 | struct i915_vma *vma = active_to_vma(ref); |
108 | |
109 | if (!i915_vma_tryget(vma)) |
110 | return -ENOENT; |
111 | |
112 | /* |
113 | * Exclude global GTT VMA from holding a GT wakeref |
114 | * while active, otherwise GPU never goes idle. |
115 | */ |
116 | if (!i915_vma_is_ggtt(vma)) { |
117 | /* |
118 | * Since we and our _retire() counterpart can be |
119 | * called asynchronously, storing a wakeref tracking |
120 | * handle inside struct i915_vma is not safe, and |
121 | * there is no other good place for that. Hence, |
122 | * use untracked variants of intel_gt_pm_get/put(). |
123 | */ |
124 | intel_gt_pm_get_untracked(gt: vma->vm->gt); |
125 | } |
126 | |
127 | return 0; |
128 | } |
129 | |
130 | static void __i915_vma_retire(struct i915_active *ref) |
131 | { |
132 | struct i915_vma *vma = active_to_vma(ref); |
133 | |
134 | if (!i915_vma_is_ggtt(vma)) { |
135 | /* |
136 | * Since we can be called from atomic contexts, |
137 | * use an async variant of intel_gt_pm_put(). |
138 | */ |
139 | intel_gt_pm_put_async_untracked(gt: vma->vm->gt); |
140 | } |
141 | |
142 | i915_vma_put(vma); |
143 | } |
144 | |
145 | static struct i915_vma * |
146 | vma_create(struct drm_i915_gem_object *obj, |
147 | struct i915_address_space *vm, |
148 | const struct i915_gtt_view *view) |
149 | { |
150 | struct i915_vma *pos = ERR_PTR(error: -E2BIG); |
151 | struct i915_vma *vma; |
152 | struct rb_node *rb, **p; |
153 | int err; |
154 | |
155 | /* The aliasing_ppgtt should never be used directly! */ |
156 | GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm); |
157 | |
158 | vma = i915_vma_alloc(); |
159 | if (vma == NULL) |
160 | return ERR_PTR(error: -ENOMEM); |
161 | |
162 | vma->ops = &vm->vma_ops; |
163 | vma->obj = obj; |
164 | vma->size = obj->base.size; |
165 | vma->display_alignment = I915_GTT_MIN_ALIGNMENT; |
166 | |
167 | i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0); |
168 | |
169 | /* Declare ourselves safe for use inside shrinkers */ |
170 | if (IS_ENABLED(CONFIG_LOCKDEP)) { |
171 | fs_reclaim_acquire(GFP_KERNEL); |
172 | might_lock(&vma->active.mutex); |
173 | fs_reclaim_release(GFP_KERNEL); |
174 | } |
175 | |
176 | INIT_LIST_HEAD(list: &vma->closed_link); |
177 | INIT_LIST_HEAD(list: &vma->obj_link); |
178 | RB_CLEAR_NODE(&vma->obj_node); |
179 | |
180 | if (view && view->type != I915_GTT_VIEW_NORMAL) { |
181 | vma->gtt_view = *view; |
182 | if (view->type == I915_GTT_VIEW_PARTIAL) { |
183 | GEM_BUG_ON(range_overflows_t(u64, |
184 | view->partial.offset, |
185 | view->partial.size, |
186 | obj->base.size >> PAGE_SHIFT)); |
187 | vma->size = view->partial.size; |
188 | vma->size <<= PAGE_SHIFT; |
189 | GEM_BUG_ON(vma->size > obj->base.size); |
190 | } else if (view->type == I915_GTT_VIEW_ROTATED) { |
191 | vma->size = intel_rotation_info_size(rot_info: &view->rotated); |
192 | vma->size <<= PAGE_SHIFT; |
193 | } else if (view->type == I915_GTT_VIEW_REMAPPED) { |
194 | vma->size = intel_remapped_info_size(rem_info: &view->remapped); |
195 | vma->size <<= PAGE_SHIFT; |
196 | } |
197 | } |
198 | |
199 | if (unlikely(vma->size > vm->total)) |
200 | goto err_vma; |
201 | |
202 | GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE)); |
203 | |
204 | err = mutex_lock_interruptible(&vm->mutex); |
205 | if (err) { |
206 | pos = ERR_PTR(error: err); |
207 | goto err_vma; |
208 | } |
209 | |
210 | vma->vm = vm; |
211 | list_add_tail(new: &vma->vm_link, head: &vm->unbound_list); |
212 | |
213 | spin_lock(lock: &obj->vma.lock); |
214 | if (i915_is_ggtt(vm)) { |
215 | if (unlikely(overflows_type(vma->size, u32))) |
216 | goto err_unlock; |
217 | |
218 | vma->fence_size = i915_gem_fence_size(i915: vm->i915, size: vma->size, |
219 | tiling: i915_gem_object_get_tiling(obj), |
220 | stride: i915_gem_object_get_stride(obj)); |
221 | if (unlikely(vma->fence_size < vma->size || /* overflow */ |
222 | vma->fence_size > vm->total)) |
223 | goto err_unlock; |
224 | |
225 | GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT)); |
226 | |
227 | vma->fence_alignment = i915_gem_fence_alignment(i915: vm->i915, size: vma->size, |
228 | tiling: i915_gem_object_get_tiling(obj), |
229 | stride: i915_gem_object_get_stride(obj)); |
230 | GEM_BUG_ON(!is_power_of_2(vma->fence_alignment)); |
231 | |
232 | __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma)); |
233 | } |
234 | |
235 | rb = NULL; |
236 | p = &obj->vma.tree.rb_node; |
237 | while (*p) { |
238 | long cmp; |
239 | |
240 | rb = *p; |
241 | pos = rb_entry(rb, struct i915_vma, obj_node); |
242 | |
243 | /* |
244 | * If the view already exists in the tree, another thread |
245 | * already created a matching vma, so return the older instance |
246 | * and dispose of ours. |
247 | */ |
248 | cmp = i915_vma_compare(vma: pos, vm, view); |
249 | if (cmp < 0) |
250 | p = &rb->rb_right; |
251 | else if (cmp > 0) |
252 | p = &rb->rb_left; |
253 | else |
254 | goto err_unlock; |
255 | } |
256 | rb_link_node(node: &vma->obj_node, parent: rb, rb_link: p); |
257 | rb_insert_color(&vma->obj_node, &obj->vma.tree); |
258 | |
259 | if (i915_vma_is_ggtt(vma)) |
260 | /* |
261 | * We put the GGTT vma at the start of the vma-list, followed |
262 | * by the ppGGTT vma. This allows us to break early when |
263 | * iterating over only the GGTT vma for an object, see |
264 | * for_each_ggtt_vma() |
265 | */ |
266 | list_add(new: &vma->obj_link, head: &obj->vma.list); |
267 | else |
268 | list_add_tail(new: &vma->obj_link, head: &obj->vma.list); |
269 | |
270 | spin_unlock(lock: &obj->vma.lock); |
271 | mutex_unlock(lock: &vm->mutex); |
272 | |
273 | return vma; |
274 | |
275 | err_unlock: |
276 | spin_unlock(lock: &obj->vma.lock); |
277 | list_del_init(entry: &vma->vm_link); |
278 | mutex_unlock(lock: &vm->mutex); |
279 | err_vma: |
280 | i915_vma_free(vma); |
281 | return pos; |
282 | } |
283 | |
284 | static struct i915_vma * |
285 | i915_vma_lookup(struct drm_i915_gem_object *obj, |
286 | struct i915_address_space *vm, |
287 | const struct i915_gtt_view *view) |
288 | { |
289 | struct rb_node *rb; |
290 | |
291 | rb = obj->vma.tree.rb_node; |
292 | while (rb) { |
293 | struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node); |
294 | long cmp; |
295 | |
296 | cmp = i915_vma_compare(vma, vm, view); |
297 | if (cmp == 0) |
298 | return vma; |
299 | |
300 | if (cmp < 0) |
301 | rb = rb->rb_right; |
302 | else |
303 | rb = rb->rb_left; |
304 | } |
305 | |
306 | return NULL; |
307 | } |
308 | |
309 | /** |
310 | * i915_vma_instance - return the singleton instance of the VMA |
311 | * @obj: parent &struct drm_i915_gem_object to be mapped |
312 | * @vm: address space in which the mapping is located |
313 | * @view: additional mapping requirements |
314 | * |
315 | * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with |
316 | * the same @view characteristics. If a match is not found, one is created. |
317 | * Once created, the VMA is kept until either the object is freed, or the |
318 | * address space is closed. |
319 | * |
320 | * Returns the vma, or an error pointer. |
321 | */ |
322 | struct i915_vma * |
323 | i915_vma_instance(struct drm_i915_gem_object *obj, |
324 | struct i915_address_space *vm, |
325 | const struct i915_gtt_view *view) |
326 | { |
327 | struct i915_vma *vma; |
328 | |
329 | GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm)); |
330 | GEM_BUG_ON(!kref_read(&vm->ref)); |
331 | |
332 | spin_lock(lock: &obj->vma.lock); |
333 | vma = i915_vma_lookup(obj, vm, view); |
334 | spin_unlock(lock: &obj->vma.lock); |
335 | |
336 | /* vma_create() will resolve the race if another creates the vma */ |
337 | if (unlikely(!vma)) |
338 | vma = vma_create(obj, vm, view); |
339 | |
340 | GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view)); |
341 | return vma; |
342 | } |
343 | |
344 | struct i915_vma_work { |
345 | struct dma_fence_work base; |
346 | struct i915_address_space *vm; |
347 | struct i915_vm_pt_stash stash; |
348 | struct i915_vma_resource *vma_res; |
349 | struct drm_i915_gem_object *obj; |
350 | struct i915_sw_dma_fence_cb cb; |
351 | unsigned int pat_index; |
352 | unsigned int flags; |
353 | }; |
354 | |
355 | static void __vma_bind(struct dma_fence_work *work) |
356 | { |
357 | struct i915_vma_work *vw = container_of(work, typeof(*vw), base); |
358 | struct i915_vma_resource *vma_res = vw->vma_res; |
359 | |
360 | /* |
361 | * We are about the bind the object, which must mean we have already |
362 | * signaled the work to potentially clear/move the pages underneath. If |
363 | * something went wrong at that stage then the object should have |
364 | * unknown_state set, in which case we need to skip the bind. |
365 | */ |
366 | if (i915_gem_object_has_unknown_state(obj: vw->obj)) |
367 | return; |
368 | |
369 | vma_res->ops->bind_vma(vma_res->vm, &vw->stash, |
370 | vma_res, vw->pat_index, vw->flags); |
371 | } |
372 | |
373 | static void __vma_release(struct dma_fence_work *work) |
374 | { |
375 | struct i915_vma_work *vw = container_of(work, typeof(*vw), base); |
376 | |
377 | if (vw->obj) |
378 | i915_gem_object_put(obj: vw->obj); |
379 | |
380 | i915_vm_free_pt_stash(vm: vw->vm, stash: &vw->stash); |
