]>
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/sched/mm.h> | |
18 | #include <linux/module.h> | |
19 | #include <linux/gfp.h> | |
20 | #include <linux/kernel_stat.h> | |
21 | #include <linux/swap.h> | |
22 | #include <linux/pagemap.h> | |
23 | #include <linux/init.h> | |
24 | #include <linux/highmem.h> | |
25 | #include <linux/vmpressure.h> | |
26 | #include <linux/vmstat.h> | |
27 | #include <linux/file.h> | |
28 | #include <linux/writeback.h> | |
29 | #include <linux/blkdev.h> | |
30 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
31 | buffer_heads_over_limit */ | |
32 | #include <linux/mm_inline.h> | |
33 | #include <linux/backing-dev.h> | |
34 | #include <linux/rmap.h> | |
35 | #include <linux/topology.h> | |
36 | #include <linux/cpu.h> | |
37 | #include <linux/cpuset.h> | |
38 | #include <linux/compaction.h> | |
39 | #include <linux/notifier.h> | |
40 | #include <linux/rwsem.h> | |
41 | #include <linux/delay.h> | |
42 | #include <linux/kthread.h> | |
43 | #include <linux/freezer.h> | |
44 | #include <linux/memcontrol.h> | |
45 | #include <linux/delayacct.h> | |
46 | #include <linux/sysctl.h> | |
47 | #include <linux/oom.h> | |
48 | #include <linux/prefetch.h> | |
49 | #include <linux/printk.h> | |
50 | #include <linux/dax.h> | |
51 | ||
52 | #include <asm/tlbflush.h> | |
53 | #include <asm/div64.h> | |
54 | ||
55 | #include <linux/swapops.h> | |
56 | #include <linux/balloon_compaction.h> | |
57 | ||
58 | #include "internal.h" | |
59 | ||
60 | #define CREATE_TRACE_POINTS | |
61 | #include <trace/events/vmscan.h> | |
62 | ||
63 | struct scan_control { | |
64 | /* How many pages shrink_list() should reclaim */ | |
65 | unsigned long nr_to_reclaim; | |
66 | ||
67 | /* This context's GFP mask */ | |
68 | gfp_t gfp_mask; | |
69 | ||
70 | /* Allocation order */ | |
71 | int order; | |
72 | ||
73 | /* | |
74 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
75 | * are scanned. | |
76 | */ | |
77 | nodemask_t *nodemask; | |
78 | ||
79 | /* | |
80 | * The memory cgroup that hit its limit and as a result is the | |
81 | * primary target of this reclaim invocation. | |
82 | */ | |
83 | struct mem_cgroup *target_mem_cgroup; | |
84 | ||
85 | /* Scan (total_size >> priority) pages at once */ | |
86 | int priority; | |
87 | ||
88 | /* The highest zone to isolate pages for reclaim from */ | |
89 | enum zone_type reclaim_idx; | |
90 | ||
91 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ | |
92 | unsigned int may_writepage:1; | |
93 | ||
94 | /* Can mapped pages be reclaimed? */ | |
95 | unsigned int may_unmap:1; | |
96 | ||
97 | /* Can pages be swapped as part of reclaim? */ | |
98 | unsigned int may_swap:1; | |
99 | ||
100 | /* | |
101 | * Cgroups are not reclaimed below their configured memory.low, | |
102 | * unless we threaten to OOM. If any cgroups are skipped due to | |
103 | * memory.low and nothing was reclaimed, go back for memory.low. | |
104 | */ | |
105 | unsigned int memcg_low_reclaim:1; | |
106 | unsigned int memcg_low_skipped:1; | |
107 | ||
108 | unsigned int hibernation_mode:1; | |
109 | ||
110 | /* One of the zones is ready for compaction */ | |
111 | unsigned int compaction_ready:1; | |
112 | ||
113 | /* Incremented by the number of inactive pages that were scanned */ | |
114 | unsigned long nr_scanned; | |
115 | ||
116 | /* Number of pages freed so far during a call to shrink_zones() */ | |
117 | unsigned long nr_reclaimed; | |
118 | }; | |
119 | ||
120 | #ifdef ARCH_HAS_PREFETCH | |
121 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
122 | do { \ | |
123 | if ((_page)->lru.prev != _base) { \ | |
124 | struct page *prev; \ | |
125 | \ | |
126 | prev = lru_to_page(&(_page->lru)); \ | |
127 | prefetch(&prev->_field); \ | |
128 | } \ | |
129 | } while (0) | |
130 | #else | |
131 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
132 | #endif | |
133 | ||
134 | #ifdef ARCH_HAS_PREFETCHW | |
135 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
136 | do { \ | |
137 | if ((_page)->lru.prev != _base) { \ | |
138 | struct page *prev; \ | |
139 | \ | |
140 | prev = lru_to_page(&(_page->lru)); \ | |
141 | prefetchw(&prev->_field); \ | |
142 | } \ | |
143 | } while (0) | |
144 | #else | |
145 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
146 | #endif | |
147 | ||
148 | /* | |
149 | * From 0 .. 100. Higher means more swappy. | |
150 | */ | |
151 | int vm_swappiness = 60; | |
152 | /* | |
153 | * The total number of pages which are beyond the high watermark within all | |
154 | * zones. | |
155 | */ | |
156 | unsigned long vm_total_pages; | |
157 | ||
158 | static LIST_HEAD(shrinker_list); | |
159 | static DECLARE_RWSEM(shrinker_rwsem); | |
160 | ||
161 | #ifdef CONFIG_MEMCG | |
162 | static bool global_reclaim(struct scan_control *sc) | |
163 | { | |
164 | return !sc->target_mem_cgroup; | |
165 | } | |
166 | ||
167 | /** | |
