1/*
2 * Copyright © 2008-2010 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 * Authors:
24 * Eric Anholt <eric@anholt.net>
25 * Chris Wilson <chris@chris-wilson.co.uuk>
26 *
27 */
28
29#include "gem/i915_gem_context.h"
30#include "gt/intel_gt.h"
31#include "gt/intel_gt_requests.h"
32
33#include "i915_drv.h"
34#include "i915_gem_evict.h"
35#include "i915_trace.h"
36
37I915_SELFTEST_DECLARE(static struct igt_evict_ctl {
38 bool fail_if_busy:1;
39} igt_evict_ctl;)
40
41static bool dying_vma(struct i915_vma *vma)
42{
43 return !kref_read(kref: &vma->obj->base.refcount);
44}
45
46static int ggtt_flush(struct i915_address_space *vm)
47{
48 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
49 struct intel_gt *gt;
50 int ret = 0;
51
52 list_for_each_entry(gt, &ggtt->gt_list, ggtt_link) {
53 /*
54 * Not everything in the GGTT is tracked via vma (otherwise we
55 * could evict as required with minimal stalling) so we are forced
56 * to idle the GPU and explicitly retire outstanding requests in
57 * the hopes that we can then remove contexts and the like only
58 * bound by their active reference.
59 */
60 ret = intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
61 if (ret)
62 return ret;
63 }
64 return ret;
65}
66
67static bool grab_vma(struct i915_vma *vma, struct i915_gem_ww_ctx *ww)
68{
69 /*
70 * We add the extra refcount so the object doesn't drop to zero until
71 * after ungrab_vma(), this way trylock is always paired with unlock.
72 */
73 if (i915_gem_object_get_rcu(obj: vma->obj)) {
74 if (!i915_gem_object_trylock(obj: vma->obj, ww)) {
75 i915_gem_object_put(obj: vma->obj);
76 return false;
77 }
78 } else {
79 /* Dead objects don't need pins */
80 atomic_and(i: ~I915_VMA_PIN_MASK, v: &vma->flags);
81 }
82
83 return true;
84}
85
86static void ungrab_vma(struct i915_vma *vma)
87{
88 if (dying_vma(vma))
89 return;
90
91 i915_gem_object_unlock(obj: vma->obj);
92 i915_gem_object_put(obj: vma->obj);
93}
94
95static bool
96mark_free(struct drm_mm_scan *scan,
97 struct i915_gem_ww_ctx *ww,
98 struct i915_vma *vma,
99 unsigned int flags,
100 struct list_head *unwind)
101{
102 if (i915_vma_is_pinned(vma))
103 return false;
104
105 if (!grab_vma(vma, ww))
106 return false;
107
108 list_add(new: &vma->evict_link, head: unwind);
109 return drm_mm_scan_add_block(scan, node: &vma->node);
110}
111
112static bool defer_evict(struct i915_vma *vma)
113{
114 if (i915_vma_is_active(vma))
115 return true;
116
117 if (i915_vma_is_scanout(vma))
118 return true;
119
120 return false;
121}
122
123/**
124 * i915_gem_evict_something - Evict vmas to make room for binding a new one
125 * @vm: address space to evict from
126 * @ww: An optional struct i915_gem_ww_ctx.
127 * @min_size: size of the desired free space
128 * @alignment: alignment constraint of the desired free space
129 * @color: color for the desired space
130 * @start: start (inclusive) of the range from which to evict objects
131 * @end: end (exclusive) of the range from which to evict objects
132 * @flags: additional flags to control the eviction algorithm
133 *
134 * This function will try to evict vmas until a free space satisfying the
135 * requirements is found. Callers must check first whether any such hole exists
136 * already before calling this function.
137 *
138 * This function is used by the object/vma binding code.
139 *
140 * Since this function is only used to free up virtual address space it only
141 * ignores pinned vmas, and not object where the backing storage itself is
142 * pinned. Hence obj->pages_pin_count does not protect against eviction.
143 *
144 * To clarify: This is for freeing up virtual address space, not for freeing
145 * memory in e.g. the shrinker.
146 */
147int
148i915_gem_evict_something(struct i915_address_space *vm,
149 struct i915_gem_ww_ctx *ww,
150 u64 min_size, u64 alignment,
151 unsigned long color,
152 u64 start, u64 end,
153 unsigned flags)
154{
155 struct drm_mm_scan scan;
156 struct list_head eviction_list;
157 struct i915_vma *vma, *next;
158 struct drm_mm_node *node;
159 enum drm_mm_insert_mode mode;
160 struct i915_vma *active;
161 struct intel_gt *gt;
162 int ret;
163
164 lockdep_assert_held(&vm->mutex);
165 trace_i915_gem_evict(vm, size: min_size, align: alignment, flags);
166
167 /*
168 * The goal is to evict objects and amalgamate space in rough LRU order.
