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 <drm/i915_drm.h>
30
31 #include "gem/i915_gem_context.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 drm_i915_private *i915)
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 i915_gem_wait_for_idle(i915,
50 I915_WAIT_INTERRUPTIBLE |
51 I915_WAIT_LOCKED,
52 MAX_SCHEDULE_TIMEOUT);
53 }
54
55 static bool
56 mark_free(struct drm_mm_scan *scan,
57 struct i915_vma *vma,
58 unsigned int flags,
59 struct list_head *unwind)
60 {
61 if (i915_vma_is_pinned(vma))
62 return false;
63
64 list_add(&vma->evict_link, unwind);
65 return drm_mm_scan_add_block(scan, &vma->node);
66 }
67
68 /**
69 * i915_gem_evict_something - Evict vmas to make room for binding a new one
70 * @vm: address space to evict from
71 * @min_size: size of the desired free space
72 * @alignment: alignment constraint of the desired free space
73 * @cache_level: cache_level for the desired space
74 * @start: start (inclusive) of the range from which to evict objects
75 * @end: end (exclusive) of the range from which to evict objects
76 * @flags: additional flags to control the eviction algorithm
77 *
78 * This function will try to evict vmas until a free space satisfying the
79 * requirements is found. Callers must check first whether any such hole exists
80 * already before calling this function.
81 *
82 * This function is used by the object/vma binding code.
83 *
84 * Since this function is only used to free up virtual address space it only
85 * ignores pinned vmas, and not object where the backing storage itself is
86 * pinned. Hence obj->pages_pin_count does not protect against eviction.
87 *
88 * To clarify: This is for freeing up virtual address space, not for freeing
89 * memory in e.g. the shrinker.
90 */
91 int
92 i915_gem_evict_something(struct i915_address_space *vm,
93 u64 min_size, u64 alignment,
94 unsigned cache_level,
95 u64 start, u64 end,
96 unsigned flags)
97 {
98 struct drm_i915_private *dev_priv = vm->i915;
99 struct drm_mm_scan scan;
100 struct list_head eviction_list;
101 struct i915_vma *vma, *next;
102 struct drm_mm_node *node;
103 enum drm_mm_insert_mode mode;
104 struct i915_vma *active;
105 int ret;
106
107 lockdep_assert_held(&vm->i915->drm.struct_mutex);
108 trace_i915_gem_evict(vm, min_size, alignment, flags);
109
110 /*
111 * The goal is to evict objects and amalgamate space in rough LRU order.
112 * Since both active and inactive objects reside on the same list,
113 * in a mix of creation and last scanned order, as we process the list
114 * we sort it into inactive/active, which keeps the active portion
115 * in a rough MRU order.
116 *
117 * The retirement sequence is thus:
118 * 1. Inactive objects (already retired, random order)
119 * 2. Active objects (will stall on unbinding, oldest scanned first)
120 */
121 mode = DRM_MM_INSERT_BEST;
122 if (flags & PIN_HIGH)
123 mode = DRM_MM_INSERT_HIGH;
124 if (flags & PIN_MAPPABLE)
125 mode = DRM_MM_INSERT_LOW;
126 drm_mm_scan_init_with_range(&scan, &vm->mm,
127 min_size, alignment, cache_level,
128 start, end, mode);
129
130 /*
131 * Retire before we search the active list. Although we have
132 * reasonable accuracy in our retirement lists, we may have
133 * a stray pin (preventing eviction) that can only be resolved by
134 * retiring.
135 */
136 if (!(flags & PIN_NONBLOCK))
137 i915_retire_requests(dev_priv);
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 /*
144 * We keep this list in a rough least-recently scanned order
145 * of active elements (inactive elements are cheap to reap).
146 * New entries are added to the end, and we move anything we
147 * scan to the end. The assumption is that the working set
148 * of applications is either steady state (and thanks to the
149 * userspace bo cache it almost always is) or volatile and
150 * frequently replaced after a frame, which are self-evicting!
151 * Given that assumption, the MRU order of the scan list is
152 * fairly static, and keeping it in least-recently scan order
153 * is suitable.
