xref: /linux/drivers/gpu/drm/i915/i915_gem_evict.c (revision 0a94608f0f7de9b1135ffea3546afe68eafef57f)
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 
37 I915_SELFTEST_DECLARE(static struct igt_evict_ctl {
38 	bool fail_if_busy:1;
39 } igt_evict_ctl;)
40 
41 static bool dying_vma(struct i915_vma *vma)
42 {
43 	return !kref_read(&vma->obj->base.refcount);
44 }
45 
46 static int ggtt_flush(struct intel_gt *gt)
47 {
48 	/*
49 	 * Not everything in the GGTT is tracked via vma (otherwise we
50 	 * could evict as required with minimal stalling) so we are forced
51 	 * to idle the GPU and explicitly retire outstanding requests in
52 	 * the hopes that we can then remove contexts and the like only
53 	 * bound by their active reference.
54 	 */
55 	return intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
56 }
57 
58 static bool grab_vma(struct i915_vma *vma, struct i915_gem_ww_ctx *ww)
59 {
60 	/*
61 	 * We add the extra refcount so the object doesn't drop to zero until
62 	 * after ungrab_vma(), this way trylock is always paired with unlock.
63 	 */
64 	if (i915_gem_object_get_rcu(vma->obj)) {
65 		if (!i915_gem_object_trylock(vma->obj, ww)) {
66 			i915_gem_object_put(vma->obj);
67 			return false;
68 		}
69 	} else {
70 		/* Dead objects don't need pins */
71 		atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
72 	}
73 
74 	return true;
75 }
76 
77 static void ungrab_vma(struct i915_vma *vma)
78 {
79 	if (dying_vma(vma))
80 		return;
81 
82 	i915_gem_object_unlock(vma->obj);
83 	i915_gem_object_put(vma->obj);
84 }
85 
86 static bool
87 mark_free(struct drm_mm_scan *scan,
88 	  struct i915_gem_ww_ctx *ww,
89 	  struct i915_vma *vma,
90 	  unsigned int flags,
91 	  struct list_head *unwind)
92 {
93 	if (i915_vma_is_pinned(vma))
94 		return false;
95 
96 	if (!grab_vma(vma, ww))
97 		return false;
98 
99 	list_add(&vma->evict_link, unwind);
100 	return drm_mm_scan_add_block(scan, &vma->node);
101 }
102 
103 static bool defer_evict(struct i915_vma *vma)
104 {
105 	if (i915_vma_is_active(vma))
106 		return true;
107 
108 	if (i915_vma_is_scanout(vma))
109 		return true;
110 
111 	return false;
112 }
113 
114 /**
115  * i915_gem_evict_something - Evict vmas to make room for binding a new one
116  * @vm: address space to evict from
117  * @ww: An optional struct i915_gem_ww_ctx.
118  * @min_size: size of the desired free space
119  * @alignment: alignment constraint of the desired free space
120  * @color: color for the desired space
121  * @start: start (inclusive) of the range from which to evict objects
122  * @end: end (exclusive) of the range from which to evict objects
123  * @flags: additional flags to control the eviction algorithm
124  *
125  * This function will try to evict vmas until a free space satisfying the
126  * requirements is found. Callers must check first whether any such hole exists
127  * already before calling this function.
128  *
129  * This function is used by the object/vma binding code.
130  *
131  * Since this function is only used to free up virtual address space it only
132  * ignores pinned vmas, and not object where the backing storage itself is
133  * pinned. Hence obj->pages_pin_count does not protect against eviction.
134  *
135  * To clarify: This is for freeing up virtual address space, not for freeing
136  * memory in e.g. the shrinker.
137  */
138 int
139 i915_gem_evict_something(struct i915_address_space *vm,
140 			 struct i915_gem_ww_ctx *ww,
141 			 u64 min_size, u64 alignment,
142 			 unsigned long color,
143 			 u64 start, u64 end,
144 			 unsigned flags)
145 {
146 	struct drm_mm_scan scan;
147 	struct list_head eviction_list;
148 	struct i915_vma *vma, *next;
149 	struct drm_mm_node *node;
150 	enum drm_mm_insert_mode mode;
151 	struct i915_vma *active;
152 	int ret;
153 
154 	lockdep_assert_held(&vm->mutex);
155 	trace_i915_gem_evict(vm, min_size, alignment, flags);
156 
157 	/*
158 	 * The goal is to evict objects and amalgamate space in rough LRU order.