381 | if (vw->vma_res) |
382 | i915_vma_resource_put(vma_res: vw->vma_res); |
383 | } |
384 | |
385 | static const struct dma_fence_work_ops bind_ops = { |
386 | .name = "bind" , |
387 | .work = __vma_bind, |
388 | .release = __vma_release, |
389 | }; |
390 | |
391 | struct i915_vma_work *i915_vma_work(void) |
392 | { |
393 | struct i915_vma_work *vw; |
394 | |
395 | vw = kzalloc(size: sizeof(*vw), GFP_KERNEL); |
396 | if (!vw) |
397 | return NULL; |
398 | |
399 | dma_fence_work_init(f: &vw->base, ops: &bind_ops); |
400 | vw->base.dma.error = -EAGAIN; /* disable the worker by default */ |
401 | |
402 | return vw; |
403 | } |
404 | |
405 | int i915_vma_wait_for_bind(struct i915_vma *vma) |
406 | { |
407 | int err = 0; |
408 | |
409 | if (rcu_access_pointer(vma->active.excl.fence)) { |
410 | struct dma_fence *fence; |
411 | |
412 | rcu_read_lock(); |
413 | fence = dma_fence_get_rcu_safe(fencep: &vma->active.excl.fence); |
414 | rcu_read_unlock(); |
415 | if (fence) { |
416 | err = dma_fence_wait(fence, intr: true); |
417 | dma_fence_put(fence); |
418 | } |
419 | } |
420 | |
421 | return err; |
422 | } |
423 | |
424 | #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) |
425 | static int i915_vma_verify_bind_complete(struct i915_vma *vma) |
426 | { |
427 | struct dma_fence *fence = i915_active_fence_get(&vma->active.excl); |
428 | int err; |
429 | |
430 | if (!fence) |
431 | return 0; |
432 | |
433 | if (dma_fence_is_signaled(fence)) |
434 | err = fence->error; |
435 | else |
436 | err = -EBUSY; |
437 | |
438 | dma_fence_put(fence); |
439 | |
440 | return err; |
441 | } |
442 | #else |
443 | #define i915_vma_verify_bind_complete(_vma) 0 |
444 | #endif |
445 | |
446 | I915_SELFTEST_EXPORT void |
447 | i915_vma_resource_init_from_vma(struct i915_vma_resource *vma_res, |
448 | struct i915_vma *vma) |
449 | { |
450 | struct drm_i915_gem_object *obj = vma->obj; |
451 | |
452 | i915_vma_resource_init(vma_res, vm: vma->vm, pages: vma->pages, page_sizes: &vma->page_sizes, |
453 | pages_rsgt: obj->mm.rsgt, readonly: i915_gem_object_is_readonly(obj), |
454 | lmem: i915_gem_object_is_lmem(obj), mr: obj->mm.region, |
455 | ops: vma->ops, private: vma->private, start: __i915_vma_offset(vma), |
456 | node_size: __i915_vma_size(vma), size: vma->size, guard: vma->guard); |
457 | } |
458 | |
459 | /** |
460 | * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space. |
461 | * @vma: VMA to map |
462 | * @pat_index: PAT index to set in PTE |
463 | * @flags: flags like global or local mapping |
464 | * @work: preallocated worker for allocating and binding the PTE |
465 | * @vma_res: pointer to a preallocated vma resource. The resource is either |
466 | * consumed or freed. |
467 | * |
468 | * DMA addresses are taken from the scatter-gather table of this object (or of |
469 | * this VMA in case of non-default GGTT views) and PTE entries set up. |
470 | * Note that DMA addresses are also the only part of the SG table we care about. |
471 | */ |
472 | int i915_vma_bind(struct i915_vma *vma, |
473 | unsigned int pat_index, |
474 | u32 flags, |
475 | struct i915_vma_work *work, |
476 | struct i915_vma_resource *vma_res) |
477 | { |
478 | u32 bind_flags; |
479 | u32 vma_flags; |
480 | int ret; |
481 | |
482 | lockdep_assert_held(&vma->vm->mutex); |
483 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); |
484 | GEM_BUG_ON(vma->size > i915_vma_size(vma)); |
485 | |
486 | if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start, |
487 | vma->node.size, |
488 | vma->vm->total))) { |
489 | i915_vma_resource_free(vma_res); |
490 | return -ENODEV; |
491 | } |
492 | |
493 | if (GEM_DEBUG_WARN_ON(!flags)) { |
494 | i915_vma_resource_free(vma_res); |
495 | return -EINVAL; |
496 | } |
497 | |
498 | bind_flags = flags; |
499 | bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; |
500 | |
501 | vma_flags = atomic_read(v: &vma->flags); |
502 | vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; |
503 | |
504 | bind_flags &= ~vma_flags; |
505 | if (bind_flags == 0) { |
506 | i915_vma_resource_free(vma_res); |
507 | return 0; |
508 | } |
509 | |
510 | GEM_BUG_ON(!atomic_read(&vma->pages_count)); |
511 | |
512 | /* Wait for or await async unbinds touching our range */ |
513 | if (work && bind_flags & vma->vm->bind_async_flags) |
514 | ret = i915_vma_resource_bind_dep_await(vm: vma->vm, |
515 | sw_fence: &work->base.chain, |
516 | first: vma->node.start, |
517 | last: vma->node.size, |
518 | intr: true, |
519 | GFP_NOWAIT | |
520 | __GFP_RETRY_MAYFAIL | |
521 | __GFP_NOWARN); |
522 | else |
523 | ret = i915_vma_resource_bind_dep_sync(vm: vma->vm, first: vma->node.start, |
524 | last: vma->node.size, intr: true); |
525 | if (ret) { |
526 | i915_vma_resource_free(vma_res); |
527 | return ret; |
528 | } |
529 | |
530 | if (vma->resource || !vma_res) { |
531 | /* Rebinding with an additional I915_VMA_*_BIND */ |
532 | GEM_WARN_ON(!vma_flags); |
533 | i915_vma_resource_free(vma_res); |
534 | } else { |
535 | i915_vma_resource_init_from_vma(vma_res, vma); |
536 | vma->resource = vma_res; |
537 | } |
538 | trace_i915_vma_bind(vma, flags: bind_flags); |
539 | if (work && bind_flags & vma->vm->bind_async_flags) { |
540 | struct dma_fence *prev; |
541 | |
542 | work->vma_res = i915_vma_resource_get(vma_res: vma->resource); |
543 | work->pat_index = pat_index; |
544 | work->flags = bind_flags; |
545 | |
546 | /* |
547 | * Note we only want to chain up to the migration fence on |
548 | * the pages (not the object itself). As we don't track that, |
549 | * yet, we have to use the exclusive fence instead. |
550 | * |
551 | * Also note that we do not want to track the async vma as |
552 | * part of the obj->resv->excl_fence as it only affects |
553 | * execution and not content or object's backing store lifetime. |
554 | */ |
555 | prev = i915_active_set_exclusive(ref: &vma->active, f: &work->base.dma); |
556 | if (prev) { |
557 | __i915_sw_fence_await_dma_fence(fence: &work->base.chain, |
558 | dma: prev, |
559 | cb: &work->cb); |
560 | dma_fence_put(fence: prev); |
561 | } |
562 | |
563 | work->base.dma.error = 0; /* enable the queue_work() */ |
564 | work->obj = i915_gem_object_get(obj: vma->obj); |
565 | } else { |
566 | ret = i915_gem_object_wait_moving_fence(obj: vma->obj, intr: true); |
567 | if (ret) { |
568 | i915_vma_resource_free(vma_res: vma->resource); |
569 | vma->resource = NULL; |
570 | |
571 | return ret; |
572 | } |
573 | vma->ops->bind_vma(vma->vm, NULL, vma->resource, pat_index, |
574 | bind_flags); |
575 | } |
576 | |
577 | atomic_or(i: bind_flags, v: &vma->flags); |
578 | return 0; |
579 | } |
580 | |
581 | void __iomem *i915_vma_pin_iomap(struct i915_vma *vma) |
582 | { |
583 | void __iomem *ptr; |
584 | int err; |
585 | |
586 | if (WARN_ON_ONCE(vma->obj->flags & I915_BO_ALLOC_GPU_ONLY)) |
587 | return IOMEM_ERR_PTR(-EINVAL); |
588 | |
589 | GEM_BUG_ON(!i915_vma_is_ggtt(vma)); |
590 | GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)); |
591 | GEM_BUG_ON(i915_vma_verify_bind_complete(vma)); |
592 | |
593 | ptr = READ_ONCE(vma->iomap); |
594 | if (ptr == NULL) { |
595 | /* |
596 | * TODO: consider just using i915_gem_object_pin_map() for lmem |
597 | * instead, which already supports mapping non-contiguous chunks |
598 | * of pages, that way we can also drop the |
599 | * I915_BO_ALLOC_CONTIGUOUS when allocating the object. |
600 | */ |
601 | if (i915_gem_object_is_lmem(obj: vma->obj)) { |
602 | ptr = i915_gem_object_lmem_io_map(obj: vma->obj, n: 0, |
603 | size: vma->obj->base.size); |
604 | } else if (i915_vma_is_map_and_fenceable(vma)) { |
605 | ptr = io_mapping_map_wc(mapping: &i915_vm_to_ggtt(vm: vma->vm)->iomap, |
606 | offset: i915_vma_offset(vma), |
607 | size: i915_vma_size(vma)); |
608 | } else { |
609 | ptr = (void __iomem *) |
610 | i915_gem_object_pin_map(obj: vma->obj, type: I915_MAP_WC); |
611 | if (IS_ERR(ptr)) { |
612 | err = PTR_ERR(ptr); |
613 | goto err; |
614 | } |
615 | ptr = page_pack_bits(ptr, 1); |
616 | } |
617 | |
618 | if (ptr == NULL) { |
619 | err = -ENOMEM; |
620 | goto err; |
621 | } |
622 | |
623 | if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) { |
624 | if (page_unmask_bits(ptr)) |
625 | __i915_gem_object_release_map(obj: vma->obj); |
626 | else |
627 | io_mapping_unmap(vaddr: ptr); |
628 | ptr = vma->iomap; |
629 | } |
630 | } |
631 | |
632 | __i915_vma_pin(vma); |
633 | |
634 | err = i915_vma_pin_fence(vma); |
635 | if (err) |
636 | goto err_unpin; |
637 | |
638 | i915_vma_set_ggtt_write(vma); |
639 | |
640 | /* NB Access through the GTT requires the device to be awake. */ |
641 | return page_mask_bits(ptr); |
642 | |
643 | err_unpin: |
644 | __i915_vma_unpin(vma); |
645 | err: |
646 | return IOMEM_ERR_PTR(err); |
647 | } |
648 | |
649 | void i915_vma_flush_writes(struct i915_vma *vma) |
650 | { |
651 | if (i915_vma_unset_ggtt_write(vma)) |
652 | intel_gt_flush_ggtt_writes(gt: vma->vm->gt); |
653 | } |
654 | |
655 | void i915_vma_unpin_iomap(struct i915_vma *vma) |
656 | { |
657 | GEM_BUG_ON(vma->iomap == NULL); |
658 | |
659 | /* XXX We keep the mapping until __i915_vma_unbind()/evict() */ |
660 | |
661 | i915_vma_flush_writes(vma); |
662 | |
