1 /* $NetBSD: i915_gem_evict.c,v 1.6 2021/12/19 12:25:37 riastradh Exp $ */
2
3 /*
4 * Copyright © 2008-2010 Intel Corporation
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
22 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
23 * IN THE SOFTWARE.
24 *
25 * Authors:
26 * Eric Anholt <eric@anholt.net>
27 * Chris Wilson <chris@chris-wilson.co.uuk>
28 *
29 */
30
31 #include <sys/cdefs.h>
32 __KERNEL_RCSID(0, "$NetBSD: i915_gem_evict.c,v 1.6 2021/12/19 12:25:37 riastradh Exp $");
33
34 #include <drm/i915_drm.h>
35
36 #include "gem/i915_gem_context.h"
37 #include "gt/intel_gt_requests.h"
38
39 #include "i915_drv.h"
40 #include "i915_trace.h"
41
42 #include <linux/nbsd-namespace.h>
43
I915_SELFTEST_DECLARE(static struct igt_evict_ctl{ bool fail_if_busy:1; } igt_evict_ctl;)44 I915_SELFTEST_DECLARE(static struct igt_evict_ctl {
45 bool fail_if_busy:1;
46 } igt_evict_ctl;)
47
48 static int ggtt_flush(struct intel_gt *gt)
49 {
50 /*
51 * Not everything in the GGTT is tracked via vma (otherwise we
52 * could evict as required with minimal stalling) so we are forced
53 * to idle the GPU and explicitly retire outstanding requests in
54 * the hopes that we can then remove contexts and the like only
55 * bound by their active reference.
56 */
57 return intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
58 }
59
60 static bool
mark_free(struct drm_mm_scan * scan,struct i915_vma * vma,unsigned int flags,struct list_head * unwind)61 mark_free(struct drm_mm_scan *scan,
62 struct i915_vma *vma,
63 unsigned int flags,
64 struct list_head *unwind)
65 {
66 if (i915_vma_is_pinned(vma))
67 return false;
68
69 list_add(&vma->evict_link, unwind);
70 return drm_mm_scan_add_block(scan, &vma->node);
71 }
72
73 /**
74 * i915_gem_evict_something - Evict vmas to make room for binding a new one
75 * @vm: address space to evict from
76 * @min_size: size of the desired free space
77 * @alignment: alignment constraint of the desired free space
78 * @color: color for the desired space
79 * @start: start (inclusive) of the range from which to evict objects
80 * @end: end (exclusive) of the range from which to evict objects
81 * @flags: additional flags to control the eviction algorithm
82 *
83 * This function will try to evict vmas until a free space satisfying the
84 * requirements is found. Callers must check first whether any such hole exists
85 * already before calling this function.
86 *
87 * This function is used by the object/vma binding code.
88 *
89 * Since this function is only used to free up virtual address space it only
90 * ignores pinned vmas, and not object where the backing storage itself is
91 * pinned. Hence obj->pages_pin_count does not protect against eviction.
92 *
93 * To clarify: This is for freeing up virtual address space, not for freeing
94 * memory in e.g. the shrinker.
95 */
96 int
i915_gem_evict_something(struct i915_address_space * vm,u64 min_size,u64 alignment,unsigned long color,u64 start,u64 end,unsigned flags)97 i915_gem_evict_something(struct i915_address_space *vm,
98 u64 min_size, u64 alignment,
99 unsigned long color,
100 u64 start, u64 end,
101 unsigned flags)
102 {
103 struct drm_mm_scan scan;
104 struct list_head eviction_list;
105 struct i915_vma *vma, *next;
106 struct drm_mm_node *node;
107 enum drm_mm_insert_mode mode;
108 struct i915_vma *active;
109 int ret;
110
111 lockdep_assert_held(&vm->mutex);
112 trace_i915_gem_evict(vm, min_size, alignment, flags);
113
114 /*
115 * The goal is to evict objects and amalgamate space in rough LRU order.
116 * Since both active and inactive objects reside on the same list,
117 * in a mix of creation and last scanned order, as we process the list
118 * we sort it into inactive/active, which keeps the active portion
119 * in a rough MRU order.