168 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
169 | * @sc: scan_control in question | |
170 | * | |
171 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
172 | * completely broken with the legacy memcg and direct stalling in | |
173 | * shrink_page_list() is used for throttling instead, which lacks all the | |
174 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
175 | * allocation and configurability. | |
176 | * | |
177 | * This function tests whether the vmscan currently in progress can assume | |
178 | * that the normal dirty throttling mechanism is operational. | |
179 | */ | |
180 | static bool sane_reclaim(struct scan_control *sc) | |
181 | { | |
182 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
183 | ||
184 | if (!memcg) | |
185 | return true; | |
186 | #ifdef CONFIG_CGROUP_WRITEBACK | |
187 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
188 | return true; | |
189 | #endif | |
190 | return false; | |
191 | } | |
192 | #else | |
193 | static bool global_reclaim(struct scan_control *sc) | |
194 | { | |
195 | return true; | |
196 | } | |
197 | ||
198 | static bool sane_reclaim(struct scan_control *sc) | |
199 | { | |
200 | return true; | |
201 | } | |
202 | #endif | |
203 | ||
204 | /* | |
205 | * This misses isolated pages which are not accounted for to save counters. | |
206 | * As the data only determines if reclaim or compaction continues, it is | |
207 | * not expected that isolated pages will be a dominating factor. | |
208 | */ | |
209 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
210 | { | |
211 | unsigned long nr; | |
212 | ||
213 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
214 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
215 | if (get_nr_swap_pages() > 0) | |
216 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
217 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
218 | ||
219 | return nr; | |
220 | } | |
221 | ||
222 | unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat) | |
223 | { | |
224 | unsigned long nr; | |
225 | ||
226 | nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) + | |
227 | node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) + | |
228 | node_page_state_snapshot(pgdat, NR_ISOLATED_FILE); | |
229 | ||
230 | if (get_nr_swap_pages() > 0) | |
231 | nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) + | |
232 | node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) + | |
233 | node_page_state_snapshot(pgdat, NR_ISOLATED_ANON); | |
234 | ||
235 | return nr; | |
236 | } | |
237 | ||
238 | /** | |
239 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
240 | * @lruvec: lru vector | |
241 | * @lru: lru to use | |
242 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
243 | */ | |
244 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) | |
245 | { | |
246 | unsigned long lru_size; | |
247 | int zid; | |
248 | ||
249 | if (!mem_cgroup_disabled()) | |
250 | lru_size = mem_cgroup_get_lru_size(lruvec, lru); | |
251 | else | |
252 | lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); | |
253 | ||
254 | for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { | |
255 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; | |
256 | unsigned long size; | |
257 | ||
258 | if (!managed_zone(zone)) | |
259 | continue; | |
260 | ||
261 | if (!mem_cgroup_disabled()) | |
262 | size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); | |
263 | else | |
264 | size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], | |
265 | NR_ZONE_LRU_BASE + lru); | |
266 | lru_size -= min(size, lru_size); | |
267 | } | |
268 | ||
269 | return lru_size; | |
270 | ||
271 | } | |
272 | ||
273 | /* | |
274 | * Add a shrinker callback to be called from the vm. | |
275 | */ | |
276 | int register_shrinker(struct shrinker *shrinker) | |
277 | { | |
278 | size_t size = sizeof(*shrinker->nr_deferred); | |
279 | ||
280 | if (shrinker->flags & SHRINKER_NUMA_AWARE) | |
281 | size *= nr_node_ids; | |
282 | ||
283 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
284 | if (!shrinker->nr_deferred) | |
285 | return -ENOMEM; | |
286 | ||
287 | down_write(&shrinker_rwsem); | |
288 | list_add_tail(&shrinker->list, &shrinker_list); | |
289 | up_write(&shrinker_rwsem); | |
290 | return 0; | |
291 | } | |
292 | EXPORT_SYMBOL(register_shrinker); | |
293 | ||
294 | /* | |
295 | * Remove one | |
296 | */ | |
297 | void unregister_shrinker(struct shrinker *shrinker) | |
298 | { | |
299 | down_write(&shrinker_rwsem); | |
300 | list_del(&shrinker->list); | |
301 | up_write(&shrinker_rwsem); | |
302 | kfree(shrinker->nr_deferred); | |
303 | } | |
304 | EXPORT_SYMBOL(unregister_shrinker); | |
305 | ||
306 | #define SHRINK_BATCH 128 | |
307 | ||
308 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, | |
309 | struct shrinker *shrinker, | |
310 | unsigned long nr_scanned, | |
311 | unsigned long nr_eligible) | |