663 | i915_vma_unpin_fence(vma); |
664 | i915_vma_unpin(vma); |
665 | } |
666 | |
667 | void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags) |
668 | { |
669 | struct i915_vma *vma; |
670 | struct drm_i915_gem_object *obj; |
671 | |
672 | vma = fetch_and_zero(p_vma); |
673 | if (!vma) |
674 | return; |
675 | |
676 | obj = vma->obj; |
677 | GEM_BUG_ON(!obj); |
678 | |
679 | i915_vma_unpin(vma); |
680 | |
681 | if (flags & I915_VMA_RELEASE_MAP) |
682 | i915_gem_object_unpin_map(obj); |
683 | |
684 | i915_gem_object_put(obj); |
685 | } |
686 | |
687 | bool i915_vma_misplaced(const struct i915_vma *vma, |
688 | u64 size, u64 alignment, u64 flags) |
689 | { |
690 | if (!drm_mm_node_allocated(node: &vma->node)) |
691 | return false; |
692 | |
693 | if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma))) |
694 | return true; |
695 | |
696 | if (i915_vma_size(vma) < size) |
697 | return true; |
698 | |
699 | GEM_BUG_ON(alignment && !is_power_of_2(alignment)); |
700 | if (alignment && !IS_ALIGNED(i915_vma_offset(vma), alignment)) |
701 | return true; |
702 | |
703 | if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma)) |
704 | return true; |
705 | |
706 | if (flags & PIN_OFFSET_BIAS && |
707 | i915_vma_offset(vma) < (flags & PIN_OFFSET_MASK)) |
708 | return true; |
709 | |
710 | if (flags & PIN_OFFSET_FIXED && |
711 | i915_vma_offset(vma) != (flags & PIN_OFFSET_MASK)) |
712 | return true; |
713 | |
714 | if (flags & PIN_OFFSET_GUARD && |
715 | vma->guard < (flags & PIN_OFFSET_MASK)) |
716 | return true; |
717 | |
718 | return false; |
719 | } |
720 | |
721 | void __i915_vma_set_map_and_fenceable(struct i915_vma *vma) |
722 | { |
723 | bool mappable, fenceable; |
724 | |
725 | GEM_BUG_ON(!i915_vma_is_ggtt(vma)); |
726 | GEM_BUG_ON(!vma->fence_size); |
727 | |
728 | fenceable = (i915_vma_size(vma) >= vma->fence_size && |
729 | IS_ALIGNED(i915_vma_offset(vma), vma->fence_alignment)); |
730 | |
731 | mappable = i915_ggtt_offset(vma) + vma->fence_size <= |
732 | i915_vm_to_ggtt(vm: vma->vm)->mappable_end; |
733 | |
734 | if (mappable && fenceable) |
735 | set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); |
736 | else |
737 | clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); |
738 | } |
739 | |
740 | bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color) |
741 | { |
742 | struct drm_mm_node *node = &vma->node; |
743 | struct drm_mm_node *other; |
744 | |
745 | /* |
746 | * On some machines we have to be careful when putting differing types |
747 | * of snoopable memory together to avoid the prefetcher crossing memory |
748 | * domains and dying. During vm initialisation, we decide whether or not |
749 | * these constraints apply and set the drm_mm.color_adjust |
750 | * appropriately. |
751 | */ |
752 | if (!i915_vm_has_cache_coloring(vm: vma->vm)) |
753 | return true; |
754 | |
755 | /* Only valid to be called on an already inserted vma */ |
756 | GEM_BUG_ON(!drm_mm_node_allocated(node)); |
757 | GEM_BUG_ON(list_empty(&node->node_list)); |
758 | |
759 | other = list_prev_entry(node, node_list); |
760 | if (i915_node_color_differs(node: other, color) && |
761 | !drm_mm_hole_follows(node: other)) |
762 | return false; |
763 | |
764 | other = list_next_entry(node, node_list); |
765 | if (i915_node_color_differs(node: other, color) && |
766 | !drm_mm_hole_follows(node)) |
767 | return false; |
768 | |
769 | return true; |
770 | } |
771 | |
772 | /** |
773 | * i915_vma_insert - finds a slot for the vma in its address space |
774 | * @vma: the vma |
775 | * @ww: An optional struct i915_gem_ww_ctx |
776 | * @size: requested size in bytes (can be larger than the VMA) |
777 | * @alignment: required alignment |
778 | * @flags: mask of PIN_* flags to use |
779 | * |
780 | * First we try to allocate some free space that meets the requirements for |
781 | * the VMA. Failiing that, if the flags permit, it will evict an old VMA, |
782 | * preferrably the oldest idle entry to make room for the new VMA. |
783 | * |
784 | * Returns: |
785 | * 0 on success, negative error code otherwise. |
786 | */ |
787 | static int |
788 | i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
789 | u64 size, u64 alignment, u64 flags) |
790 | { |
791 | unsigned long color, guard; |
792 | u64 start, end; |
793 | int ret; |
794 | |
795 | GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); |
796 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); |
797 | GEM_BUG_ON(hweight64(flags & (PIN_OFFSET_GUARD | PIN_OFFSET_FIXED | PIN_OFFSET_BIAS)) > 1); |
798 | |
799 | size = max(size, vma->size); |
800 | alignment = max_t(typeof(alignment), alignment, vma->display_alignment); |
801 | if (flags & PIN_MAPPABLE) { |
802 | size = max_t(typeof(size), size, vma->fence_size); |
803 | alignment = max_t(typeof(alignment), |
804 | alignment, vma->fence_alignment); |
805 | } |
806 | |
807 | GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); |
808 | GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT)); |
809 | GEM_BUG_ON(!is_power_of_2(alignment)); |
810 | |
811 | guard = vma->guard; /* retain guard across rebinds */ |
812 | if (flags & PIN_OFFSET_GUARD) { |
813 | GEM_BUG_ON(overflows_type(flags & PIN_OFFSET_MASK, u32)); |
814 | guard = max_t(u32, guard, flags & PIN_OFFSET_MASK); |
815 | } |
816 | /* |
817 | * As we align the node upon insertion, but the hardware gets |
818 | * node.start + guard, the easiest way to make that work is |
819 | * to make the guard a multiple of the alignment size. |
820 | */ |
821 | guard = ALIGN(guard, alignment); |
822 | |
823 | start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0; |
824 | GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); |
825 | |
826 | end = vma->vm->total; |
827 | if (flags & PIN_MAPPABLE) |
828 | end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end); |
829 | if (flags & PIN_ZONE_4G) |
830 | end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE); |
831 | GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); |
832 | |
833 | alignment = max(alignment, i915_vm_obj_min_alignment(vma->vm, vma->obj)); |
834 | |
835 | /* |
836 | * If binding the object/GGTT view requires more space than the entire |
837 | * aperture has, reject it early before evicting everything in a vain |
838 | * attempt to find space. |
839 | */ |
840 | if (size > end - 2 * guard) { |
841 | drm_dbg(vma->obj->base.dev, |
842 | "Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n" , |
843 | size, flags & PIN_MAPPABLE ? "mappable" : "total" , end); |
844 | return -ENOSPC; |
845 | } |
846 | |
847 | color = 0; |
848 | |
849 | if (i915_vm_has_cache_coloring(vm: vma->vm)) |
850 | color = vma->obj->pat_index; |
851 | |
852 | if (flags & PIN_OFFSET_FIXED) { |
853 | u64 offset = flags & PIN_OFFSET_MASK; |
854 | if (!IS_ALIGNED(offset, alignment) || |
855 | range_overflows(offset, size, end)) |
856 | return -EINVAL; |
857 | /* |
858 | * The caller knows not of the guard added by others and |
859 | * requests for the offset of the start of its buffer |
860 | * to be fixed, which may not be the same as the position |
861 | * of the vma->node due to the guard pages. |
862 | */ |
863 | if (offset < guard || offset + size > end - guard) |
864 | return -ENOSPC; |
865 | |
866 | ret = i915_gem_gtt_reserve(vm: vma->vm, ww, node: &vma->node, |
867 | size: size + 2 * guard, |
868 | offset: offset - guard, |
869 | color, flags); |
870 | if (ret) |
871 | return ret; |
872 | } else { |
873 | size += 2 * guard; |
874 | /* |
875 | * We only support huge gtt pages through the 48b PPGTT, |
876 | * however we also don't want to force any alignment for |
877 | * objects which need to be tightly packed into the low 32bits. |
878 | * |
879 | * Note that we assume that GGTT are limited to 4GiB for the |
880 | * forseeable future. See also i915_ggtt_offset(). |
881 | */ |
882 | if (upper_32_bits(end - 1) && |
883 | vma->page_sizes.sg > I915_GTT_PAGE_SIZE && |
884 | !HAS_64K_PAGES(vma->vm->i915)) { |
885 | /* |
886 | * We can't mix 64K and 4K PTEs in the same page-table |
887 | * (2M block), and so to avoid the ugliness and |
888 | * complexity of coloring we opt for just aligning 64K |
889 | * objects to 2M. |
890 | */ |
891 | u64 page_alignment = |
892 | rounddown_pow_of_two(vma->page_sizes.sg | |
893 | I915_GTT_PAGE_SIZE_2M); |
894 | |
895 | /* |
896 | * Check we don't expand for the limited Global GTT |
897 | * (mappable aperture is even more precious!). This |
898 | * also checks that we exclude the aliasing-ppgtt. |
899 | */ |
900 | GEM_BUG_ON(i915_vma_is_ggtt(vma)); |
901 | |
902 | alignment = max(alignment, page_alignment); |
903 | |
904 | if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) |
905 | size = round_up(size, I915_GTT_PAGE_SIZE_2M); |
906 | } |
907 | |
908 | ret = i915_gem_gtt_insert(vm: vma->vm, ww, node: &vma->node, |
909 | size, alignment, color, |
910 | start, end, flags); |
911 | if (ret) |
912 | return ret; |
913 | |
914 | GEM_BUG_ON(vma->node.start < start); |
915 | GEM_BUG_ON(vma->node.start + vma->node.size > end); |
916 | } |
917 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); |
918 | GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color)); |
919 | |
