]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/mm/vmscan.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
5 | * | |
6 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
7 | * kswapd added: 7.1.96 sct | |
8 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
9 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
10 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
11 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
12 | */ | |
13 | ||
14 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
15 | ||
16 | #include <linux/mm.h> | |
17 | #include <linux/module.h> | |
18 | #include <linux/gfp.h> | |
19 | #include <linux/kernel_stat.h> | |
20 | #include <linux/swap.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/init.h> | |
23 | #include <linux/highmem.h> | |
24 | #include <linux/vmpressure.h> | |
25 | #include <linux/vmstat.h> | |
26 | #include <linux/file.h> | |
27 | #include <linux/writeback.h> | |
28 | #include <linux/blkdev.h> | |
29 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
30 | buffer_heads_over_limit */ | |
31 | #include <linux/mm_inline.h> | |
32 | #include <linux/backing-dev.h> | |
33 | #include <linux/rmap.h> | |
34 | #include <linux/topology.h> | |
35 | #include <linux/cpu.h> | |
36 | #include <linux/cpuset.h> | |
37 | #include <linux/compaction.h> | |
38 | #include <linux/notifier.h> | |
39 | #include <linux/rwsem.h> | |
40 | #include <linux/delay.h> | |
41 | #include <linux/kthread.h> | |
42 | #include <linux/freezer.h> | |
43 | #include <linux/memcontrol.h> | |
44 | #include <linux/delayacct.h> | |
45 | #include <linux/sysctl.h> | |
46 | #include <linux/oom.h> | |
47 | #include <linux/prefetch.h> | |
48 | #include <linux/printk.h> | |
49 | ||
50 | #include <asm/tlbflush.h> | |
51 | #include <asm/div64.h> | |
52 | ||
53 | #include <linux/swapops.h> | |
54 | #include <linux/balloon_compaction.h> | |
55 | ||
56 | #include "internal.h" | |
57 | ||
58 | #define CREATE_TRACE_POINTS | |
59 | #include <trace/events/vmscan.h> | |
60 | ||
61 | struct scan_control { | |
62 | /* How many pages shrink_list() should reclaim */ | |
63 | unsigned long nr_to_reclaim; | |
64 | ||
65 | /* This context's GFP mask */ | |
66 | gfp_t gfp_mask; | |
67 | ||
68 | /* Allocation order */ | |
69 | int order; | |
70 | ||
71 | /* | |
72 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
73 | * are scanned. | |
74 | */ | |
75 | nodemask_t *nodemask; | |
76 | ||
77 | /* | |
78 | * The memory cgroup that hit its limit and as a result is the | |
79 | * primary target of this reclaim invocation. | |
80 | */ | |
81 | struct mem_cgroup *target_mem_cgroup; | |
82 | ||
83 | /* Scan (total_size >> priority) pages at once */ | |
84 | int priority; | |
85 | ||
86 | unsigned int may_writepage:1; | |
87 | ||
88 | /* Can mapped pages be reclaimed? */ | |
89 | unsigned int may_unmap:1; | |
90 | ||
91 | /* Can pages be swapped as part of reclaim? */ | |
92 | unsigned int may_swap:1; | |
93 | ||
94 | /* Can cgroups be reclaimed below their normal consumption range? */ | |
95 | unsigned int may_thrash:1; | |
96 | ||
97 | unsigned int hibernation_mode:1; | |
98 | ||
99 | /* One of the zones is ready for compaction */ | |
100 | unsigned int compaction_ready:1; | |
101 | ||
102 | /* Incremented by the number of inactive pages that were scanned */ | |
103 | unsigned long nr_scanned; | |
104 | ||
105 | /* Number of pages freed so far during a call to shrink_zones() */ | |
106 | unsigned long nr_reclaimed; | |
107 | }; | |
108 | ||
109 | #ifdef ARCH_HAS_PREFETCH | |
110 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
111 | do { \ | |
112 | if ((_page)->lru.prev != _base) { \ | |
113 | struct page *prev; \ | |
114 | \ | |
115 | prev = lru_to_page(&(_page->lru)); \ | |
116 | prefetch(&prev->_field); \ | |
117 | } \ | |
118 | } while (0) | |
119 | #else | |
120 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
121 | #endif | |
122 | ||
123 | #ifdef ARCH_HAS_PREFETCHW | |
124 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
125 | do { \ | |
126 | if ((_page)->lru.prev != _base) { \ | |
127 | struct page *prev; \ | |
128 | \ | |
129 | prev = lru_to_page(&(_page->lru)); \ | |
130 | prefetchw(&prev->_field); \ | |
131 | } \ | |
132 | } while (0) | |
133 | #else | |
134 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
135 | #endif | |
136 | ||
137 | /* | |
138 | * From 0 .. 100. Higher means more swappy. | |
139 | */ | |
140 | int vm_swappiness = 60; | |
141 | /* | |
142 | * The total number of pages which are beyond the high watermark within all | |
143 | * zones. | |
144 | */ | |
145 | unsigned long vm_total_pages; | |
146 | ||
147 | static LIST_HEAD(shrinker_list); | |
148 | static DECLARE_RWSEM(shrinker_rwsem); | |
149 | ||
150 | #ifdef CONFIG_MEMCG | |
151 | static bool global_reclaim(struct scan_control *sc) | |
152 | { | |
153 | return !sc->target_mem_cgroup; | |
154 | } | |
155 | ||
156 | /** | |
157 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
158 | * @sc: scan_control in question | |
159 | * | |
160 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
161 | * completely broken with the legacy memcg and direct stalling in | |
162 | * shrink_page_list() is used for throttling instead, which lacks all the | |
163 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
164 | * allocation and configurability. | |
165 | * | |
166 | * This function tests whether the vmscan currently in progress can assume | |
167 | * that the normal dirty throttling mechanism is operational. | |
168 | */ | |
169 | static bool sane_reclaim(struct scan_control *sc) | |
170 | { | |
171 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
172 | ||
173 | if (!memcg) | |
174 | return true; | |
175 | #ifdef CONFIG_CGROUP_WRITEBACK | |
176 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
177 | return true; | |
178 | #endif | |
179 | return false; | |
180 | } | |
181 | #else | |
182 | static bool global_reclaim(struct scan_control *sc) | |
183 | { | |
184 | return true; | |
185 | } | |
186 | ||
187 | static bool sane_reclaim(struct scan_control *sc) | |
188 | { | |
189 | return true; | |
190 | } | |
191 | #endif | |
192 | ||
193 | static unsigned long zone_reclaimable_pages(struct zone *zone) | |
194 | { | |
195 | unsigned long nr; | |
196 | ||
197 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
198 | zone_page_state(zone, NR_INACTIVE_FILE) + | |
199 | zone_page_state(zone, NR_ISOLATED_FILE); | |
200 | ||
201 | if (get_nr_swap_pages() > 0) | |
202 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
203 | zone_page_state(zone, NR_INACTIVE_ANON) + | |
204 | zone_page_state(zone, NR_ISOLATED_ANON); | |
205 | ||
206 | return nr; | |
207 | } | |
208 | ||
209 | bool zone_reclaimable(struct zone *zone) | |
210 | { | |
211 | return zone_page_state(zone, NR_PAGES_SCANNED) < | |
212 | zone_reclaimable_pages(zone) * 6; | |
213 | } | |
214 | ||
215 | static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) | |
216 | { | |
217 | if (!mem_cgroup_disabled()) | |
218 | return mem_cgroup_get_lru_size(lruvec, lru); | |
219 | ||
220 | return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); | |
221 | } | |
222 | ||
223 | /* | |
224 | * Add a shrinker callback to be called from the vm. | |
225 | */ | |
226 | int register_shrinker(struct shrinker *shrinker) | |
227 | { | |
228 | size_t size = sizeof(*shrinker->nr_deferred); | |
229 | ||
230 | /* | |
231 | * If we only have one possible node in the system anyway, save | |
232 | * ourselves the trouble and disable NUMA aware behavior. This way we | |
233 | * will save memory and some small loop time later. | |
234 | */ | |
235 | if (nr_node_ids == 1) | |
236 | shrinker->flags &= ~SHRINKER_NUMA_AWARE; | |
237 | ||
238 | if (shrinker->flags & SHRINKER_NUMA_AWARE) | |
239 | size *= nr_node_ids; | |
240 | ||
241 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
242 | if (!shrinker->nr_deferred) | |
243 | return -ENOMEM; | |
244 | ||
245 | down_write(&shrinker_rwsem); | |
246 | list_add_tail(&shrinker->list, &shrinker_list); | |
247 | up_write(&shrinker_rwsem); | |
248 | return 0; | |
249 | } | |
250 | EXPORT_SYMBOL(register_shrinker); | |
251 | ||
252 | /* | |
253 | * Remove one | |
254 | */ | |
255 | void unregister_shrinker(struct shrinker *shrinker) | |
256 | { | |
257 | down_write(&shrinker_rwsem); | |
258 | list_del(&shrinker->list); | |
259 | up_write(&shrinker_rwsem); | |
260 | kfree(shrinker->nr_deferred); | |
261 | } | |
262 | EXPORT_SYMBOL(unregister_shrinker); | |
263 | ||
264 | #define SHRINK_BATCH 128 | |
265 | ||
266 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, | |
267 | struct shrinker *shrinker, | |
268 | unsigned long nr_scanned, | |
269 | unsigned long nr_eligible) | |
270 | { | |
271 | unsigned long freed = 0; | |
272 | unsigned long long delta; | |
273 | long total_scan; | |
274 | long freeable; | |
275 | long nr; | |
276 | long new_nr; | |
277 | int nid = shrinkctl->nid; | |
278 | long batch_size = shrinker->batch ? shrinker->batch | |
279 | : SHRINK_BATCH; | |
280 | ||
281 | freeable = shrinker->count_objects(shrinker, shrinkctl); | |
282 | if (freeable == 0) | |
283 | return 0; | |
284 | ||
285 | /* | |
286 | * copy the current shrinker scan count into a local variable | |
287 | * and zero it so that other concurrent shrinker invocations | |
288 | * don't also do this scanning work. | |
289 | */ | |
290 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
291 | ||
292 | total_scan = nr; | |
293 | delta = (4 * nr_scanned) / shrinker->seeks; | |
294 | delta *= freeable; | |
295 | do_div(delta, nr_eligible + 1); | |
296 | total_scan += delta; | |
297 | if (total_scan < 0) { | |
298 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", | |
299 | shrinker->scan_objects, total_scan); | |
300 | total_scan = freeable; | |
301 | } | |
302 | ||
303 | /* | |
304 | * We need to avoid excessive windup on filesystem shrinkers | |
305 | * due to large numbers of GFP_NOFS allocations causing the | |
306 | * shrinkers to return -1 all the time. This results in a large | |
307 | * nr being built up so when a shrink that can do some work | |
308 | * comes along it empties the entire cache due to nr >>> | |
309 | * freeable. This is bad for sustaining a working set in | |
310 | * memory. | |
311 | * | |
312 | * Hence only allow the shrinker to scan the entire cache when | |
313 | * a large delta change is calculated directly. | |
314 | */ | |
315 | if (delta < freeable / 4) | |
316 | total_scan = min(total_scan, freeable / 2); | |
317 | ||
318 | /* | |
319 | * Avoid risking looping forever due to too large nr value: | |
320 | * never try to free more than twice the estimate number of | |
321 | * freeable entries. | |
322 | */ | |
323 | if (total_scan > freeable * 2) | |
324 | total_scan = freeable * 2; | |
325 | ||
326 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
327 | nr_scanned, nr_eligible, | |
328 | freeable, delta, total_scan); | |
329 | ||
330 | /* | |
331 | * Normally, we should not scan less than batch_size objects in one | |
332 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
333 | * than batch_size objects in total and we are really tight on memory, | |
334 | * we will try to reclaim all available objects, otherwise we can end | |
335 | * up failing allocations although there are plenty of reclaimable | |
336 | * objects spread over several slabs with usage less than the | |
337 | * batch_size. | |
338 | * | |
339 | * We detect the "tight on memory" situations by looking at the total | |
340 | * number of objects we want to scan (total_scan). If it is greater | |
341 | * than the total number of objects on slab (freeable), we must be | |
342 | * scanning at high prio and therefore should try to reclaim as much as | |
343 | * possible. | |
344 | */ | |
345 | while (total_scan >= batch_size || | |
346 | total_scan >= freeable) { | |
347 | unsigned long ret; | |
348 | unsigned long nr_to_scan = min(batch_size, total_scan); | |
349 | ||
350 | shrinkctl->nr_to_scan = nr_to_scan; | |
351 | ret = shrinker->scan_objects(shrinker, shrinkctl); | |
352 | if (ret == SHRINK_STOP) | |
353 | break; | |
354 | freed += ret; | |
355 | ||
356 | count_vm_events(SLABS_SCANNED, nr_to_scan); | |
357 | total_scan -= nr_to_scan; | |
358 | ||
359 | cond_resched(); | |
360 | } | |
361 | ||
362 | /* | |
363 | * move the unused scan count back into the shrinker in a | |
364 | * manner that handles concurrent updates. If we exhausted the | |
365 | * scan, there is no need to do an update. | |
366 | */ | |
367 | if (total_scan > 0) | |
368 | new_nr = atomic_long_add_return(total_scan, | |
369 | &shrinker->nr_deferred[nid]); | |
370 | else | |
371 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
372 | ||
373 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); | |
374 | return freed; | |
375 | } | |
376 | ||
377 | /** | |
378 | * shrink_slab - shrink slab caches | |
379 | * @gfp_mask: allocation context | |
380 | * @nid: node whose slab caches to target | |
381 | * @memcg: memory cgroup whose slab caches to target | |
382 | * @nr_scanned: pressure numerator | |
383 | * @nr_eligible: pressure denominator | |
384 | * | |
385 | * Call the shrink functions to age shrinkable caches. | |
386 | * | |
387 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, | |
388 | * unaware shrinkers will receive a node id of 0 instead. | |
389 | * | |
390 | * @memcg specifies the memory cgroup to target. If it is not NULL, | |
391 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan | |
392 | * objects from the memory cgroup specified. Otherwise all shrinkers | |
393 | * are called, and memcg aware shrinkers are supposed to scan the | |
394 | * global list then. | |
395 | * | |
396 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of | |
397 | * the available objects should be scanned. Page reclaim for example | |
398 | * passes the number of pages scanned and the number of pages on the | |
399 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned | |
400 | * when it encountered mapped pages. The ratio is further biased by | |
401 | * the ->seeks setting of the shrink function, which indicates the | |
402 | * cost to recreate an object relative to that of an LRU page. | |
403 | * | |
404 | * Returns the number of reclaimed slab objects. | |
405 | */ | |
406 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, | |
407 | struct mem_cgroup *memcg, | |
408 | unsigned long nr_scanned, | |
409 | unsigned long nr_eligible) | |
410 | { | |
411 | struct shrinker *shrinker; | |
412 | unsigned long freed = 0; | |
413 | ||
414 | if (memcg && !memcg_kmem_is_active(memcg)) | |
415 | return 0; | |
416 | ||
417 | if (nr_scanned == 0) | |
418 | nr_scanned = SWAP_CLUSTER_MAX; | |
419 | ||
420 | if (!down_read_trylock(&shrinker_rwsem)) { | |
421 | /* | |
422 | * If we would return 0, our callers would understand that we | |
423 | * have nothing else to shrink and give up trying. By returning | |
424 | * 1 we keep it going and assume we'll be able to shrink next | |
425 | * time. | |
426 | */ | |
427 | freed = 1; | |
428 | goto out; | |
429 | } | |
430 | ||
431 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
432 | struct shrink_control sc = { | |
433 | .gfp_mask = gfp_mask, | |
434 | .nid = nid, | |
435 | .memcg = memcg, | |
436 | }; | |
437 | ||
438 | if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
439 | continue; | |
440 | ||
441 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) | |
442 | sc.nid = 0; | |
443 | ||
444 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); | |
445 | } | |
446 | ||
447 | up_read(&shrinker_rwsem); | |
448 | out: | |
449 | cond_resched(); | |
450 | return freed; | |
451 | } | |
452 | ||
453 | void drop_slab_node(int nid) | |
454 | { | |
455 | unsigned long freed; | |
456 | ||
457 | do { | |
458 | struct mem_cgroup *memcg = NULL; | |
459 | ||
460 | freed = 0; | |
461 | do { | |
462 | freed += shrink_slab(GFP_KERNEL, nid, memcg, | |
463 | 1000, 1000); | |
464 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
465 | } while (freed > 10); | |
466 | } | |
467 | ||
468 | void drop_slab(void) | |
469 | { | |
470 | int nid; | |
471 | ||
472 | for_each_online_node(nid) | |
473 | drop_slab_node(nid); | |
474 | } | |
475 | ||
476 | static inline int is_page_cache_freeable(struct page *page) | |
477 | { | |
478 | /* | |
479 | * A freeable page cache page is referenced only by the caller | |
480 | * that isolated the page, the page cache radix tree and | |
481 | * optional buffer heads at page->private. | |
482 | */ | |
483 | return page_count(page) - page_has_private(page) == 2; | |
484 | } | |
485 | ||
486 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) | |
487 | { | |
488 | if (current->flags & PF_SWAPWRITE) | |
489 | return 1; | |
490 | if (!inode_write_congested(inode)) | |
491 | return 1; | |
492 | if (inode_to_bdi(inode) == current->backing_dev_info) | |
493 | return 1; | |
494 | return 0; | |
495 | } | |
496 | ||
497 | /* | |
498 | * We detected a synchronous write error writing a page out. Probably | |
499 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
500 | * fsync(), msync() or close(). | |
501 | * | |
502 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
503 | * prevents it from being freed up. But we have a ref on the page and once | |
504 | * that page is locked, the mapping is pinned. | |
505 | * | |
506 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
507 | * __GFP_FS. | |
508 | */ | |
509 | static void handle_write_error(struct address_space *mapping, | |
510 | struct page *page, int error) | |
511 | { | |
512 | lock_page(page); | |
513 | if (page_mapping(page) == mapping) | |
514 | mapping_set_error(mapping, error); | |
515 | unlock_page(page); | |
516 | } | |
517 | ||
518 | /* possible outcome of pageout() */ | |
519 | typedef enum { | |
520 | /* failed to write page out, page is locked */ | |
521 | PAGE_KEEP, | |
522 | /* move page to the active list, page is locked */ | |
523 | PAGE_ACTIVATE, | |
524 | /* page has been sent to the disk successfully, page is unlocked */ | |