120 *
121 * The retirement sequence is thus:
122 * 1. Inactive objects (already retired, random order)
123 * 2. Active objects (will stall on unbinding, oldest scanned first)
124 */
125 mode = DRM_MM_INSERT_BEST;
126 if (flags & PIN_HIGH)
127 mode = DRM_MM_INSERT_HIGH;
128 if (flags & PIN_MAPPABLE)
129 mode = DRM_MM_INSERT_LOW;
130 drm_mm_scan_init_with_range(&scan, &vm->mm,
131 min_size, alignment, color,
132 start, end, mode);
133
134 intel_gt_retire_requests(vm->gt);
135
136 search_again:
137 active = NULL;
138 INIT_LIST_HEAD(&eviction_list);
139 list_for_each_entry_safe(vma, next, &vm->bound_list, vm_link) {
140 /*
141 * We keep this list in a rough least-recently scanned order
142 * of active elements (inactive elements are cheap to reap).
143 * New entries are added to the end, and we move anything we
144 * scan to the end. The assumption is that the working set
145 * of applications is either steady state (and thanks to the
146 * userspace bo cache it almost always is) or volatile and
147 * frequently replaced after a frame, which are self-evicting!
148 * Given that assumption, the MRU order of the scan list is
149 * fairly static, and keeping it in least-recently scan order
150 * is suitable.
151 *
152 * To notice when we complete one full cycle, we record the
153 * first active element seen, before moving it to the tail.
154 */
155 if (i915_vma_is_active(vma)) {
156 if (vma == active) {
157 if (flags & PIN_NONBLOCK)
158 break;
159
160 active = ERR_PTR(-EAGAIN);
161 }
162
163 if (active != ERR_PTR(-EAGAIN)) {
164 if (!active)
165 active = vma;
166
167 list_move_tail(&vma->vm_link, &vm->bound_list);
168 continue;
169 }
170 }
171
172 if (mark_free(&scan, vma, flags, &eviction_list))
173 goto found;
174 }
175
176 /* Nothing found, clean up and bail out! */
177 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
178 ret = drm_mm_scan_remove_block(&scan, &vma->node);
179 BUG_ON(ret);
180 }
181
182 /*
183 * Can we unpin some objects such as idle hw contents,
184 * or pending flips? But since only the GGTT has global entries
185 * such as scanouts, rinbuffers and contexts, we can skip the
186 * purge when inspecting per-process local address spaces.
187 */
188 if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK)
189 return -ENOSPC;
190
191 /*
192 * Not everything in the GGTT is tracked via VMA using
193 * i915_vma_move_to_active(), otherwise we could evict as required
194 * with minimal stalling. Instead we are forced to idle the GPU and
195 * explicitly retire outstanding requests which will then remove
196 * the pinning for active objects such as contexts and ring,
197 * enabling us to evict them on the next iteration.
198 *
199 * To ensure that all user contexts are evictable, we perform
200 * a switch to the perma-pinned kernel context. This all also gives
201 * us a termination condition, when the last retired context is
202 * the kernel's there is no more we can evict.
203 */
204 if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy))
205 return -EBUSY;
206
207 ret = ggtt_flush(vm->gt);
208 if (ret)
209 return ret;
210
211 cond_resched();
212
213 flags |= PIN_NONBLOCK;
214 goto search_again;
215
216 found:
217 /* drm_mm doesn't allow any other other operations while
218 * scanning, therefore store to-be-evicted objects on a
219 * temporary list and take a reference for all before
220 * calling unbind (which may remove the active reference
221 * of any of our objects, thus corrupting the list).
222 */
223 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
224 if (drm_mm_scan_remove_block(&scan, &vma->node))
225 __i915_vma_pin(vma);
226 else
227 list_del(&vma->evict_link);
228 }
229
230 /* Unbinding will emit any required flushes */
231 ret = 0;
232 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
233 __i915_vma_unpin(vma);
234 if (ret == 0)
235 ret = __i915_vma_unbind(vma);
236 }
237
238 while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) {
239 vma = container_of(node, struct i915_vma, node);
240 ret = __i915_vma_unbind(vma);
241 }
242
243 return ret;
244 }
245
246 /**
247 * i915_gem_evict_for_vma - Evict vmas to make room for binding a new one
248 * @vm: address space to evict from
249 * @target: range (and color) to evict for
250 * @flags: additional flags to control the eviction algorithm
251 *
252 * This function will try to evict vmas that overlap the target node.
253 *
254 * To clarify: This is for freeing up virtual address space, not for freeing
255 * memory in e.g. the shrinker.
256 */
i915_gem_evict_for_node(struct i915_address_space * vm,struct drm_mm_node * target,unsigned int flags)257 int i915_gem_evict_for_node(struct i915_address_space *vm,
258 struct drm_mm_node *target,
259 unsigned int flags)
260 {
261 LIST_HEAD(eviction_list);
262 struct drm_mm_node *node;
263 u64 start = target->start;
264 u64 end = start + target->size;
265 struct i915_vma *vma, *next;
266 int ret = 0;
267
268 lockdep_assert_held(&vm->mutex);
269 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
270 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
271
272 trace_i915_gem_evict_node(vm, target, flags);
273
274 /*
275 * Retire before we search the active list. Although we have
276 * reasonable accuracy in our retirement lists, we may have
277 * a stray pin (preventing eviction) that can only be resolved by
278 * retiring.