312 | { | |
313 | unsigned long freed = 0; | |
314 | unsigned long long delta; | |
315 | long total_scan; | |
316 | long freeable; | |
317 | long nr; | |
318 | long new_nr; | |
319 | int nid = shrinkctl->nid; | |
320 | long batch_size = shrinker->batch ? shrinker->batch | |
321 | : SHRINK_BATCH; | |
322 | long scanned = 0, next_deferred; | |
323 | ||
324 | freeable = shrinker->count_objects(shrinker, shrinkctl); | |
325 | if (freeable == 0) | |
326 | return 0; | |
327 | ||
328 | /* | |
329 | * copy the current shrinker scan count into a local variable | |
330 | * and zero it so that other concurrent shrinker invocations | |
331 | * don't also do this scanning work. | |
332 | */ | |
333 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
334 | ||
335 | total_scan = nr; | |
336 | delta = (4 * nr_scanned) / shrinker->seeks; | |
337 | delta *= freeable; | |
338 | do_div(delta, nr_eligible + 1); | |
339 | total_scan += delta; | |
340 | if (total_scan < 0) { | |
341 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", | |
342 | shrinker->scan_objects, total_scan); | |
343 | total_scan = freeable; | |
344 | next_deferred = nr; | |
345 | } else | |
346 | next_deferred = total_scan; | |
347 | ||
348 | /* | |
349 | * We need to avoid excessive windup on filesystem shrinkers | |
350 | * due to large numbers of GFP_NOFS allocations causing the | |
351 | * shrinkers to return -1 all the time. This results in a large | |
352 | * nr being built up so when a shrink that can do some work | |
353 | * comes along it empties the entire cache due to nr >>> | |
354 | * freeable. This is bad for sustaining a working set in | |
355 | * memory. | |
356 | * | |
357 | * Hence only allow the shrinker to scan the entire cache when | |
358 | * a large delta change is calculated directly. | |
359 | */ | |
360 | if (delta < freeable / 4) | |
361 | total_scan = min(total_scan, freeable / 2); | |
362 | ||
363 | /* | |
364 | * Avoid risking looping forever due to too large nr value: | |
365 | * never try to free more than twice the estimate number of | |
366 | * freeable entries. | |
367 | */ | |
368 | if (total_scan > freeable * 2) | |
369 | total_scan = freeable * 2; | |
370 | ||
371 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
372 | nr_scanned, nr_eligible, | |
373 | freeable, delta, total_scan); | |
374 | ||
375 | /* | |
376 | * Normally, we should not scan less than batch_size objects in one | |
377 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
378 | * than batch_size objects in total and we are really tight on memory, | |
379 | * we will try to reclaim all available objects, otherwise we can end | |
380 | * up failing allocations although there are plenty of reclaimable | |
381 | * objects spread over several slabs with usage less than the | |
382 | * batch_size. | |
383 | * | |
384 | * We detect the "tight on memory" situations by looking at the total | |
385 | * number of objects we want to scan (total_scan). If it is greater | |
386 | * than the total number of objects on slab (freeable), we must be | |
387 | * scanning at high prio and therefore should try to reclaim as much as | |
388 | * possible. | |
389 | */ | |
390 | while (total_scan >= batch_size || | |
391 | total_scan >= freeable) { | |
392 | unsigned long ret; | |
393 | unsigned long nr_to_scan = min(batch_size, total_scan); | |
394 | ||
395 | shrinkctl->nr_to_scan = nr_to_scan; | |
396 | ret = shrinker->scan_objects(shrinker, shrinkctl); | |
397 | if (ret == SHRINK_STOP) | |
398 | break; | |
399 | freed += ret; | |
400 | ||
401 | count_vm_events(SLABS_SCANNED, nr_to_scan); | |
402 | total_scan -= nr_to_scan; | |
403 | scanned += nr_to_scan; | |
404 | ||
405 | cond_resched(); | |
406 | } | |
407 | ||
408 | if (next_deferred >= scanned) | |
409 | next_deferred -= scanned; | |
410 | else | |
411 | next_deferred = 0; | |
412 | /* | |
413 | * move the unused scan count back into the shrinker in a | |
414 | * manner that handles concurrent updates. If we exhausted the | |
415 | * scan, there is no need to do an update. | |
416 | */ | |
417 | if (next_deferred > 0) | |
418 | new_nr = atomic_long_add_return(next_deferred, | |
419 | &shrinker->nr_deferred[nid]); | |
420 | else | |
421 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
422 | ||
423 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); | |
424 | return freed; | |
425 | } | |
426 | ||
427 | /** | |
428 | * shrink_slab - shrink slab caches | |
429 | * @gfp_mask: allocation context | |
430 | * @nid: node whose slab caches to target | |
431 | * @memcg: memory cgroup whose slab caches to target | |
432 | * @nr_scanned: pressure numerator | |
433 | * @nr_eligible: pressure denominator | |
434 | * | |
435 | * Call the shrink functions to age shrinkable caches. | |
436 | * | |
437 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, | |