525 | PAGE_SUCCESS, | |
526 | /* page is clean and locked */ | |
527 | PAGE_CLEAN, | |
528 | } pageout_t; | |
529 | ||
530 | /* | |
531 | * pageout is called by shrink_page_list() for each dirty page. | |
532 | * Calls ->writepage(). | |
533 | */ | |
534 | static pageout_t pageout(struct page *page, struct address_space *mapping, | |
535 | struct scan_control *sc) | |
536 | { | |
537 | /* | |
538 | * If the page is dirty, only perform writeback if that write | |
539 | * will be non-blocking. To prevent this allocation from being | |
540 | * stalled by pagecache activity. But note that there may be | |
541 | * stalls if we need to run get_block(). We could test | |
542 | * PagePrivate for that. | |
543 | * | |
544 | * If this process is currently in __generic_file_write_iter() against | |
545 | * this page's queue, we can perform writeback even if that | |
546 | * will block. | |
547 | * | |
548 | * If the page is swapcache, write it back even if that would | |
549 | * block, for some throttling. This happens by accident, because | |
550 | * swap_backing_dev_info is bust: it doesn't reflect the | |
551 | * congestion state of the swapdevs. Easy to fix, if needed. | |
552 | */ | |
553 | if (!is_page_cache_freeable(page)) | |
554 | return PAGE_KEEP; | |
555 | if (!mapping) { | |
556 | /* | |
557 | * Some data journaling orphaned pages can have | |
558 | * page->mapping == NULL while being dirty with clean buffers. | |
559 | */ | |
560 | if (page_has_private(page)) { | |
561 | if (try_to_free_buffers(page)) { | |
562 | ClearPageDirty(page); | |
563 | pr_info("%s: orphaned page\n", __func__); | |
564 | return PAGE_CLEAN; | |
565 | } | |
566 | } | |
567 | return PAGE_KEEP; | |
568 | } | |
569 | if (mapping->a_ops->writepage == NULL) | |
570 | return PAGE_ACTIVATE; | |
571 | if (!may_write_to_inode(mapping->host, sc)) | |
572 | return PAGE_KEEP; | |
573 | ||
574 | if (clear_page_dirty_for_io(page)) { | |
575 | int res; | |
576 | struct writeback_control wbc = { | |
577 | .sync_mode = WB_SYNC_NONE, | |
578 | .nr_to_write = SWAP_CLUSTER_MAX, | |
579 | .range_start = 0, | |
580 | .range_end = LLONG_MAX, | |
581 | .for_reclaim = 1, | |
582 | }; | |
583 | ||
584 | SetPageReclaim(page); | |
585 | res = mapping->a_ops->writepage(page, &wbc); | |
586 | if (res < 0) | |
587 | handle_write_error(mapping, page, res); | |
588 | if (res == AOP_WRITEPAGE_ACTIVATE) { | |
589 | ClearPageReclaim(page); | |
590 | return PAGE_ACTIVATE; | |
591 | } | |
592 | ||
593 | if (!PageWriteback(page)) { | |
594 | /* synchronous write or broken a_ops? */ | |
595 | ClearPageReclaim(page); | |
596 | } | |
597 | trace_mm_vmscan_writepage(page); | |
598 | inc_zone_page_state(page, NR_VMSCAN_WRITE); | |
599 | return PAGE_SUCCESS; | |
600 | } | |
601 | ||
602 | return PAGE_CLEAN; | |
603 | } | |
604 | ||
605 | /* | |
606 | * Same as remove_mapping, but if the page is removed from the mapping, it | |
607 | * gets returned with a refcount of 0. | |
608 | */ | |
609 | static int __remove_mapping(struct address_space *mapping, struct page *page, | |
610 | bool reclaimed) | |
611 | { | |
612 | unsigned long flags; | |
613 | struct mem_cgroup *memcg; | |
614 | ||
615 | BUG_ON(!PageLocked(page)); | |
616 | BUG_ON(mapping != page_mapping(page)); | |
617 | ||
618 | memcg = mem_cgroup_begin_page_stat(page); | |
619 | spin_lock_irqsave(&mapping->tree_lock, flags); | |
620 | /* | |
621 | * The non racy check for a busy page. | |
622 | * | |
623 | * Must be careful with the order of the tests. When someone has | |
624 | * a ref to the page, it may be possible that they dirty it then | |
625 | * drop the reference. So if PageDirty is tested before page_count | |
626 | * here, then the following race may occur: | |
627 | * | |
628 | * get_user_pages(&page); | |
629 | * [user mapping goes away] | |
630 | * write_to(page); | |
631 | * !PageDirty(page) [good] | |
632 | * SetPageDirty(page); | |
633 | * put_page(page); | |
634 | * !page_count(page) [good, discard it] | |
635 | * | |
636 | * [oops, our write_to data is lost] | |
637 | * | |
638 | * Reversing the order of the tests ensures such a situation cannot | |
639 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
640 | * load is not satisfied before that of page->_count. | |
641 | * | |
642 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
643 | * and thus under tree_lock, then this ordering is not required. | |
644 | */ | |
645 | if (!page_freeze_refs(page, 2)) | |
646 | goto cannot_free; | |
647 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ | |
648 | if (unlikely(PageDirty(page))) { | |
649 | page_unfreeze_refs(page, 2); | |
650 | goto cannot_free; | |
651 | } | |
652 | ||
653 | if (PageSwapCache(page)) { | |
654 | swp_entry_t swap = { .val = page_private(page) }; | |
655 | mem_cgroup_swapout(page, swap); | |
656 | __delete_from_swap_cache(page); | |
657 | spin_unlock_irqrestore(&mapping->tree_lock, flags); | |
658 | mem_cgroup_end_page_stat(memcg); | |
659 | swapcache_free(swap); | |
660 | } else { | |
661 | void (*freepage)(struct page *); | |
662 | void *shadow = NULL; | |
663 | ||
664 | freepage = mapping->a_ops->freepage; | |
665 | /* | |
666 | * Remember a shadow entry for reclaimed file cache in | |
667 | * order to detect refaults, thus thrashing, later on. | |
668 | * | |
669 | * But don't store shadows in an address space that is | |
670 | * already exiting. This is not just an optizimation, | |
671 | * inode reclaim needs to empty out the radix tree or | |
672 | * the nodes are lost. Don't plant shadows behind its | |
673 | * back. | |
674 | */ | |
675 | if (reclaimed && page_is_file_cache(page) && | |
676 | !mapping_exiting(mapping)) | |
677 | shadow = workingset_eviction(mapping, page); | |
678 | __delete_from_page_cache(page, shadow, memcg); | |
679 | spin_unlock_irqrestore(&mapping->tree_lock, flags); | |
680 | mem_cgroup_end_page_stat(memcg); | |
681 | ||
682 | if (freepage != NULL) | |
683 | freepage(page); | |
684 | } | |
685 | ||
686 | return 1; | |
687 | ||
688 | cannot_free: | |
689 | spin_unlock_irqrestore(&mapping->tree_lock, flags); | |
690 | mem_cgroup_end_page_stat(memcg); | |
691 | return 0; | |
692 | } | |
693 | ||
694 | /* | |
695 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
696 | * someone else has a ref on the page, abort and return 0. If it was | |
697 | * successfully detached, return 1. Assumes the caller has a single ref on | |
698 | * this page. | |
699 | */ | |
700 | int remove_mapping(struct address_space *mapping, struct page *page) | |
701 | { | |
702 | if (__remove_mapping(mapping, page, false)) { | |
703 | /* | |
704 | * Unfreezing the refcount with 1 rather than 2 effectively | |
705 | * drops the pagecache ref for us without requiring another | |
706 | * atomic operation. | |
707 | */ | |
708 | page_unfreeze_refs(page, 1); | |
709 | return 1; | |
710 | } | |
711 | return 0; | |
712 | } | |
713 | ||
714 | /** | |
715 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
716 | * @page: page to be put back to appropriate lru list | |
717 | * | |
718 | * Add previously isolated @page to appropriate LRU list. | |
719 | * Page may still be unevictable for other reasons. | |
720 | * | |
721 | * lru_lock must not be held, interrupts must be enabled. | |
722 | */ | |
723 | void putback_lru_page(struct page *page) | |
724 | { | |
725 | bool is_unevictable; | |
726 | int was_unevictable = PageUnevictable(page); | |
727 | ||
728 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
729 | ||
730 | redo: | |
731 | ClearPageUnevictable(page); | |
732 | ||
733 | if (page_evictable(page)) { | |
734 | /* | |
735 | * For evictable pages, we can use the cache. | |
736 | * In event of a race, worst case is we end up with an | |
737 | * unevictable page on [in]active list. | |
738 | * We know how to handle that. | |
739 | */ | |
740 | is_unevictable = false; | |
741 | lru_cache_add(page); | |
742 | } else { | |
743 | /* | |
744 | * Put unevictable pages directly on zone's unevictable | |
745 | * list. | |
746 | */ | |
747 | is_unevictable = true; | |
748 | add_page_to_unevictable_list(page); | |
749 | /* | |
750 | * When racing with an mlock or AS_UNEVICTABLE clearing | |
751 | * (page is unlocked) make sure that if the other thread | |
752 | * does not observe our setting of PG_lru and fails | |
753 | * isolation/check_move_unevictable_pages, | |
754 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move | |
755 | * the page back to the evictable list. | |
756 | * | |
757 | * The other side is TestClearPageMlocked() or shmem_lock(). | |
758 | */ | |
759 | smp_mb(); | |
760 | } | |
761 | ||
762 | /* | |
763 | * page's status can change while we move it among lru. If an evictable | |
764 | * page is on unevictable list, it never be freed. To avoid that, | |
765 | * check after we added it to the list, again. | |
766 | */ | |
767 | if (is_unevictable && page_evictable(page)) { | |
768 | if (!isolate_lru_page(page)) { | |
769 | put_page(page); | |
770 | goto redo; | |
771 | } | |
772 | /* This means someone else dropped this page from LRU | |
773 | * So, it will be freed or putback to LRU again. There is | |
774 | * nothing to do here. | |
775 | */ | |
776 | } | |
777 | ||
778 | if (was_unevictable && !is_unevictable) | |
779 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
780 | else if (!was_unevictable && is_unevictable) | |
781 | count_vm_event(UNEVICTABLE_PGCULLED); | |
782 | ||
783 | put_page(page); /* drop ref from isolate */ | |
784 | } | |
785 | ||
786 | enum page_references { | |
787 | PAGEREF_RECLAIM, | |
788 | PAGEREF_RECLAIM_CLEAN, | |
789 | PAGEREF_KEEP, | |
790 | PAGEREF_ACTIVATE, | |
791 | }; | |
792 | ||
793 | static enum page_references page_check_references(struct page *page, | |
794 | struct scan_control *sc) | |
795 | { | |
796 | int referenced_ptes, referenced_page; | |
797 | unsigned long vm_flags; | |
798 | ||
799 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, | |
800 | &vm_flags); | |
801 | referenced_page = TestClearPageReferenced(page); | |
802 | ||
803 | /* | |
804 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
805 | * move the page to the unevictable list. | |
806 | */ | |
807 | if (vm_flags & VM_LOCKED) | |
808 | return PAGEREF_RECLAIM; | |
809 | ||
810 | if (referenced_ptes) { | |
811 | if (PageSwapBacked(page)) | |
812 | return PAGEREF_ACTIVATE; | |
813 | /* | |
814 | * All mapped pages start out with page table | |
815 | * references from the instantiating fault, so we need | |
816 | * to look twice if a mapped file page is used more | |
817 | * than once. | |
818 | * | |
819 | * Mark it and spare it for another trip around the | |
820 | * inactive list. Another page table reference will | |
821 | * lead to its activation. | |
822 | * | |
823 | * Note: the mark is set for activated pages as well | |
824 | * so that recently deactivated but used pages are | |
825 | * quickly recovered. | |
826 | */ | |
827 | SetPageReferenced(page); | |
828 | ||
829 | if (referenced_page || referenced_ptes > 1) | |
830 | return PAGEREF_ACTIVATE; | |
831 | ||
832 | /* | |
833 | * Activate file-backed executable pages after first usage. | |
834 | */ | |
835 | if (vm_flags & VM_EXEC) | |
836 | return PAGEREF_ACTIVATE; | |
837 | ||
838 | return PAGEREF_KEEP; | |
839 | } | |
840 | ||
841 | /* Reclaim if clean, defer dirty pages to writeback */ | |
842 | if (referenced_page && !PageSwapBacked(page)) | |
843 | return PAGEREF_RECLAIM_CLEAN; | |
844 | ||
845 | return PAGEREF_RECLAIM; | |
846 | } | |
847 | ||
848 | /* Check if a page is dirty or under writeback */ | |
849 | static void page_check_dirty_writeback(struct page *page, | |
850 | bool *dirty, bool *writeback) | |
851 | { | |
852 | struct address_space *mapping; | |
853 | ||
854 | /* | |
855 | * Anonymous pages are not handled by flushers and must be written | |
856 | * from reclaim context. Do not stall reclaim based on them | |
857 | */ | |
858 | if (!page_is_file_cache(page)) { | |
859 | *dirty = false; | |
860 | *writeback = false; | |
861 | return; | |
862 | } | |
863 | ||
864 | /* By default assume that the page flags are accurate */ | |
865 | *dirty = PageDirty(page); | |
866 | *writeback = PageWriteback(page); | |
867 | ||
868 | /* Verify dirty/writeback state if the filesystem supports it */ | |
869 | if (!page_has_private(page)) | |
870 | return; | |
871 | ||
872 | mapping = page_mapping(page); | |
873 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
874 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
875 | } | |
876 | ||
877 | /* | |
878 | * shrink_page_list() returns the number of reclaimed pages | |
879 | */ | |
880 | static unsigned long shrink_page_list(struct list_head *page_list, | |
881 | struct zone *zone, | |
882 | struct scan_control *sc, | |
883 | enum ttu_flags ttu_flags, | |
884 | unsigned long *ret_nr_dirty, | |
885 | unsigned long *ret_nr_unqueued_dirty, | |
886 | unsigned long *ret_nr_congested, | |
887 | unsigned long *ret_nr_writeback, | |
888 | unsigned long *ret_nr_immediate, | |
889 | bool force_reclaim) | |
890 | { | |
891 | LIST_HEAD(ret_pages); | |
892 | LIST_HEAD(free_pages); | |
893 | int pgactivate = 0; | |
894 | unsigned long nr_unqueued_dirty = 0; | |
895 | unsigned long nr_dirty = 0; | |
896 | unsigned long nr_congested = 0; | |
897 | unsigned long nr_reclaimed = 0; | |
898 | unsigned long nr_writeback = 0; | |
899 | unsigned long nr_immediate = 0; | |
900 | ||
901 | cond_resched(); | |
902 | ||
903 | while (!list_empty(page_list)) { | |
904 | struct address_space *mapping; | |
905 | struct page *page; | |
906 | int may_enter_fs; | |
907 | enum page_references references = PAGEREF_RECLAIM_CLEAN; | |
908 | bool dirty, writeback; | |
909 | ||
910 | cond_resched(); | |
911 | ||
912 | page = lru_to_page(page_list); | |
913 | list_del(&page->lru); | |
914 | ||
915 | if (!trylock_page(page)) | |
916 | goto keep; | |
917 | ||
918 | VM_BUG_ON_PAGE(PageActive(page), page); | |
919 | VM_BUG_ON_PAGE(page_zone(page) != zone, page); | |
920 | ||
921 | sc->nr_scanned++; | |
922 | ||
923 | if (unlikely(!page_evictable(page))) | |
924 | goto cull_mlocked; | |
925 | ||
926 | if (!sc->may_unmap && page_mapped(page)) | |
927 | goto keep_locked; | |
928 | ||
929 | /* Double the slab pressure for mapped and swapcache pages */ | |
930 | if (page_mapped(page) || PageSwapCache(page)) | |
931 | sc->nr_scanned++; | |
932 | ||
933 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || | |
934 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
935 | ||
936 | /* | |
937 | * The number of dirty pages determines if a zone is marked | |
938 | * reclaim_congested which affects wait_iff_congested. kswapd | |
939 | * will stall and start writing pages if the tail of the LRU | |
940 | * is all dirty unqueued pages. | |
941 | */ | |
942 | page_check_dirty_writeback(page, &dirty, &writeback); | |
943 | if (dirty || writeback) | |
944 | nr_dirty++; | |
945 | ||
946 | if (dirty && !writeback) | |
947 | nr_unqueued_dirty++; | |
948 | ||
949 | /* | |
950 | * Treat this page as congested if the underlying BDI is or if | |
951 | * pages are cycling through the LRU so quickly that the | |
952 | * pages marked for immediate reclaim are making it to the | |
953 | * end of the LRU a second time. | |
954 | */ | |
955 | mapping = page_mapping(page); | |
956 | if (((dirty || writeback) && mapping && | |
957 | inode_write_congested(mapping->host)) || | |
958 | (writeback && PageReclaim(page))) | |
959 | nr_congested++; | |
960 | ||
961 | /* | |
962 | * If a page at the tail of the LRU is under writeback, there | |
963 | * are three cases to consider. | |
964 | * | |
965 | * 1) If reclaim is encountering an excessive number of pages | |
966 | * under writeback and this page is both under writeback and | |
967 | * PageReclaim then it indicates that pages are being queued | |
968 | * for IO but are being recycled through the LRU before the | |
969 | * IO can complete. Waiting on the page itself risks an | |
970 | * indefinite stall if it is impossible to writeback the | |
971 | * page due to IO error or disconnected storage so instead | |
972 | * note that the LRU is being scanned too quickly and the | |
973 | * caller can stall after page list has been processed. | |
974 | * | |
975 | * 2) Global or new memcg reclaim encounters a page that is | |
976 | * not marked for immediate reclaim, or the caller does not | |
977 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
978 | * not to fs). In this case mark the page for immediate | |
979 | * reclaim and continue scanning. | |
980 | * | |
981 | * Require may_enter_fs because we would wait on fs, which | |
982 | * may not have submitted IO yet. And the loop driver might | |
983 | * enter reclaim, and deadlock if it waits on a page for | |
984 | * which it is needed to do the write (loop masks off | |
985 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
986 | * would probably show more reasons. | |
987 | * | |
988 | * 3) Legacy memcg encounters a page that is already marked | |
989 | * PageReclaim. memcg does not have any dirty pages | |
990 | * throttling so we could easily OOM just because too many | |
991 | * pages are in writeback and there is nothing else to | |
992 | * reclaim. Wait for the writeback to complete. | |
993 | */ | |
994 | if (PageWriteback(page)) { | |
995 | /* Case 1 above */ | |
996 | if (current_is_kswapd() && | |
997 | PageReclaim(page) && | |
998 | test_bit(ZONE_WRITEBACK, &zone->flags)) { | |
999 | nr_immediate++; | |
1000 | goto keep_locked; | |
1001 | ||
1002 | /* Case 2 above */ | |
1003 | } else if (sane_reclaim(sc) || | |
1004 | !PageReclaim(page) || !may_enter_fs) { | |
1005 | /* | |
1006 | * This is slightly racy - end_page_writeback() | |
1007 | * might have just cleared PageReclaim, then | |
1008 | * setting PageReclaim here end up interpreted | |
1009 | * as PageReadahead - but that does not matter | |
1010 | * enough to care. What we do want is for this | |
1011 | * page to have PageReclaim set next time memcg | |
1012 | * reclaim reaches the tests above, so it will | |