154 *
155 * To notice when we complete one full cycle, we record the
156 * first active element seen, before moving it to the tail.
157 */
158 if (i915_vma_is_active(vma)) {
159 if (vma == active) {
160 if (flags & PIN_NONBLOCK)
161 break;
162
163 active = ERR_PTR(-EAGAIN);
164 }
165
166 if (active != ERR_PTR(-EAGAIN)) {
167 if (!active)
168 active = vma;
169
170 list_move_tail(&vma->vm_link, &vm->bound_list);
171 continue;
172 }
173 }
174
175 if (mark_free(&scan, vma, flags, &eviction_list))
176 goto found;
177 }
178
179 /* Nothing found, clean up and bail out! */
180 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
181 ret = drm_mm_scan_remove_block(&scan, &vma->node);
182 BUG_ON(ret);
183 }
184
185 /*
186 * Can we unpin some objects such as idle hw contents,
187 * or pending flips? But since only the GGTT has global entries
188 * such as scanouts, rinbuffers and contexts, we can skip the
189 * purge when inspecting per-process local address spaces.
190 */
191 if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK)
192 return -ENOSPC;
193
194 /*
195 * Not everything in the GGTT is tracked via VMA using
196 * i915_vma_move_to_active(), otherwise we could evict as required
197 * with minimal stalling. Instead we are forced to idle the GPU and
198 * explicitly retire outstanding requests which will then remove
199 * the pinning for active objects such as contexts and ring,
200 * enabling us to evict them on the next iteration.
201 *
202 * To ensure that all user contexts are evictable, we perform
203 * a switch to the perma-pinned kernel context. This all also gives
204 * us a termination condition, when the last retired context is
205 * the kernel's there is no more we can evict.
206 */
207 if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy))
208 return -EBUSY;
209
210 ret = ggtt_flush(dev_priv);
211 if (ret)
212 return ret;
213
214 cond_resched();
215
216 flags |= PIN_NONBLOCK;
217 goto search_again;
218
219 found:
220 /* drm_mm doesn't allow any other other operations while
221 * scanning, therefore store to-be-evicted objects on a
222 * temporary list and take a reference for all before
223 * calling unbind (which may remove the active reference
224 * of any of our objects, thus corrupting the list).
225 */
226 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
227 if (drm_mm_scan_remove_block(&scan, &vma->node))
228 __i915_vma_pin(vma);
229 else
230 list_del(&vma->evict_link);
231 }
232
233 /* Unbinding will emit any required flushes */
234 ret = 0;
235 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
236 __i915_vma_unpin(vma);
237 if (ret == 0)
238 ret = i915_vma_unbind(vma);
239 }
240
241 while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) {
242 vma = container_of(node, struct i915_vma, node);
243 ret = i915_vma_unbind(vma);
244 }
245
246 return ret;
247 }
248
249 /**
250 * i915_gem_evict_for_vma - Evict vmas to make room for binding a new one
251 * @vm: address space to evict from
252 * @target: range (and color) to evict for
253 * @flags: additional flags to control the eviction algorithm
254 *
255 * This function will try to evict vmas that overlap the target node.
256 *
257 * To clarify: This is for freeing up virtual address space, not for freeing
258 * memory in e.g. the shrinker.
259 */
260 int i915_gem_evict_for_node(struct i915_address_space *vm,
261 struct drm_mm_node *target,
262 unsigned int flags)
263 {
264 LIST_HEAD(eviction_list);
265 struct drm_mm_node *node;
266 u64 start = target->start;
267 u64 end = start + target->size;
268 struct i915_vma *vma, *next;
269 bool check_color;
270 int ret = 0;
271
272 lockdep_assert_held(&vm->i915->drm.struct_mutex);
273 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
274 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
275
276 trace_i915_gem_evict_node(vm, target, flags);
277
278 /* Retire before we search the active list. Although we have
279 * reasonable accuracy in our retirement lists, we may have
280 * a stray pin (preventing eviction) that can only be resolved by
281 * retiring.