169 * Since both active and inactive objects reside on the same list,
170 * in a mix of creation and last scanned order, as we process the list
171 * we sort it into inactive/active, which keeps the active portion
172 * in a rough MRU order.
173 *
174 * The retirement sequence is thus:
175 * 1. Inactive objects (already retired, random order)
176 * 2. Active objects (will stall on unbinding, oldest scanned first)
177 */
178 mode = DRM_MM_INSERT_BEST;
179 if (flags & PIN_HIGH)
180 mode = DRM_MM_INSERT_HIGH;
181 if (flags & PIN_MAPPABLE)
182 mode = DRM_MM_INSERT_LOW;
183 drm_mm_scan_init_with_range(scan: &scan, mm: &vm->mm,
184 size: min_size, alignment, color,
185 start, end, mode);
186
187 if (i915_is_ggtt(vm)) {
188 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
189
190 list_for_each_entry(gt, &ggtt->gt_list, ggtt_link)
191 intel_gt_retire_requests(gt);
192 } else {
193 intel_gt_retire_requests(gt: vm->gt);
194 }
195
196search_again:
197 active = NULL;
198 INIT_LIST_HEAD(list: &eviction_list);
199 list_for_each_entry_safe(vma, next, &vm->bound_list, vm_link) {
200 if (vma == active) { /* now seen this vma twice */
201 if (flags & PIN_NONBLOCK)
202 break;
203
204 active = ERR_PTR(error: -EAGAIN);
205 }
206
207 /*
208 * We keep this list in a rough least-recently scanned order
209 * of active elements (inactive elements are cheap to reap).
210 * New entries are added to the end, and we move anything we
211 * scan to the end. The assumption is that the working set
212 * of applications is either steady state (and thanks to the
213 * userspace bo cache it almost always is) or volatile and
214 * frequently replaced after a frame, which are self-evicting!
215 * Given that assumption, the MRU order of the scan list is
216 * fairly static, and keeping it in least-recently scan order
217 * is suitable.
218 *
219 * To notice when we complete one full cycle, we record the
220 * first active element seen, before moving it to the tail.
221 */
222 if (active != ERR_PTR(error: -EAGAIN) && defer_evict(vma)) {
223 if (!active)
224 active = vma;
225
226 list_move_tail(list: &vma->vm_link, head: &vm->bound_list);
227 continue;
228 }
229
230 if (mark_free(scan: &scan, ww, vma, flags, unwind: &eviction_list))
231 goto found;
232 }
233
234 /* Nothing found, clean up and bail out! */
235 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
236 ret = drm_mm_scan_remove_block(scan: &scan, node: &vma->node);
237 BUG_ON(ret);
238 ungrab_vma(vma);
239 }
240
241 /*
242 * Can we unpin some objects such as idle hw contents,
243 * or pending flips? But since only the GGTT has global entries
244 * such as scanouts, rinbuffers and contexts, we can skip the
245 * purge when inspecting per-process local address spaces.
246 */
247 if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK)
248 return -ENOSPC;
249
250 /*
251 * Not everything in the GGTT is tracked via VMA using
252 * i915_vma_move_to_active(), otherwise we could evict as required
253 * with minimal stalling. Instead we are forced to idle the GPU and
254 * explicitly retire outstanding requests which will then remove
255 * the pinning for active objects such as contexts and ring,
256 * enabling us to evict them on the next iteration.
257 *
258 * To ensure that all user contexts are evictable, we perform
259 * a switch to the perma-pinned kernel context. This all also gives
260 * us a termination condition, when the last retired context is
261 * the kernel's there is no more we can evict.
262 */
263 if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy))
264 return -EBUSY;
265
266 ret = ggtt_flush(vm);
267 if (ret)
268 return ret;
269
270 cond_resched();
271
272 flags |= PIN_NONBLOCK;
273 goto search_again;
274
275found:
276 /* drm_mm doesn't allow any other other operations while
277 * scanning, therefore store to-be-evicted objects on a
278 * temporary list and take a reference for all before
279 * calling unbind (which may remove the active reference
280 * of any of our objects, thus corrupting the list).