438 | * unaware shrinkers will receive a node id of 0 instead. | |
439 | * | |
440 | * @memcg specifies the memory cgroup to target. If it is not NULL, | |
441 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan | |
442 | * objects from the memory cgroup specified. Otherwise, only unaware | |
443 | * shrinkers are called. | |
444 | * | |
445 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of | |
446 | * the available objects should be scanned. Page reclaim for example | |
447 | * passes the number of pages scanned and the number of pages on the | |
448 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned | |
449 | * when it encountered mapped pages. The ratio is further biased by | |
450 | * the ->seeks setting of the shrink function, which indicates the | |
451 | * cost to recreate an object relative to that of an LRU page. | |
452 | * | |
453 | * Returns the number of reclaimed slab objects. | |
454 | */ | |
455 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, | |
456 | struct mem_cgroup *memcg, | |
457 | unsigned long nr_scanned, | |
458 | unsigned long nr_eligible) | |
459 | { | |
460 | struct shrinker *shrinker; | |
461 | unsigned long freed = 0; | |
462 | ||
463 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) | |
464 | return 0; | |
465 | ||
466 | if (nr_scanned == 0) | |
467 | nr_scanned = SWAP_CLUSTER_MAX; | |
468 | ||
469 | if (!down_read_trylock(&shrinker_rwsem)) { | |
470 | /* | |
471 | * If we would return 0, our callers would understand that we | |
472 | * have nothing else to shrink and give up trying. By returning | |
473 | * 1 we keep it going and assume we'll be able to shrink next | |
474 | * time. | |
475 | */ | |
476 | freed = 1; | |
477 | goto out; | |
478 | } | |
479 | ||
480 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
481 | struct shrink_control sc = { | |
482 | .gfp_mask = gfp_mask, | |
483 | .nid = nid, | |
484 | .memcg = memcg, | |
485 | }; | |
486 | ||
487 | /* | |
488 | * If kernel memory accounting is disabled, we ignore | |
489 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers | |
490 | * passing NULL for memcg. | |
491 | */ | |
492 | if (memcg_kmem_enabled() && | |
493 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
494 | continue; | |
495 | ||
496 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) | |
497 | sc.nid = 0; | |
498 | ||
499 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); | |
500 | } | |
501 | ||
502 | up_read(&shrinker_rwsem); | |
503 | out: | |
504 | cond_resched(); | |
505 | return freed; | |
506 | } | |
507 | ||
508 | void drop_slab_node(int nid) | |
509 | { | |
510 | unsigned long freed; | |
511 | ||
512 | do { | |
513 | struct mem_cgroup *memcg = NULL; | |
514 | ||
515 | freed = 0; | |
516 | do { | |
517 | freed += shrink_slab(GFP_KERNEL, nid, memcg, | |
518 | 1000, 1000); | |
519 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
520 | } while (freed > 10); | |
521 | } | |
522 | ||
523 | void drop_slab(void) | |
524 | { | |
525 | int nid; | |
526 | ||
527 | for_each_online_node(nid) | |
528 | drop_slab_node(nid); | |
529 | } | |
530 | ||
531 | static inline int is_page_cache_freeable(struct page *page) | |
532 | { | |
533 | /* | |
534 | * A freeable page cache page is referenced only by the caller | |
535 | * that isolated the page, the page cache radix tree and | |
536 | * optional buffer heads at page->private. | |
537 | */ | |
538 | return page_count(page) - page_has_private(page) == 2; | |
539 | } | |
540 | ||
541 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) | |
542 | { | |
543 | if (current->flags & PF_SWAPWRITE) | |
544 | return 1; | |
545 | if (!inode_write_congested(inode)) | |
546 | return 1; | |
547 | if (inode_to_bdi(inode) == current->backing_dev_info) | |
548 | return 1; | |
549 | return 0; | |
550 | } | |
551 | ||
552 | /* | |
553 | * We detected a synchronous write error writing a page out. Probably | |
554 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
555 | * fsync(), msync() or close(). | |
556 | * | |
557 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
558 | * prevents it from being freed up. But we have a ref on the page and once | |
559 | * that page is locked, the mapping is pinned. | |
560 | * | |
561 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
562 | * __GFP_FS. | |
563 | */ | |
564 | static void handle_write_error(struct address_space *mapping, | |
565 | struct page *page, int error) | |
566 | { | |
567 | lock_page(page); | |
568 | if (page_mapping(page) == mapping) | |
569 | mapping_set_error(mapping, error); | |
570 | unlock_page(page); | |
571 | } | |
572 | ||
573 | /* possible outcome of pageout() */ | |
574 | typedef enum { | |
575 | /* failed to write page out, page is locked */ | |
576 | PAGE_KEEP, | |
577 | /* move page to the active list, page is locked */ | |
578 | PAGE_ACTIVATE, | |
579 | /* page has been sent to the disk successfully, page is unlocked */ | |