1013 | * then wait_on_page_writeback() to avoid OOM; | |
1014 | * and it's also appropriate in global reclaim. | |
1015 | */ | |
1016 | SetPageReclaim(page); | |
1017 | nr_writeback++; | |
1018 | goto keep_locked; | |
1019 | ||
1020 | /* Case 3 above */ | |
1021 | } else { | |
1022 | unlock_page(page); | |
1023 | wait_on_page_writeback(page); | |
1024 | /* then go back and try same page again */ | |
1025 | list_add_tail(&page->lru, page_list); | |
1026 | continue; | |
1027 | } | |
1028 | } | |
1029 | ||
1030 | if (!force_reclaim) | |
1031 | references = page_check_references(page, sc); | |
1032 | ||
1033 | switch (references) { | |
1034 | case PAGEREF_ACTIVATE: | |
1035 | goto activate_locked; | |
1036 | case PAGEREF_KEEP: | |
1037 | goto keep_locked; | |
1038 | case PAGEREF_RECLAIM: | |
1039 | case PAGEREF_RECLAIM_CLEAN: | |
1040 | ; /* try to reclaim the page below */ | |
1041 | } | |
1042 | ||
1043 | /* | |
1044 | * Anonymous process memory has backing store? | |
1045 | * Try to allocate it some swap space here. | |
1046 | */ | |
1047 | if (PageAnon(page) && !PageSwapCache(page)) { | |
1048 | if (!(sc->gfp_mask & __GFP_IO)) | |
1049 | goto keep_locked; | |
1050 | if (!add_to_swap(page, page_list)) | |
1051 | goto activate_locked; | |
1052 | may_enter_fs = 1; | |
1053 | ||
1054 | /* Adding to swap updated mapping */ | |
1055 | mapping = page_mapping(page); | |
1056 | } | |
1057 | ||
1058 | /* | |
1059 | * The page is mapped into the page tables of one or more | |
1060 | * processes. Try to unmap it here. | |
1061 | */ | |
1062 | if (page_mapped(page) && mapping) { | |
1063 | switch (try_to_unmap(page, | |
1064 | ttu_flags|TTU_BATCH_FLUSH)) { | |
1065 | case SWAP_FAIL: | |
1066 | goto activate_locked; | |
1067 | case SWAP_AGAIN: | |
1068 | goto keep_locked; | |
1069 | case SWAP_MLOCK: | |
1070 | goto cull_mlocked; | |
1071 | case SWAP_SUCCESS: | |
1072 | ; /* try to free the page below */ | |
1073 | } | |
1074 | } | |
1075 | ||
1076 | if (PageDirty(page)) { | |
1077 | /* | |
1078 | * Only kswapd can writeback filesystem pages to | |
1079 | * avoid risk of stack overflow but only writeback | |
1080 | * if many dirty pages have been encountered. | |
1081 | */ | |
1082 | if (page_is_file_cache(page) && | |
1083 | (!current_is_kswapd() || | |
1084 | !test_bit(ZONE_DIRTY, &zone->flags))) { | |
1085 | /* | |
1086 | * Immediately reclaim when written back. | |
1087 | * Similar in principal to deactivate_page() | |
1088 | * except we already have the page isolated | |
1089 | * and know it's dirty | |
1090 | */ | |
1091 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
1092 | SetPageReclaim(page); | |
1093 | ||
1094 | goto keep_locked; | |
1095 | } | |
1096 | ||
1097 | if (references == PAGEREF_RECLAIM_CLEAN) | |
1098 | goto keep_locked; | |
1099 | if (!may_enter_fs) | |
1100 | goto keep_locked; | |
1101 | if (!sc->may_writepage) | |
1102 | goto keep_locked; | |
1103 | ||
1104 | /* | |
1105 | * Page is dirty. Flush the TLB if a writable entry | |
1106 | * potentially exists to avoid CPU writes after IO | |
1107 | * starts and then write it out here. | |
1108 | */ | |
1109 | try_to_unmap_flush_dirty(); | |
1110 | switch (pageout(page, mapping, sc)) { | |
1111 | case PAGE_KEEP: | |
1112 | goto keep_locked; | |
1113 | case PAGE_ACTIVATE: | |
1114 | goto activate_locked; | |
1115 | case PAGE_SUCCESS: | |
1116 | if (PageWriteback(page)) | |
1117 | goto keep; | |
1118 | if (PageDirty(page)) | |
1119 | goto keep; | |
1120 | ||
1121 | /* | |
1122 | * A synchronous write - probably a ramdisk. Go | |
1123 | * ahead and try to reclaim the page. | |
1124 | */ | |
1125 | if (!trylock_page(page)) | |
1126 | goto keep; | |
1127 | if (PageDirty(page) || PageWriteback(page)) | |
1128 | goto keep_locked; | |
1129 | mapping = page_mapping(page); | |
1130 | case PAGE_CLEAN: | |
1131 | ; /* try to free the page below */ | |
1132 | } | |
1133 | } | |
1134 | ||
1135 | /* | |
1136 | * If the page has buffers, try to free the buffer mappings | |
1137 | * associated with this page. If we succeed we try to free | |
1138 | * the page as well. | |
1139 | * | |
1140 | * We do this even if the page is PageDirty(). | |
1141 | * try_to_release_page() does not perform I/O, but it is | |
1142 | * possible for a page to have PageDirty set, but it is actually | |
1143 | * clean (all its buffers are clean). This happens if the | |
1144 | * buffers were written out directly, with submit_bh(). ext3 | |
1145 | * will do this, as well as the blockdev mapping. | |
1146 | * try_to_release_page() will discover that cleanness and will | |
1147 | * drop the buffers and mark the page clean - it can be freed. | |
1148 | * | |
1149 | * Rarely, pages can have buffers and no ->mapping. These are | |
1150 | * the pages which were not successfully invalidated in | |
1151 | * truncate_complete_page(). We try to drop those buffers here | |
1152 | * and if that worked, and the page is no longer mapped into | |
1153 | * process address space (page_count == 1) it can be freed. | |
1154 | * Otherwise, leave the page on the LRU so it is swappable. | |
1155 | */ | |
1156 | if (page_has_private(page)) { | |
1157 | if (!try_to_release_page(page, sc->gfp_mask)) | |
1158 | goto activate_locked; | |
1159 | if (!mapping && page_count(page) == 1) { | |
1160 | unlock_page(page); | |
1161 | if (put_page_testzero(page)) | |
1162 | goto free_it; | |
1163 | else { | |
1164 | /* | |
1165 | * rare race with speculative reference. | |
1166 | * the speculative reference will free | |
1167 | * this page shortly, so we may | |
1168 | * increment nr_reclaimed here (and | |
1169 | * leave it off the LRU). | |
1170 | */ | |
1171 | nr_reclaimed++; | |
1172 | continue; | |
1173 | } | |
1174 | } | |
1175 | } | |
1176 | ||
1177 | if (!mapping || !__remove_mapping(mapping, page, true)) | |
1178 | goto keep_locked; | |
1179 | ||
1180 | /* | |
1181 | * At this point, we have no other references and there is | |
1182 | * no way to pick any more up (removed from LRU, removed | |
1183 | * from pagecache). Can use non-atomic bitops now (and | |
1184 | * we obviously don't have to worry about waking up a process | |
1185 | * waiting on the page lock, because there are no references. | |
1186 | */ | |
1187 | __clear_page_locked(page); | |
1188 | free_it: | |
1189 | nr_reclaimed++; | |
1190 | ||
1191 | /* | |
1192 | * Is there need to periodically free_page_list? It would | |
1193 | * appear not as the counts should be low | |
1194 | */ | |
1195 | list_add(&page->lru, &free_pages); | |
1196 | continue; | |
1197 | ||
1198 | cull_mlocked: | |
1199 | if (PageSwapCache(page)) | |
1200 | try_to_free_swap(page); | |
1201 | unlock_page(page); | |
1202 | list_add(&page->lru, &ret_pages); | |
1203 | continue; | |
1204 | ||
1205 | activate_locked: | |
1206 | /* Not a candidate for swapping, so reclaim swap space. */ | |
1207 | if (PageSwapCache(page) && vm_swap_full()) | |
1208 | try_to_free_swap(page); | |
1209 | VM_BUG_ON_PAGE(PageActive(page), page); | |
1210 | SetPageActive(page); | |
1211 | pgactivate++; | |
1212 | keep_locked: | |
1213 | unlock_page(page); | |
1214 | keep: | |
1215 | list_add(&page->lru, &ret_pages); | |
1216 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); | |
1217 | } | |
1218 | ||
1219 | mem_cgroup_uncharge_list(&free_pages); | |
1220 | try_to_unmap_flush(); | |
1221 | free_hot_cold_page_list(&free_pages, true); | |
1222 | ||
1223 | list_splice(&ret_pages, page_list); | |
1224 | count_vm_events(PGACTIVATE, pgactivate); | |
1225 | ||
1226 | *ret_nr_dirty += nr_dirty; | |
1227 | *ret_nr_congested += nr_congested; | |
1228 | *ret_nr_unqueued_dirty += nr_unqueued_dirty; | |
1229 | *ret_nr_writeback += nr_writeback; | |
1230 | *ret_nr_immediate += nr_immediate; | |
1231 | return nr_reclaimed; | |
1232 | } | |
1233 | ||
1234 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, | |
1235 | struct list_head *page_list) | |
1236 | { | |
1237 | struct scan_control sc = { | |
1238 | .gfp_mask = GFP_KERNEL, | |
1239 | .priority = DEF_PRIORITY, | |
1240 | .may_unmap = 1, | |
1241 | }; | |
1242 | unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; | |
1243 | struct page *page, *next; | |
1244 | LIST_HEAD(clean_pages); | |
1245 | ||
1246 | list_for_each_entry_safe(page, next, page_list, lru) { | |
1247 | if (page_is_file_cache(page) && !PageDirty(page) && | |
1248 | !isolated_balloon_page(page)) { | |
1249 | ClearPageActive(page); | |
1250 | list_move(&page->lru, &clean_pages); | |
1251 | } | |
1252 | } | |
1253 | ||
1254 | ret = shrink_page_list(&clean_pages, zone, &sc, | |
1255 | TTU_UNMAP|TTU_IGNORE_ACCESS, | |
1256 | &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); | |
1257 | list_splice(&clean_pages, page_list); | |
1258 | mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); | |
1259 | return ret; | |
1260 | } | |
1261 | ||
1262 | /* | |
1263 | * Attempt to remove the specified page from its LRU. Only take this page | |
1264 | * if it is of the appropriate PageActive status. Pages which are being | |
1265 | * freed elsewhere are also ignored. | |
1266 | * | |
1267 | * page: page to consider | |
1268 | * mode: one of the LRU isolation modes defined above | |
1269 | * | |
1270 | * returns 0 on success, -ve errno on failure. | |
1271 | */ | |
1272 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) | |
1273 | { | |
1274 | int ret = -EINVAL; | |
1275 | ||
1276 | /* Only take pages on the LRU. */ | |
1277 | if (!PageLRU(page)) | |
1278 | return ret; | |
1279 | ||
1280 | /* Compaction should not handle unevictable pages but CMA can do so */ | |
1281 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
1282 | return ret; | |
1283 | ||
1284 | ret = -EBUSY; | |
1285 | ||
1286 | /* | |
1287 | * To minimise LRU disruption, the caller can indicate that it only | |
1288 | * wants to isolate pages it will be able to operate on without | |
1289 | * blocking - clean pages for the most part. | |
1290 | * | |
1291 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1292 | * is used by reclaim when it is cannot write to backing storage | |
1293 | * | |
1294 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1295 | * that it is possible to migrate without blocking | |
1296 | */ | |
1297 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1298 | /* All the caller can do on PageWriteback is block */ | |
1299 | if (PageWriteback(page)) | |
1300 | return ret; | |
1301 | ||
1302 | if (PageDirty(page)) { | |
1303 | struct address_space *mapping; | |
1304 | ||
1305 | /* ISOLATE_CLEAN means only clean pages */ | |
1306 | if (mode & ISOLATE_CLEAN) | |
1307 | return ret; | |
1308 | ||
1309 | /* | |
1310 | * Only pages without mappings or that have a | |
1311 | * ->migratepage callback are possible to migrate | |
1312 | * without blocking | |
1313 | */ | |
1314 | mapping = page_mapping(page); | |
1315 | if (mapping && !mapping->a_ops->migratepage) | |
1316 | return ret; | |
1317 | } | |
1318 | } | |
1319 | ||
1320 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) | |
1321 | return ret; | |
1322 | ||
1323 | if (likely(get_page_unless_zero(page))) { | |
1324 | /* | |
1325 | * Be careful not to clear PageLRU until after we're | |
1326 | * sure the page is not being freed elsewhere -- the | |
1327 | * page release code relies on it. | |
1328 | */ | |
1329 | ClearPageLRU(page); | |
1330 | ret = 0; | |
1331 | } | |
1332 | ||
1333 | return ret; | |
1334 | } | |
1335 | ||
1336 | /* | |
1337 | * zone->lru_lock is heavily contended. Some of the functions that | |
1338 | * shrink the lists perform better by taking out a batch of pages | |
1339 | * and working on them outside the LRU lock. | |
1340 | * | |
1341 | * For pagecache intensive workloads, this function is the hottest | |
1342 | * spot in the kernel (apart from copy_*_user functions). | |
1343 | * | |
1344 | * Appropriate locks must be held before calling this function. | |
1345 | * | |
1346 | * @nr_to_scan: The number of pages to look through on the list. | |
1347 | * @lruvec: The LRU vector to pull pages from. | |
1348 | * @dst: The temp list to put pages on to. | |
1349 | * @nr_scanned: The number of pages that were scanned. | |
1350 | * @sc: The scan_control struct for this reclaim session | |
1351 | * @mode: One of the LRU isolation modes | |
1352 | * @lru: LRU list id for isolating | |
1353 | * | |
1354 | * returns how many pages were moved onto *@dst. | |
1355 | */ | |
1356 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, | |
1357 | struct lruvec *lruvec, struct list_head *dst, | |
1358 | unsigned long *nr_scanned, struct scan_control *sc, | |
1359 | isolate_mode_t mode, enum lru_list lru) | |
1360 | { | |
1361 | struct list_head *src = &lruvec->lists[lru]; | |
1362 | unsigned long nr_taken = 0; | |
1363 | unsigned long scan; | |
1364 | ||
1365 | for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan && | |
1366 | !list_empty(src); scan++) { | |
1367 | struct page *page; | |
1368 | int nr_pages; | |
1369 | ||
1370 | page = lru_to_page(src); | |
1371 | prefetchw_prev_lru_page(page, src, flags); | |
1372 | ||
1373 | VM_BUG_ON_PAGE(!PageLRU(page), page); | |
1374 | ||
1375 | switch (__isolate_lru_page(page, mode)) { | |
1376 | case 0: | |
1377 | nr_pages = hpage_nr_pages(page); | |
1378 | mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); | |
1379 | list_move(&page->lru, dst); | |
1380 | nr_taken += nr_pages; | |
1381 | break; | |
1382 | ||
1383 | case -EBUSY: | |
1384 | /* else it is being freed elsewhere */ | |
1385 | list_move(&page->lru, src); | |
1386 | continue; | |
1387 | ||
1388 | default: | |
1389 | BUG(); | |
1390 | } | |
1391 | } | |
1392 | ||
1393 | *nr_scanned = scan; | |
1394 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, | |
1395 | nr_taken, mode, is_file_lru(lru)); | |
1396 | return nr_taken; | |
1397 | } | |
1398 | ||
1399 | /** | |
1400 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1401 | * @page: page to isolate from its LRU list | |
1402 | * | |
1403 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1404 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1405 | * | |
1406 | * Returns 0 if the page was removed from an LRU list. | |
1407 | * Returns -EBUSY if the page was not on an LRU list. | |
1408 | * | |
1409 | * The returned page will have PageLRU() cleared. If it was found on | |
1410 | * the active list, it will have PageActive set. If it was found on | |
1411 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1412 | * may need to be cleared by the caller before letting the page go. | |
1413 | * | |
1414 | * The vmstat statistic corresponding to the list on which the page was | |
1415 | * found will be decremented. | |
1416 | * | |
1417 | * Restrictions: | |
1418 | * (1) Must be called with an elevated refcount on the page. This is a | |
1419 | * fundamentnal difference from isolate_lru_pages (which is called | |
1420 | * without a stable reference). | |
1421 | * (2) the lru_lock must not be held. | |
1422 | * (3) interrupts must be enabled. | |
1423 | */ | |
1424 | int isolate_lru_page(struct page *page) | |
1425 | { | |
1426 | int ret = -EBUSY; | |
1427 | ||
1428 | VM_BUG_ON_PAGE(!page_count(page), page); | |
1429 | VM_BUG_ON_PAGE(PageTail(page), page); | |
1430 | ||
1431 | if (PageLRU(page)) { | |
1432 | struct zone *zone = page_zone(page); | |
1433 | struct lruvec *lruvec; | |
1434 | ||
1435 | spin_lock_irq(&zone->lru_lock); | |
1436 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1437 | if (PageLRU(page)) { | |
1438 | int lru = page_lru(page); | |
1439 | get_page(page); | |
1440 | ClearPageLRU(page); | |
1441 | del_page_from_lru_list(page, lruvec, lru); | |
1442 | ret = 0; | |
1443 | } | |
1444 | spin_unlock_irq(&zone->lru_lock); | |
1445 | } | |
1446 | return ret; | |
1447 | } | |
1448 | ||
1449 | /* | |
1450 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and | |
1451 | * then get resheduled. When there are massive number of tasks doing page | |
1452 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1453 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1454 | * unnecessary swapping, thrashing and OOM. | |
1455 | */ | |
1456 | static int too_many_isolated(struct zone *zone, int file, | |
1457 | struct scan_control *sc) | |
1458 | { | |
1459 | unsigned long inactive, isolated; | |
1460 | ||
1461 | if (current_is_kswapd()) | |
1462 | return 0; | |
1463 | ||
1464 | if (!sane_reclaim(sc)) | |
1465 | return 0; | |
1466 | ||
1467 | if (file) { | |
1468 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1469 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1470 | } else { | |
1471 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1472 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1473 | } | |
1474 | ||
1475 | /* | |
1476 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1477 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1478 | * deadlock. | |
1479 | */ | |
1480 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) | |
1481 | inactive >>= 3; | |
1482 | ||
1483 | return isolated > inactive; | |
1484 | } | |
1485 | ||
1486 | static noinline_for_stack void | |
1487 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) | |
1488 | { | |
1489 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1490 | struct zone *zone = lruvec_zone(lruvec); | |
1491 | LIST_HEAD(pages_to_free); | |
1492 | ||
1493 | /* | |
1494 | * Put back any unfreeable pages. | |
1495 | */ | |
1496 | while (!list_empty(page_list)) { | |
1497 | struct page *page = lru_to_page(page_list); | |
1498 | int lru; | |
1499 | ||
1500 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
1501 | list_del(&page->lru); | |
1502 | if (unlikely(!page_evictable(page))) { | |
1503 | spin_unlock_irq(&zone->lru_lock); | |
1504 | putback_lru_page(page); | |
1505 | spin_lock_irq(&zone->lru_lock); | |
1506 | continue; | |
1507 | } | |
1508 | ||
1509 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1510 | ||
1511 | SetPageLRU(page); | |
1512 | lru = page_lru(page); | |
1513 | add_page_to_lru_list(page, lruvec, lru); | |
1514 | ||
1515 | if (is_active_lru(lru)) { | |
1516 | int file = is_file_lru(lru); | |