159 	 * Since both active and inactive objects reside on the same list,
160 	 * in a mix of creation and last scanned order, as we process the list
161 	 * we sort it into inactive/active, which keeps the active portion
162 	 * in a rough MRU order.
163 	 *
164 	 * The retirement sequence is thus:
165 	 *   1. Inactive objects (already retired, random order)
166 	 *   2. Active objects (will stall on unbinding, oldest scanned first)
167 	 */
168 	mode = DRM_MM_INSERT_BEST;
169 	if (flags & PIN_HIGH)
170 		mode = DRM_MM_INSERT_HIGH;
171 	if (flags & PIN_MAPPABLE)
172 		mode = DRM_MM_INSERT_LOW;
173 	drm_mm_scan_init_with_range(&scan, &vm->mm,
174 				    min_size, alignment, color,
175 				    start, end, mode);
176 
177 	intel_gt_retire_requests(vm->gt);
178 
179 search_again:
180 	active = NULL;
181 	INIT_LIST_HEAD(&eviction_list);
182 	list_for_each_entry_safe(vma, next, &vm->bound_list, vm_link) {
183 		if (vma == active) { /* now seen this vma twice */
184 			if (flags & PIN_NONBLOCK)
185 				break;
186 
187 			active = ERR_PTR(-EAGAIN);
188 		}
189 
190 		/*
191 		 * We keep this list in a rough least-recently scanned order
192 		 * of active elements (inactive elements are cheap to reap).
193 		 * New entries are added to the end, and we move anything we
194 		 * scan to the end. The assumption is that the working set
195 		 * of applications is either steady state (and thanks to the
196 		 * userspace bo cache it almost always is) or volatile and
197 		 * frequently replaced after a frame, which are self-evicting!
198 		 * Given that assumption, the MRU order of the scan list is
199 		 * fairly static, and keeping it in least-recently scan order
200 		 * is suitable.
201 		 *
202 		 * To notice when we complete one full cycle, we record the
203 		 * first active element seen, before moving it to the tail.
204 		 */
205 		if (active != ERR_PTR(-EAGAIN) && defer_evict(vma)) {
206 			if (!active)
207 				active = vma;
208 
209 			list_move_tail(&vma->vm_link, &vm->bound_list);
210 			continue;
211 		}
212 
213 		if (mark_free(&scan, ww, vma, flags, &eviction_list))
214 			goto found;
215 	}
216 
217 	/* Nothing found, clean up and bail out! */
218 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
219 		ret = drm_mm_scan_remove_block(&scan, &vma->node);
220 		BUG_ON(ret);
221 		ungrab_vma(vma);
222 	}
223 
224 	/*
225 	 * Can we unpin some objects such as idle hw contents,
226 	 * or pending flips? But since only the GGTT has global entries
227 	 * such as scanouts, rinbuffers and contexts, we can skip the
228 	 * purge when inspecting per-process local address spaces.
229 	 */
230 	if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK)
231 		return -ENOSPC;
232 
233 	/*
234 	 * Not everything in the GGTT is tracked via VMA using
235 	 * i915_vma_move_to_active(), otherwise we could evict as required
236 	 * with minimal stalling. Instead we are forced to idle the GPU and
237 	 * explicitly retire outstanding requests which will then remove
238 	 * the pinning for active objects such as contexts and ring,
239 	 * enabling us to evict them on the next iteration.
240 	 *
241 	 * To ensure that all user contexts are evictable, we perform
242 	 * a switch to the perma-pinned kernel context. This all also gives
243 	 * us a termination condition, when the last retired context is
244 	 * the kernel's there is no more we can evict.
245 	 */
246 	if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy))
247 		return -EBUSY;
248 
249 	ret = ggtt_flush(vm->gt);
250 	if (ret)
251 		return ret;
252 
253 	cond_resched();
254 
255 	flags |= PIN_NONBLOCK;
256 	goto search_again;
257 
258 found:
259 	/* drm_mm doesn't allow any other other operations while
260 	 * scanning, therefore store to-be-evicted objects on a
261 	 * temporary list and take a reference for all before
262 	 * calling unbind (which may remove the active reference
263 	 * of any of our objects, thus corrupting the list).