580 | PAGE_SUCCESS, | |
581 | /* page is clean and locked */ | |
582 | PAGE_CLEAN, | |
583 | } pageout_t; | |
584 | ||
585 | /* | |
586 | * pageout is called by shrink_page_list() for each dirty page. | |
587 | * Calls ->writepage(). | |
588 | */ | |
589 | static pageout_t pageout(struct page *page, struct address_space *mapping, | |
590 | struct scan_control *sc) | |
591 | { | |
592 | /* | |
593 | * If the page is dirty, only perform writeback if that write | |
594 | * will be non-blocking. To prevent this allocation from being | |
595 | * stalled by pagecache activity. But note that there may be | |
596 | * stalls if we need to run get_block(). We could test | |
597 | * PagePrivate for that. | |
598 | * | |
599 | * If this process is currently in __generic_file_write_iter() against | |
600 | * this page's queue, we can perform writeback even if that | |
601 | * will block. | |
602 | * | |
603 | * If the page is swapcache, write it back even if that would | |
604 | * block, for some throttling. This happens by accident, because | |
605 | * swap_backing_dev_info is bust: it doesn't reflect the | |
606 | * congestion state of the swapdevs. Easy to fix, if needed. | |
607 | */ | |
608 | if (!is_page_cache_freeable(page)) | |
609 | return PAGE_KEEP; | |
610 | if (!mapping) { | |
611 | /* | |
612 | * Some data journaling orphaned pages can have | |
613 | * page->mapping == NULL while being dirty with clean buffers. | |
614 | */ | |
615 | if (page_has_private(page)) { | |
616 | if (try_to_free_buffers(page)) { | |
617 | ClearPageDirty(page); | |
618 | pr_info("%s: orphaned page\n", __func__); | |
619 | return PAGE_CLEAN; | |
620 | } | |
621 | } | |
622 | return PAGE_KEEP; | |
623 | } | |
624 | if (mapping->a_ops->writepage == NULL) | |
625 | return PAGE_ACTIVATE; | |
626 | if (!may_write_to_inode(mapping->host, sc)) | |
627 | return PAGE_KEEP; | |
628 | ||
629 | if (clear_page_dirty_for_io(page)) { | |
630 | int res; | |
631 | struct writeback_control wbc = { | |
632 | .sync_mode = WB_SYNC_NONE, | |
633 | .nr_to_write = SWAP_CLUSTER_MAX, | |
634 | .range_start = 0, | |
635 | .range_end = LLONG_MAX, | |
636 | .for_reclaim = 1, | |
637 | }; | |
638 | ||
639 | SetPageReclaim(page); | |
640 | res = mapping->a_ops->writepage(page, &wbc); | |
641 | if (res < 0) | |
642 | handle_write_error(mapping, page, res); | |
643 | if (res == AOP_WRITEPAGE_ACTIVATE) { | |
644 | ClearPageReclaim(page); | |
645 | return PAGE_ACTIVATE; | |
646 | } | |
647 | ||
648 | if (!PageWriteback(page)) { | |
649 | /* synchronous write or broken a_ops? */ | |
650 | ClearPageReclaim(page); | |
651 | } | |
652 | trace_mm_vmscan_writepage(page); | |
653 | inc_node_page_state(page, NR_VMSCAN_WRITE); | |
654 | return PAGE_SUCCESS; | |
655 | } | |
656 | ||
657 | return PAGE_CLEAN; | |
658 | } | |
659 | ||
660 | /* | |
661 | * Same as remove_mapping, but if the page is removed from the mapping, it | |
662 | * gets returned with a refcount of 0. | |
663 | */ | |
664 | static int __remove_mapping(struct address_space *mapping, struct page *page, | |
665 | bool reclaimed) | |
666 | { | |
667 | unsigned long flags; | |
668 | ||
669 | BUG_ON(!PageLocked(page)); | |
670 | BUG_ON(mapping != page_mapping(page)); | |
671 | ||
672 | spin_lock_irqsave(&mapping->tree_lock, flags); | |
673 | /* | |
674 | * The non racy check for a busy page. | |
675 | * | |
676 | * Must be careful with the order of the tests. When someone has | |
677 | * a ref to the page, it may be possible that they dirty it then | |
678 | * drop the reference. So if PageDirty is tested before page_count | |
679 | * here, then the following race may occur: | |
680 | * | |
681 | * get_user_pages(&page); | |
682 | * [user mapping goes away] | |
683 | * write_to(page); | |
684 | * !PageDirty(page) [good] | |
685 | * SetPageDirty(page); | |
686 | * put_page(page); | |
687 | * !page_count(page) [good, discard it] | |
688 | * | |
689 | * [oops, our write_to data is lost] | |
690 | * | |
691 | * Reversing the order of the tests ensures such a situation cannot | |
692 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
693 | * load is not satisfied before that of page->_refcount. | |
694 | * | |
695 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
696 | * and thus under tree_lock, then this ordering is not required. | |
697 | */ | |
698 | if (!page_ref_freeze(page, 2)) | |
699 | goto cannot_free; | |
700 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ | |
701 | if (unlikely(PageDirty(page))) { | |
702 | page_ref_unfreeze(page, 2); | |
703 | goto cannot_free; | |
704 | } | |
705 | ||
706 | if (PageSwapCache(page)) { | |
707 | swp_entry_t swap = { .val = page_private(page) }; | |
708 | mem_cgroup_swapout(page, swap); | |
709 | __delete_from_swap_cache(page); | |
710 | spin_unlock_irqrestore(&mapping->tree_lock, flags); | |