1517 | int numpages = hpage_nr_pages(page); | |
1518 | reclaim_stat->recent_rotated[file] += numpages; | |
1519 | } | |
1520 | if (put_page_testzero(page)) { | |
1521 | __ClearPageLRU(page); | |
1522 | __ClearPageActive(page); | |
1523 | del_page_from_lru_list(page, lruvec, lru); | |
1524 | ||
1525 | if (unlikely(PageCompound(page))) { | |
1526 | spin_unlock_irq(&zone->lru_lock); | |
1527 | mem_cgroup_uncharge(page); | |
1528 | (*get_compound_page_dtor(page))(page); | |
1529 | spin_lock_irq(&zone->lru_lock); | |
1530 | } else | |
1531 | list_add(&page->lru, &pages_to_free); | |
1532 | } | |
1533 | } | |
1534 | ||
1535 | /* | |
1536 | * To save our caller's stack, now use input list for pages to free. | |
1537 | */ | |
1538 | list_splice(&pages_to_free, page_list); | |
1539 | } | |
1540 | ||
1541 | /* | |
1542 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1543 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1544 | * In that case we should only throttle if the backing device it is | |
1545 | * writing to is congested. In other cases it is safe to throttle. | |
1546 | */ | |
1547 | static int current_may_throttle(void) | |
1548 | { | |
1549 | return !(current->flags & PF_LESS_THROTTLE) || | |
1550 | current->backing_dev_info == NULL || | |
1551 | bdi_write_congested(current->backing_dev_info); | |
1552 | } | |
1553 | ||
1554 | /* | |
1555 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number | |
1556 | * of reclaimed pages | |
1557 | */ | |
1558 | static noinline_for_stack unsigned long | |
1559 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, | |
1560 | struct scan_control *sc, enum lru_list lru) | |
1561 | { | |
1562 | LIST_HEAD(page_list); | |
1563 | unsigned long nr_scanned; | |
1564 | unsigned long nr_reclaimed = 0; | |
1565 | unsigned long nr_taken; | |
1566 | unsigned long nr_dirty = 0; | |
1567 | unsigned long nr_congested = 0; | |
1568 | unsigned long nr_unqueued_dirty = 0; | |
1569 | unsigned long nr_writeback = 0; | |
1570 | unsigned long nr_immediate = 0; | |
1571 | isolate_mode_t isolate_mode = 0; | |
1572 | int file = is_file_lru(lru); | |
1573 | struct zone *zone = lruvec_zone(lruvec); | |
1574 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1575 | ||
1576 | while (unlikely(too_many_isolated(zone, file, sc))) { | |
1577 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
1578 | ||
1579 | /* We are about to die and free our memory. Return now. */ | |
1580 | if (fatal_signal_pending(current)) | |
1581 | return SWAP_CLUSTER_MAX; | |
1582 | } | |
1583 | ||
1584 | lru_add_drain(); | |
1585 | ||
1586 | if (!sc->may_unmap) | |
1587 | isolate_mode |= ISOLATE_UNMAPPED; | |
1588 | if (!sc->may_writepage) | |
1589 | isolate_mode |= ISOLATE_CLEAN; | |
1590 | ||
1591 | spin_lock_irq(&zone->lru_lock); | |
1592 | ||
1593 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, | |
1594 | &nr_scanned, sc, isolate_mode, lru); | |
1595 | ||
1596 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1597 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1598 | ||
1599 | if (global_reclaim(sc)) { | |
1600 | __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); | |
1601 | if (current_is_kswapd()) | |
1602 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); | |
1603 | else | |
1604 | __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); | |
1605 | } | |
1606 | spin_unlock_irq(&zone->lru_lock); | |
1607 | ||
1608 | if (nr_taken == 0) | |
1609 | return 0; | |
1610 | ||
1611 | nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, | |
1612 | &nr_dirty, &nr_unqueued_dirty, &nr_congested, | |
1613 | &nr_writeback, &nr_immediate, | |
1614 | false); | |
1615 | ||
1616 | spin_lock_irq(&zone->lru_lock); | |
1617 | ||
1618 | reclaim_stat->recent_scanned[file] += nr_taken; | |
1619 | ||
1620 | if (global_reclaim(sc)) { | |
1621 | if (current_is_kswapd()) | |
1622 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1623 | nr_reclaimed); | |
1624 | else | |
1625 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1626 | nr_reclaimed); | |
1627 | } | |
1628 | ||
1629 | putback_inactive_pages(lruvec, &page_list); | |
1630 | ||
1631 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); | |
1632 | ||
1633 | spin_unlock_irq(&zone->lru_lock); | |
1634 | ||
1635 | mem_cgroup_uncharge_list(&page_list); | |
1636 | free_hot_cold_page_list(&page_list, true); | |
1637 | ||
1638 | /* | |
1639 | * If reclaim is isolating dirty pages under writeback, it implies | |
1640 | * that the long-lived page allocation rate is exceeding the page | |
1641 | * laundering rate. Either the global limits are not being effective | |
1642 | * at throttling processes due to the page distribution throughout | |
1643 | * zones or there is heavy usage of a slow backing device. The | |
1644 | * only option is to throttle from reclaim context which is not ideal | |
1645 | * as there is no guarantee the dirtying process is throttled in the | |
1646 | * same way balance_dirty_pages() manages. | |
1647 | * | |
1648 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number | |
1649 | * of pages under pages flagged for immediate reclaim and stall if any | |
1650 | * are encountered in the nr_immediate check below. | |
1651 | */ | |
1652 | if (nr_writeback && nr_writeback == nr_taken) | |
1653 | set_bit(ZONE_WRITEBACK, &zone->flags); | |
1654 | ||
1655 | /* | |
1656 | * Legacy memcg will stall in page writeback so avoid forcibly | |
1657 | * stalling here. | |
1658 | */ | |
1659 | if (sane_reclaim(sc)) { | |
1660 | /* | |
1661 | * Tag a zone as congested if all the dirty pages scanned were | |
1662 | * backed by a congested BDI and wait_iff_congested will stall. | |
1663 | */ | |
1664 | if (nr_dirty && nr_dirty == nr_congested) | |
1665 | set_bit(ZONE_CONGESTED, &zone->flags); | |
1666 | ||
1667 | /* | |
1668 | * If dirty pages are scanned that are not queued for IO, it | |
1669 | * implies that flushers are not keeping up. In this case, flag | |
1670 | * the zone ZONE_DIRTY and kswapd will start writing pages from | |
1671 | * reclaim context. | |
1672 | */ | |
1673 | if (nr_unqueued_dirty == nr_taken) | |
1674 | set_bit(ZONE_DIRTY, &zone->flags); | |
1675 | ||
1676 | /* | |
1677 | * If kswapd scans pages marked marked for immediate | |
1678 | * reclaim and under writeback (nr_immediate), it implies | |
1679 | * that pages are cycling through the LRU faster than | |
1680 | * they are written so also forcibly stall. | |
1681 | */ | |
1682 | if (nr_immediate && current_may_throttle()) | |
1683 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
1684 | } | |
1685 | ||
1686 | /* | |
1687 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1688 | * is congested. Allow kswapd to continue until it starts encountering | |
1689 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1690 | */ | |
1691 | if (!sc->hibernation_mode && !current_is_kswapd() && | |
1692 | current_may_throttle()) | |
1693 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); | |
1694 | ||
1695 | trace_mm_vmscan_lru_shrink_inactive(zone, nr_scanned, nr_reclaimed, | |
1696 | sc->priority, file); | |
1697 | return nr_reclaimed; | |
1698 | } | |
1699 | ||
1700 | /* | |
1701 | * This moves pages from the active list to the inactive list. | |
1702 | * | |
1703 | * We move them the other way if the page is referenced by one or more | |
1704 | * processes, from rmap. | |
1705 | * | |
1706 | * If the pages are mostly unmapped, the processing is fast and it is | |
1707 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1708 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1709 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1710 | * this, so instead we remove the pages from the LRU while processing them. | |
1711 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1712 | * nobody will play with that bit on a non-LRU page. | |
1713 | * | |
1714 | * The downside is that we have to touch page->_count against each page. | |
1715 | * But we had to alter page->flags anyway. | |
1716 | */ | |
1717 | ||
1718 | static void move_active_pages_to_lru(struct lruvec *lruvec, | |
1719 | struct list_head *list, | |
1720 | struct list_head *pages_to_free, | |
1721 | enum lru_list lru) | |
1722 | { | |
1723 | struct zone *zone = lruvec_zone(lruvec); | |
1724 | unsigned long pgmoved = 0; | |
1725 | struct page *page; | |
1726 | int nr_pages; | |
1727 | ||
1728 | while (!list_empty(list)) { | |
1729 | page = lru_to_page(list); | |
1730 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1731 | ||
1732 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
1733 | SetPageLRU(page); | |
1734 | ||
1735 | nr_pages = hpage_nr_pages(page); | |
1736 | mem_cgroup_update_lru_size(lruvec, lru, nr_pages); | |
1737 | list_move(&page->lru, &lruvec->lists[lru]); | |
1738 | pgmoved += nr_pages; | |
1739 | ||
1740 | if (put_page_testzero(page)) { | |
1741 | __ClearPageLRU(page); | |
1742 | __ClearPageActive(page); | |
1743 | del_page_from_lru_list(page, lruvec, lru); | |
1744 | ||
1745 | if (unlikely(PageCompound(page))) { | |
1746 | spin_unlock_irq(&zone->lru_lock); | |
1747 | mem_cgroup_uncharge(page); | |
1748 | (*get_compound_page_dtor(page))(page); | |
1749 | spin_lock_irq(&zone->lru_lock); | |
1750 | } else | |
1751 | list_add(&page->lru, pages_to_free); | |
1752 | } | |
1753 | } | |
1754 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1755 | if (!is_active_lru(lru)) | |
1756 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1757 | } | |
1758 | ||
1759 | static void shrink_active_list(unsigned long nr_to_scan, | |
1760 | struct lruvec *lruvec, | |
1761 | struct scan_control *sc, | |
1762 | enum lru_list lru) | |
1763 | { | |
1764 | unsigned long nr_taken; | |
1765 | unsigned long nr_scanned; | |
1766 | unsigned long vm_flags; | |
1767 | LIST_HEAD(l_hold); /* The pages which were snipped off */ | |
1768 | LIST_HEAD(l_active); | |
1769 | LIST_HEAD(l_inactive); | |
1770 | struct page *page; | |
1771 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1772 | unsigned long nr_rotated = 0; | |
1773 | isolate_mode_t isolate_mode = 0; | |
1774 | int file = is_file_lru(lru); | |
1775 | struct zone *zone = lruvec_zone(lruvec); | |
1776 | ||
1777 | lru_add_drain(); | |
1778 | ||
1779 | if (!sc->may_unmap) | |
1780 | isolate_mode |= ISOLATE_UNMAPPED; | |
1781 | if (!sc->may_writepage) | |
1782 | isolate_mode |= ISOLATE_CLEAN; | |
1783 | ||
1784 | spin_lock_irq(&zone->lru_lock); | |
1785 | ||
1786 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, | |
1787 | &nr_scanned, sc, isolate_mode, lru); | |
1788 | if (global_reclaim(sc)) | |
1789 | __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); | |
1790 | ||
1791 | reclaim_stat->recent_scanned[file] += nr_taken; | |
1792 | ||
1793 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); | |
1794 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1795 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1796 | spin_unlock_irq(&zone->lru_lock); | |
1797 | ||
1798 | while (!list_empty(&l_hold)) { | |
1799 | cond_resched(); | |
1800 | page = lru_to_page(&l_hold); | |
1801 | list_del(&page->lru); | |
1802 | ||
1803 | if (unlikely(!page_evictable(page))) { | |
1804 | putback_lru_page(page); | |
1805 | continue; | |
1806 | } | |
1807 | ||
1808 | if (unlikely(buffer_heads_over_limit)) { | |
1809 | if (page_has_private(page) && trylock_page(page)) { | |
1810 | if (page_has_private(page)) | |
1811 | try_to_release_page(page, 0); | |
1812 | unlock_page(page); | |
1813 | } | |
1814 | } | |
1815 | ||
1816 | if (page_referenced(page, 0, sc->target_mem_cgroup, | |
1817 | &vm_flags)) { | |
1818 | nr_rotated += hpage_nr_pages(page); | |
1819 | /* | |
1820 | * Identify referenced, file-backed active pages and | |
1821 | * give them one more trip around the active list. So | |
1822 | * that executable code get better chances to stay in | |
1823 | * memory under moderate memory pressure. Anon pages | |
1824 | * are not likely to be evicted by use-once streaming | |
1825 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1826 | * so we ignore them here. | |
1827 | */ | |
1828 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { | |
1829 | list_add(&page->lru, &l_active); | |
1830 | continue; | |
1831 | } | |
1832 | } | |
1833 | ||
1834 | ClearPageActive(page); /* we are de-activating */ | |
1835 | list_add(&page->lru, &l_inactive); | |
1836 | } | |
1837 | ||
1838 | /* | |
1839 | * Move pages back to the lru list. | |
1840 | */ | |
1841 | spin_lock_irq(&zone->lru_lock); | |
1842 | /* | |
1843 | * Count referenced pages from currently used mappings as rotated, | |
1844 | * even though only some of them are actually re-activated. This | |
1845 | * helps balance scan pressure between file and anonymous pages in | |
1846 | * get_scan_count. | |
1847 | */ | |
1848 | reclaim_stat->recent_rotated[file] += nr_rotated; | |
1849 | ||
1850 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); | |
1851 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
1852 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); | |
1853 | spin_unlock_irq(&zone->lru_lock); | |
1854 | ||
1855 | mem_cgroup_uncharge_list(&l_hold); | |
1856 | free_hot_cold_page_list(&l_hold, true); | |
1857 | } | |
1858 | ||
1859 | #ifdef CONFIG_SWAP | |
1860 | static bool inactive_anon_is_low_global(struct zone *zone) | |
1861 | { | |
1862 | unsigned long active, inactive; | |
1863 | ||
1864 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1865 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1866 | ||
1867 | return inactive * zone->inactive_ratio < active; | |
1868 | } | |
1869 | ||
1870 | /** | |
1871 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
1872 | * @lruvec: LRU vector to check | |
1873 | * | |
1874 | * Returns true if the zone does not have enough inactive anon pages, | |
1875 | * meaning some active anon pages need to be deactivated. | |
1876 | */ | |
1877 | static bool inactive_anon_is_low(struct lruvec *lruvec) | |
1878 | { | |
1879 | /* | |
1880 | * If we don't have swap space, anonymous page deactivation | |
1881 | * is pointless. | |
1882 | */ | |
1883 | if (!total_swap_pages) | |
1884 | return false; | |
1885 | ||
1886 | if (!mem_cgroup_disabled()) | |
1887 | return mem_cgroup_inactive_anon_is_low(lruvec); | |
1888 | ||
1889 | return inactive_anon_is_low_global(lruvec_zone(lruvec)); | |
1890 | } | |
1891 | #else | |
1892 | static inline bool inactive_anon_is_low(struct lruvec *lruvec) | |
1893 | { | |
1894 | return false; | |
1895 | } | |
1896 | #endif | |
1897 | ||
1898 | /** | |
1899 | * inactive_file_is_low - check if file pages need to be deactivated | |
1900 | * @lruvec: LRU vector to check | |
1901 | * | |
1902 | * When the system is doing streaming IO, memory pressure here | |
1903 | * ensures that active file pages get deactivated, until more | |
1904 | * than half of the file pages are on the inactive list. | |
1905 | * | |
1906 | * Once we get to that situation, protect the system's working | |
1907 | * set from being evicted by disabling active file page aging. | |
1908 | * | |
1909 | * This uses a different ratio than the anonymous pages, because | |
1910 | * the page cache uses a use-once replacement algorithm. | |
1911 | */ | |
1912 | static bool inactive_file_is_low(struct lruvec *lruvec) | |
1913 | { | |
1914 | unsigned long inactive; | |
1915 | unsigned long active; | |
1916 | ||
1917 | inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
1918 | active = get_lru_size(lruvec, LRU_ACTIVE_FILE); | |
1919 | ||
1920 | return active > inactive; | |
1921 | } | |
1922 | ||
1923 | static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) | |
1924 | { | |
1925 | if (is_file_lru(lru)) | |
1926 | return inactive_file_is_low(lruvec); | |
1927 | else | |
1928 | return inactive_anon_is_low(lruvec); | |
1929 | } | |
1930 | ||
1931 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, | |
1932 | struct lruvec *lruvec, struct scan_control *sc) | |
1933 | { | |
1934 | if (is_active_lru(lru)) { | |
1935 | if (inactive_list_is_low(lruvec, lru)) | |
1936 | shrink_active_list(nr_to_scan, lruvec, sc, lru); | |
1937 | return 0; | |
1938 | } | |
1939 | ||
1940 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); | |
1941 | } | |
1942 | ||
1943 | enum scan_balance { | |
1944 | SCAN_EQUAL, | |
1945 | SCAN_FRACT, | |
1946 | SCAN_ANON, | |
1947 | SCAN_FILE, | |
1948 | }; | |
1949 | ||
1950 | /* | |
1951 | * Determine how aggressively the anon and file LRU lists should be | |
1952 | * scanned. The relative value of each set of LRU lists is determined | |
1953 | * by looking at the fraction of the pages scanned we did rotate back | |
1954 | * onto the active list instead of evict. | |
1955 | * | |
1956 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan | |
1957 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
1958 | */ | |
1959 | static void get_scan_count(struct lruvec *lruvec, int swappiness, | |
1960 | struct scan_control *sc, unsigned long *nr, | |
1961 | unsigned long *lru_pages) | |
1962 | { | |
1963 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1964 | u64 fraction[2]; | |
1965 | u64 denominator = 0; /* gcc */ | |
1966 | struct zone *zone = lruvec_zone(lruvec); | |
1967 | unsigned long anon_prio, file_prio; | |
1968 | enum scan_balance scan_balance; | |
1969 | unsigned long anon, file; | |
1970 | bool force_scan = false; | |
1971 | unsigned long ap, fp; | |
1972 | enum lru_list lru; | |
1973 | bool some_scanned; | |
1974 | int pass; | |
1975 | ||
1976 | /* | |
1977 | * If the zone or memcg is small, nr[l] can be 0. This | |
1978 | * results in no scanning on this priority and a potential | |
1979 | * priority drop. Global direct reclaim can go to the next | |
1980 | * zone and tends to have no problems. Global kswapd is for | |
1981 | * zone balancing and it needs to scan a minimum amount. When | |
1982 | * reclaiming for a memcg, a priority drop can cause high | |
1983 | * latencies, so it's better to scan a minimum amount there as | |
1984 | * well. | |
1985 | */ | |
1986 | if (current_is_kswapd()) { | |
1987 | if (!zone_reclaimable(zone)) | |