264 	 */
265 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
266 		if (drm_mm_scan_remove_block(&scan, &vma->node)) {
267 			__i915_vma_pin(vma);
268 		} else {
269 			list_del(&vma->evict_link);
270 			ungrab_vma(vma);
271 		}
272 	}
273 
274 	/* Unbinding will emit any required flushes */
275 	ret = 0;
276 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
277 		__i915_vma_unpin(vma);
278 		if (ret == 0)
279 			ret = __i915_vma_unbind(vma);
280 		ungrab_vma(vma);
281 	}
282 
283 	while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) {
284 		vma = container_of(node, struct i915_vma, node);
285 
286 		/* If we find any non-objects (!vma), we cannot evict them */
287 		if (vma->node.color != I915_COLOR_UNEVICTABLE &&
288 		    grab_vma(vma, ww)) {
289 			ret = __i915_vma_unbind(vma);
290 			ungrab_vma(vma);
291 		} else {
292 			ret = -ENOSPC;
293 		}
294 	}
295 
296 	return ret;
297 }
298 
299 /**
300  * i915_gem_evict_for_node - Evict vmas to make room for binding a new one
301  * @vm: address space to evict from
302  * @ww: An optional struct i915_gem_ww_ctx.
303  * @target: range (and color) to evict for
304  * @flags: additional flags to control the eviction algorithm
305  *
306  * This function will try to evict vmas that overlap the target node.
307  *
308  * To clarify: This is for freeing up virtual address space, not for freeing
309  * memory in e.g. the shrinker.
310  */
311 int i915_gem_evict_for_node(struct i915_address_space *vm,
312 			    struct i915_gem_ww_ctx *ww,
313 			    struct drm_mm_node *target,
314 			    unsigned int flags)
315 {
316 	LIST_HEAD(eviction_list);
317 	struct drm_mm_node *node;
318 	u64 start = target->start;
319 	u64 end = start + target->size;
320 	struct i915_vma *vma, *next;
321 	int ret = 0;
322 
323 	lockdep_assert_held(&vm->mutex);
324 	GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
325 	GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
326 
327 	trace_i915_gem_evict_node(vm, target, flags);
328 
329 	/*
330 	 * Retire before we search the active list. Although we have
331 	 * reasonable accuracy in our retirement lists, we may have
332 	 * a stray pin (preventing eviction) that can only be resolved by
333 	 * retiring.
334 	 */
335 	intel_gt_retire_requests(vm->gt);
336 
337 	if (i915_vm_has_cache_coloring(vm)) {
338 		/* Expand search to cover neighbouring guard pages (or lack!) */
339 		if (start)
340 			start -= I915_GTT_PAGE_SIZE;
341 
342 		/* Always look at the page afterwards to avoid the end-of-GTT */
343 		end += I915_GTT_PAGE_SIZE;
344 	}
345 	GEM_BUG_ON(start >= end);
346 
347 	drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
348 		/* If we find any non-objects (!vma), we cannot evict them */
349 		if (node->color == I915_COLOR_UNEVICTABLE) {
350 			ret = -ENOSPC;
351 			break;
352 		}
353 
354 		GEM_BUG_ON(!drm_mm_node_allocated(node));
355 		vma = container_of(node, typeof(*vma), node);
356 
357 		/*
358 		 * If we are using coloring to insert guard pages between
359 		 * different cache domains within the address space, we have
360 		 * to check whether the objects on either side of our range
361 		 * abutt and conflict. If they are in conflict, then we evict
362 		 * those as well to make room for our guard pages.
363 		 */
364 		if (i915_vm_has_cache_coloring(vm)) {
365 			if (node->start + node->size == target->start) {
366 				if (node->color == target->color)
367 					continue;
368 			}
369 			if (node->start == target->start + target->size) {
370 				if (node->color == target->color)
371 					continue;
372 			}
373 		}
374 
375 		if (i915_vma_is_pinned(vma)) {
376 			ret = -ENOSPC;
377 			break;
378 		}
379 
380 		if (flags & PIN_NONBLOCK && i915_vma_is_active(vma)) {
381 			ret = -ENOSPC;
382 			break;
383 		}
384 
385 		if (!grab_vma(vma, ww)) {
386 			ret = -ENOSPC;
387 			break;
388 		}
389 
390 		/*
391 		 * Never show fear in the face of dragons!