920 | list_move_tail(list: &vma->vm_link, head: &vma->vm->bound_list); |
921 | vma->guard = guard; |
922 | |
923 | return 0; |
924 | } |
925 | |
926 | static void |
927 | i915_vma_detach(struct i915_vma *vma) |
928 | { |
929 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); |
930 | GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); |
931 | |
932 | /* |
933 | * And finally now the object is completely decoupled from this |
934 | * vma, we can drop its hold on the backing storage and allow |
935 | * it to be reaped by the shrinker. |
936 | */ |
937 | list_move_tail(list: &vma->vm_link, head: &vma->vm->unbound_list); |
938 | } |
939 | |
940 | static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) |
941 | { |
942 | unsigned int bound; |
943 | |
944 | bound = atomic_read(v: &vma->flags); |
945 | |
946 | if (flags & PIN_VALIDATE) { |
947 | flags &= I915_VMA_BIND_MASK; |
948 | |
949 | return (flags & bound) == flags; |
950 | } |
951 | |
952 | /* with the lock mandatory for unbind, we don't race here */ |
953 | flags &= I915_VMA_BIND_MASK; |
954 | do { |
955 | if (unlikely(flags & ~bound)) |
956 | return false; |
957 | |
958 | if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) |
959 | return false; |
960 | |
961 | GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0); |
962 | } while (!atomic_try_cmpxchg(v: &vma->flags, old: &bound, new: bound + 1)); |
963 | |
964 | return true; |
965 | } |
966 | |
967 | static struct scatterlist * |
968 | rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset, |
969 | unsigned int width, unsigned int height, |
970 | unsigned int src_stride, unsigned int dst_stride, |
971 | struct sg_table *st, struct scatterlist *sg) |
972 | { |
973 | unsigned int column, row; |
974 | pgoff_t src_idx; |
975 | |
976 | for (column = 0; column < width; column++) { |
977 | unsigned int left; |
978 | |
979 | src_idx = src_stride * (height - 1) + column + offset; |
980 | for (row = 0; row < height; row++) { |
981 | st->nents++; |
982 | /* |
983 | * We don't need the pages, but need to initialize |
984 | * the entries so the sg list can be happily traversed. |
985 | * The only thing we need are DMA addresses. |
986 | */ |
987 | sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, offset: 0); |
988 | sg_dma_address(sg) = |
989 | i915_gem_object_get_dma_address(obj, src_idx); |
990 | sg_dma_len(sg) = I915_GTT_PAGE_SIZE; |
991 | sg = sg_next(sg); |
992 | src_idx -= src_stride; |
993 | } |
994 | |
995 | left = (dst_stride - height) * I915_GTT_PAGE_SIZE; |
996 | |
997 | if (!left) |
998 | continue; |
999 | |
1000 | st->nents++; |
1001 | |
1002 | /* |
1003 | * The DE ignores the PTEs for the padding tiles, the sg entry |
1004 | * here is just a conenience to indicate how many padding PTEs |
1005 | * to insert at this spot. |
1006 | */ |
1007 | sg_set_page(sg, NULL, len: left, offset: 0); |
1008 | sg_dma_address(sg) = 0; |
1009 | sg_dma_len(sg) = left; |
1010 | sg = sg_next(sg); |
1011 | } |
1012 | |
1013 | return sg; |
1014 | } |
1015 | |
1016 | static noinline struct sg_table * |
1017 | intel_rotate_pages(struct intel_rotation_info *rot_info, |
1018 | struct drm_i915_gem_object *obj) |
1019 | { |
1020 | unsigned int size = intel_rotation_info_size(rot_info); |
1021 | struct drm_i915_private *i915 = to_i915(dev: obj->base.dev); |
1022 | struct sg_table *st; |
1023 | struct scatterlist *sg; |
1024 | int ret = -ENOMEM; |
1025 | int i; |
1026 | |
1027 | /* Allocate target SG list. */ |
1028 | st = kmalloc(size: sizeof(*st), GFP_KERNEL); |
1029 | if (!st) |
1030 | goto err_st_alloc; |
1031 | |
1032 | ret = sg_alloc_table(st, size, GFP_KERNEL); |
1033 | if (ret) |
1034 | goto err_sg_alloc; |
1035 | |
1036 | st->nents = 0; |
1037 | sg = st->sgl; |
1038 | |
1039 | for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) |
1040 | sg = rotate_pages(obj, offset: rot_info->plane[i].offset, |
1041 | width: rot_info->plane[i].width, height: rot_info->plane[i].height, |
1042 | src_stride: rot_info->plane[i].src_stride, |
1043 | dst_stride: rot_info->plane[i].dst_stride, |
1044 | st, sg); |
1045 | |
1046 | return st; |
1047 | |
1048 | err_sg_alloc: |
1049 | kfree(objp: st); |
1050 | err_st_alloc: |
1051 | |
1052 | drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n" , |
1053 | obj->base.size, rot_info->plane[0].width, |
1054 | rot_info->plane[0].height, size); |
1055 | |
1056 | return ERR_PTR(error: ret); |
1057 | } |
1058 | |
1059 | static struct scatterlist * |
1060 | add_padding_pages(unsigned int count, |
1061 | struct sg_table *st, struct scatterlist *sg) |
1062 | { |
1063 | st->nents++; |
1064 | |
1065 | /* |
1066 | * The DE ignores the PTEs for the padding tiles, the sg entry |
1067 | * here is just a convenience to indicate how many padding PTEs |
1068 | * to insert at this spot. |
1069 | */ |
1070 | sg_set_page(sg, NULL, len: count * I915_GTT_PAGE_SIZE, offset: 0); |
1071 | sg_dma_address(sg) = 0; |
1072 | sg_dma_len(sg) = count * I915_GTT_PAGE_SIZE; |
1073 | sg = sg_next(sg); |
1074 | |
1075 | return sg; |
1076 | } |
1077 | |
1078 | static struct scatterlist * |
1079 | remap_tiled_color_plane_pages(struct drm_i915_gem_object *obj, |
1080 | unsigned long offset, unsigned int alignment_pad, |
1081 | unsigned int width, unsigned int height, |
1082 | unsigned int src_stride, unsigned int dst_stride, |
1083 | struct sg_table *st, struct scatterlist *sg, |
1084 | unsigned int *gtt_offset) |
1085 | { |
1086 | unsigned int row; |
1087 | |
1088 | if (!width || !height) |
1089 | return sg; |
1090 | |
1091 | if (alignment_pad) |
1092 | sg = add_padding_pages(count: alignment_pad, st, sg); |
1093 | |
1094 | for (row = 0; row < height; row++) { |
1095 | unsigned int left = width * I915_GTT_PAGE_SIZE; |
1096 | |
1097 | while (left) { |
1098 | dma_addr_t addr; |
1099 | unsigned int length; |
1100 | |
1101 | /* |
1102 | * We don't need the pages, but need to initialize |
1103 | * the entries so the sg list can be happily traversed. |
1104 | * The only thing we need are DMA addresses. |
1105 | */ |
1106 | |
1107 | addr = i915_gem_object_get_dma_address_len(obj, offset, &length); |
1108 | |
1109 | length = min(left, length); |
1110 | |
1111 | st->nents++; |
1112 | |
1113 | sg_set_page(sg, NULL, len: length, offset: 0); |
1114 | sg_dma_address(sg) = addr; |
1115 | sg_dma_len(sg) = length; |
1116 | sg = sg_next(sg); |
1117 | |
1118 | offset += length / I915_GTT_PAGE_SIZE; |
1119 | left -= length; |
1120 | } |
1121 | |
1122 | offset += src_stride - width; |
1123 | |
1124 | left = (dst_stride - width) * I915_GTT_PAGE_SIZE; |
1125 | |
1126 | if (!left) |
1127 | continue; |
1128 | |
1129 | sg = add_padding_pages(count: left >> PAGE_SHIFT, st, sg); |
1130 | } |
1131 | |
1132 | *gtt_offset += alignment_pad + dst_stride * height; |
1133 | |
1134 | return sg; |
1135 | } |
1136 | |
1137 | static struct scatterlist * |
1138 | remap_contiguous_pages(struct drm_i915_gem_object *obj, |
1139 | pgoff_t obj_offset, |
1140 | unsigned int count, |
1141 | struct sg_table *st, struct scatterlist *sg) |
1142 | { |
1143 | struct scatterlist *iter; |
1144 | unsigned int offset; |
1145 | |
1146 | iter = i915_gem_object_get_sg_dma(obj, obj_offset, &offset); |
1147 | GEM_BUG_ON(!iter); |
1148 | |
1149 | do { |
1150 | unsigned int len; |
1151 | |
1152 | len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT), |
1153 | count << PAGE_SHIFT); |
1154 | sg_set_page(sg, NULL, len, offset: 0); |
1155 | sg_dma_address(sg) = |
1156 | sg_dma_address(iter) + (offset << PAGE_SHIFT); |
1157 | sg_dma_len(sg) = len; |
1158 | |
1159 | st->nents++; |
1160 | count -= len >> PAGE_SHIFT; |
1161 | if (count == 0) |
1162 | return sg; |
1163 | |
1164 | sg = __sg_next(sg); |
1165 | iter = __sg_next(sg: iter); |
1166 | offset = 0; |
1167 | } while (1); |
1168 | } |
1169 | |
1170 | static struct scatterlist * |
1171 | remap_linear_color_plane_pages(struct drm_i915_gem_object *obj, |
1172 | pgoff_t obj_offset, unsigned int alignment_pad, |
1173 | unsigned int size, |
1174 | struct sg_table *st, struct scatterlist *sg, |
1175 | unsigned int *gtt_offset) |
1176 | { |
1177 | if (!size) |
1178 | return sg; |
1179 | |
1180 | if (alignment_pad) |
1181 | sg = add_padding_pages(count: alignment_pad, st, sg); |
1182 | |
1183 | sg = remap_contiguous_pages(obj, obj_offset, count: size, st, sg); |
1184 | sg = sg_next(sg); |
1185 | |
1186 | *gtt_offset += alignment_pad + size; |
1187 | |
1188 | return sg; |
1189 | } |
1190 | |
1191 | static struct scatterlist * |
1192 | remap_color_plane_pages(const struct intel_remapped_info *rem_info, |
1193 | struct drm_i915_gem_object *obj, |
1194 | int color_plane, |
1195 | struct sg_table *st, struct scatterlist *sg, |
1196 | unsigned int *gtt_offset) |
1197 | { |
1198 | unsigned int alignment_pad = 0; |
1199 | |
1200 | if (rem_info->plane_alignment) |
1201 | alignment_pad = ALIGN(*gtt_offset, rem_info->plane_alignment) - *gtt_offset; |
1202 | |
1203 | if (rem_info->plane[color_plane].linear) |
1204 | sg = remap_linear_color_plane_pages(obj, |
1205 | obj_offset: rem_info->plane[color_plane].offset, |
1206 | alignment_pad, |
1207 | size: rem_info->plane[color_plane].size, |
1208 | st, sg, |
1209 | gtt_offset); |