1988 | force_scan = true; | |
1989 | if (!mem_cgroup_lruvec_online(lruvec)) | |
1990 | force_scan = true; | |
1991 | } | |
1992 | if (!global_reclaim(sc)) | |
1993 | force_scan = true; | |
1994 | ||
1995 | /* If we have no swap space, do not bother scanning anon pages. */ | |
1996 | if (!sc->may_swap || (get_nr_swap_pages() <= 0)) { | |
1997 | scan_balance = SCAN_FILE; | |
1998 | goto out; | |
1999 | } | |
2000 | ||
2001 | /* | |
2002 | * Global reclaim will swap to prevent OOM even with no | |
2003 | * swappiness, but memcg users want to use this knob to | |
2004 | * disable swapping for individual groups completely when | |
2005 | * using the memory controller's swap limit feature would be | |
2006 | * too expensive. | |
2007 | */ | |
2008 | if (!global_reclaim(sc) && !swappiness) { | |
2009 | scan_balance = SCAN_FILE; | |
2010 | goto out; | |
2011 | } | |
2012 | ||
2013 | /* | |
2014 | * Do not apply any pressure balancing cleverness when the | |
2015 | * system is close to OOM, scan both anon and file equally | |
2016 | * (unless the swappiness setting disagrees with swapping). | |
2017 | */ | |
2018 | if (!sc->priority && swappiness) { | |
2019 | scan_balance = SCAN_EQUAL; | |
2020 | goto out; | |
2021 | } | |
2022 | ||
2023 | /* | |
2024 | * Prevent the reclaimer from falling into the cache trap: as | |
2025 | * cache pages start out inactive, every cache fault will tip | |
2026 | * the scan balance towards the file LRU. And as the file LRU | |
2027 | * shrinks, so does the window for rotation from references. | |
2028 | * This means we have a runaway feedback loop where a tiny | |
2029 | * thrashing file LRU becomes infinitely more attractive than | |
2030 | * anon pages. Try to detect this based on file LRU size. | |
2031 | */ | |
2032 | if (global_reclaim(sc)) { | |
2033 | unsigned long zonefile; | |
2034 | unsigned long zonefree; | |
2035 | ||
2036 | zonefree = zone_page_state(zone, NR_FREE_PAGES); | |
2037 | zonefile = zone_page_state(zone, NR_ACTIVE_FILE) + | |
2038 | zone_page_state(zone, NR_INACTIVE_FILE); | |
2039 | ||
2040 | if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) { | |
2041 | scan_balance = SCAN_ANON; | |
2042 | goto out; | |
2043 | } | |
2044 | } | |
2045 | ||
2046 | /* | |
2047 | * If there is enough inactive page cache, i.e. if the size of the | |
2048 | * inactive list is greater than that of the active list *and* the | |
2049 | * inactive list actually has some pages to scan on this priority, we | |
2050 | * do not reclaim anything from the anonymous working set right now. | |
2051 | * Without the second condition we could end up never scanning an | |
2052 | * lruvec even if it has plenty of old anonymous pages unless the | |
2053 | * system is under heavy pressure. | |
2054 | */ | |
2055 | if (!inactive_file_is_low(lruvec) && | |
2056 | get_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) { | |
2057 | scan_balance = SCAN_FILE; | |
2058 | goto out; | |
2059 | } | |
2060 | ||
2061 | scan_balance = SCAN_FRACT; | |
2062 | ||
2063 | /* | |
2064 | * With swappiness at 100, anonymous and file have the same priority. | |
2065 | * This scanning priority is essentially the inverse of IO cost. | |
2066 | */ | |
2067 | anon_prio = swappiness; | |
2068 | file_prio = 200 - anon_prio; | |
2069 | ||
2070 | /* | |
2071 | * OK, so we have swap space and a fair amount of page cache | |
2072 | * pages. We use the recently rotated / recently scanned | |
2073 | * ratios to determine how valuable each cache is. | |
2074 | * | |
2075 | * Because workloads change over time (and to avoid overflow) | |
2076 | * we keep these statistics as a floating average, which ends | |
2077 | * up weighing recent references more than old ones. | |
2078 | * | |
2079 | * anon in [0], file in [1] | |
2080 | */ | |
2081 | ||
2082 | anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + | |
2083 | get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
2084 | file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
2085 | get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
2086 | ||
2087 | spin_lock_irq(&zone->lru_lock); | |
2088 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { | |
2089 | reclaim_stat->recent_scanned[0] /= 2; | |
2090 | reclaim_stat->recent_rotated[0] /= 2; | |
2091 | } | |
2092 | ||
2093 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { | |
2094 | reclaim_stat->recent_scanned[1] /= 2; | |
2095 | reclaim_stat->recent_rotated[1] /= 2; | |
2096 | } | |
2097 | ||
2098 | /* | |
2099 | * The amount of pressure on anon vs file pages is inversely | |
2100 | * proportional to the fraction of recently scanned pages on | |
2101 | * each list that were recently referenced and in active use. | |
2102 | */ | |
2103 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); | |
2104 | ap /= reclaim_stat->recent_rotated[0] + 1; | |
2105 | ||
2106 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); | |
2107 | fp /= reclaim_stat->recent_rotated[1] + 1; | |
2108 | spin_unlock_irq(&zone->lru_lock); | |
2109 | ||
2110 | fraction[0] = ap; | |
2111 | fraction[1] = fp; | |
2112 | denominator = ap + fp + 1; | |
2113 | out: | |
2114 | some_scanned = false; | |
2115 | /* Only use force_scan on second pass. */ | |
2116 | for (pass = 0; !some_scanned && pass < 2; pass++) { | |
2117 | *lru_pages = 0; | |
2118 | for_each_evictable_lru(lru) { | |
2119 | int file = is_file_lru(lru); | |
2120 | unsigned long size; | |
2121 | unsigned long scan; | |
2122 | ||
2123 | size = get_lru_size(lruvec, lru); | |
2124 | scan = size >> sc->priority; | |
2125 | ||
2126 | if (!scan && pass && force_scan) | |
2127 | scan = min(size, SWAP_CLUSTER_MAX); | |
2128 | ||
2129 | switch (scan_balance) { | |
2130 | case SCAN_EQUAL: | |
2131 | /* Scan lists relative to size */ | |
2132 | break; | |
2133 | case SCAN_FRACT: | |
2134 | /* | |
2135 | * Scan types proportional to swappiness and | |
2136 | * their relative recent reclaim efficiency. | |
2137 | */ | |
2138 | scan = div64_u64(scan * fraction[file], | |
2139 | denominator); | |
2140 | break; | |
2141 | case SCAN_FILE: | |
2142 | case SCAN_ANON: | |
2143 | /* Scan one type exclusively */ | |
2144 | if ((scan_balance == SCAN_FILE) != file) { | |
2145 | size = 0; | |
2146 | scan = 0; | |
2147 | } | |
2148 | break; | |
2149 | default: | |
2150 | /* Look ma, no brain */ | |
2151 | BUG(); | |
2152 | } | |
2153 | ||
2154 | *lru_pages += size; | |
2155 | nr[lru] = scan; | |
2156 | ||
2157 | /* | |
2158 | * Skip the second pass and don't force_scan, | |
2159 | * if we found something to scan. | |
2160 | */ | |
2161 | some_scanned |= !!scan; | |
2162 | } | |
2163 | } | |
2164 | } | |
2165 | ||
2166 | #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH | |
2167 | static void init_tlb_ubc(void) | |
2168 | { | |
2169 | /* | |
2170 | * This deliberately does not clear the cpumask as it's expensive | |
2171 | * and unnecessary. If there happens to be data in there then the | |
2172 | * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and | |
2173 | * then will be cleared. | |
2174 | */ | |
2175 | current->tlb_ubc.flush_required = false; | |
2176 | } | |
2177 | #else | |
2178 | static inline void init_tlb_ubc(void) | |
2179 | { | |
2180 | } | |
2181 | #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ | |
2182 | ||
2183 | /* | |
2184 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
2185 | */ | |
2186 | static void shrink_lruvec(struct lruvec *lruvec, int swappiness, | |
2187 | struct scan_control *sc, unsigned long *lru_pages) | |
2188 | { | |
2189 | unsigned long nr[NR_LRU_LISTS]; | |
2190 | unsigned long targets[NR_LRU_LISTS]; | |
2191 | unsigned long nr_to_scan; | |
2192 | enum lru_list lru; | |
2193 | unsigned long nr_reclaimed = 0; | |
2194 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2195 | struct blk_plug plug; | |
2196 | bool scan_adjusted; | |
2197 | ||
2198 | get_scan_count(lruvec, swappiness, sc, nr, lru_pages); | |
2199 | ||
2200 | /* Record the original scan target for proportional adjustments later */ | |
2201 | memcpy(targets, nr, sizeof(nr)); | |
2202 | ||
2203 | /* | |
2204 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2205 | * event that can occur when there is little memory pressure e.g. | |
2206 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2207 | * when the requested number of pages are reclaimed when scanning at | |
2208 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2209 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2210 | * do a batch of work at once. For memcg reclaim one check is made to | |
2211 | * abort proportional reclaim if either the file or anon lru has already | |
2212 | * dropped to zero at the first pass. | |
2213 | */ | |
2214 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2215 | sc->priority == DEF_PRIORITY); | |
2216 | ||
2217 | init_tlb_ubc(); | |
2218 | ||
2219 | blk_start_plug(&plug); | |
2220 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2221 | nr[LRU_INACTIVE_FILE]) { | |
2222 | unsigned long nr_anon, nr_file, percentage; | |
2223 | unsigned long nr_scanned; | |
2224 | ||
2225 | for_each_evictable_lru(lru) { | |
2226 | if (nr[lru]) { | |
2227 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2228 | nr[lru] -= nr_to_scan; | |
2229 | ||
2230 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2231 | lruvec, sc); | |
2232 | } | |
2233 | } | |
2234 | ||
2235 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) | |
2236 | continue; | |
2237 | ||
2238 | /* | |
2239 | * For kswapd and memcg, reclaim at least the number of pages | |
2240 | * requested. Ensure that the anon and file LRUs are scanned | |
2241 | * proportionally what was requested by get_scan_count(). We | |
2242 | * stop reclaiming one LRU and reduce the amount scanning | |
2243 | * proportional to the original scan target. | |
2244 | */ | |
2245 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2246 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2247 | ||
2248 | /* | |
2249 | * It's just vindictive to attack the larger once the smaller | |
2250 | * has gone to zero. And given the way we stop scanning the | |
2251 | * smaller below, this makes sure that we only make one nudge | |
2252 | * towards proportionality once we've got nr_to_reclaim. | |
2253 | */ | |
2254 | if (!nr_file || !nr_anon) | |
2255 | break; | |
2256 | ||
2257 | if (nr_file > nr_anon) { | |
2258 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2259 | targets[LRU_ACTIVE_ANON] + 1; | |
2260 | lru = LRU_BASE; | |
2261 | percentage = nr_anon * 100 / scan_target; | |
2262 | } else { | |
2263 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2264 | targets[LRU_ACTIVE_FILE] + 1; | |
2265 | lru = LRU_FILE; | |
2266 | percentage = nr_file * 100 / scan_target; | |
2267 | } | |
2268 | ||
2269 | /* Stop scanning the smaller of the LRU */ | |
2270 | nr[lru] = 0; | |
2271 | nr[lru + LRU_ACTIVE] = 0; | |
2272 | ||
2273 | /* | |
2274 | * Recalculate the other LRU scan count based on its original | |
2275 | * scan target and the percentage scanning already complete | |
2276 | */ | |
2277 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2278 | nr_scanned = targets[lru] - nr[lru]; | |
2279 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2280 | nr[lru] -= min(nr[lru], nr_scanned); | |
2281 | ||
2282 | lru += LRU_ACTIVE; | |
2283 | nr_scanned = targets[lru] - nr[lru]; | |
2284 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2285 | nr[lru] -= min(nr[lru], nr_scanned); | |
2286 | ||
2287 | scan_adjusted = true; | |
2288 | } | |
2289 | blk_finish_plug(&plug); | |
2290 | sc->nr_reclaimed += nr_reclaimed; | |
2291 | ||
2292 | /* | |
2293 | * Even if we did not try to evict anon pages at all, we want to | |
2294 | * rebalance the anon lru active/inactive ratio. | |
2295 | */ | |
2296 | if (inactive_anon_is_low(lruvec)) | |
2297 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
2298 | sc, LRU_ACTIVE_ANON); | |
2299 | ||
2300 | throttle_vm_writeout(sc->gfp_mask); | |
2301 | } | |
2302 | ||
2303 | /* Use reclaim/compaction for costly allocs or under memory pressure */ | |
2304 | static bool in_reclaim_compaction(struct scan_control *sc) | |
2305 | { | |
2306 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && | |
2307 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || | |
2308 | sc->priority < DEF_PRIORITY - 2)) | |
2309 | return true; | |
2310 | ||
2311 | return false; | |
2312 | } | |
2313 | ||
2314 | /* | |
2315 | * Reclaim/compaction is used for high-order allocation requests. It reclaims | |
2316 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2317 | * true if more pages should be reclaimed such that when the page allocator | |
2318 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2319 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
2320 | */ | |
2321 | static inline bool should_continue_reclaim(struct zone *zone, | |
2322 | unsigned long nr_reclaimed, | |
2323 | unsigned long nr_scanned, | |
2324 | struct scan_control *sc) | |
2325 | { | |
2326 | unsigned long pages_for_compaction; | |
2327 | unsigned long inactive_lru_pages; | |
2328 | ||
2329 | /* If not in reclaim/compaction mode, stop */ | |
2330 | if (!in_reclaim_compaction(sc)) | |
2331 | return false; | |
2332 | ||
2333 | /* Consider stopping depending on scan and reclaim activity */ | |
2334 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2335 | /* | |
2336 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2337 | * full LRU list has been scanned and we are still failing | |
2338 | * to reclaim pages. This full LRU scan is potentially | |
2339 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2340 | */ | |
2341 | if (!nr_reclaimed && !nr_scanned) | |
2342 | return false; | |
2343 | } else { | |
2344 | /* | |
2345 | * For non-__GFP_REPEAT allocations which can presumably | |
2346 | * fail without consequence, stop if we failed to reclaim | |
2347 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2348 | * pages that were scanned. This will return to the | |
2349 | * caller faster at the risk reclaim/compaction and | |
2350 | * the resulting allocation attempt fails | |
2351 | */ | |
2352 | if (!nr_reclaimed) | |
2353 | return false; | |
2354 | } | |
2355 | ||
2356 | /* | |
2357 | * If we have not reclaimed enough pages for compaction and the | |
2358 | * inactive lists are large enough, continue reclaiming | |
2359 | */ | |
2360 | pages_for_compaction = (2UL << sc->order); | |
2361 | inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); | |
2362 | if (get_nr_swap_pages() > 0) | |
2363 | inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); | |
2364 | if (sc->nr_reclaimed < pages_for_compaction && | |
2365 | inactive_lru_pages > pages_for_compaction) | |
2366 | return true; | |
2367 | ||
2368 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
2369 | switch (compaction_suitable(zone, sc->order, 0, 0)) { | |
2370 | case COMPACT_PARTIAL: | |
2371 | case COMPACT_CONTINUE: | |
2372 | return false; | |
2373 | default: | |
2374 | return true; | |
2375 | } | |
2376 | } | |
2377 | ||
2378 | static bool shrink_zone(struct zone *zone, struct scan_control *sc, | |
2379 | bool is_classzone) | |
2380 | { | |
2381 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
2382 | unsigned long nr_reclaimed, nr_scanned; | |
2383 | bool reclaimable = false; | |
2384 | ||
2385 | do { | |
2386 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2387 | struct mem_cgroup_reclaim_cookie reclaim = { | |
2388 | .zone = zone, | |
2389 | .priority = sc->priority, | |
2390 | }; | |
2391 | unsigned long zone_lru_pages = 0; | |
2392 | struct mem_cgroup *memcg; | |
2393 | ||
2394 | nr_reclaimed = sc->nr_reclaimed; | |
2395 | nr_scanned = sc->nr_scanned; | |
2396 | ||
2397 | memcg = mem_cgroup_iter(root, NULL, &reclaim); | |
2398 | do { | |
2399 | unsigned long lru_pages; | |
2400 | unsigned long reclaimed; | |
2401 | unsigned long scanned; | |
2402 | struct lruvec *lruvec; | |
2403 | int swappiness; | |
2404 | ||
2405 | if (mem_cgroup_low(root, memcg)) { | |
2406 | if (!sc->may_thrash) | |
2407 | continue; | |
2408 | mem_cgroup_events(memcg, MEMCG_LOW, 1); | |
2409 | } | |
2410 | ||
2411 | lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
2412 | swappiness = mem_cgroup_swappiness(memcg); | |
2413 | reclaimed = sc->nr_reclaimed; | |
2414 | scanned = sc->nr_scanned; | |
2415 | ||
2416 | shrink_lruvec(lruvec, swappiness, sc, &lru_pages); | |
2417 | zone_lru_pages += lru_pages; | |
2418 | ||
2419 | if (memcg && is_classzone) | |
2420 | shrink_slab(sc->gfp_mask, zone_to_nid(zone), | |
2421 | memcg, sc->nr_scanned - scanned, | |
2422 | lru_pages); | |
2423 | ||
2424 | /* Record the group's reclaim efficiency */ | |
2425 | vmpressure(sc->gfp_mask, memcg, false, | |
2426 | sc->nr_scanned - scanned, | |
2427 | sc->nr_reclaimed - reclaimed); | |
2428 | ||
2429 | /* | |
2430 | * Direct reclaim and kswapd have to scan all memory | |
2431 | * cgroups to fulfill the overall scan target for the | |
2432 | * zone. | |
2433 | * | |
2434 | * Limit reclaim, on the other hand, only cares about | |
2435 | * nr_to_reclaim pages to be reclaimed and it will | |
2436 | * retry with decreasing priority if one round over the | |
2437 | * whole hierarchy is not sufficient. | |
2438 | */ | |
2439 | if (!global_reclaim(sc) && | |
2440 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
2441 | mem_cgroup_iter_break(root, memcg); | |
2442 | break; | |
2443 | } | |
2444 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); | |
2445 | ||
2446 | /* | |
2447 | * Shrink the slab caches in the same proportion that | |
2448 | * the eligible LRU pages were scanned. | |
2449 | */ | |
2450 | if (global_reclaim(sc) && is_classzone) | |
2451 | shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL, | |
2452 | sc->nr_scanned - nr_scanned, | |
2453 | zone_lru_pages); | |
2454 | ||
2455 | if (reclaim_state) { | |
2456 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2457 | reclaim_state->reclaimed_slab = 0; | |
2458 | } | |
2459 | ||
2460 | /* Record the subtree's reclaim efficiency */ | |