711 | put_swap_page(page, swap); | |
712 | } else { | |
713 | void (*freepage)(struct page *); | |
714 | void *shadow = NULL; | |
715 | ||
716 | freepage = mapping->a_ops->freepage; | |
717 | /* | |
718 | * Remember a shadow entry for reclaimed file cache in | |
719 | * order to detect refaults, thus thrashing, later on. | |
720 | * | |
721 | * But don't store shadows in an address space that is | |
722 | * already exiting. This is not just an optizimation, | |
723 | * inode reclaim needs to empty out the radix tree or | |
724 | * the nodes are lost. Don't plant shadows behind its | |
725 | * back. | |
726 | * | |
727 | * We also don't store shadows for DAX mappings because the | |
728 | * only page cache pages found in these are zero pages | |
729 | * covering holes, and because we don't want to mix DAX | |
730 | * exceptional entries and shadow exceptional entries in the | |
731 | * same page_tree. | |
732 | */ | |
733 | if (reclaimed && page_is_file_cache(page) && | |
734 | !mapping_exiting(mapping) && !dax_mapping(mapping)) | |
735 | shadow = workingset_eviction(mapping, page); | |
736 | __delete_from_page_cache(page, shadow); | |
737 | spin_unlock_irqrestore(&mapping->tree_lock, flags); | |
738 | ||
739 | if (freepage != NULL) | |
740 | freepage(page); | |
741 | } | |
742 | ||
743 | return 1; | |
744 | ||
745 | cannot_free: | |
746 | spin_unlock_irqrestore(&mapping->tree_lock, flags); | |
747 | return 0; | |
748 | } | |
749 | ||
750 | /* | |
751 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
752 | * someone else has a ref on the page, abort and return 0. If it was | |
753 | * successfully detached, return 1. Assumes the caller has a single ref on | |
754 | * this page. | |
755 | */ | |
756 | int remove_mapping(struct address_space *mapping, struct page *page) | |
757 | { | |
758 | if (__remove_mapping(mapping, page, false)) { | |
759 | /* | |
760 | * Unfreezing the refcount with 1 rather than 2 effectively | |
761 | * drops the pagecache ref for us without requiring another | |
762 | * atomic operation. | |
763 | */ | |
764 | page_ref_unfreeze(page, 1); | |
765 | return 1; | |
766 | } | |
767 | return 0; | |
768 | } | |
769 | ||
770 | /** | |
771 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
772 | * @page: page to be put back to appropriate lru list | |
773 | * | |
774 | * Add previously isolated @page to appropriate LRU list. | |
775 | * Page may still be unevictable for other reasons. | |
776 | * | |
777 | * lru_lock must not be held, interrupts must be enabled. | |
778 | */ | |
779 | void putback_lru_page(struct page *page) | |
780 | { | |
781 | bool is_unevictable; | |
782 | int was_unevictable = PageUnevictable(page); | |
783 | ||
784 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
785 | ||
786 | redo: | |
787 | ClearPageUnevictable(page); | |
788 | ||
789 | if (page_evictable(page)) { | |
790 | /* | |
791 | * For evictable pages, we can use the cache. | |
792 | * In event of a race, worst case is we end up with an | |
793 | * unevictable page on [in]active list. | |
794 | * We know how to handle that. | |
795 | */ | |
796 | is_unevictable = false; | |
797 | lru_cache_add(page); | |
798 | } else { | |
799 | /* | |
800 | * Put unevictable pages directly on zone's unevictable | |
801 | * list. | |
802 | */ | |
803 | is_unevictable = true; | |
804 | add_page_to_unevictable_list(page); | |
805 | /* | |
806 | * When racing with an mlock or AS_UNEVICTABLE clearing | |
807 | * (page is unlocked) make sure that if the other thread | |
808 | * does not observe our setting of PG_lru and fails | |
809 | * isolation/check_move_unevictable_pages, | |
810 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move | |
811 | * the page back to the evictable list. | |
812 | * | |
813 | * The other side is TestClearPageMlocked() or shmem_lock(). | |
814 | */ | |
815 | smp_mb(); | |
816 | } | |
817 | ||
818 | /* | |
819 | * page's status can change while we move it among lru. If an evictable | |
820 | * page is on unevictable list, it never be freed. To avoid that, | |
821 | * check after we added it to the list, again. | |
822 | */ | |
823 | if (is_unevictable && page_evictable(page)) { | |
824 | if (!isolate_lru_page(page)) { | |
825 | put_page(page); | |
826 | goto redo; | |
827 | } | |
828 | /* This means someone else dropped this page from LRU | |
829 | * So, it will be freed or putback to LRU again. There is | |
830 | * nothing to do here. | |
831 | */ | |
832 | } | |
833 | ||
834 | if (was_unevictable && !is_unevictable) | |
835 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
836 | else if (!was_unevictable && is_unevictable) | |
837 | count_vm_event(UNEVICTABLE_PGCULLED); | |
838 | ||
839 | put_page(page); /* drop ref from isolate */ | |
840 | } | |
841 | ||
842 | enum page_references { | |
843 | PAGEREF_RECLAIM, | |
844 | PAGEREF_RECLAIM_CLEAN, | |
845 | PAGEREF_KEEP, | |