1210 | |
1211 | else |
1212 | sg = remap_tiled_color_plane_pages(obj, |
1213 | offset: rem_info->plane[color_plane].offset, |
1214 | alignment_pad, |
1215 | width: rem_info->plane[color_plane].width, |
1216 | height: rem_info->plane[color_plane].height, |
1217 | src_stride: rem_info->plane[color_plane].src_stride, |
1218 | dst_stride: rem_info->plane[color_plane].dst_stride, |
1219 | st, sg, |
1220 | gtt_offset); |
1221 | |
1222 | return sg; |
1223 | } |
1224 | |
1225 | static noinline struct sg_table * |
1226 | intel_remap_pages(struct intel_remapped_info *rem_info, |
1227 | struct drm_i915_gem_object *obj) |
1228 | { |
1229 | unsigned int size = intel_remapped_info_size(rem_info); |
1230 | struct drm_i915_private *i915 = to_i915(dev: obj->base.dev); |
1231 | struct sg_table *st; |
1232 | struct scatterlist *sg; |
1233 | unsigned int gtt_offset = 0; |
1234 | int ret = -ENOMEM; |
1235 | int i; |
1236 | |
1237 | /* Allocate target SG list. */ |
1238 | st = kmalloc(size: sizeof(*st), GFP_KERNEL); |
1239 | if (!st) |
1240 | goto err_st_alloc; |
1241 | |
1242 | ret = sg_alloc_table(st, size, GFP_KERNEL); |
1243 | if (ret) |
1244 | goto err_sg_alloc; |
1245 | |
1246 | st->nents = 0; |
1247 | sg = st->sgl; |
1248 | |
1249 | for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) |
1250 | sg = remap_color_plane_pages(rem_info, obj, color_plane: i, st, sg, gtt_offset: >t_offset); |
1251 | |
1252 | i915_sg_trim(orig_st: st); |
1253 | |
1254 | return st; |
1255 | |
1256 | err_sg_alloc: |
1257 | kfree(objp: st); |
1258 | err_st_alloc: |
1259 | |
1260 | drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n" , |
1261 | obj->base.size, rem_info->plane[0].width, |
1262 | rem_info->plane[0].height, size); |
1263 | |
1264 | return ERR_PTR(error: ret); |
1265 | } |
1266 | |
1267 | static noinline struct sg_table * |
1268 | intel_partial_pages(const struct i915_gtt_view *view, |
1269 | struct drm_i915_gem_object *obj) |
1270 | { |
1271 | struct sg_table *st; |
1272 | struct scatterlist *sg; |
1273 | unsigned int count = view->partial.size; |
1274 | int ret = -ENOMEM; |
1275 | |
1276 | st = kmalloc(size: sizeof(*st), GFP_KERNEL); |
1277 | if (!st) |
1278 | goto err_st_alloc; |
1279 | |
1280 | ret = sg_alloc_table(st, count, GFP_KERNEL); |
1281 | if (ret) |
1282 | goto err_sg_alloc; |
1283 | |
1284 | st->nents = 0; |
1285 | |
1286 | sg = remap_contiguous_pages(obj, obj_offset: view->partial.offset, count, st, sg: st->sgl); |
1287 | |
1288 | sg_mark_end(sg); |
1289 | i915_sg_trim(orig_st: st); /* Drop any unused tail entries. */ |
1290 | |
1291 | return st; |
1292 | |
1293 | err_sg_alloc: |
1294 | kfree(objp: st); |
1295 | err_st_alloc: |
1296 | return ERR_PTR(error: ret); |
1297 | } |
1298 | |
1299 | static int |
1300 | __i915_vma_get_pages(struct i915_vma *vma) |
1301 | { |
1302 | struct sg_table *pages; |
1303 | |
1304 | /* |
1305 | * The vma->pages are only valid within the lifespan of the borrowed |
1306 | * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so |
1307 | * must be the vma->pages. A simple rule is that vma->pages must only |
1308 | * be accessed when the obj->mm.pages are pinned. |
1309 | */ |
1310 | GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj)); |
1311 | |
1312 | switch (vma->gtt_view.type) { |
1313 | default: |
1314 | GEM_BUG_ON(vma->gtt_view.type); |
1315 | fallthrough; |
1316 | case I915_GTT_VIEW_NORMAL: |
1317 | pages = vma->obj->mm.pages; |
1318 | break; |
1319 | |
1320 | case I915_GTT_VIEW_ROTATED: |
1321 | pages = |
1322 | intel_rotate_pages(rot_info: &vma->gtt_view.rotated, obj: vma->obj); |
1323 | break; |
1324 | |
1325 | case I915_GTT_VIEW_REMAPPED: |
1326 | pages = |
1327 | intel_remap_pages(rem_info: &vma->gtt_view.remapped, obj: vma->obj); |
1328 | break; |
1329 | |
1330 | case I915_GTT_VIEW_PARTIAL: |
1331 | pages = intel_partial_pages(view: &vma->gtt_view, obj: vma->obj); |
1332 | break; |
1333 | } |
1334 | |
1335 | if (IS_ERR(ptr: pages)) { |
1336 | drm_err(&vma->vm->i915->drm, |
1337 | "Failed to get pages for VMA view type %u (%ld)!\n" , |
1338 | vma->gtt_view.type, PTR_ERR(pages)); |
1339 | return PTR_ERR(ptr: pages); |
1340 | } |
1341 | |
1342 | vma->pages = pages; |
1343 | |
1344 | return 0; |
1345 | } |
1346 | |
1347 | I915_SELFTEST_EXPORT int i915_vma_get_pages(struct i915_vma *vma) |
1348 | { |
1349 | int err; |
1350 | |
1351 | if (atomic_add_unless(v: &vma->pages_count, a: 1, u: 0)) |
1352 | return 0; |
1353 | |
1354 | err = i915_gem_object_pin_pages(obj: vma->obj); |
1355 | if (err) |
1356 | return err; |
1357 | |
1358 | err = __i915_vma_get_pages(vma); |
1359 | if (err) |
1360 | goto err_unpin; |
1361 | |
1362 | vma->page_sizes = vma->obj->mm.page_sizes; |
1363 | atomic_inc(v: &vma->pages_count); |
1364 | |
1365 | return 0; |
1366 | |
1367 | err_unpin: |
1368 | __i915_gem_object_unpin_pages(obj: vma->obj); |
1369 | |
1370 | return err; |
1371 | } |
1372 | |
1373 | void vma_invalidate_tlb(struct i915_address_space *vm, u32 *tlb) |
1374 | { |
1375 | struct intel_gt *gt; |
1376 | int id; |
1377 | |
1378 | if (!tlb) |
1379 | return; |
1380 | |
1381 | /* |
1382 | * Before we release the pages that were bound by this vma, we |
1383 | * must invalidate all the TLBs that may still have a reference |
1384 | * back to our physical address. It only needs to be done once, |
1385 | * so after updating the PTE to point away from the pages, record |
1386 | * the most recent TLB invalidation seqno, and if we have not yet |
1387 | * flushed the TLBs upon release, perform a full invalidation. |
1388 | */ |
1389 | for_each_gt(gt, vm->i915, id) |
1390 | WRITE_ONCE(tlb[id], |
1391 | intel_gt_next_invalidate_tlb_full(gt)); |
1392 | } |
1393 | |
1394 | static void __vma_put_pages(struct i915_vma *vma, unsigned int count) |
1395 | { |
1396 | /* We allocate under vma_get_pages, so beware the shrinker */ |
1397 | GEM_BUG_ON(atomic_read(&vma->pages_count) < count); |
1398 | |
1399 | if (atomic_sub_return(i: count, v: &vma->pages_count) == 0) { |
1400 | if (vma->pages != vma->obj->mm.pages) { |
1401 | sg_free_table(vma->pages); |
1402 | kfree(objp: vma->pages); |
1403 | } |
1404 | vma->pages = NULL; |
1405 | |
1406 | i915_gem_object_unpin_pages(obj: vma->obj); |
1407 | } |
1408 | } |
1409 | |
1410 | I915_SELFTEST_EXPORT void i915_vma_put_pages(struct i915_vma *vma) |
1411 | { |
1412 | if (atomic_add_unless(v: &vma->pages_count, a: -1, u: 1)) |
1413 | return; |
1414 | |
1415 | __vma_put_pages(vma, count: 1); |
1416 | } |
1417 | |
1418 | static void vma_unbind_pages(struct i915_vma *vma) |
1419 | { |
1420 | unsigned int count; |
1421 | |
1422 | lockdep_assert_held(&vma->vm->mutex); |
1423 | |
1424 | /* The upper portion of pages_count is the number of bindings */ |
1425 | count = atomic_read(v: &vma->pages_count); |
1426 | count >>= I915_VMA_PAGES_BIAS; |
1427 | GEM_BUG_ON(!count); |
1428 | |
1429 | __vma_put_pages(vma, count: count | count << I915_VMA_PAGES_BIAS); |
1430 | } |
1431 | |
1432 | int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
1433 | u64 size, u64 alignment, u64 flags) |
1434 | { |
1435 | struct i915_vma_work *work = NULL; |
1436 | struct dma_fence *moving = NULL; |
1437 | struct i915_vma_resource *vma_res = NULL; |
1438 | intel_wakeref_t wakeref; |
1439 | unsigned int bound; |
1440 | int err; |
1441 | |
1442 | assert_vma_held(vma); |
1443 | GEM_BUG_ON(!ww); |
1444 | |
1445 | BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND); |
1446 | BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND); |
1447 | |
1448 | GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL))); |
1449 | |
1450 | /* First try and grab the pin without rebinding the vma */ |
1451 | if (try_qad_pin(vma, flags)) |
1452 | return 0; |
1453 | |
1454 | err = i915_vma_get_pages(vma); |
1455 | if (err) |
1456 | return err; |
1457 | |
1458 | /* |
1459 | * In case of a global GTT, we must hold a runtime-pm wakeref |
1460 | * while global PTEs are updated. In other cases, we hold |
1461 | * the rpm reference while the VMA is active. Since runtime |
1462 | * resume may require allocations, which are forbidden inside |
1463 | * vm->mutex, get the first rpm wakeref outside of the mutex. |
1464 | */ |
1465 | wakeref = intel_runtime_pm_get(rpm: &vma->vm->i915->runtime_pm); |
1466 | |
1467 | if (flags & vma->vm->bind_async_flags) { |
1468 | /* lock VM */ |
1469 | err = i915_vm_lock_objects(vm: vma->vm, ww); |
1470 | if (err) |
1471 | goto err_rpm; |
1472 | |
1473 | work = i915_vma_work(); |
1474 | if (!work) { |
1475 | err = -ENOMEM; |
1476 | goto err_rpm; |
1477 | } |
1478 | |
1479 | work->vm = vma->vm; |
1480 | |
1481 | err = i915_gem_object_get_moving_fence(obj: vma->obj, fence: &moving); |
1482 | if (err) |
1483 | goto err_rpm; |
1484 | |
1485 | dma_fence_work_chain(f: &work->base, signal: moving); |
1486 | |
1487 | /* Allocate enough page directories to used PTE */ |
1488 | if (vma->vm->allocate_va_range) { |
1489 | err = i915_vm_alloc_pt_stash(vm: vma->vm, |
1490 | stash: &work->stash, |
1491 | size: vma->size); |
1492 | if (err) |
1493 | goto err_fence; |
1494 | |
1495 | err = i915_vm_map_pt_stash(vm: vma->vm, stash: &work->stash); |
1496 | if (err) |
1497 | goto err_fence; |