2461 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
2462 | sc->nr_scanned - nr_scanned, | |
2463 | sc->nr_reclaimed - nr_reclaimed); | |
2464 | ||
2465 | if (sc->nr_reclaimed - nr_reclaimed) | |
2466 | reclaimable = true; | |
2467 | ||
2468 | } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, | |
2469 | sc->nr_scanned - nr_scanned, sc)); | |
2470 | ||
2471 | return reclaimable; | |
2472 | } | |
2473 | ||
2474 | /* | |
2475 | * Returns true if compaction should go ahead for a high-order request, or | |
2476 | * the high-order allocation would succeed without compaction. | |
2477 | */ | |
2478 | static inline bool compaction_ready(struct zone *zone, int order) | |
2479 | { | |
2480 | unsigned long balance_gap, watermark; | |
2481 | bool watermark_ok; | |
2482 | ||
2483 | /* | |
2484 | * Compaction takes time to run and there are potentially other | |
2485 | * callers using the pages just freed. Continue reclaiming until | |
2486 | * there is a buffer of free pages available to give compaction | |
2487 | * a reasonable chance of completing and allocating the page | |
2488 | */ | |
2489 | balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( | |
2490 | zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); | |
2491 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << order); | |
2492 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0); | |
2493 | ||
2494 | /* | |
2495 | * If compaction is deferred, reclaim up to a point where | |
2496 | * compaction will have a chance of success when re-enabled | |
2497 | */ | |
2498 | if (compaction_deferred(zone, order)) | |
2499 | return watermark_ok; | |
2500 | ||
2501 | /* | |
2502 | * If compaction is not ready to start and allocation is not likely | |
2503 | * to succeed without it, then keep reclaiming. | |
2504 | */ | |
2505 | if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED) | |
2506 | return false; | |
2507 | ||
2508 | return watermark_ok; | |
2509 | } | |
2510 | ||
2511 | /* | |
2512 | * This is the direct reclaim path, for page-allocating processes. We only | |
2513 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2514 | * request. | |
2515 | * | |
2516 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). | |
2517 | * Because: | |
2518 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order | |
2519 | * allocation or | |
2520 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones | |
2521 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2522 | * zone defense algorithm. | |
2523 | * | |
2524 | * If a zone is deemed to be full of pinned pages then just give it a light | |
2525 | * scan then give up on it. | |
2526 | * | |
2527 | * Returns true if a zone was reclaimable. | |
2528 | */ | |
2529 | static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) | |
2530 | { | |
2531 | struct zoneref *z; | |
2532 | struct zone *zone; | |
2533 | unsigned long nr_soft_reclaimed; | |
2534 | unsigned long nr_soft_scanned; | |
2535 | gfp_t orig_mask; | |
2536 | enum zone_type requested_highidx = gfp_zone(sc->gfp_mask); | |
2537 | bool reclaimable = false; | |
2538 | ||
2539 | /* | |
2540 | * If the number of buffer_heads in the machine exceeds the maximum | |
2541 | * allowed level, force direct reclaim to scan the highmem zone as | |
2542 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2543 | */ | |
2544 | orig_mask = sc->gfp_mask; | |
2545 | if (buffer_heads_over_limit) | |
2546 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2547 | ||
2548 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2549 | requested_highidx, sc->nodemask) { | |
2550 | enum zone_type classzone_idx; | |
2551 | ||
2552 | if (!populated_zone(zone)) | |
2553 | continue; | |
2554 | ||
2555 | classzone_idx = requested_highidx; | |
2556 | while (!populated_zone(zone->zone_pgdat->node_zones + | |
2557 | classzone_idx)) | |
2558 | classzone_idx--; | |
2559 | ||
2560 | /* | |
2561 | * Take care memory controller reclaiming has small influence | |
2562 | * to global LRU. | |
2563 | */ | |
2564 | if (global_reclaim(sc)) { | |
2565 | if (!cpuset_zone_allowed(zone, | |
2566 | GFP_KERNEL | __GFP_HARDWALL)) | |
2567 | continue; | |
2568 | ||
2569 | if (sc->priority != DEF_PRIORITY && | |
2570 | !zone_reclaimable(zone)) | |
2571 | continue; /* Let kswapd poll it */ | |
2572 | ||
2573 | /* | |
2574 | * If we already have plenty of memory free for | |
2575 | * compaction in this zone, don't free any more. | |
2576 | * Even though compaction is invoked for any | |
2577 | * non-zero order, only frequent costly order | |
2578 | * reclamation is disruptive enough to become a | |
2579 | * noticeable problem, like transparent huge | |
2580 | * page allocations. | |
2581 | */ | |
2582 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2583 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
2584 | zonelist_zone_idx(z) <= requested_highidx && | |
2585 | compaction_ready(zone, sc->order)) { | |
2586 | sc->compaction_ready = true; | |
2587 | continue; | |
2588 | } | |
2589 | ||
2590 | /* | |
2591 | * This steals pages from memory cgroups over softlimit | |
2592 | * and returns the number of reclaimed pages and | |
2593 | * scanned pages. This works for global memory pressure | |
2594 | * and balancing, not for a memcg's limit. | |
2595 | */ | |
2596 | nr_soft_scanned = 0; | |
2597 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2598 | sc->order, sc->gfp_mask, | |
2599 | &nr_soft_scanned); | |
2600 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2601 | sc->nr_scanned += nr_soft_scanned; | |
2602 | if (nr_soft_reclaimed) | |
2603 | reclaimable = true; | |
2604 | /* need some check for avoid more shrink_zone() */ | |
2605 | } | |
2606 | ||
2607 | if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx)) | |
2608 | reclaimable = true; | |
2609 | ||
2610 | if (global_reclaim(sc) && | |
2611 | !reclaimable && zone_reclaimable(zone)) | |
2612 | reclaimable = true; | |
2613 | } | |
2614 | ||
2615 | /* | |
2616 | * Restore to original mask to avoid the impact on the caller if we | |
2617 | * promoted it to __GFP_HIGHMEM. | |
2618 | */ | |
2619 | sc->gfp_mask = orig_mask; | |
2620 | ||
2621 | return reclaimable; | |
2622 | } | |
2623 | ||
2624 | /* | |
2625 | * This is the main entry point to direct page reclaim. | |
2626 | * | |
2627 | * If a full scan of the inactive list fails to free enough memory then we | |
2628 | * are "out of memory" and something needs to be killed. | |
2629 | * | |
2630 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2631 | * high - the zone may be full of dirty or under-writeback pages, which this | |
2632 | * caller can't do much about. We kick the writeback threads and take explicit | |
2633 | * naps in the hope that some of these pages can be written. But if the | |
2634 | * allocating task holds filesystem locks which prevent writeout this might not | |
2635 | * work, and the allocation attempt will fail. | |
2636 | * | |
2637 | * returns: 0, if no pages reclaimed | |
2638 | * else, the number of pages reclaimed | |
2639 | */ | |
2640 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, | |
2641 | struct scan_control *sc) | |
2642 | { | |
2643 | int initial_priority = sc->priority; | |
2644 | unsigned long total_scanned = 0; | |
2645 | unsigned long writeback_threshold; | |
2646 | bool zones_reclaimable; | |
2647 | retry: | |
2648 | delayacct_freepages_start(); | |
2649 | ||
2650 | if (global_reclaim(sc)) | |
2651 | count_vm_event(ALLOCSTALL); | |
2652 | ||
2653 | do { | |
2654 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, | |
2655 | sc->priority); | |
2656 | sc->nr_scanned = 0; | |
2657 | zones_reclaimable = shrink_zones(zonelist, sc); | |
2658 | ||
2659 | total_scanned += sc->nr_scanned; | |
2660 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) | |
2661 | break; | |
2662 | ||
2663 | if (sc->compaction_ready) | |
2664 | break; | |
2665 | ||
2666 | /* | |
2667 | * If we're getting trouble reclaiming, start doing | |
2668 | * writepage even in laptop mode. | |
2669 | */ | |
2670 | if (sc->priority < DEF_PRIORITY - 2) | |
2671 | sc->may_writepage = 1; | |
2672 | ||
2673 | /* | |
2674 | * Try to write back as many pages as we just scanned. This | |
2675 | * tends to cause slow streaming writers to write data to the | |
2676 | * disk smoothly, at the dirtying rate, which is nice. But | |
2677 | * that's undesirable in laptop mode, where we *want* lumpy | |
2678 | * writeout. So in laptop mode, write out the whole world. | |
2679 | */ | |
2680 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; | |
2681 | if (total_scanned > writeback_threshold) { | |
2682 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, | |
2683 | WB_REASON_TRY_TO_FREE_PAGES); | |
2684 | sc->may_writepage = 1; | |
2685 | } | |
2686 | } while (--sc->priority >= 0); | |
2687 | ||
2688 | delayacct_freepages_end(); | |
2689 | ||
2690 | if (sc->nr_reclaimed) | |
2691 | return sc->nr_reclaimed; | |
2692 | ||
2693 | /* Aborted reclaim to try compaction? don't OOM, then */ | |
2694 | if (sc->compaction_ready) | |
2695 | return 1; | |
2696 | ||
2697 | /* Untapped cgroup reserves? Don't OOM, retry. */ | |
2698 | if (!sc->may_thrash) { | |
2699 | sc->priority = initial_priority; | |
2700 | sc->may_thrash = 1; | |
2701 | goto retry; | |
2702 | } | |
2703 | ||
2704 | /* Any of the zones still reclaimable? Don't OOM. */ | |
2705 | if (zones_reclaimable) | |
2706 | return 1; | |
2707 | ||
2708 | return 0; | |
2709 | } | |
2710 | ||
2711 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) | |
2712 | { | |
2713 | struct zone *zone; | |
2714 | unsigned long pfmemalloc_reserve = 0; | |
2715 | unsigned long free_pages = 0; | |
2716 | int i; | |
2717 | bool wmark_ok; | |
2718 | ||
2719 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2720 | zone = &pgdat->node_zones[i]; | |
2721 | if (!populated_zone(zone) || | |
2722 | zone_reclaimable_pages(zone) == 0) | |
2723 | continue; | |
2724 | ||
2725 | pfmemalloc_reserve += min_wmark_pages(zone); | |
2726 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2727 | } | |
2728 | ||
2729 | /* If there are no reserves (unexpected config) then do not throttle */ | |
2730 | if (!pfmemalloc_reserve) | |
2731 | return true; | |
2732 | ||
2733 | wmark_ok = free_pages > pfmemalloc_reserve / 2; | |
2734 | ||
2735 | /* kswapd must be awake if processes are being throttled */ | |
2736 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2737 | pgdat->classzone_idx = min(pgdat->classzone_idx, | |
2738 | (enum zone_type)ZONE_NORMAL); | |
2739 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2740 | } | |
2741 | ||
2742 | return wmark_ok; | |
2743 | } | |
2744 | ||
2745 | /* | |
2746 | * Throttle direct reclaimers if backing storage is backed by the network | |
2747 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2748 | * depleted. kswapd will continue to make progress and wake the processes | |
2749 | * when the low watermark is reached. | |
2750 | * | |
2751 | * Returns true if a fatal signal was delivered during throttling. If this | |
2752 | * happens, the page allocator should not consider triggering the OOM killer. | |
2753 | */ | |
2754 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, | |
2755 | nodemask_t *nodemask) | |
2756 | { | |
2757 | struct zoneref *z; | |
2758 | struct zone *zone; | |
2759 | pg_data_t *pgdat = NULL; | |
2760 | ||
2761 | /* | |
2762 | * Kernel threads should not be throttled as they may be indirectly | |
2763 | * responsible for cleaning pages necessary for reclaim to make forward | |
2764 | * progress. kjournald for example may enter direct reclaim while | |
2765 | * committing a transaction where throttling it could forcing other | |
2766 | * processes to block on log_wait_commit(). | |
2767 | */ | |
2768 | if (current->flags & PF_KTHREAD) | |
2769 | goto out; | |
2770 | ||
2771 | /* | |
2772 | * If a fatal signal is pending, this process should not throttle. | |
2773 | * It should return quickly so it can exit and free its memory | |
2774 | */ | |
2775 | if (fatal_signal_pending(current)) | |
2776 | goto out; | |
2777 | ||
2778 | /* | |
2779 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2780 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2781 | * GFP_KERNEL will be required for allocating network buffers when | |
2782 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2783 | * | |
2784 | * Throttling is based on the first usable node and throttled processes | |
2785 | * wait on a queue until kswapd makes progress and wakes them. There | |
2786 | * is an affinity then between processes waking up and where reclaim | |
2787 | * progress has been made assuming the process wakes on the same node. | |
2788 | * More importantly, processes running on remote nodes will not compete | |
2789 | * for remote pfmemalloc reserves and processes on different nodes | |
2790 | * should make reasonable progress. | |
2791 | */ | |
2792 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2793 | gfp_zone(gfp_mask), nodemask) { | |
2794 | if (zone_idx(zone) > ZONE_NORMAL) | |
2795 | continue; | |
2796 | ||
2797 | /* Throttle based on the first usable node */ | |
2798 | pgdat = zone->zone_pgdat; | |
2799 | if (pfmemalloc_watermark_ok(pgdat)) | |
2800 | goto out; | |
2801 | break; | |
2802 | } | |
2803 | ||
2804 | /* If no zone was usable by the allocation flags then do not throttle */ | |
2805 | if (!pgdat) | |
2806 | goto out; | |
2807 | ||
2808 | /* Account for the throttling */ | |
2809 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2810 | ||
2811 | /* | |
2812 | * If the caller cannot enter the filesystem, it's possible that it | |
2813 | * is due to the caller holding an FS lock or performing a journal | |
2814 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2815 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2816 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2817 | * second before continuing. | |
2818 | */ | |
2819 | if (!(gfp_mask & __GFP_FS)) { | |
2820 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2821 | pfmemalloc_watermark_ok(pgdat), HZ); | |
2822 | ||
2823 | goto check_pending; | |
2824 | } | |
2825 | ||
2826 | /* Throttle until kswapd wakes the process */ | |
2827 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2828 | pfmemalloc_watermark_ok(pgdat)); | |
2829 | ||
2830 | check_pending: | |
2831 | if (fatal_signal_pending(current)) | |
2832 | return true; | |
2833 | ||
2834 | out: | |
2835 | return false; | |
2836 | } | |
2837 | ||
2838 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, | |
2839 | gfp_t gfp_mask, nodemask_t *nodemask) | |
2840 | { | |
2841 | unsigned long nr_reclaimed; | |
2842 | struct scan_control sc = { | |
2843 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2844 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), | |
2845 | .order = order, | |
2846 | .nodemask = nodemask, | |
2847 | .priority = DEF_PRIORITY, | |
2848 | .may_writepage = !laptop_mode, | |
2849 | .may_unmap = 1, | |
2850 | .may_swap = 1, | |
2851 | }; | |
2852 | ||
2853 | /* | |
2854 | * Do not enter reclaim if fatal signal was delivered while throttled. | |
2855 | * 1 is returned so that the page allocator does not OOM kill at this | |
2856 | * point. | |
2857 | */ | |
2858 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) | |
2859 | return 1; | |
2860 | ||
2861 | trace_mm_vmscan_direct_reclaim_begin(order, | |
2862 | sc.may_writepage, | |
2863 | gfp_mask); | |
2864 | ||
2865 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); | |
2866 | ||
2867 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2868 | ||
2869 | return nr_reclaimed; | |
2870 | } | |
2871 | ||
2872 | #ifdef CONFIG_MEMCG | |
2873 | ||
2874 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, | |
2875 | gfp_t gfp_mask, bool noswap, | |
2876 | struct zone *zone, | |
2877 | unsigned long *nr_scanned) | |
2878 | { | |
2879 | struct scan_control sc = { | |
2880 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2881 | .target_mem_cgroup = memcg, | |
2882 | .may_writepage = !laptop_mode, | |
2883 | .may_unmap = 1, | |
2884 | .may_swap = !noswap, | |
2885 | }; | |
2886 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
2887 | int swappiness = mem_cgroup_swappiness(memcg); | |
2888 | unsigned long lru_pages; | |
2889 | ||
2890 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
2891 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
2892 | ||
2893 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, | |
2894 | sc.may_writepage, | |
2895 | sc.gfp_mask); | |
2896 | ||
2897 | /* | |
2898 | * NOTE: Although we can get the priority field, using it | |
2899 | * here is not a good idea, since it limits the pages we can scan. | |
2900 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2901 | * will pick up pages from other mem cgroup's as well. We hack | |
2902 | * the priority and make it zero. | |
2903 | */ | |
2904 | shrink_lruvec(lruvec, swappiness, &sc, &lru_pages); | |
2905 | ||
2906 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2907 | ||
2908 | *nr_scanned = sc.nr_scanned; | |
2909 | return sc.nr_reclaimed; | |
2910 | } | |
2911 | ||
2912 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, | |
2913 | unsigned long nr_pages, | |
2914 | gfp_t gfp_mask, | |
2915 | bool may_swap) | |
2916 | { | |
2917 | struct zonelist *zonelist; | |
2918 | unsigned long nr_reclaimed; | |
2919 | int nid; | |
2920 | struct scan_control sc = { | |
2921 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), | |
2922 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
2923 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2924 | .target_mem_cgroup = memcg, | |
2925 | .priority = DEF_PRIORITY, | |
2926 | .may_writepage = !laptop_mode, | |
2927 | .may_unmap = 1, | |
2928 | .may_swap = may_swap, | |
2929 | }; | |
2930 | ||
2931 | /* | |
2932 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2933 | * take care of from where we get pages. So the node where we start the | |
2934 | * scan does not need to be the current node. | |
2935 | */ | |