846 | PAGEREF_ACTIVATE, | |
847 | }; | |
848 | ||
849 | static enum page_references page_check_references(struct page *page, | |
850 | struct scan_control *sc) | |
851 | { | |
852 | int referenced_ptes, referenced_page; | |
853 | unsigned long vm_flags; | |
854 | ||
855 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, | |
856 | &vm_flags); | |
857 | referenced_page = TestClearPageReferenced(page); | |
858 | ||
859 | /* | |
860 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
861 | * move the page to the unevictable list. | |
862 | */ | |
863 | if (vm_flags & VM_LOCKED) | |
864 | return PAGEREF_RECLAIM; | |
865 | ||
866 | if (referenced_ptes) { | |
867 | if (PageSwapBacked(page)) | |
868 | return PAGEREF_ACTIVATE; | |
869 | /* | |
870 | * All mapped pages start out with page table | |
871 | * references from the instantiating fault, so we need | |
872 | * to look twice if a mapped file page is used more | |
873 | * than once. | |
874 | * | |
875 | * Mark it and spare it for another trip around the | |
876 | * inactive list. Another page table reference will | |
877 | * lead to its activation. | |
878 | * | |
879 | * Note: the mark is set for activated pages as well | |
880 | * so that recently deactivated but used pages are | |
881 | * quickly recovered. | |
882 | */ | |
883 | SetPageReferenced(page); | |
884 | ||
885 | if (referenced_page || referenced_ptes > 1) | |
886 | return PAGEREF_ACTIVATE; | |
887 | ||
888 | /* | |
889 | * Activate file-backed executable pages after first usage. | |
890 | */ | |
891 | if (vm_flags & VM_EXEC) | |
892 | return PAGEREF_ACTIVATE; | |
893 | ||
894 | return PAGEREF_KEEP; | |
895 | } | |
896 | ||
897 | /* Reclaim if clean, defer dirty pages to writeback */ | |
898 | if (referenced_page && !PageSwapBacked(page)) | |
899 | return PAGEREF_RECLAIM_CLEAN; | |
900 | ||
901 | return PAGEREF_RECLAIM; | |
902 | } | |
903 | ||
904 | /* Check if a page is dirty or under writeback */ | |
905 | static void page_check_dirty_writeback(struct page *page, | |
906 | bool *dirty, bool *writeback) | |
907 | { | |
908 | struct address_space *mapping; | |
909 | ||
910 | /* | |
911 | * Anonymous pages are not handled by flushers and must be written | |
912 | * from reclaim context. Do not stall reclaim based on them | |
913 | */ | |
914 | if (!page_is_file_cache(page) || | |
915 | (PageAnon(page) && !PageSwapBacked(page))) { | |
916 | *dirty = false; | |
917 | *writeback = false; | |
918 | return; | |
919 | } | |
920 | ||
921 | /* By default assume that the page flags are accurate */ | |
922 | *dirty = PageDirty(page); | |
923 | *writeback = PageWriteback(page); | |
924 | ||
925 | /* Verify dirty/writeback state if the filesystem supports it */ | |
926 | if (!page_has_private(page)) | |
927 | return; | |
928 | ||
929 | mapping = page_mapping(page); | |
930 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
931 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
932 | } | |
933 | ||
934 | struct reclaim_stat { | |
935 | unsigned nr_dirty; | |
936 | unsigned nr_unqueued_dirty; | |
937 | unsigned nr_congested; | |
938 | unsigned nr_writeback; | |
939 | unsigned nr_immediate; | |
940 | unsigned nr_activate; | |
941 | unsigned nr_ref_keep; | |
942 | unsigned nr_unmap_fail; | |
943 | }; | |
944 | ||
945 | /* | |
946 | * shrink_page_list() returns the number of reclaimed pages | |
947 | */ | |
948 | static unsigned long shrink_page_list(struct list_head *page_list, | |
949 | struct pglist_data *pgdat, | |
950 | struct scan_control *sc, | |
951 | enum ttu_flags ttu_flags, | |
952 | struct reclaim_stat *stat, | |
953 | bool force_reclaim) | |
954 | { | |
955 | LIST_HEAD(ret_pages); | |
956 | LIST_HEAD(free_pages); | |
957 | int pgactivate = 0; | |
958 | unsigned nr_unqueued_dirty = 0; | |
959 | unsigned nr_dirty = 0; | |
960 | unsigned nr_congested = 0; | |
961 | unsigned nr_reclaimed = 0; | |
962 | unsigned nr_writeback = 0; | |
963 | unsigned nr_immediate = 0; | |
964 | unsigned nr_ref_keep = 0; | |
965 | unsigned nr_unmap_fail = 0; | |
966 | ||
967 | cond_resched(); | |
968 | ||
969 | while (!list_empty(page_list)) { | |
970 | struct address_space *mapping; | |
971 | struct page *page; | |
972 | int may_enter_fs; | |
973 | enum page_references references = PAGEREF_RECLAIM_CLEAN; | |
974 | bool dirty, writeback; | |
975 | ||
976 | cond_resched(); | |
977 | ||
978 | page = lru_to_page(page_list); | |
979 | list_del(&page->lru); | |
980 | ||
981 | if (!trylock_page(page)) | |
982 | goto keep; | |
983 | ||
984 | VM_BUG_ON_PAGE(PageActive(page), page); | |
985 | ||
986 | sc->nr_scanned++; | |
987 | ||
988 | if (unlikely(!page_evictable(page))) | |
989 | goto activate_locked; | |
990 | ||
991 | if (!sc->may_unmap && page_mapped(page)) | |
992 | goto keep_locked; | |
993 | ||
994 | /* Double the slab pressure for mapped and swapcache pages */ | |