279 */
280 intel_gt_retire_requests(vm->gt);
281
282 if (i915_vm_has_cache_coloring(vm)) {
283 /* Expand search to cover neighbouring guard pages (or lack!) */
284 if (start)
285 start -= I915_GTT_PAGE_SIZE;
286
287 /* Always look at the page afterwards to avoid the end-of-GTT */
288 end += I915_GTT_PAGE_SIZE;
289 }
290 GEM_BUG_ON(start >= end);
291
292 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
293 /* If we find any non-objects (!vma), we cannot evict them */
294 if (node->color == I915_COLOR_UNEVICTABLE) {
295 ret = -ENOSPC;
296 break;
297 }
298
299 GEM_BUG_ON(!drm_mm_node_allocated(node));
300 vma = container_of(node, typeof(*vma), node);
301
302 /* If we are using coloring to insert guard pages between
303 * different cache domains within the address space, we have
304 * to check whether the objects on either side of our range
305 * abutt and conflict. If they are in conflict, then we evict
306 * those as well to make room for our guard pages.
307 */
308 if (i915_vm_has_cache_coloring(vm)) {
309 if (node->start + node->size == target->start) {
310 if (node->color == target->color)
311 continue;
312 }
313 if (node->start == target->start + target->size) {
314 if (node->color == target->color)
315 continue;
316 }
317 }
318
319 if (flags & PIN_NONBLOCK &&
320 (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))) {
321 ret = -ENOSPC;
322 break;
323 }
324
325 /* Overlap of objects in the same batch? */
326 if (i915_vma_is_pinned(vma)) {
327 ret = -ENOSPC;
328 if (vma->exec_flags &&
329 *vma->exec_flags & EXEC_OBJECT_PINNED)
330 ret = -EINVAL;
331 break;
332 }
333
334 /* Never show fear in the face of dragons!
335 *
336 * We cannot directly remove this node from within this
337 * iterator and as with i915_gem_evict_something() we employ
338 * the vma pin_count in order to prevent the action of
339 * unbinding one vma from freeing (by dropping its active
340 * reference) another in our eviction list.
341 */
342 __i915_vma_pin(vma);
343 list_add(&vma->evict_link, &eviction_list);
344 }
345
346 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
347 __i915_vma_unpin(vma);
348 if (ret == 0)
349 ret = __i915_vma_unbind(vma);
350 }
351
352 return ret;
353 }
354
355 /**
356 * i915_gem_evict_vm - Evict all idle vmas from a vm
357 * @vm: Address space to cleanse
358 *
359 * This function evicts all vmas from a vm.
360 *
361 * This is used by the execbuf code as a last-ditch effort to defragment the
362 * address space.
363 *
364 * To clarify: This is for freeing up virtual address space, not for freeing
365 * memory in e.g. the shrinker.
366 */
i915_gem_evict_vm(struct i915_address_space * vm)367 int i915_gem_evict_vm(struct i915_address_space *vm)
368 {
369 int ret = 0;
370
371 lockdep_assert_held(&vm->mutex);
372 trace_i915_gem_evict_vm(vm);
373
374 /* Switch back to the default context in order to unpin
375 * the existing context objects. However, such objects only
376 * pin themselves inside the global GTT and performing the
377 * switch otherwise is ineffective.
378 */
379 if (i915_is_ggtt(vm)) {
380 ret = ggtt_flush(vm->gt);
381 if (ret)
382 return ret;
383 }
384
385 do {
386 struct i915_vma *vma, *vn;
387 LIST_HEAD(eviction_list);
388
389 list_for_each_entry(vma, &vm->bound_list, vm_link) {
390 if (i915_vma_is_pinned(vma))
391 continue;
392
393 __i915_vma_pin(vma);
394 list_add(&vma->evict_link, &eviction_list);
395 }
396 if (list_empty(&eviction_list))
397 break;
398
399 ret = 0;
400 list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) {
401 __i915_vma_unpin(vma);
402 if (ret == 0)
403 ret = __i915_vma_unbind(vma);
404 if (ret != -EINTR) /* "Get me out of here!" */
405 ret = 0;
406 }
407 } while (ret == 0);
408
409 return ret;
410 }
411
412 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
413 #include "selftests/i915_gem_evict.c"
414 #endif
415