2936 | nid = mem_cgroup_select_victim_node(memcg); | |
2937 | ||
2938 | zonelist = NODE_DATA(nid)->node_zonelists; | |
2939 | ||
2940 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2941 | sc.may_writepage, | |
2942 | sc.gfp_mask); | |
2943 | ||
2944 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); | |
2945 | ||
2946 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2947 | ||
2948 | return nr_reclaimed; | |
2949 | } | |
2950 | #endif | |
2951 | ||
2952 | static void age_active_anon(struct zone *zone, struct scan_control *sc) | |
2953 | { | |
2954 | struct mem_cgroup *memcg; | |
2955 | ||
2956 | if (!total_swap_pages) | |
2957 | return; | |
2958 | ||
2959 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2960 | do { | |
2961 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
2962 | ||
2963 | if (inactive_anon_is_low(lruvec)) | |
2964 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
2965 | sc, LRU_ACTIVE_ANON); | |
2966 | ||
2967 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2968 | } while (memcg); | |
2969 | } | |
2970 | ||
2971 | static bool zone_balanced(struct zone *zone, int order, | |
2972 | unsigned long balance_gap, int classzone_idx) | |
2973 | { | |
2974 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + | |
2975 | balance_gap, classzone_idx)) | |
2976 | return false; | |
2977 | ||
2978 | if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone, | |
2979 | order, 0, classzone_idx) == COMPACT_SKIPPED) | |
2980 | return false; | |
2981 | ||
2982 | return true; | |
2983 | } | |
2984 | ||
2985 | /* | |
2986 | * pgdat_balanced() is used when checking if a node is balanced. | |
2987 | * | |
2988 | * For order-0, all zones must be balanced! | |
2989 | * | |
2990 | * For high-order allocations only zones that meet watermarks and are in a | |
2991 | * zone allowed by the callers classzone_idx are added to balanced_pages. The | |
2992 | * total of balanced pages must be at least 25% of the zones allowed by | |
2993 | * classzone_idx for the node to be considered balanced. Forcing all zones to | |
2994 | * be balanced for high orders can cause excessive reclaim when there are | |
2995 | * imbalanced zones. | |
2996 | * The choice of 25% is due to | |
2997 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2998 | * reasonable sized machine | |
2999 | * o On all other machines, the top zone must be at least a reasonable | |
3000 | * percentage of the middle zones. For example, on 32-bit x86, highmem | |
3001 | * would need to be at least 256M for it to be balance a whole node. | |
3002 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
3003 | * to balance a node on its own. These seemed like reasonable ratios. | |
3004 | */ | |
3005 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
3006 | { | |
3007 | unsigned long managed_pages = 0; | |
3008 | unsigned long balanced_pages = 0; | |
3009 | int i; | |
3010 | ||
3011 | /* Check the watermark levels */ | |
3012 | for (i = 0; i <= classzone_idx; i++) { | |
3013 | struct zone *zone = pgdat->node_zones + i; | |
3014 | ||
3015 | if (!populated_zone(zone)) | |
3016 | continue; | |
3017 | ||
3018 | managed_pages += zone->managed_pages; | |
3019 | ||
3020 | /* | |
3021 | * A special case here: | |
3022 | * | |
3023 | * balance_pgdat() skips over all_unreclaimable after | |
3024 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
3025 | * they must be considered balanced here as well! | |
3026 | */ | |
3027 | if (!zone_reclaimable(zone)) { | |
3028 | balanced_pages += zone->managed_pages; | |
3029 | continue; | |
3030 | } | |
3031 | ||
3032 | if (zone_balanced(zone, order, 0, i)) | |
3033 | balanced_pages += zone->managed_pages; | |
3034 | else if (!order) | |
3035 | return false; | |
3036 | } | |
3037 | ||
3038 | if (order) | |
3039 | return balanced_pages >= (managed_pages >> 2); | |
3040 | else | |
3041 | return true; | |
3042 | } | |
3043 | ||
3044 | /* | |
3045 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3046 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3047 | * | |
3048 | * Returns true if kswapd is ready to sleep | |
3049 | */ | |
3050 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, | |
3051 | int classzone_idx) | |
3052 | { | |
3053 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ | |
3054 | if (remaining) | |
3055 | return false; | |
3056 | ||
3057 | /* | |
3058 | * The throttled processes are normally woken up in balance_pgdat() as | |
3059 | * soon as pfmemalloc_watermark_ok() is true. But there is a potential | |
3060 | * race between when kswapd checks the watermarks and a process gets | |
3061 | * throttled. There is also a potential race if processes get | |
3062 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3063 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3064 | * the wake up checks. If kswapd is going to sleep, no process should | |
3065 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3066 | * the wake up is premature, processes will wake kswapd and get | |
3067 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3068 | * that here we are under prepare_to_wait(). | |
3069 | */ | |
3070 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) | |
3071 | wake_up_all(&pgdat->pfmemalloc_wait); | |
3072 | ||
3073 | return pgdat_balanced(pgdat, order, classzone_idx); | |
3074 | } | |
3075 | ||
3076 | /* | |
3077 | * kswapd shrinks the zone by the number of pages required to reach | |
3078 | * the high watermark. | |
3079 | * | |
3080 | * Returns true if kswapd scanned at least the requested number of pages to | |
3081 | * reclaim or if the lack of progress was due to pages under writeback. | |
3082 | * This is used to determine if the scanning priority needs to be raised. | |
3083 | */ | |
3084 | static bool kswapd_shrink_zone(struct zone *zone, | |
3085 | int classzone_idx, | |
3086 | struct scan_control *sc, | |
3087 | unsigned long *nr_attempted) | |
3088 | { | |
3089 | int testorder = sc->order; | |
3090 | unsigned long balance_gap; | |
3091 | bool lowmem_pressure; | |
3092 | ||
3093 | /* Reclaim above the high watermark. */ | |
3094 | sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone)); | |
3095 | ||
3096 | /* | |
3097 | * Kswapd reclaims only single pages with compaction enabled. Trying | |
3098 | * too hard to reclaim until contiguous free pages have become | |
3099 | * available can hurt performance by evicting too much useful data | |
3100 | * from memory. Do not reclaim more than needed for compaction. | |
3101 | */ | |
3102 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && | |
3103 | compaction_suitable(zone, sc->order, 0, classzone_idx) | |
3104 | != COMPACT_SKIPPED) | |
3105 | testorder = 0; | |
3106 | ||
3107 | /* | |
3108 | * We put equal pressure on every zone, unless one zone has way too | |
3109 | * many pages free already. The "too many pages" is defined as the | |
3110 | * high wmark plus a "gap" where the gap is either the low | |
3111 | * watermark or 1% of the zone, whichever is smaller. | |
3112 | */ | |
3113 | balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( | |
3114 | zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); | |
3115 | ||
3116 | /* | |
3117 | * If there is no low memory pressure or the zone is balanced then no | |
3118 | * reclaim is necessary | |
3119 | */ | |
3120 | lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone)); | |
3121 | if (!lowmem_pressure && zone_balanced(zone, testorder, | |
3122 | balance_gap, classzone_idx)) | |
3123 | return true; | |
3124 | ||
3125 | shrink_zone(zone, sc, zone_idx(zone) == classzone_idx); | |
3126 | ||
3127 | /* Account for the number of pages attempted to reclaim */ | |
3128 | *nr_attempted += sc->nr_to_reclaim; | |
3129 | ||
3130 | clear_bit(ZONE_WRITEBACK, &zone->flags); | |
3131 | ||
3132 | /* | |
3133 | * If a zone reaches its high watermark, consider it to be no longer | |
3134 | * congested. It's possible there are dirty pages backed by congested | |
3135 | * BDIs but as pressure is relieved, speculatively avoid congestion | |
3136 | * waits. | |
3137 | */ | |
3138 | if (zone_reclaimable(zone) && | |
3139 | zone_balanced(zone, testorder, 0, classzone_idx)) { | |
3140 | clear_bit(ZONE_CONGESTED, &zone->flags); | |
3141 | clear_bit(ZONE_DIRTY, &zone->flags); | |
3142 | } | |
3143 | ||
3144 | return sc->nr_scanned >= sc->nr_to_reclaim; | |
3145 | } | |
3146 | ||
3147 | /* | |
3148 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
3149 | * they are all at high_wmark_pages(zone). | |
3150 | * | |
3151 | * Returns the final order kswapd was reclaiming at | |
3152 | * | |
3153 | * There is special handling here for zones which are full of pinned pages. | |
3154 | * This can happen if the pages are all mlocked, or if they are all used by | |
3155 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
3156 | * What we do is to detect the case where all pages in the zone have been | |
3157 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
3158 | * dead and from now on, only perform a short scan. Basically we're polling | |
3159 | * the zone for when the problem goes away. | |
3160 | * | |
3161 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
3162 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is | |
3163 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
3164 | * lower zones regardless of the number of free pages in the lower zones. This | |
3165 | * interoperates with the page allocator fallback scheme to ensure that aging | |
3166 | * of pages is balanced across the zones. | |
3167 | */ | |
3168 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, | |
3169 | int *classzone_idx) | |
3170 | { | |
3171 | int i; | |
3172 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ | |
3173 | unsigned long nr_soft_reclaimed; | |
3174 | unsigned long nr_soft_scanned; | |
3175 | struct scan_control sc = { | |
3176 | .gfp_mask = GFP_KERNEL, | |
3177 | .order = order, | |
3178 | .priority = DEF_PRIORITY, | |
3179 | .may_writepage = !laptop_mode, | |
3180 | .may_unmap = 1, | |
3181 | .may_swap = 1, | |
3182 | }; | |
3183 | count_vm_event(PAGEOUTRUN); | |
3184 | ||
3185 | do { | |
3186 | unsigned long nr_attempted = 0; | |
3187 | bool raise_priority = true; | |
3188 | bool pgdat_needs_compaction = (order > 0); | |
3189 | ||
3190 | sc.nr_reclaimed = 0; | |
3191 | ||
3192 | /* | |
3193 | * Scan in the highmem->dma direction for the highest | |
3194 | * zone which needs scanning | |
3195 | */ | |
3196 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
3197 | struct zone *zone = pgdat->node_zones + i; | |
3198 | ||
3199 | if (!populated_zone(zone)) | |
3200 | continue; | |
3201 | ||
3202 | if (sc.priority != DEF_PRIORITY && | |
3203 | !zone_reclaimable(zone)) | |
3204 | continue; | |
3205 | ||
3206 | /* | |
3207 | * Do some background aging of the anon list, to give | |
3208 | * pages a chance to be referenced before reclaiming. | |
3209 | */ | |
3210 | age_active_anon(zone, &sc); | |
3211 | ||
3212 | /* | |
3213 | * If the number of buffer_heads in the machine | |
3214 | * exceeds the maximum allowed level and this node | |
3215 | * has a highmem zone, force kswapd to reclaim from | |
3216 | * it to relieve lowmem pressure. | |
3217 | */ | |
3218 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
3219 | end_zone = i; | |
3220 | break; | |
3221 | } | |
3222 | ||
3223 | if (!zone_balanced(zone, order, 0, 0)) { | |
3224 | end_zone = i; | |
3225 | break; | |
3226 | } else { | |
3227 | /* | |
3228 | * If balanced, clear the dirty and congested | |
3229 | * flags | |
3230 | */ | |
3231 | clear_bit(ZONE_CONGESTED, &zone->flags); | |
3232 | clear_bit(ZONE_DIRTY, &zone->flags); | |
3233 | } | |
3234 | } | |
3235 | ||
3236 | if (i < 0) | |
3237 | goto out; | |
3238 | ||
3239 | for (i = 0; i <= end_zone; i++) { | |
3240 | struct zone *zone = pgdat->node_zones + i; | |
3241 | ||
3242 | if (!populated_zone(zone)) | |
3243 | continue; | |
3244 | ||
3245 | /* | |
3246 | * If any zone is currently balanced then kswapd will | |
3247 | * not call compaction as it is expected that the | |
3248 | * necessary pages are already available. | |
3249 | */ | |
3250 | if (pgdat_needs_compaction && | |
3251 | zone_watermark_ok(zone, order, | |
3252 | low_wmark_pages(zone), | |
3253 | *classzone_idx, 0)) | |
3254 | pgdat_needs_compaction = false; | |
3255 | } | |
3256 | ||
3257 | /* | |
3258 | * If we're getting trouble reclaiming, start doing writepage | |
3259 | * even in laptop mode. | |
3260 | */ | |
3261 | if (sc.priority < DEF_PRIORITY - 2) | |
3262 | sc.may_writepage = 1; | |
3263 | ||
3264 | /* | |
3265 | * Now scan the zone in the dma->highmem direction, stopping | |
3266 | * at the last zone which needs scanning. | |
3267 | * | |
3268 | * We do this because the page allocator works in the opposite | |
3269 | * direction. This prevents the page allocator from allocating | |
3270 | * pages behind kswapd's direction of progress, which would | |
3271 | * cause too much scanning of the lower zones. | |
3272 | */ | |
3273 | for (i = 0; i <= end_zone; i++) { | |
3274 | struct zone *zone = pgdat->node_zones + i; | |
3275 | ||
3276 | if (!populated_zone(zone)) | |
3277 | continue; | |
3278 | ||
3279 | if (sc.priority != DEF_PRIORITY && | |
3280 | !zone_reclaimable(zone)) | |
3281 | continue; | |
3282 | ||
3283 | sc.nr_scanned = 0; | |
3284 | ||
3285 | nr_soft_scanned = 0; | |
3286 | /* | |
3287 | * Call soft limit reclaim before calling shrink_zone. | |
3288 | */ | |
3289 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
3290 | order, sc.gfp_mask, | |
3291 | &nr_soft_scanned); | |
3292 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3293 | ||
3294 | /* | |
3295 | * There should be no need to raise the scanning | |
3296 | * priority if enough pages are already being scanned | |
3297 | * that that high watermark would be met at 100% | |
3298 | * efficiency. | |
3299 | */ | |
3300 | if (kswapd_shrink_zone(zone, end_zone, | |
3301 | &sc, &nr_attempted)) | |
3302 | raise_priority = false; | |
3303 | } | |
3304 | ||
3305 | /* | |
3306 | * If the low watermark is met there is no need for processes | |
3307 | * to be throttled on pfmemalloc_wait as they should not be | |
3308 | * able to safely make forward progress. Wake them | |
3309 | */ | |
3310 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
3311 | pfmemalloc_watermark_ok(pgdat)) | |
3312 | wake_up_all(&pgdat->pfmemalloc_wait); | |
3313 | ||
3314 | /* | |
3315 | * Fragmentation may mean that the system cannot be rebalanced | |
3316 | * for high-order allocations in all zones. If twice the | |
3317 | * allocation size has been reclaimed and the zones are still | |
3318 | * not balanced then recheck the watermarks at order-0 to | |
3319 | * prevent kswapd reclaiming excessively. Assume that a | |
3320 | * process requested a high-order can direct reclaim/compact. | |
3321 | */ | |
3322 | if (order && sc.nr_reclaimed >= 2UL << order) | |
3323 | order = sc.order = 0; | |
3324 | ||
3325 | /* Check if kswapd should be suspending */ | |
3326 | if (try_to_freeze() || kthread_should_stop()) | |
3327 | break; | |
3328 | ||
3329 | /* | |
3330 | * Compact if necessary and kswapd is reclaiming at least the | |
3331 | * high watermark number of pages as requsted | |
3332 | */ | |
3333 | if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted) | |
3334 | compact_pgdat(pgdat, order); | |
3335 | ||
3336 | /* | |
3337 | * Raise priority if scanning rate is too low or there was no | |
3338 | * progress in reclaiming pages | |
3339 | */ | |
3340 | if (raise_priority || !sc.nr_reclaimed) | |
3341 | sc.priority--; | |
3342 | } while (sc.priority >= 1 && | |
3343 | !pgdat_balanced(pgdat, order, *classzone_idx)); | |
3344 | ||
3345 | out: | |
3346 | /* | |
3347 | * Return the order we were reclaiming at so prepare_kswapd_sleep() | |
3348 | * makes a decision on the order we were last reclaiming at. However, | |
3349 | * if another caller entered the allocator slow path while kswapd | |
3350 | * was awake, order will remain at the higher level | |
3351 | */ | |
3352 | *classzone_idx = end_zone; | |
3353 | return order; | |
3354 | } | |
3355 | ||
3356 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) | |
3357 | { | |
3358 | long remaining = 0; | |
3359 | DEFINE_WAIT(wait); | |
3360 | ||
3361 | if (freezing(current) || kthread_should_stop()) | |
3362 | return; | |
3363 | ||
3364 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3365 | ||
3366 | /* Try to sleep for a short interval */ | |
3367 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { | |
3368 | remaining = schedule_timeout(HZ/10); | |
3369 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3370 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3371 | } | |
3372 | ||
3373 | /* | |
3374 | * After a short sleep, check if it was a premature sleep. If not, then | |
3375 | * go fully to sleep until explicitly woken up. | |
3376 | */ | |
3377 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { | |
3378 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); | |
3379 | ||
3380 | /* | |
3381 | * vmstat counters are not perfectly accurate and the estimated | |
3382 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3383 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3384 | * watermarks being breached while under pressure, we reduce the | |
3385 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3386 | * them before going back to sleep. | |
3387 | */ | |
3388 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
3389 | ||
3390 | /* | |
3391 | * Compaction records what page blocks it recently failed to | |
3392 | * isolate pages from and skips them in the future scanning. | |
3393 | * When kswapd is going to sleep, it is reasonable to assume | |
3394 | * that pages and compaction may succeed so reset the cache. | |
3395 | */ | |
3396 | reset_isolation_suitable(pgdat); | |
3397 | ||
3398 | if (!kthread_should_stop()) | |
3399 | schedule(); | |
3400 | ||
3401 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); | |
3402 | } else { | |
3403 | if (remaining) | |
