xref: /linux/drivers/gpu/drm/i915/i915_gem_evict.c (revision bdd1a21b52557ea8f61d0a5dc2f77151b576eb70)
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_trace.h"
35 
36 I915_SELFTEST_DECLARE(static struct igt_evict_ctl {
37 	bool fail_if_busy:1;
38 } igt_evict_ctl;)
39 
40 static int ggtt_flush(struct intel_gt *gt)
41 {
42 	/*
43 	 * Not everything in the GGTT is tracked via vma (otherwise we
44 	 * could evict as required with minimal stalling) so we are forced
45 	 * to idle the GPU and explicitly retire outstanding requests in
46 	 * the hopes that we can then remove contexts and the like only
47 	 * bound by their active reference.
48 	 */
49 	return intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
50 }
51 
52 static bool
53 mark_free(struct drm_mm_scan *scan,
54 	  struct i915_vma *vma,
55 	  unsigned int flags,
56 	  struct list_head *unwind)
57 {
58 	if (i915_vma_is_pinned(vma))
59 		return false;
60 
61 	list_add(&vma->evict_link, unwind);
62 	return drm_mm_scan_add_block(scan, &vma->node);
63 }
64 
65 static bool defer_evict(struct i915_vma *vma)
66 {
67 	if (i915_vma_is_active(vma))
68 		return true;
69 
70 	if (i915_vma_is_scanout(vma))
71 		return true;
72 
73 	return false;
74 }
75 
76 /**
77  * i915_gem_evict_something - Evict vmas to make room for binding a new one
78  * @vm: address space to evict from
79  * @min_size: size of the desired free space
80  * @alignment: alignment constraint of the desired free space
81  * @color: color for the desired space
82  * @start: start (inclusive) of the range from which to evict objects
83  * @end: end (exclusive) of the range from which to evict objects
84  * @flags: additional flags to control the eviction algorithm
85  *
86  * This function will try to evict vmas until a free space satisfying the
87  * requirements is found. Callers must check first whether any such hole exists
88  * already before calling this function.
89  *
90  * This function is used by the object/vma binding code.
91  *
92  * Since this function is only used to free up virtual address space it only
93  * ignores pinned vmas, and not object where the backing storage itself is
94  * pinned. Hence obj->pages_pin_count does not protect against eviction.
95  *
96  * To clarify: This is for freeing up virtual address space, not for freeing
97  * memory in e.g. the shrinker.
98  */
99 int
100 i915_gem_evict_something(struct i915_address_space *vm,
101 			 u64 min_size, u64 alignment,
102 			 unsigned long color,
103 			 u64 start, u64 end,
104 			 unsigned flags)
105 {
106 	struct drm_mm_scan scan;
107 	struct list_head eviction_list;
108 	struct i915_vma *vma, *next;
109 	struct drm_mm_node *node;
110 	enum drm_mm_insert_mode mode;
111 	struct i915_vma *active;
112 	int ret;
113 
114 	lockdep_assert_held(&vm->mutex);
115 	trace_i915_gem_evict(vm, min_size, alignment, flags);
116 
117 	/*
118 	 * The goal is to evict objects and amalgamate space in rough LRU order.
119 	 * Since both active and inactive objects reside on the same list,
120 	 * in a mix of creation and last scanned order, as we process the list
121 	 * we sort it into inactive/active, which keeps the active portion
122 	 * in a rough MRU order.
123 	 *
124 	 * The retirement sequence is thus:
125 	 *   1. Inactive objects (already retired, random order)
126 	 *   2. Active objects (will stall on unbinding, oldest scanned first)
127 	 */
128 	mode = DRM_MM_INSERT_BEST;
129 	if (flags & PIN_HIGH)
130 		mode = DRM_MM_INSERT_HIGH;
131 	if (flags & PIN_MAPPABLE)
132 		mode = DRM_MM_INSERT_LOW;
133 	drm_mm_scan_init_with_range(&scan, &vm->mm,
134 				    min_size, alignment, color,
135 				    start, end, mode);
136 
137 	intel_gt_retire_requests(vm->gt);
138 
139 search_again:
140 	active = NULL;
141 	INIT_LIST_HEAD(&eviction_list);
142 	list_for_each_entry_safe(vma, next, &vm->bound_list, vm_link) {
143 		if (vma == active) { /* now seen this vma twice */
144 			if (flags & PIN_NONBLOCK)
145 				break;
146 
147 			active = ERR_PTR(-EAGAIN);
148 		}
149 
150 		/*
151 		 * We keep this list in a rough least-recently scanned order
152 		 * of active elements (inactive elements are cheap to reap).