1498 | } |
1499 | } |
1500 | |
1501 | vma_res = i915_vma_resource_alloc(); |
1502 | if (IS_ERR(ptr: vma_res)) { |
1503 | err = PTR_ERR(ptr: vma_res); |
1504 | goto err_fence; |
1505 | } |
1506 | |
1507 | /* |
1508 | * Differentiate between user/kernel vma inside the aliasing-ppgtt. |
1509 | * |
1510 | * We conflate the Global GTT with the user's vma when using the |
1511 | * aliasing-ppgtt, but it is still vitally important to try and |
1512 | * keep the use cases distinct. For example, userptr objects are |
1513 | * not allowed inside the Global GTT as that will cause lock |
1514 | * inversions when we have to evict them the mmu_notifier callbacks - |
1515 | * but they are allowed to be part of the user ppGTT which can never |
1516 | * be mapped. As such we try to give the distinct users of the same |
1517 | * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt |
1518 | * and i915_ppgtt separate]. |
1519 | * |
1520 | * NB this may cause us to mask real lock inversions -- while the |
1521 | * code is safe today, lockdep may not be able to spot future |
1522 | * transgressions. |
1523 | */ |
1524 | err = mutex_lock_interruptible_nested(lock: &vma->vm->mutex, |
1525 | subclass: !(flags & PIN_GLOBAL)); |
1526 | if (err) |
1527 | goto err_vma_res; |
1528 | |
1529 | /* No more allocations allowed now we hold vm->mutex */ |
1530 | |
1531 | if (unlikely(i915_vma_is_closed(vma))) { |
1532 | err = -ENOENT; |
1533 | goto err_unlock; |
1534 | } |
1535 | |
1536 | bound = atomic_read(v: &vma->flags); |
1537 | if (unlikely(bound & I915_VMA_ERROR)) { |
1538 | err = -ENOMEM; |
1539 | goto err_unlock; |
1540 | } |
1541 | |
1542 | if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) { |
1543 | err = -EAGAIN; /* pins are meant to be fairly temporary */ |
1544 | goto err_unlock; |
1545 | } |
1546 | |
1547 | if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) { |
1548 | if (!(flags & PIN_VALIDATE)) |
1549 | __i915_vma_pin(vma); |
1550 | goto err_unlock; |
1551 | } |
1552 | |
1553 | err = i915_active_acquire(ref: &vma->active); |
1554 | if (err) |
1555 | goto err_unlock; |
1556 | |
1557 | if (!(bound & I915_VMA_BIND_MASK)) { |
1558 | err = i915_vma_insert(vma, ww, size, alignment, flags); |
1559 | if (err) |
1560 | goto err_active; |
1561 | |
1562 | if (i915_is_ggtt(vma->vm)) |
1563 | __i915_vma_set_map_and_fenceable(vma); |
1564 | } |
1565 | |
1566 | GEM_BUG_ON(!vma->pages); |
1567 | err = i915_vma_bind(vma, |
1568 | pat_index: vma->obj->pat_index, |
1569 | flags, work, vma_res); |
1570 | vma_res = NULL; |
1571 | if (err) |
1572 | goto err_remove; |
1573 | |
1574 | /* There should only be at most 2 active bindings (user, global) */ |
1575 | GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound); |
1576 | atomic_add(I915_VMA_PAGES_ACTIVE, v: &vma->pages_count); |
1577 | list_move_tail(list: &vma->vm_link, head: &vma->vm->bound_list); |
1578 | |
1579 | if (!(flags & PIN_VALIDATE)) { |
1580 | __i915_vma_pin(vma); |
1581 | GEM_BUG_ON(!i915_vma_is_pinned(vma)); |
1582 | } |
1583 | GEM_BUG_ON(!i915_vma_is_bound(vma, flags)); |
1584 | GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags)); |
1585 | |
1586 | err_remove: |
1587 | if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) { |
1588 | i915_vma_detach(vma); |
1589 | drm_mm_remove_node(node: &vma->node); |
1590 | } |
1591 | err_active: |
1592 | i915_active_release(ref: &vma->active); |
1593 | err_unlock: |
1594 | mutex_unlock(lock: &vma->vm->mutex); |
1595 | err_vma_res: |
1596 | i915_vma_resource_free(vma_res); |
1597 | err_fence: |
1598 | if (work) |
1599 | dma_fence_work_commit_imm(f: &work->base); |
1600 | err_rpm: |
1601 | intel_runtime_pm_put(rpm: &vma->vm->i915->runtime_pm, wref: wakeref); |
1602 | |
1603 | if (moving) |
1604 | dma_fence_put(fence: moving); |
1605 | |
1606 | i915_vma_put_pages(vma); |
1607 | return err; |
1608 | } |
1609 | |
1610 | static void flush_idle_contexts(struct intel_gt *gt) |
1611 | { |
1612 | struct intel_engine_cs *engine; |
1613 | enum intel_engine_id id; |
1614 | |
1615 | for_each_engine(engine, gt, id) |
1616 | intel_engine_flush_barriers(engine); |
1617 | |
1618 | intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT); |
1619 | } |
1620 | |
1621 | static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
1622 | u32 align, unsigned int flags) |
1623 | { |
1624 | struct i915_address_space *vm = vma->vm; |
1625 | struct intel_gt *gt; |
1626 | struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm); |
1627 | int err; |
1628 | |
1629 | do { |
1630 | err = i915_vma_pin_ww(vma, ww, size: 0, alignment: align, flags: flags | PIN_GLOBAL); |
1631 | |
1632 | if (err != -ENOSPC) { |
1633 | if (!err) { |
1634 | err = i915_vma_wait_for_bind(vma); |
1635 | if (err) |
1636 | i915_vma_unpin(vma); |
1637 | } |
1638 | return err; |
1639 | } |
1640 | |
1641 | /* Unlike i915_vma_pin, we don't take no for an answer! */ |
1642 | list_for_each_entry(gt, &ggtt->gt_list, ggtt_link) |
1643 | flush_idle_contexts(gt); |
1644 | if (mutex_lock_interruptible(&vm->mutex) == 0) { |
1645 | /* |
1646 | * We pass NULL ww here, as we don't want to unbind |
1647 | * locked objects when called from execbuf when pinning |
1648 | * is removed. This would probably regress badly. |
1649 | */ |
1650 | i915_gem_evict_vm(vm, NULL, NULL); |
1651 | mutex_unlock(lock: &vm->mutex); |
1652 | } |
1653 | } while (1); |
1654 | } |
1655 | |
1656 | int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
1657 | u32 align, unsigned int flags) |
1658 | { |
1659 | struct i915_gem_ww_ctx _ww; |
1660 | int err; |
1661 | |
1662 | GEM_BUG_ON(!i915_vma_is_ggtt(vma)); |
1663 | |
1664 | if (ww) |
1665 | return __i915_ggtt_pin(vma, ww, align, flags); |
1666 | |
1667 | lockdep_assert_not_held(&vma->obj->base.resv->lock.base); |
1668 | |
1669 | for_i915_gem_ww(&_ww, err, true) { |
1670 | err = i915_gem_object_lock(obj: vma->obj, ww: &_ww); |
1671 | if (!err) |
1672 | err = __i915_ggtt_pin(vma, ww: &_ww, align, flags); |
1673 | } |
1674 | |
1675 | return err; |
1676 | } |
1677 | |
1678 | /** |
1679 | * i915_ggtt_clear_scanout - Clear scanout flag for all objects ggtt vmas |
1680 | * @obj: i915 GEM object |
1681 | * This function clears scanout flags for objects ggtt vmas. These flags are set |
1682 | * when object is pinned for display use and this function to clear them all is |
1683 | * targeted to be called by frontbuffer tracking code when the frontbuffer is |
1684 | * about to be released. |
1685 | */ |
1686 | void i915_ggtt_clear_scanout(struct drm_i915_gem_object *obj) |
1687 | { |
1688 | struct i915_vma *vma; |
1689 | |
1690 | spin_lock(lock: &obj->vma.lock); |
1691 | for_each_ggtt_vma(vma, obj) { |
1692 | i915_vma_clear_scanout(vma); |
1693 | vma->display_alignment = I915_GTT_MIN_ALIGNMENT; |
1694 | } |
1695 | spin_unlock(lock: &obj->vma.lock); |
1696 | } |
1697 | |
1698 | static void __vma_close(struct i915_vma *vma, struct intel_gt *gt) |
1699 | { |
1700 | /* |
1701 | * We defer actually closing, unbinding and destroying the VMA until |
1702 | * the next idle point, or if the object is freed in the meantime. By |
1703 | * postponing the unbind, we allow for it to be resurrected by the |
1704 | * client, avoiding the work required to rebind the VMA. This is |
1705 | * advantageous for DRI, where the client/server pass objects |
1706 | * between themselves, temporarily opening a local VMA to the |
1707 | * object, and then closing it again. The same object is then reused |
1708 | * on the next frame (or two, depending on the depth of the swap queue) |
1709 | * causing us to rebind the VMA once more. This ends up being a lot |
1710 | * of wasted work for the steady state. |
1711 | */ |
1712 | GEM_BUG_ON(i915_vma_is_closed(vma)); |
1713 | list_add(new: &vma->closed_link, head: >->closed_vma); |
1714 | } |
1715 | |
1716 | void i915_vma_close(struct i915_vma *vma) |
1717 | { |
1718 | struct intel_gt *gt = vma->vm->gt; |
1719 | unsigned long flags; |
1720 | |
1721 | if (i915_vma_is_ggtt(vma)) |
1722 | return; |
1723 | |
1724 | GEM_BUG_ON(!atomic_read(&vma->open_count)); |
1725 | if (atomic_dec_and_lock_irqsave(&vma->open_count, |
1726 | >->closed_lock, |
1727 | flags)) { |
1728 | __vma_close(vma, gt); |
1729 | spin_unlock_irqrestore(lock: >->closed_lock, flags); |
1730 | } |
1731 | } |
1732 | |
1733 | static void __i915_vma_remove_closed(struct i915_vma *vma) |
1734 | { |
1735 | list_del_init(entry: &vma->closed_link); |
1736 | } |
1737 | |
1738 | void i915_vma_reopen(struct i915_vma *vma) |
1739 | { |
1740 | struct intel_gt *gt = vma->vm->gt; |
1741 | |
1742 | spin_lock_irq(lock: >->closed_lock); |
1743 | if (i915_vma_is_closed(vma)) |
1744 | __i915_vma_remove_closed(vma); |
1745 | spin_unlock_irq(lock: >->closed_lock); |
1746 | } |
1747 | |
1748 | static void force_unbind(struct i915_vma *vma) |
1749 | { |
1750 | if (!drm_mm_node_allocated(node: &vma->node)) |
1751 | return; |
1752 | |
1753 | atomic_and(i: ~I915_VMA_PIN_MASK, v: &vma->flags); |
1754 | WARN_ON(__i915_vma_unbind(vma)); |
1755 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); |
1756 | } |
1757 | |
1758 | static void release_references(struct i915_vma *vma, struct intel_gt *gt, |