3404 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3405 | else | |
3406 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3407 | } | |
3408 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3409 | } | |
3410 | ||
3411 | /* | |
3412 | * The background pageout daemon, started as a kernel thread | |
3413 | * from the init process. | |
3414 | * | |
3415 | * This basically trickles out pages so that we have _some_ | |
3416 | * free memory available even if there is no other activity | |
3417 | * that frees anything up. This is needed for things like routing | |
3418 | * etc, where we otherwise might have all activity going on in | |
3419 | * asynchronous contexts that cannot page things out. | |
3420 | * | |
3421 | * If there are applications that are active memory-allocators | |
3422 | * (most normal use), this basically shouldn't matter. | |
3423 | */ | |
3424 | static int kswapd(void *p) | |
3425 | { | |
3426 | unsigned long order, new_order; | |
3427 | unsigned balanced_order; | |
3428 | int classzone_idx, new_classzone_idx; | |
3429 | int balanced_classzone_idx; | |
3430 | pg_data_t *pgdat = (pg_data_t*)p; | |
3431 | struct task_struct *tsk = current; | |
3432 | ||
3433 | struct reclaim_state reclaim_state = { | |
3434 | .reclaimed_slab = 0, | |
3435 | }; | |
3436 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
3437 | ||
3438 | lockdep_set_current_reclaim_state(GFP_KERNEL); | |
3439 | ||
3440 | if (!cpumask_empty(cpumask)) | |
3441 | set_cpus_allowed_ptr(tsk, cpumask); | |
3442 | current->reclaim_state = &reclaim_state; | |
3443 | ||
3444 | /* | |
3445 | * Tell the memory management that we're a "memory allocator", | |
3446 | * and that if we need more memory we should get access to it | |
3447 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3448 | * never get caught in the normal page freeing logic. | |
3449 | * | |
3450 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3451 | * you need a small amount of memory in order to be able to | |
3452 | * page out something else, and this flag essentially protects | |
3453 | * us from recursively trying to free more memory as we're | |
3454 | * trying to free the first piece of memory in the first place). | |
3455 | */ | |
3456 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; | |
3457 | set_freezable(); | |
3458 | ||
3459 | order = new_order = 0; | |
3460 | balanced_order = 0; | |
3461 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; | |
3462 | balanced_classzone_idx = classzone_idx; | |
3463 | for ( ; ; ) { | |
3464 | bool ret; | |
3465 | ||
3466 | /* | |
3467 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
3468 | * new request of a similar or harder type will succeed soon | |
3469 | * so consider going to sleep on the basis we reclaimed at | |
3470 | */ | |
3471 | if (balanced_classzone_idx >= new_classzone_idx && | |
3472 | balanced_order == new_order) { | |
3473 | new_order = pgdat->kswapd_max_order; | |
3474 | new_classzone_idx = pgdat->classzone_idx; | |
3475 | pgdat->kswapd_max_order = 0; | |
3476 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3477 | } | |
3478 | ||
3479 | if (order < new_order || classzone_idx > new_classzone_idx) { | |
3480 | /* | |
3481 | * Don't sleep if someone wants a larger 'order' | |
3482 | * allocation or has tigher zone constraints | |
3483 | */ | |
3484 | order = new_order; | |
3485 | classzone_idx = new_classzone_idx; | |
3486 | } else { | |
3487 | kswapd_try_to_sleep(pgdat, balanced_order, | |
3488 | balanced_classzone_idx); | |
3489 | order = pgdat->kswapd_max_order; | |
3490 | classzone_idx = pgdat->classzone_idx; | |
3491 | new_order = order; | |
3492 | new_classzone_idx = classzone_idx; | |
3493 | pgdat->kswapd_max_order = 0; | |
3494 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3495 | } | |
3496 | ||
3497 | ret = try_to_freeze(); | |
3498 | if (kthread_should_stop()) | |
3499 | break; | |
3500 | ||
3501 | /* | |
3502 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3503 | * after returning from the refrigerator | |
3504 | */ | |
3505 | if (!ret) { | |
3506 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
3507 | balanced_classzone_idx = classzone_idx; | |
3508 | balanced_order = balance_pgdat(pgdat, order, | |
3509 | &balanced_classzone_idx); | |
3510 | } | |
3511 | } | |
3512 | ||
3513 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); | |
3514 | current->reclaim_state = NULL; | |
3515 | lockdep_clear_current_reclaim_state(); | |
3516 | ||
3517 | return 0; | |
3518 | } | |
3519 | ||
3520 | /* | |
3521 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3522 | */ | |
3523 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) | |
3524 | { | |
3525 | pg_data_t *pgdat; | |
3526 | ||
3527 | if (!populated_zone(zone)) | |
3528 | return; | |
3529 | ||
3530 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) | |
3531 | return; | |
3532 | pgdat = zone->zone_pgdat; | |
3533 | if (pgdat->kswapd_max_order < order) { | |
3534 | pgdat->kswapd_max_order = order; | |
3535 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); | |
3536 | } | |
3537 | if (!waitqueue_active(&pgdat->kswapd_wait)) | |
3538 | return; | |
3539 | if (zone_balanced(zone, order, 0, 0)) | |
3540 | return; | |
3541 | ||
3542 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
3543 | wake_up_interruptible(&pgdat->kswapd_wait); | |
3544 | } | |
3545 | ||
3546 | #ifdef CONFIG_HIBERNATION | |
3547 | /* | |
3548 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of | |
3549 | * freed pages. | |
3550 | * | |
3551 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3552 | * LRU order by reclaiming preferentially | |
3553 | * inactive > active > active referenced > active mapped | |
3554 | */ | |
3555 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) | |
3556 | { | |
3557 | struct reclaim_state reclaim_state; | |
3558 | struct scan_control sc = { | |
3559 | .nr_to_reclaim = nr_to_reclaim, | |
3560 | .gfp_mask = GFP_HIGHUSER_MOVABLE, | |
3561 | .priority = DEF_PRIORITY, | |
3562 | .may_writepage = 1, | |
3563 | .may_unmap = 1, | |
3564 | .may_swap = 1, | |
3565 | .hibernation_mode = 1, | |
3566 | }; | |
3567 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
3568 | struct task_struct *p = current; | |
3569 | unsigned long nr_reclaimed; | |
3570 | ||
3571 | p->flags |= PF_MEMALLOC; | |
3572 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3573 | reclaim_state.reclaimed_slab = 0; | |
3574 | p->reclaim_state = &reclaim_state; | |
3575 | ||
3576 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); | |
3577 | ||
3578 | p->reclaim_state = NULL; | |
3579 | lockdep_clear_current_reclaim_state(); | |
3580 | p->flags &= ~PF_MEMALLOC; | |
3581 | ||
3582 | return nr_reclaimed; | |
3583 | } | |
3584 | #endif /* CONFIG_HIBERNATION */ | |
3585 | ||
3586 | /* It's optimal to keep kswapds on the same CPUs as their memory, but | |
3587 | not required for correctness. So if the last cpu in a node goes | |
3588 | away, we get changed to run anywhere: as the first one comes back, | |
3589 | restore their cpu bindings. */ | |
3590 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, | |
3591 | void *hcpu) | |
3592 | { | |
3593 | int nid; | |
3594 | ||
3595 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { | |
3596 | for_each_node_state(nid, N_MEMORY) { | |
3597 | pg_data_t *pgdat = NODE_DATA(nid); | |
3598 | const struct cpumask *mask; | |
3599 | ||
3600 | mask = cpumask_of_node(pgdat->node_id); | |
3601 | ||
3602 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) | |
3603 | /* One of our CPUs online: restore mask */ | |
3604 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
3605 | } | |
3606 | } | |
3607 | return NOTIFY_OK; | |
3608 | } | |
3609 | ||
3610 | /* | |
3611 | * This kswapd start function will be called by init and node-hot-add. | |
3612 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3613 | */ | |
3614 | int kswapd_run(int nid) | |
3615 | { | |
3616 | pg_data_t *pgdat = NODE_DATA(nid); | |
3617 | int ret = 0; | |
3618 | ||
3619 | if (pgdat->kswapd) | |
3620 | return 0; | |
3621 | ||
3622 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3623 | if (IS_ERR(pgdat->kswapd)) { | |
3624 | /* failure at boot is fatal */ | |
3625 | BUG_ON(system_state == SYSTEM_BOOTING); | |
3626 | pr_err("Failed to start kswapd on node %d\n", nid); | |
3627 | ret = PTR_ERR(pgdat->kswapd); | |
3628 | pgdat->kswapd = NULL; | |
3629 | } | |
3630 | return ret; | |
3631 | } | |
3632 | ||
3633 | /* | |
3634 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
3635 | * hold mem_hotplug_begin/end(). | |
3636 | */ | |
3637 | void kswapd_stop(int nid) | |
3638 | { | |
3639 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3640 | ||
3641 | if (kswapd) { | |
3642 | kthread_stop(kswapd); | |
3643 | NODE_DATA(nid)->kswapd = NULL; | |
3644 | } | |
3645 | } | |
3646 | ||
3647 | static int __init kswapd_init(void) | |
3648 | { | |
3649 | int nid; | |
3650 | ||
3651 | swap_setup(); | |
3652 | for_each_node_state(nid, N_MEMORY) | |
3653 | kswapd_run(nid); | |
3654 | hotcpu_notifier(cpu_callback, 0); | |
3655 | return 0; | |
3656 | } | |
3657 | ||
3658 | module_init(kswapd_init) | |
3659 | ||
3660 | #ifdef CONFIG_NUMA | |
3661 | /* | |
3662 | * Zone reclaim mode | |
3663 | * | |
3664 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3665 | * the watermarks. | |
3666 | */ | |
3667 | int zone_reclaim_mode __read_mostly; | |
3668 | ||
3669 | #define RECLAIM_OFF 0 | |
3670 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ | |
3671 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ | |
3672 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ | |
3673 | ||
3674 | /* | |
3675 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3676 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3677 | * a zone. | |
3678 | */ | |
3679 | #define ZONE_RECLAIM_PRIORITY 4 | |
3680 | ||
3681 | /* | |
3682 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3683 | * occur. | |
3684 | */ | |
3685 | int sysctl_min_unmapped_ratio = 1; | |
3686 | ||
3687 | /* | |
3688 | * If the number of slab pages in a zone grows beyond this percentage then | |
3689 | * slab reclaim needs to occur. | |
3690 | */ | |
3691 | int sysctl_min_slab_ratio = 5; | |
3692 | ||
3693 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) | |
3694 | { | |
3695 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3696 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3697 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3698 | ||
3699 | /* | |
3700 | * It's possible for there to be more file mapped pages than | |
3701 | * accounted for by the pages on the file LRU lists because | |
3702 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3703 | */ | |
3704 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3705 | } | |
3706 | ||
3707 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3708 | static unsigned long zone_pagecache_reclaimable(struct zone *zone) | |
3709 | { | |
3710 | unsigned long nr_pagecache_reclaimable; | |
3711 | unsigned long delta = 0; | |
3712 | ||
3713 | /* | |
3714 | * If RECLAIM_UNMAP is set, then all file pages are considered | |
3715 | * potentially reclaimable. Otherwise, we have to worry about | |
3716 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3717 | * a better estimate | |
3718 | */ | |
3719 | if (zone_reclaim_mode & RECLAIM_UNMAP) | |
3720 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3721 | else | |
3722 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3723 | ||
3724 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3725 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3726 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3727 | ||
3728 | /* Watch for any possible underflows due to delta */ | |
3729 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3730 | delta = nr_pagecache_reclaimable; | |
3731 | ||
3732 | return nr_pagecache_reclaimable - delta; | |
3733 | } | |
3734 | ||
3735 | /* | |
3736 | * Try to free up some pages from this zone through reclaim. | |
3737 | */ | |
3738 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3739 | { | |
3740 | /* Minimum pages needed in order to stay on node */ | |
3741 | const unsigned long nr_pages = 1 << order; | |
3742 | struct task_struct *p = current; | |
3743 | struct reclaim_state reclaim_state; | |
3744 | struct scan_control sc = { | |
3745 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), | |
3746 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), | |
3747 | .order = order, | |
3748 | .priority = ZONE_RECLAIM_PRIORITY, | |
3749 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
3750 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP), | |
3751 | .may_swap = 1, | |
3752 | }; | |
3753 | ||
3754 | cond_resched(); | |
3755 | /* | |
3756 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP | |
3757 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3758 | * and RECLAIM_UNMAP. | |
3759 | */ | |
3760 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
3761 | lockdep_set_current_reclaim_state(gfp_mask); | |
3762 | reclaim_state.reclaimed_slab = 0; | |
3763 | p->reclaim_state = &reclaim_state; | |
3764 | ||
3765 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { | |
3766 | /* | |
3767 | * Free memory by calling shrink zone with increasing | |
3768 | * priorities until we have enough memory freed. | |
3769 | */ | |
3770 | do { | |
3771 | shrink_zone(zone, &sc, true); | |
3772 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); | |
3773 | } | |
3774 | ||
3775 | p->reclaim_state = NULL; | |
3776 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); | |
3777 | lockdep_clear_current_reclaim_state(); | |
3778 | return sc.nr_reclaimed >= nr_pages; | |
3779 | } | |
3780 | ||
3781 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3782 | { | |
3783 | int node_id; | |
3784 | int ret; | |
3785 | ||
3786 | /* | |
3787 | * Zone reclaim reclaims unmapped file backed pages and | |
3788 | * slab pages if we are over the defined limits. | |
3789 | * | |
3790 | * A small portion of unmapped file backed pages is needed for | |
3791 | * file I/O otherwise pages read by file I/O will be immediately | |
3792 | * thrown out if the zone is overallocated. So we do not reclaim | |
3793 | * if less than a specified percentage of the zone is used by | |
3794 | * unmapped file backed pages. | |
3795 | */ | |
3796 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && | |
3797 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
3798 | return ZONE_RECLAIM_FULL; | |
3799 | ||
3800 | if (!zone_reclaimable(zone)) | |
3801 | return ZONE_RECLAIM_FULL; | |
3802 | ||
3803 | /* | |
3804 | * Do not scan if the allocation should not be delayed. | |
3805 | */ | |
3806 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) | |
3807 | return ZONE_RECLAIM_NOSCAN; | |
3808 | ||
3809 | /* | |
3810 | * Only run zone reclaim on the local zone or on zones that do not | |
3811 | * have associated processors. This will favor the local processor | |
3812 | * over remote processors and spread off node memory allocations | |
3813 | * as wide as possible. | |
3814 | */ | |
3815 | node_id = zone_to_nid(zone); | |
3816 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) | |
3817 | return ZONE_RECLAIM_NOSCAN; | |
3818 | ||
3819 | if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags)) | |
3820 | return ZONE_RECLAIM_NOSCAN; | |
3821 | ||
3822 | ret = __zone_reclaim(zone, gfp_mask, order); | |
3823 | clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags); | |
3824 | ||
3825 | if (!ret) | |
3826 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3827 | ||
3828 | return ret; | |
3829 | } | |
3830 | #endif | |
3831 | ||
3832 | /* | |
3833 | * page_evictable - test whether a page is evictable | |
3834 | * @page: the page to test | |
3835 | * | |
3836 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
3837 | * lists vs unevictable list. | |
3838 | * | |
3839 | * Reasons page might not be evictable: | |
3840 | * (1) page's mapping marked unevictable | |
3841 | * (2) page is part of an mlocked VMA | |
3842 | * | |
3843 | */ | |
3844 | int page_evictable(struct page *page) | |
3845 | { | |
3846 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); | |
3847 | } | |
3848 | ||
3849 | #ifdef CONFIG_SHMEM | |
3850 | /** | |
3851 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list | |
3852 | * @pages: array of pages to check | |
3853 | * @nr_pages: number of pages to check | |
3854 | * | |
3855 | * Checks pages for evictability and moves them to the appropriate lru list. | |
3856 | * | |
3857 | * This function is only used for SysV IPC SHM_UNLOCK. | |
3858 | */ | |
3859 | void check_move_unevictable_pages(struct page **pages, int nr_pages) | |
3860 | { | |
3861 | struct lruvec *lruvec; | |
3862 | struct zone *zone = NULL; | |
3863 | int pgscanned = 0; | |
3864 | int pgrescued = 0; | |
3865 | int i; | |
3866 | ||
3867 | for (i = 0; i < nr_pages; i++) { | |
3868 | struct page *page = pages[i]; | |
3869 | struct zone *pagezone; | |
3870 | ||
3871 | pgscanned++; | |
3872 | pagezone = page_zone(page); | |
3873 | if (pagezone != zone) { | |
3874 | if (zone) | |
3875 | spin_unlock_irq(&zone->lru_lock); | |
3876 | zone = pagezone; | |
3877 | spin_lock_irq(&zone->lru_lock); | |
3878 | } | |
3879 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
3880 | ||
3881 | if (!PageLRU(page) || !PageUnevictable(page)) | |
3882 | continue; | |
3883 | ||
3884 | if (page_evictable(page)) { | |
3885 | enum lru_list lru = page_lru_base_type(page); | |
3886 | ||
3887 | VM_BUG_ON_PAGE(PageActive(page), page); | |
3888 | ClearPageUnevictable(page); | |
3889 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); | |
3890 | add_page_to_lru_list(page, lruvec, lru); | |
3891 | pgrescued++; | |
3892 | } | |
3893 | } | |
3894 | ||
3895 | if (zone) { | |
3896 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3897 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3898 | spin_unlock_irq(&zone->lru_lock); | |
3899 | } | |
3900 | } | |
3901 | #endif /* CONFIG_SHMEM */ |