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/drmP.h>
30 #include <drm/i915_drm.h>
31
32 #include "i915_drv.h"
33 #include "intel_drv.h"
34 #include "i915_trace.h"
35
ggtt_is_idle(struct drm_i915_private * i915)36 static bool ggtt_is_idle(struct drm_i915_private *i915)
37 {
38 struct intel_engine_cs *engine;
39 enum intel_engine_id id;
40
41 if (i915->gt.active_requests)
42 return false;
43
44 for_each_engine(engine, i915, id) {
45 if (engine->last_retired_context != i915->kernel_context)
46 return false;
47 }
48
49 return true;
50 }
51
ggtt_flush(struct drm_i915_private * i915)52 static int ggtt_flush(struct drm_i915_private *i915)
53 {
54 int err;
55
56 /* Not everything in the GGTT is tracked via vma (otherwise we
57 * could evict as required with minimal stalling) so we are forced
58 * to idle the GPU and explicitly retire outstanding requests in
59 * the hopes that we can then remove contexts and the like only
60 * bound by their active reference.
61 */
62 err = i915_gem_switch_to_kernel_context(i915);
63 if (err)
64 return err;
65
66 err = i915_gem_wait_for_idle(i915,
67 I915_WAIT_INTERRUPTIBLE |
68 I915_WAIT_LOCKED);
69 if (err)
70 return err;
71
72 return 0;
73 }
74
75 static bool
mark_free(struct drm_mm_scan * scan,struct i915_vma * vma,unsigned int flags,struct list_head * unwind)76 mark_free(struct drm_mm_scan *scan,
77 struct i915_vma *vma,
78 unsigned int flags,
79 struct list_head *unwind)
80 {
81 if (i915_vma_is_pinned(vma))
82 return false;
83
84 if (flags & PIN_NONFAULT && !list_empty(&vma->obj->userfault_link))
85 return false;
86
87 list_add(&vma->evict_link, unwind);
88 return drm_mm_scan_add_block(scan, &vma->node);
89 }
90
91 /**
92 * i915_gem_evict_something - Evict vmas to make room for binding a new one
93 * @vm: address space to evict from
94 * @min_size: size of the desired free space
95 * @alignment: alignment constraint of the desired free space
96 * @cache_level: cache_level for the desired space
97 * @start: start (inclusive) of the range from which to evict objects
98 * @end: end (exclusive) of the range from which to evict objects
99 * @flags: additional flags to control the eviction algorithm
100 *
101 * This function will try to evict vmas until a free space satisfying the
102 * requirements is found. Callers must check first whether any such hole exists
103 * already before calling this function.
104 *
105 * This function is used by the object/vma binding code.
106 *
107 * Since this function is only used to free up virtual address space it only
108 * ignores pinned vmas, and not object where the backing storage itself is
109 * pinned. Hence obj->pages_pin_count does not protect against eviction.
110 *
111 * To clarify: This is for freeing up virtual address space, not for freeing
112 * memory in e.g. the shrinker.
113 */
114 int
i915_gem_evict_something(struct i915_address_space * vm,u64 min_size,u64 alignment,unsigned cache_level,u64 start,u64 end,unsigned flags)115 i915_gem_evict_something(struct i915_address_space *vm,
116 u64 min_size, u64 alignment,
117 unsigned cache_level,
118 u64 start, u64 end,
119 unsigned flags)
120 {
121 struct drm_i915_private *dev_priv = vm->i915;
122 struct drm_mm_scan scan;
123 struct list_head eviction_list;
124 struct list_head *phases[] = {
125 &vm->inactive_list,
126 &vm->active_list,
127 NULL,
128 }, **phase;
129 struct i915_vma *vma, *next;
130 struct drm_mm_node *node;
131 enum drm_mm_insert_mode mode;
132 int ret;
133
134 lockdep_assert_held(&vm->i915->drm.struct_mutex);
135 trace_i915_gem_evict(vm, min_size, alignment, flags);
136
137 /*
138 * The goal is to evict objects and amalgamate space in LRU order.
139 * The oldest idle objects reside on the inactive list, which is in
140 * retirement order. The next objects to retire are those in flight,
141 * on the active list, again in retirement order.
142 *
143 * The retirement sequence is thus:
144 * 1. Inactive objects (already retired)
145 * 2. Active objects (will stall on unbinding)
146 *
147 * On each list, the oldest objects lie at the HEAD with the freshest
148 * object on the TAIL.
149 */
150 mode = DRM_MM_INSERT_BEST;
151 if (flags & PIN_HIGH)
152 mode = DRM_MM_INSERT_HIGH;
153 if (flags & PIN_MAPPABLE)
154 mode = DRM_MM_INSERT_LOW;
155 drm_mm_scan_init_with_range(&scan, &vm->mm,
156 min_size, alignment, cache_level,
157 start, end, mode);
158
159 /*
160 * Retire before we search the active list. Although we have
161 * reasonable accuracy in our retirement lists, we may have
162 * a stray pin (preventing eviction) that can only be resolved by
163 * retiring.
164 */
165 if (!(flags & PIN_NONBLOCK))
166 i915_gem_retire_requests(dev_priv);
167 else
168 phases[1] = NULL;
169
170 search_again:
171 INIT_LIST_HEAD(&eviction_list);
172 phase = phases;
173 do {
174 list_for_each_entry(vma, *phase, vm_link)
175 if (mark_free(&scan, vma, flags, &eviction_list))
176 goto found;
177 } while (*++phase);
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 (!ggtt_is_idle(dev_priv)) {
208 ret = ggtt_flush(dev_priv);
209 if (ret)
210 return ret;
211
212 goto search_again;
213 }
214
215 /*
216 * If we still have pending pageflip completions, drop
217 * back to userspace to give our workqueues time to
218 * acquire our locks and unpin the old scanouts.