281 */
282 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
283 if (drm_mm_scan_remove_block(scan: &scan, node: &vma->node)) {
284 __i915_vma_pin(vma);
285 } else {
286 list_del(entry: &vma->evict_link);
287 ungrab_vma(vma);
288 }
289 }
290
291 /* Unbinding will emit any required flushes */
292 ret = 0;
293 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
294 __i915_vma_unpin(vma);
295 if (ret == 0)
296 ret = __i915_vma_unbind(vma);
297 ungrab_vma(vma);
298 }
299
300 while (ret == 0 && (node = drm_mm_scan_color_evict(scan: &scan))) {
301 vma = container_of(node, struct i915_vma, node);
302
303 /* If we find any non-objects (!vma), we cannot evict them */
304 if (vma->node.color != I915_COLOR_UNEVICTABLE &&
305 grab_vma(vma, ww)) {
306 ret = __i915_vma_unbind(vma);
307 ungrab_vma(vma);
308 } else {
309 ret = -ENOSPC;
310 }
311 }
312
313 return ret;
314}
315
316/**
317 * i915_gem_evict_for_node - Evict vmas to make room for binding a new one
318 * @vm: address space to evict from
319 * @ww: An optional struct i915_gem_ww_ctx.
320 * @target: range (and color) to evict for
321 * @flags: additional flags to control the eviction algorithm
322 *
323 * This function will try to evict vmas that overlap the target node.
324 *
325 * To clarify: This is for freeing up virtual address space, not for freeing
326 * memory in e.g. the shrinker.
327 */
328int i915_gem_evict_for_node(struct i915_address_space *vm,
329 struct i915_gem_ww_ctx *ww,
330 struct drm_mm_node *target,
331 unsigned int flags)
332{
333 LIST_HEAD(eviction_list);
334 struct drm_mm_node *node;
335 u64 start = target->start;
336 u64 end = start + target->size;
337 struct i915_vma *vma, *next;
338 int ret = 0;
339
340 lockdep_assert_held(&vm->mutex);
341 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
342 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
343
344 trace_i915_gem_evict_node(vm, node: target, flags);
345
346 /*
347 * Retire before we search the active list. Although we have
348 * reasonable accuracy in our retirement lists, we may have
349 * a stray pin (preventing eviction) that can only be resolved by
350 * retiring.
351 */
352 if (i915_is_ggtt(vm)) {
353 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
354 struct intel_gt *gt;
355
356 list_for_each_entry(gt, &ggtt->gt_list, ggtt_link)
357 intel_gt_retire_requests(gt);
358 } else {
359 intel_gt_retire_requests(gt: vm->gt);
360 }
361
362 if (i915_vm_has_cache_coloring(vm)) {
363 /* Expand search to cover neighbouring guard pages (or lack!) */
364 if (start)
365 start -= I915_GTT_PAGE_SIZE;
366
367 /* Always look at the page afterwards to avoid the end-of-GTT */
368 end += I915_GTT_PAGE_SIZE;
369 }
370 GEM_BUG_ON(start >= end);
371
372 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
373 /* If we find any non-objects (!vma), we cannot evict them */
374 if (node->color == I915_COLOR_UNEVICTABLE) {
375 ret = -ENOSPC;
376 break;
377 }
378
379 GEM_BUG_ON(!drm_mm_node_allocated(node));
380 vma = container_of(node, typeof(*vma), node);
381
382 /*
383 * If we are using coloring to insert guard pages between
384 * different cache domains within the address space, we have
385 * to check whether the objects on either side of our range
386 * abutt and conflict. If they are in conflict, then we evict
387 * those as well to make room for our guard pages.
388 */
389 if (i915_vm_has_cache_coloring(vm)) {
390 if (node->start + node->size == target->start) {
391 if (node->color == target->color)
392 continue;
393 }
394 if (node->start == target->start + target->size) {
395 if (node->color == target->color)
396 continue;
397 }
398 }
399
400 if (i915_vma_is_pinned(vma)) {
401 ret = -ENOSPC;
402 break;
403 }
404
405 if (flags & PIN_NONBLOCK && i915_vma_is_active(vma)) {
406 ret = -ENOSPC;
407 break;
408 }
409
410 if (!grab_vma(vma, ww)) {
411 ret = -ENOSPC;
412 break;
413 }
414
415 /*
416 * Never show fear in the face of dragons!
417 *
418 * We cannot directly remove this node from within this
419 * iterator and as with i915_gem_evict_something() we employ
420 * the vma pin_count in order to prevent the action of
421 * unbinding one vma from freeing (by dropping its active
422 * reference) another in our eviction list.