392 		 *
393 		 * We cannot directly remove this node from within this
394 		 * iterator and as with i915_gem_evict_something() we employ
395 		 * the vma pin_count in order to prevent the action of
396 		 * unbinding one vma from freeing (by dropping its active
397 		 * reference) another in our eviction list.
398 		 */
399 		__i915_vma_pin(vma);
400 		list_add(&vma->evict_link, &eviction_list);
401 	}
402 
403 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
404 		__i915_vma_unpin(vma);
405 		if (ret == 0)
406 			ret = __i915_vma_unbind(vma);
407 
408 		ungrab_vma(vma);
409 	}
410 
411 	return ret;
412 }
413 
414 /**
415  * i915_gem_evict_vm - Evict all idle vmas from a vm
416  * @vm: Address space to cleanse
417  * @ww: An optional struct i915_gem_ww_ctx. If not NULL, i915_gem_evict_vm
418  * will be able to evict vma's locked by the ww as well.
419  *
420  * This function evicts all vmas from a vm.
421  *
422  * This is used by the execbuf code as a last-ditch effort to defragment the
423  * address space.
424  *
425  * To clarify: This is for freeing up virtual address space, not for freeing
426  * memory in e.g. the shrinker.
427  */
428 int i915_gem_evict_vm(struct i915_address_space *vm, struct i915_gem_ww_ctx *ww)
429 {
430 	int ret = 0;
431 
432 	lockdep_assert_held(&vm->mutex);
433 	trace_i915_gem_evict_vm(vm);
434 
435 	/* Switch back to the default context in order to unpin
436 	 * the existing context objects. However, such objects only
437 	 * pin themselves inside the global GTT and performing the
438 	 * switch otherwise is ineffective.
439 	 */
440 	if (i915_is_ggtt(vm)) {
441 		ret = ggtt_flush(vm->gt);
442 		if (ret)
443 			return ret;
444 	}
445 
446 	do {
447 		struct i915_vma *vma, *vn;
448 		LIST_HEAD(eviction_list);
449 		LIST_HEAD(locked_eviction_list);
450 
451 		list_for_each_entry(vma, &vm->bound_list, vm_link) {
452 			if (i915_vma_is_pinned(vma))
453 				continue;
454 
455 			/*
456 			 * If we already own the lock, trylock fails. In case
457 			 * the resv is shared among multiple objects, we still
458 			 * need the object ref.
459 			 */
460 			if (dying_vma(vma) ||
461 			    (ww && (dma_resv_locking_ctx(vma->obj->base.resv) == &ww->ctx))) {
462 				__i915_vma_pin(vma);
463 				list_add(&vma->evict_link, &locked_eviction_list);
464 				continue;
465 			}
466 
467 			if (!i915_gem_object_trylock(vma->obj, ww))
468 				continue;
469 
470 			__i915_vma_pin(vma);
471 			list_add(&vma->evict_link, &eviction_list);
472 		}
473 		if (list_empty(&eviction_list) && list_empty(&locked_eviction_list))
474 			break;
475 
476 		ret = 0;
477 		/* Unbind locked objects first, before unlocking the eviction_list */
478 		list_for_each_entry_safe(vma, vn, &locked_eviction_list, evict_link) {
479 			__i915_vma_unpin(vma);
480 
481 			if (ret == 0)
482 				ret = __i915_vma_unbind(vma);
483 			if (ret != -EINTR) /* "Get me out of here!" */
484 				ret = 0;
485 		}
486 
487 		list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) {
488 			__i915_vma_unpin(vma);
489 			if (ret == 0)
490 				ret = __i915_vma_unbind(vma);
491 			if (ret != -EINTR) /* "Get me out of here!" */
492 				ret = 0;
493 
494 			i915_gem_object_unlock(vma->obj);
495 		}
496 	} while (ret == 0);
497 
498 	return ret;
499 }
500 
501 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
502 #include "selftests/i915_gem_evict.c"
503 #endif
504