219 */
220 return intel_has_pending_fb_unpin(dev_priv) ? -EAGAIN : -ENOSPC;
221
222 found:
223 /* drm_mm doesn't allow any other other operations while
224 * scanning, therefore store to-be-evicted objects on a
225 * temporary list and take a reference for all before
226 * calling unbind (which may remove the active reference
227 * of any of our objects, thus corrupting the list).
228 */
229 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
230 if (drm_mm_scan_remove_block(&scan, &vma->node))
231 __i915_vma_pin(vma);
232 else
233 list_del(&vma->evict_link);
234 }
235
236 /* Unbinding will emit any required flushes */
237 ret = 0;
238 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
239 __i915_vma_unpin(vma);
240 if (ret == 0)
241 ret = i915_vma_unbind(vma);
242 }
243
244 while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) {
245 vma = container_of(node, struct i915_vma, node);
246 ret = i915_vma_unbind(vma);
247 }
248
249 return ret;
250 }
251
252 /**
253 * i915_gem_evict_for_vma - 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 */
i915_gem_evict_for_node(struct i915_address_space * vm,struct drm_mm_node * target,unsigned int flags)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 bool check_color;
273 int ret = 0;
274
275 lockdep_assert_held(&vm->i915->drm.struct_mutex);
276 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
277 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
278
279 trace_i915_gem_evict_node(vm, target, flags);
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 if (!(flags & PIN_NONBLOCK))
287 i915_gem_retire_requests(vm->i915);
288
289 check_color = vm->mm.color_adjust;
290 if (check_color) {
291 /* Expand search to cover neighbouring guard pages (or lack!) */
292 if (start)
293 start -= I915_GTT_PAGE_SIZE;
294
295 /* Always look at the page afterwards to avoid the end-of-GTT */
296 end += I915_GTT_PAGE_SIZE;
297 }
298 GEM_BUG_ON(start >= end);
299
300 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
301 /* If we find any non-objects (!vma), we cannot evict them */
302 if (node->color == I915_COLOR_UNEVICTABLE) {
303 ret = -ENOSPC;
304 break;
305 }
306
307 GEM_BUG_ON(!node->allocated);
308 vma = container_of(node, typeof(*vma), node);
309
310 /* If we are using coloring to insert guard pages between
311 * different cache domains within the address space, we have
312 * to check whether the objects on either side of our range
313 * abutt and conflict. If they are in conflict, then we evict
314 * those as well to make room for our guard pages.
315 */
316 if (check_color) {
317 if (node->start + node->size == target->start) {
318 if (node->color == target->color)
319 continue;
320 }
321 if (node->start == target->start + target->size) {
322 if (node->color == target->color)
323 continue;
324 }
325 }
326
327 if (flags & PIN_NONBLOCK &&
328 (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))) {
329 ret = -ENOSPC;
330 break;
331 }
332
333 /* Overlap of objects in the same batch? */
334 if (i915_vma_is_pinned(vma)) {
335 ret = -ENOSPC;
336 if (vma->exec_flags &&
337 *vma->exec_flags & EXEC_OBJECT_PINNED)
338 ret = -EINVAL;
339 break;
340 }
341
342 /* Never show fear in the face of dragons!
343 *
344 * We cannot directly remove this node from within this
345 * iterator and as with i915_gem_evict_something() we employ
346 * the vma pin_count in order to prevent the action of
347 * unbinding one vma from freeing (by dropping its active
348 * reference) another in our eviction list.
349 */
350 __i915_vma_pin(vma);
351 list_add(&vma->evict_link, &eviction_list);
352 }
353
354 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
355 __i915_vma_unpin(vma);
356 if (ret == 0)
357 ret = i915_vma_unbind(vma);
358 }
359
360 return ret;
361 }
362
363 /**
364 * i915_gem_evict_vm - Evict all idle vmas from a vm
365 * @vm: Address space to cleanse
366 *
367 * This function evicts all vmas from a vm.
368 *
369 * This is used by the execbuf code as a last-ditch effort to defragment the
370 * address space.
371 *
372 * To clarify: This is for freeing up virtual address space, not for freeing
373 * memory in e.g. the shrinker.
374 */
i915_gem_evict_vm(struct i915_address_space * vm)375 int i915_gem_evict_vm(struct i915_address_space *vm)
376 {
377 struct list_head *phases[] = {
378 &vm->inactive_list,
379 &vm->active_list,
380 NULL
381 }, **phase;
382 struct list_head eviction_list;
383 struct i915_vma *vma, *next;
384 int ret;
385
386 lockdep_assert_held(&vm->i915->drm.struct_mutex);
387 trace_i915_gem_evict_vm(vm);
388
389 /* Switch back to the default context in order to unpin
390 * the existing context objects. However, such objects only
391 * pin themselves inside the global GTT and performing the
392 * switch otherwise is ineffective.
393 */
394 if (i915_is_ggtt(vm)) {
395 ret = ggtt_flush(vm->i915);
396 if (ret)
397 return ret;
398 }
399
400 INIT_LIST_HEAD(&eviction_list);
401 phase = phases;
402 do {
403 list_for_each_entry(vma, *phase, vm_link) {
404 if (i915_vma_is_pinned(vma))
405 continue;
406
407 __i915_vma_pin(vma);
408 list_add(&vma->evict_link, &eviction_list);
409 }
410 } while (*++phase);
411
412 ret = 0;
413 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
414 __i915_vma_unpin(vma);
415 if (ret == 0)
416 ret = i915_vma_unbind(vma);
417 }
418 return ret;
419 }
420
421 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
422 #include "selftests/i915_gem_evict.c"
423 #endif
424