282 */
283 if (!(flags & PIN_NONBLOCK))
284 i915_retire_requests(vm->i915);
285
286 check_color = vm->mm.color_adjust;
287 if (check_color) {
288 /* Expand search to cover neighbouring guard pages (or lack!) */
289 if (start)
290 start -= I915_GTT_PAGE_SIZE;
291
292 /* Always look at the page afterwards to avoid the end-of-GTT */
293 end += I915_GTT_PAGE_SIZE;
294 }
295 GEM_BUG_ON(start >= end);
296
297 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
298 /* If we find any non-objects (!vma), we cannot evict them */
299 if (node->color == I915_COLOR_UNEVICTABLE) {
300 ret = -ENOSPC;
301 break;
302 }
303
304 GEM_BUG_ON(!node->allocated);
305 vma = container_of(node, typeof(*vma), node);
306
307 /* If we are using coloring to insert guard pages between
308 * different cache domains within the address space, we have
309 * to check whether the objects on either side of our range
310 * abutt and conflict. If they are in conflict, then we evict
311 * those as well to make room for our guard pages.
312 */
313 if (check_color) {
314 if (node->start + node->size == target->start) {
315 if (node->color == target->color)
316 continue;
317 }
318 if (node->start == target->start + target->size) {
319 if (node->color == target->color)
320 continue;
321 }
322 }
323
324 if (flags & PIN_NONBLOCK &&
325 (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))) {
326 ret = -ENOSPC;
327 break;
328 }
329
330 /* Overlap of objects in the same batch? */
331 if (i915_vma_is_pinned(vma)) {
332 ret = -ENOSPC;
333 if (vma->exec_flags &&
334 *vma->exec_flags & EXEC_OBJECT_PINNED)
335 ret = -EINVAL;
336 break;
337 }
338
339 /* Never show fear in the face of dragons!
340 *
341 * We cannot directly remove this node from within this
342 * iterator and as with i915_gem_evict_something() we employ
343 * the vma pin_count in order to prevent the action of
344 * unbinding one vma from freeing (by dropping its active
345 * reference) another in our eviction list.
346 */
347 __i915_vma_pin(vma);
348 list_add(&vma->evict_link, &eviction_list);
349 }
350
351 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
352 __i915_vma_unpin(vma);
353 if (ret == 0)
354 ret = i915_vma_unbind(vma);
355 }
356
357 return ret;
358 }
359
360 /**
361 * i915_gem_evict_vm - Evict all idle vmas from a vm
362 * @vm: Address space to cleanse
363 *
364 * This function evicts all vmas from a vm.
365 *
366 * This is used by the execbuf code as a last-ditch effort to defragment the
367 * address space.
368 *
369 * To clarify: This is for freeing up virtual address space, not for freeing
370 * memory in e.g. the shrinker.
371 */
372 int i915_gem_evict_vm(struct i915_address_space *vm)
373 {
374 struct list_head eviction_list;
375 struct i915_vma *vma, *next;
376 int ret;
377
378 lockdep_assert_held(&vm->i915->drm.struct_mutex);
379 trace_i915_gem_evict_vm(vm);
380
381 /* Switch back to the default context in order to unpin
382 * the existing context objects. However, such objects only
383 * pin themselves inside the global GTT and performing the
384 * switch otherwise is ineffective.
385 */
386 if (i915_is_ggtt(vm)) {
387 ret = ggtt_flush(vm->i915);
388 if (ret)
389 return ret;
390 }
391
392 INIT_LIST_HEAD(&eviction_list);
393 mutex_lock(&vm->mutex);
394 list_for_each_entry(vma, &vm->bound_list, vm_link) {
395 if (i915_vma_is_pinned(vma))
396 continue;
397
398 __i915_vma_pin(vma);
399 list_add(&vma->evict_link, &eviction_list);
400 }
401 mutex_unlock(&vm->mutex);
402
403 ret = 0;
404 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
405 __i915_vma_unpin(vma);
406 if (ret == 0)
407 ret = i915_vma_unbind(vma);
408 }
409 return ret;
410 }
411
412 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
413 #include "selftests/i915_gem_evict.c"
414 #endif