423 */
424 __i915_vma_pin(vma);
425 list_add(new: &vma->evict_link, head: &eviction_list);
426 }
427
428 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
429 __i915_vma_unpin(vma);
430 if (ret == 0)
431 ret = __i915_vma_unbind(vma);
432
433 ungrab_vma(vma);
434 }
435
436 return ret;
437}
438
439/**
440 * i915_gem_evict_vm - Evict all idle vmas from a vm
441 * @vm: Address space to cleanse
442 * @ww: An optional struct i915_gem_ww_ctx. If not NULL, i915_gem_evict_vm
443 * will be able to evict vma's locked by the ww as well.
444 * @busy_bo: Optional pointer to struct drm_i915_gem_object. If not NULL, then
445 * in the event i915_gem_evict_vm() is unable to trylock an object for eviction,
446 * then @busy_bo will point to it. -EBUSY is also returned. The caller must drop
447 * the vm->mutex, before trying again to acquire the contended lock. The caller
448 * also owns a reference to the object.
449 *
450 * This function evicts all vmas from a vm.
451 *
452 * This is used by the execbuf code as a last-ditch effort to defragment the
453 * address space.
454 *
455 * To clarify: This is for freeing up virtual address space, not for freeing
456 * memory in e.g. the shrinker.
457 */
458int i915_gem_evict_vm(struct i915_address_space *vm, struct i915_gem_ww_ctx *ww,
459 struct drm_i915_gem_object **busy_bo)
460{
461 int ret = 0;
462
463 lockdep_assert_held(&vm->mutex);
464 trace_i915_gem_evict_vm(vm);
465
466 /* Switch back to the default context in order to unpin
467 * the existing context objects. However, such objects only
468 * pin themselves inside the global GTT and performing the
469 * switch otherwise is ineffective.
470 */
471 if (i915_is_ggtt(vm)) {
472 ret = ggtt_flush(vm);
473 if (ret)
474 return ret;
475 }
476
477 do {
478 struct i915_vma *vma, *vn;
479 LIST_HEAD(eviction_list);
480 LIST_HEAD(locked_eviction_list);
481
482 list_for_each_entry(vma, &vm->bound_list, vm_link) {
483 if (i915_vma_is_pinned(vma))
484 continue;
485
486 /*
487 * If we already own the lock, trylock fails. In case
488 * the resv is shared among multiple objects, we still
489 * need the object ref.
490 */
491 if (!i915_gem_object_get_rcu(obj: vma->obj) ||
492 (ww && (dma_resv_locking_ctx(obj: vma->obj->base.resv) == &ww->ctx))) {
493 __i915_vma_pin(vma);
494 list_add(new: &vma->evict_link, head: &locked_eviction_list);
495 continue;
496 }
497
498 if (!i915_gem_object_trylock(obj: vma->obj, ww)) {
499 if (busy_bo) {
500 *busy_bo = vma->obj; /* holds ref */
501 ret = -EBUSY;
502 break;
503 }
504 i915_gem_object_put(obj: vma->obj);
505 continue;
506 }
507
508 __i915_vma_pin(vma);
509 list_add(new: &vma->evict_link, head: &eviction_list);
510 }
511 if (list_empty(head: &eviction_list) && list_empty(head: &locked_eviction_list))
512 break;
513
514 /* Unbind locked objects first, before unlocking the eviction_list */
515 list_for_each_entry_safe(vma, vn, &locked_eviction_list, evict_link) {
516 __i915_vma_unpin(vma);
517
518 if (ret == 0) {
519 ret = __i915_vma_unbind(vma);
520 if (ret != -EINTR) /* "Get me out of here!" */
521 ret = 0;
522 }
523 if (!dying_vma(vma))
524 i915_gem_object_put(obj: vma->obj);
525 }
526
527 list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) {
528 __i915_vma_unpin(vma);
529 if (ret == 0) {
530 ret = __i915_vma_unbind(vma);
531 if (ret != -EINTR) /* "Get me out of here!" */
532 ret = 0;
533 }
534
535 i915_gem_object_unlock(obj: vma->obj);
536 i915_gem_object_put(obj: vma->obj);
537 }
538 } while (ret == 0);
539
540 return ret;
541}
542
543#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
544#include "selftests/i915_gem_evict.c"
545#endif
546

source code of linux/drivers/gpu/drm/i915/i915_gem_evict.c