153 		 * New entries are added to the end, and we move anything we
154 		 * scan to the end. The assumption is that the working set
155 		 * of applications is either steady state (and thanks to the
156 		 * userspace bo cache it almost always is) or volatile and
157 		 * frequently replaced after a frame, which are self-evicting!
158 		 * Given that assumption, the MRU order of the scan list is
159 		 * fairly static, and keeping it in least-recently scan order
160 		 * is suitable.
161 		 *
162 		 * To notice when we complete one full cycle, we record the
163 		 * first active element seen, before moving it to the tail.
164 		 */
165 		if (active != ERR_PTR(-EAGAIN) && defer_evict(vma)) {
166 			if (!active)
167 				active = vma;
168 
169 			list_move_tail(&vma->vm_link, &vm->bound_list);
170 			continue;
171 		}
172 
173 		if (mark_free(&scan, vma, flags, &eviction_list))
174 			goto found;
175 	}
176 
177 	/* Nothing found, clean up and bail out! */
178 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
179 		ret = drm_mm_scan_remove_block(&scan, &vma->node);
180 		BUG_ON(ret);
181 	}
182 
183 	/*
184 	 * Can we unpin some objects such as idle hw contents,
185 	 * or pending flips? But since only the GGTT has global entries
186 	 * such as scanouts, rinbuffers and contexts, we can skip the
187 	 * purge when inspecting per-process local address spaces.
188 	 */
189 	if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK)
190 		return -ENOSPC;
191 
192 	/*
193 	 * Not everything in the GGTT is tracked via VMA using
194 	 * i915_vma_move_to_active(), otherwise we could evict as required
195 	 * with minimal stalling. Instead we are forced to idle the GPU and
196 	 * explicitly retire outstanding requests which will then remove
197 	 * the pinning for active objects such as contexts and ring,
198 	 * enabling us to evict them on the next iteration.
199 	 *
200 	 * To ensure that all user contexts are evictable, we perform
201 	 * a switch to the perma-pinned kernel context. This all also gives
202 	 * us a termination condition, when the last retired context is
203 	 * the kernel's there is no more we can evict.
204 	 */
205 	if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy))
206 		return -EBUSY;
207 
208 	ret = ggtt_flush(vm->gt);
209 	if (ret)
210 		return ret;
211 
212 	cond_resched();
213 
214 	flags |= PIN_NONBLOCK;
215 	goto search_again;
216 
217 found:
218 	/* drm_mm doesn't allow any other other operations while
219 	 * scanning, therefore store to-be-evicted objects on a
220 	 * temporary list and take a reference for all before
221 	 * calling unbind (which may remove the active reference
222 	 * of any of our objects, thus corrupting the list).
223 	 */
224 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
225 		if (drm_mm_scan_remove_block(&scan, &vma->node))
226 			__i915_vma_pin(vma);
227 		else
228 			list_del(&vma->evict_link);
229 	}
230 
231 	/* Unbinding will emit any required flushes */
232 	ret = 0;
233 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
234 		__i915_vma_unpin(vma);
235 		if (ret == 0)
236 			ret = __i915_vma_unbind(vma);
237 	}
238 
239 	while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) {
240 		vma = container_of(node, struct i915_vma, node);
241 
242 		/* If we find any non-objects (!vma), we cannot evict them */
243 		if (vma->node.color != I915_COLOR_UNEVICTABLE)
244 			ret = __i915_vma_unbind(vma);
245 		else
246 			ret = -ENOSPC; /* XXX search failed, try again? */
247 	}
248 
249 	return ret;
250 }
251 
252 /**
253  * i915_gem_evict_for_node - Evict vmas to make room for binding a new one
254  * @vm: address space to evict from
255  * @target: range (and color) to evict for
256  * @flags: additional flags to control the eviction algorithm
257  *
258  * This function will try to evict vmas that overlap the target node.
259  *
260  * To clarify: This is for freeing up virtual address space, not for freeing
261  * memory in e.g. the shrinker.
262  */
263 int i915_gem_evict_for_node(struct i915_address_space *vm,
264 			    struct drm_mm_node *target,
265 			    unsigned int flags)
266 {
267 	LIST_HEAD(eviction_list);
268 	struct drm_mm_node *node;
269 	u64 start = target->start;
270 	u64 end = start + target->size;
271 	struct i915_vma *vma, *next;
272 	int ret = 0;
273 
274 	lockdep_assert_held(&vm->mutex);
275 	GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
276 	GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
277 
278 	trace_i915_gem_evict_node(vm, target, flags);
279 
280 	/*
281 	 * Retire before we search the active list. Although we have
282 	 * reasonable accuracy in our retirement lists, we may have
283 	 * a stray pin (preventing eviction) that can only be resolved by
284 	 * retiring.