1759 | bool vm_ddestroy) |
1760 | { |
1761 | struct drm_i915_gem_object *obj = vma->obj; |
1762 | |
1763 | GEM_BUG_ON(i915_vma_is_active(vma)); |
1764 | |
1765 | spin_lock(lock: &obj->vma.lock); |
1766 | list_del(entry: &vma->obj_link); |
1767 | if (!RB_EMPTY_NODE(&vma->obj_node)) |
1768 | rb_erase(&vma->obj_node, &obj->vma.tree); |
1769 | |
1770 | spin_unlock(lock: &obj->vma.lock); |
1771 | |
1772 | spin_lock_irq(lock: >->closed_lock); |
1773 | __i915_vma_remove_closed(vma); |
1774 | spin_unlock_irq(lock: >->closed_lock); |
1775 | |
1776 | if (vm_ddestroy) |
1777 | i915_vm_resv_put(vm: vma->vm); |
1778 | |
1779 | /* Wait for async active retire */ |
1780 | i915_active_wait(ref: &vma->active); |
1781 | i915_active_fini(ref: &vma->active); |
1782 | GEM_WARN_ON(vma->resource); |
1783 | i915_vma_free(vma); |
1784 | } |
1785 | |
1786 | /* |
1787 | * i915_vma_destroy_locked - Remove all weak reference to the vma and put |
1788 | * the initial reference. |
1789 | * |
1790 | * This function should be called when it's decided the vma isn't needed |
1791 | * anymore. The caller must assure that it doesn't race with another lookup |
1792 | * plus destroy, typically by taking an appropriate reference. |
1793 | * |
1794 | * Current callsites are |
1795 | * - __i915_gem_object_pages_fini() |
1796 | * - __i915_vm_close() - Blocks the above function by taking a reference on |
1797 | * the object. |
1798 | * - __i915_vma_parked() - Blocks the above functions by taking a reference |
1799 | * on the vm and a reference on the object. Also takes the object lock so |
1800 | * destruction from __i915_vma_parked() can be blocked by holding the |
1801 | * object lock. Since the object lock is only allowed from within i915 with |
1802 | * an object refcount, holding the object lock also implicitly blocks the |
1803 | * vma freeing from __i915_gem_object_pages_fini(). |
1804 | * |
1805 | * Because of locks taken during destruction, a vma is also guaranteed to |
1806 | * stay alive while the following locks are held if it was looked up while |
1807 | * holding one of the locks: |
1808 | * - vm->mutex |
1809 | * - obj->vma.lock |
1810 | * - gt->closed_lock |
1811 | */ |
1812 | void i915_vma_destroy_locked(struct i915_vma *vma) |
1813 | { |
1814 | lockdep_assert_held(&vma->vm->mutex); |
1815 | |
1816 | force_unbind(vma); |
1817 | list_del_init(entry: &vma->vm_link); |
1818 | release_references(vma, gt: vma->vm->gt, vm_ddestroy: false); |
1819 | } |
1820 | |
1821 | void i915_vma_destroy(struct i915_vma *vma) |
1822 | { |
1823 | struct intel_gt *gt; |
1824 | bool vm_ddestroy; |
1825 | |
1826 | mutex_lock(&vma->vm->mutex); |
1827 | force_unbind(vma); |
1828 | list_del_init(entry: &vma->vm_link); |
1829 | vm_ddestroy = vma->vm_ddestroy; |
1830 | vma->vm_ddestroy = false; |
1831 | |
1832 | /* vma->vm may be freed when releasing vma->vm->mutex. */ |
1833 | gt = vma->vm->gt; |
1834 | mutex_unlock(lock: &vma->vm->mutex); |
1835 | release_references(vma, gt, vm_ddestroy); |
1836 | } |
1837 | |
1838 | void i915_vma_parked(struct intel_gt *gt) |
1839 | { |
1840 | struct i915_vma *vma, *next; |
1841 | LIST_HEAD(closed); |
1842 | |
1843 | spin_lock_irq(lock: >->closed_lock); |
1844 | list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) { |
1845 | struct drm_i915_gem_object *obj = vma->obj; |
1846 | struct i915_address_space *vm = vma->vm; |
1847 | |
1848 | /* XXX All to avoid keeping a reference on i915_vma itself */ |
1849 | |
1850 | if (!kref_get_unless_zero(kref: &obj->base.refcount)) |
1851 | continue; |
1852 | |
1853 | if (!i915_vm_tryget(vm)) { |
1854 | i915_gem_object_put(obj); |
1855 | continue; |
1856 | } |
1857 | |
1858 | list_move(list: &vma->closed_link, head: &closed); |
1859 | } |
1860 | spin_unlock_irq(lock: >->closed_lock); |
1861 | |
1862 | /* As the GT is held idle, no vma can be reopened as we destroy them */ |
1863 | list_for_each_entry_safe(vma, next, &closed, closed_link) { |
1864 | struct drm_i915_gem_object *obj = vma->obj; |
1865 | struct i915_address_space *vm = vma->vm; |
1866 | |
1867 | if (i915_gem_object_trylock(obj, NULL)) { |
1868 | INIT_LIST_HEAD(list: &vma->closed_link); |
1869 | i915_vma_destroy(vma); |
1870 | i915_gem_object_unlock(obj); |
1871 | } else { |
1872 | /* back you go.. */ |
1873 | spin_lock_irq(lock: >->closed_lock); |
1874 | list_add(new: &vma->closed_link, head: >->closed_vma); |
1875 | spin_unlock_irq(lock: >->closed_lock); |
1876 | } |
1877 | |
1878 | i915_gem_object_put(obj); |
1879 | i915_vm_put(vm); |
1880 | } |
1881 | } |
1882 | |
1883 | static void __i915_vma_iounmap(struct i915_vma *vma) |
1884 | { |
1885 | GEM_BUG_ON(i915_vma_is_pinned(vma)); |
1886 | |
1887 | if (vma->iomap == NULL) |
1888 | return; |
1889 | |
1890 | if (page_unmask_bits(vma->iomap)) |
1891 | __i915_gem_object_release_map(obj: vma->obj); |
1892 | else |
1893 | io_mapping_unmap(vaddr: vma->iomap); |
1894 | vma->iomap = NULL; |
1895 | } |
1896 | |
1897 | void i915_vma_revoke_mmap(struct i915_vma *vma) |
1898 | { |
1899 | struct drm_vma_offset_node *node; |
1900 | u64 vma_offset; |
1901 | |
1902 | if (!i915_vma_has_userfault(vma)) |
1903 | return; |
1904 | |
1905 | GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma)); |
1906 | GEM_BUG_ON(!vma->obj->userfault_count); |
1907 | |
1908 | node = &vma->mmo->vma_node; |
1909 | vma_offset = vma->gtt_view.partial.offset << PAGE_SHIFT; |
1910 | unmap_mapping_range(mapping: vma->vm->i915->drm.anon_inode->i_mapping, |
1911 | holebegin: drm_vma_node_offset_addr(node) + vma_offset, |
1912 | holelen: vma->size, |
1913 | even_cows: 1); |
1914 | |
1915 | i915_vma_unset_userfault(vma); |
1916 | if (!--vma->obj->userfault_count) |
1917 | list_del(entry: &vma->obj->userfault_link); |
1918 | } |
1919 | |
1920 | static int |
1921 | __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma) |
1922 | { |
1923 | return __i915_request_await_exclusive(rq, active: &vma->active); |
1924 | } |
1925 | |
1926 | static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq) |
1927 | { |
1928 | int err; |
1929 | |
1930 | /* Wait for the vma to be bound before we start! */ |
1931 | err = __i915_request_await_bind(rq, vma); |
1932 | if (err) |
1933 | return err; |
1934 | |
1935 | return i915_active_add_request(ref: &vma->active, rq); |
1936 | } |
1937 | |
1938 | int _i915_vma_move_to_active(struct i915_vma *vma, |
1939 | struct i915_request *rq, |
1940 | struct dma_fence *fence, |
1941 | unsigned int flags) |
1942 | { |
1943 | struct drm_i915_gem_object *obj = vma->obj; |
1944 | int err; |
1945 | |
1946 | assert_object_held(obj); |
1947 | |
1948 | GEM_BUG_ON(!vma->pages); |
1949 | |
1950 | if (!(flags & __EXEC_OBJECT_NO_REQUEST_AWAIT)) { |
1951 | err = i915_request_await_object(to: rq, obj: vma->obj, write: flags & EXEC_OBJECT_WRITE); |
1952 | if (unlikely(err)) |
1953 | return err; |
1954 | } |
1955 | err = __i915_vma_move_to_active(vma, rq); |
1956 | if (unlikely(err)) |
1957 | return err; |
1958 | |
1959 | /* |
1960 | * Reserve fences slot early to prevent an allocation after preparing |
1961 | * the workload and associating fences with dma_resv. |
1962 | */ |
1963 | if (fence && !(flags & __EXEC_OBJECT_NO_RESERVE)) { |
1964 | struct dma_fence *curr; |
1965 | int idx; |
1966 | |
1967 | dma_fence_array_for_each(curr, idx, fence) |
1968 | ; |
1969 | err = dma_resv_reserve_fences(obj: vma->obj->base.resv, num_fences: idx); |
1970 | if (unlikely(err)) |
1971 | return err; |
1972 | } |
1973 | |
1974 | if (flags & EXEC_OBJECT_WRITE) { |
1975 | struct intel_frontbuffer *front; |
1976 | |
1977 | front = i915_gem_object_get_frontbuffer(obj); |
1978 | if (unlikely(front)) { |
1979 | if (intel_frontbuffer_invalidate(front, origin: ORIGIN_CS)) |
1980 | i915_active_add_request(ref: &front->write, rq); |
1981 | intel_frontbuffer_put(front); |
1982 | } |
1983 | } |
1984 | |
1985 | if (fence) { |
1986 | struct dma_fence *curr; |
1987 | enum dma_resv_usage usage; |
1988 | int idx; |
1989 | |
1990 | if (flags & EXEC_OBJECT_WRITE) { |
1991 | usage = DMA_RESV_USAGE_WRITE; |
1992 | obj->write_domain = I915_GEM_DOMAIN_RENDER; |
1993 | obj->read_domains = 0; |
1994 | } else { |
1995 | usage = DMA_RESV_USAGE_READ; |
1996 | obj->write_domain = 0; |
1997 | } |
1998 | |
1999 | dma_fence_array_for_each(curr, idx, fence) |
2000 | dma_resv_add_fence(obj: vma->obj->base.resv, fence: curr, usage); |
2001 | } |
2002 | |
2003 | if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence) |
2004 | i915_active_add_request(ref: &vma->fence->active, rq); |
2005 | |
2006 | obj->read_domains |= I915_GEM_GPU_DOMAINS; |
2007 | obj->mm.dirty = true; |
2008 | |
2009 | GEM_BUG_ON(!i915_vma_is_active(vma)); |
2010 | return 0; |
2011 | } |
2012 | |
2013 | struct dma_fence *__i915_vma_evict(struct i915_vma *vma, bool async) |
2014 | { |
2015 | struct i915_vma_resource *vma_res = vma->resource; |
2016 | struct dma_fence *unbind_fence; |
2017 | |
2018 | GEM_BUG_ON(i915_vma_is_pinned(vma)); |
2019 | assert_vma_held_evict(vma); |
2020 | |
2021 | if (i915_vma_is_map_and_fenceable(vma)) { |
2022 | /* Force a pagefault for domain tracking on next user access */ |
2023 | i915_vma_revoke_mmap(vma); |
2024 | |
2025 | /* |