285 	 */
286 	intel_gt_retire_requests(vm->gt);
287 
288 	if (i915_vm_has_cache_coloring(vm)) {
289 		/* Expand search to cover neighbouring guard pages (or lack!) */
290 		if (start)
291 			start -= I915_GTT_PAGE_SIZE;
292 
293 		/* Always look at the page afterwards to avoid the end-of-GTT */
294 		end += I915_GTT_PAGE_SIZE;
295 	}
296 	GEM_BUG_ON(start >= end);
297 
298 	drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
299 		/* If we find any non-objects (!vma), we cannot evict them */
300 		if (node->color == I915_COLOR_UNEVICTABLE) {
301 			ret = -ENOSPC;
302 			break;
303 		}
304 
305 		GEM_BUG_ON(!drm_mm_node_allocated(node));
306 		vma = container_of(node, typeof(*vma), node);
307 
308 		/*
309 		 * If we are using coloring to insert guard pages between
310 		 * different cache domains within the address space, we have
311 		 * to check whether the objects on either side of our range
312 		 * abutt and conflict. If they are in conflict, then we evict
313 		 * those as well to make room for our guard pages.
314 		 */
315 		if (i915_vm_has_cache_coloring(vm)) {
316 			if (node->start + node->size == target->start) {
317 				if (node->color == target->color)
318 					continue;
319 			}
320 			if (node->start == target->start + target->size) {
321 				if (node->color == target->color)
322 					continue;
323 			}
324 		}
325 
326 		if (i915_vma_is_pinned(vma)) {
327 			ret = -ENOSPC;
328 			break;
329 		}
330 
331 		if (flags & PIN_NONBLOCK && i915_vma_is_active(vma)) {
332 			ret = -ENOSPC;
333 			break;
334 		}
335 
336 		/*
337 		 * Never show fear in the face of dragons!
338 		 *
339 		 * We cannot directly remove this node from within this
340 		 * iterator and as with i915_gem_evict_something() we employ
341 		 * the vma pin_count in order to prevent the action of
342 		 * unbinding one vma from freeing (by dropping its active
343 		 * reference) another in our eviction list.
344 		 */
345 		__i915_vma_pin(vma);
346 		list_add(&vma->evict_link, &eviction_list);
347 	}
348 
349 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
350 		__i915_vma_unpin(vma);
351 		if (ret == 0)
352 			ret = __i915_vma_unbind(vma);
353 	}
354 
355 	return ret;
356 }
357 
358 /**
359  * i915_gem_evict_vm - Evict all idle vmas from a vm
360  * @vm: Address space to cleanse
361  *
362  * This function evicts all vmas from a vm.
363  *
364  * This is used by the execbuf code as a last-ditch effort to defragment the
365  * address space.
366  *
367  * To clarify: This is for freeing up virtual address space, not for freeing
368  * memory in e.g. the shrinker.
369  */
370 int i915_gem_evict_vm(struct i915_address_space *vm)
371 {
372 	int ret = 0;
373 
374 	lockdep_assert_held(&vm->mutex);
375 	trace_i915_gem_evict_vm(vm);
376 
377 	/* Switch back to the default context in order to unpin
378 	 * the existing context objects. However, such objects only
379 	 * pin themselves inside the global GTT and performing the
380 	 * switch otherwise is ineffective.
381 	 */
382 	if (i915_is_ggtt(vm)) {
383 		ret = ggtt_flush(vm->gt);
384 		if (ret)
385 			return ret;
386 	}
387 
388 	do {
389 		struct i915_vma *vma, *vn;
390 		LIST_HEAD(eviction_list);
391 
392 		list_for_each_entry(vma, &vm->bound_list, vm_link) {
393 			if (i915_vma_is_pinned(vma))
394 				continue;
395 
396 			__i915_vma_pin(vma);
397 			list_add(&vma->evict_link, &eviction_list);
398 		}
399 		if (list_empty(&eviction_list))
400 			break;
401 
402 		ret = 0;
403 		list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) {
404 			__i915_vma_unpin(vma);
405 			if (ret == 0)
406 				ret = __i915_vma_unbind(vma);
407 			if (ret != -EINTR) /* "Get me out of here!" */
408 				ret = 0;
409 		}
410 	} while (ret == 0);
411 
412 	return ret;
413 }
414 
415 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
416 #include "selftests/i915_gem_evict.c"
417 #endif
418