2026 | * Check that we have flushed all writes through the GGTT |
2027 | * before the unbind, other due to non-strict nature of those |
2028 | * indirect writes they may end up referencing the GGTT PTE |
2029 | * after the unbind. |
2030 | * |
2031 | * Note that we may be concurrently poking at the GGTT_WRITE |
2032 | * bit from set-domain, as we mark all GGTT vma associated |
2033 | * with an object. We know this is for another vma, as we |
2034 | * are currently unbinding this one -- so if this vma will be |
2035 | * reused, it will be refaulted and have its dirty bit set |
2036 | * before the next write. |
2037 | */ |
2038 | i915_vma_flush_writes(vma); |
2039 | |
2040 | /* release the fence reg _after_ flushing */ |
2041 | i915_vma_revoke_fence(vma); |
2042 | |
2043 | clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); |
2044 | } |
2045 | |
2046 | __i915_vma_iounmap(vma); |
2047 | |
2048 | GEM_BUG_ON(vma->fence); |
2049 | GEM_BUG_ON(i915_vma_has_userfault(vma)); |
2050 | |
2051 | /* Object backend must be async capable. */ |
2052 | GEM_WARN_ON(async && !vma->resource->bi.pages_rsgt); |
2053 | |
2054 | /* If vm is not open, unbind is a nop. */ |
2055 | vma_res->needs_wakeref = i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND) && |
2056 | kref_read(kref: &vma->vm->ref); |
2057 | vma_res->skip_pte_rewrite = !kref_read(kref: &vma->vm->ref) || |
2058 | vma->vm->skip_pte_rewrite; |
2059 | trace_i915_vma_unbind(vma); |
2060 | |
2061 | if (async) |
2062 | unbind_fence = i915_vma_resource_unbind(vma_res, |
2063 | tlb: vma->obj->mm.tlb); |
2064 | else |
2065 | unbind_fence = i915_vma_resource_unbind(vma_res, NULL); |
2066 | |
2067 | vma->resource = NULL; |
2068 | |
2069 | atomic_and(i: ~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE), |
2070 | v: &vma->flags); |
2071 | |
2072 | i915_vma_detach(vma); |
2073 | |
2074 | if (!async) { |
2075 | if (unbind_fence) { |
2076 | dma_fence_wait(fence: unbind_fence, intr: false); |
2077 | dma_fence_put(fence: unbind_fence); |
2078 | unbind_fence = NULL; |
2079 | } |
2080 | vma_invalidate_tlb(vm: vma->vm, tlb: vma->obj->mm.tlb); |
2081 | } |
2082 | |
2083 | /* |
2084 | * Binding itself may not have completed until the unbind fence signals, |
2085 | * so don't drop the pages until that happens, unless the resource is |
2086 | * async_capable. |
2087 | */ |
2088 | |
2089 | vma_unbind_pages(vma); |
2090 | return unbind_fence; |
2091 | } |
2092 | |
2093 | int __i915_vma_unbind(struct i915_vma *vma) |
2094 | { |
2095 | int ret; |
2096 | |
2097 | lockdep_assert_held(&vma->vm->mutex); |
2098 | assert_vma_held_evict(vma); |
2099 | |
2100 | if (!drm_mm_node_allocated(node: &vma->node)) |
2101 | return 0; |
2102 | |
2103 | if (i915_vma_is_pinned(vma)) { |
2104 | vma_print_allocator(vma, reason: "is pinned" ); |
2105 | return -EAGAIN; |
2106 | } |
2107 | |
2108 | /* |
2109 | * After confirming that no one else is pinning this vma, wait for |
2110 | * any laggards who may have crept in during the wait (through |
2111 | * a residual pin skipping the vm->mutex) to complete. |
2112 | */ |
2113 | ret = i915_vma_sync(vma); |
2114 | if (ret) |
2115 | return ret; |
2116 | |
2117 | GEM_BUG_ON(i915_vma_is_active(vma)); |
2118 | __i915_vma_evict(vma, async: false); |
2119 | |
2120 | drm_mm_remove_node(node: &vma->node); /* pairs with i915_vma_release() */ |
2121 | return 0; |
2122 | } |
2123 | |
2124 | static struct dma_fence *__i915_vma_unbind_async(struct i915_vma *vma) |
2125 | { |
2126 | struct dma_fence *fence; |
2127 | |
2128 | lockdep_assert_held(&vma->vm->mutex); |
2129 | |
2130 | if (!drm_mm_node_allocated(node: &vma->node)) |
2131 | return NULL; |
2132 | |
2133 | if (i915_vma_is_pinned(vma) || |
2134 | &vma->obj->mm.rsgt->table != vma->resource->bi.pages) |
2135 | return ERR_PTR(error: -EAGAIN); |
2136 | |
2137 | /* |
2138 | * We probably need to replace this with awaiting the fences of the |
2139 | * object's dma_resv when the vma active goes away. When doing that |
2140 | * we need to be careful to not add the vma_resource unbind fence |
2141 | * immediately to the object's dma_resv, because then unbinding |
2142 | * the next vma from the object, in case there are many, will |
2143 | * actually await the unbinding of the previous vmas, which is |
2144 | * undesirable. |
2145 | */ |
2146 | if (i915_sw_fence_await_active(fence: &vma->resource->chain, ref: &vma->active, |
2147 | I915_ACTIVE_AWAIT_EXCL | |
2148 | I915_ACTIVE_AWAIT_ACTIVE) < 0) { |
2149 | return ERR_PTR(error: -EBUSY); |
2150 | } |
2151 | |
2152 | fence = __i915_vma_evict(vma, async: true); |
2153 | |
2154 | drm_mm_remove_node(node: &vma->node); /* pairs with i915_vma_release() */ |
2155 | |
2156 | return fence; |
2157 | } |
2158 | |
2159 | int i915_vma_unbind(struct i915_vma *vma) |
2160 | { |
2161 | struct i915_address_space *vm = vma->vm; |
2162 | intel_wakeref_t wakeref = 0; |
2163 | int err; |
2164 | |
2165 | assert_object_held_shared(obj: vma->obj); |
2166 | |
2167 | /* Optimistic wait before taking the mutex */ |
2168 | err = i915_vma_sync(vma); |
2169 | if (err) |
2170 | return err; |
2171 | |
2172 | if (!drm_mm_node_allocated(node: &vma->node)) |
2173 | return 0; |
2174 | |
2175 | if (i915_vma_is_pinned(vma)) { |
2176 | vma_print_allocator(vma, reason: "is pinned" ); |
2177 | return -EAGAIN; |
2178 | } |
2179 | |
2180 | if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) |
2181 | /* XXX not always required: nop_clear_range */ |
2182 | wakeref = intel_runtime_pm_get(rpm: &vm->i915->runtime_pm); |
2183 | |
2184 | err = mutex_lock_interruptible_nested(lock: &vma->vm->mutex, subclass: !wakeref); |
2185 | if (err) |
2186 | goto out_rpm; |
2187 | |
2188 | err = __i915_vma_unbind(vma); |
2189 | mutex_unlock(lock: &vm->mutex); |
2190 | |
2191 | out_rpm: |
2192 | if (wakeref) |
2193 | intel_runtime_pm_put(rpm: &vm->i915->runtime_pm, wref: wakeref); |
2194 | return err; |
2195 | } |
2196 | |
2197 | int i915_vma_unbind_async(struct i915_vma *vma, bool trylock_vm) |
2198 | { |
2199 | struct drm_i915_gem_object *obj = vma->obj; |
2200 | struct i915_address_space *vm = vma->vm; |
2201 | intel_wakeref_t wakeref = 0; |
2202 | struct dma_fence *fence; |
2203 | int err; |
2204 | |
2205 | /* |
2206 | * We need the dma-resv lock since we add the |
2207 | * unbind fence to the dma-resv object. |
2208 | */ |
2209 | assert_object_held(obj); |
2210 | |
2211 | if (!drm_mm_node_allocated(node: &vma->node)) |
2212 | return 0; |
2213 | |
2214 | if (i915_vma_is_pinned(vma)) { |
2215 | vma_print_allocator(vma, reason: "is pinned" ); |
2216 | return -EAGAIN; |
2217 | } |
2218 | |
2219 | if (!obj->mm.rsgt) |
2220 | return -EBUSY; |
2221 | |
2222 | err = dma_resv_reserve_fences(obj: obj->base.resv, num_fences: 2); |
2223 | if (err) |
2224 | return -EBUSY; |
2225 | |
2226 | /* |
2227 | * It would be great if we could grab this wakeref from the |
2228 | * async unbind work if needed, but we can't because it uses |
2229 | * kmalloc and it's in the dma-fence signalling critical path. |
2230 | */ |
2231 | if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) |
2232 | wakeref = intel_runtime_pm_get(rpm: &vm->i915->runtime_pm); |
2233 | |
2234 | if (trylock_vm && !mutex_trylock(lock: &vm->mutex)) { |
2235 | err = -EBUSY; |
2236 | goto out_rpm; |
2237 | } else if (!trylock_vm) { |
2238 | err = mutex_lock_interruptible_nested(lock: &vm->mutex, subclass: !wakeref); |
2239 | if (err) |
2240 | goto out_rpm; |
2241 | } |
2242 | |
2243 | fence = __i915_vma_unbind_async(vma); |
2244 | mutex_unlock(lock: &vm->mutex); |
2245 | if (IS_ERR_OR_NULL(ptr: fence)) { |
2246 | err = PTR_ERR_OR_ZERO(ptr: fence); |
2247 | goto out_rpm; |
2248 | } |
2249 | |
2250 | dma_resv_add_fence(obj: obj->base.resv, fence, usage: DMA_RESV_USAGE_READ); |
2251 | dma_fence_put(fence); |
2252 | |
2253 | out_rpm: |
2254 | if (wakeref) |
2255 | intel_runtime_pm_put(rpm: &vm->i915->runtime_pm, wref: wakeref); |
2256 | return err; |
2257 | } |
2258 | |
2259 | int i915_vma_unbind_unlocked(struct i915_vma *vma) |
2260 | { |
2261 | int err; |
2262 | |
2263 | i915_gem_object_lock(obj: vma->obj, NULL); |
2264 | err = i915_vma_unbind(vma); |
2265 | i915_gem_object_unlock(obj: vma->obj); |
2266 | |
2267 | return err; |
2268 | } |
2269 | |
2270 | struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma) |
2271 | { |
2272 | i915_gem_object_make_unshrinkable(obj: vma->obj); |
2273 | return vma; |
2274 | } |
2275 | |
2276 | void i915_vma_make_shrinkable(struct i915_vma *vma) |
2277 | { |
2278 | i915_gem_object_make_shrinkable(obj: vma->obj); |
2279 | } |
2280 | |
2281 | void i915_vma_make_purgeable(struct i915_vma *vma) |
2282 | { |
2283 | i915_gem_object_make_purgeable(obj: vma->obj); |
2284 | } |
2285 | |
2286 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) |
2287 | #include "selftests/i915_vma.c" |
2288 | #endif |
2289 | |
2290 | void i915_vma_module_exit(void) |
2291 | { |
2292 | kmem_cache_destroy(s: slab_vmas); |
2293 | } |
2294 | |
2295 | int __init i915_vma_module_init(void) |
2296 | { |
2297 | slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN); |
2298 | if (!slab_vmas) |
2299 | return -ENOMEM; |
2300 | |
2301 | return 0; |
2302 | } |
2303 | |