995 | if ((page_mapped(page) || PageSwapCache(page)) && | |
996 | !(PageAnon(page) && !PageSwapBacked(page))) | |
997 | sc->nr_scanned++; | |
998 | ||
999 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || | |
1000 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1001 | ||
1002 | /* | |
1003 | * The number of dirty pages determines if a zone is marked | |
1004 | * reclaim_congested which affects wait_iff_congested. kswapd | |
1005 | * will stall and start writing pages if the tail of the LRU | |
1006 | * is all dirty unqueued pages. | |
1007 | */ | |
1008 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1009 | if (dirty || writeback) | |
1010 | nr_dirty++; | |
1011 | ||
1012 | if (dirty && !writeback) | |
1013 | nr_unqueued_dirty++; | |
1014 | ||
1015 | /* | |
1016 | * Treat this page as congested if the underlying BDI is or if | |
1017 | * pages are cycling through the LRU so quickly that the | |
1018 | * pages marked for immediate reclaim are making it to the | |
1019 | * end of the LRU a second time. | |
1020 | */ | |
1021 | mapping = page_mapping(page); | |
1022 | if (((dirty || writeback) && mapping && | |
1023 | inode_write_congested(mapping->host)) || | |
1024 | (writeback && PageReclaim(page))) | |
1025 | nr_congested++; | |
1026 | ||
1027 | /* | |
1028 | * If a page at the tail of the LRU is under writeback, there | |
1029 | * are three cases to consider. | |
1030 | * | |
1031 | * 1) If reclaim is encountering an excessive number of pages | |
1032 | * under writeback and this page is both under writeback and | |
1033 | * PageReclaim then it indicates that pages are being queued | |
1034 | * for IO but are being recycled through the LRU before the | |
1035 | * IO can complete. Waiting on the page itself risks an | |
1036 | * indefinite stall if it is impossible to writeback the | |
1037 | * page due to IO error or disconnected storage so instead | |
1038 | * note that the LRU is being scanned too quickly and the | |
1039 | * caller can stall after page list has been processed. | |
1040 | * | |
1041 | * 2) Global or new memcg reclaim encounters a page that is | |
1042 | * not marked for immediate reclaim, or the caller does not | |
1043 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1044 | * not to fs). In this case mark the page for immediate | |
1045 | * reclaim and continue scanning. | |
1046 | * | |
1047 | * Require may_enter_fs because we would wait on fs, which | |
1048 | * may not have submitted IO yet. And the loop driver might | |
1049 | * enter reclaim, and deadlock if it waits on a page for | |
1050 | * which it is needed to do the write (loop masks off | |
1051 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1052 | * would probably show more reasons. | |
1053 | * | |
1054 | * 3) Legacy memcg encounters a page that is already marked | |
1055 | * PageReclaim. memcg does not have any dirty pages | |
1056 | * throttling so we could easily OOM just because too many | |
1057 | * pages are in writeback and there is nothing else to | |
1058 | * reclaim. Wait for the writeback to complete. | |
1059 | * | |
1060 | * In cases 1) and 2) we activate the pages to get them out of | |
1061 | * the way while we continue scanning for clean pages on the | |
1062 | * inactive list and refilling from the active list. The | |
1063 | * observation here is that waiting for disk writes is more | |
1064 | * expensive than potentially causing reloads down the line. | |
1065 | * Since they're marked for immediate reclaim, they won't put | |
1066 | * memory pressure on the cache working set any longer than it | |
1067 | * takes to write them to disk. | |
1068 | */ | |
1069 | if (PageWriteback(page)) { | |
1070 | /* Case 1 above */ | |
1071 | if (current_is_kswapd() && | |
1072 | PageReclaim(page) && | |
1073 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { | |
1074 | nr_immediate++; | |
1075 | goto activate_locked; | |
1076 | ||
1077 | /* Case 2 above */ | |
1078 | } else if (sane_reclaim(sc) || | |
1079 | !PageReclaim(page) || !may_enter_fs) { | |
1080 | /* | |
1081 | * This is slightly racy - end_page_writeback() | |
1082 | * might have just cleared PageReclaim, then | |
1083 | * setting PageReclaim here end up interpreted | |
1084 | * as PageReadahead - but that does not matter | |
1085 | * enough to care. What we do want is for this | |
1086 | * page to have PageReclaim set next time memcg | |
1087 | * reclaim reaches the tests above, so it will | |
1088 | * then wait_on_page_writeback() to avoid OOM; | |
1089 | * and it's also appropriate in global reclaim. | |
1090 | */ | |
1091 | SetPageReclaim(page); | |
1092 | nr_writeback++; | |
1093 | goto activate_locked; | |
1094 | ||
1095 | /* Case 3 above */ | |
1096 | } else { | |
1097 | unlock_page(page); | |
1098 | wait_on_page_writeback(page); | |
1099 | /* then go back and try same page again */ | |
1100 | list_add_tail(&page->lru, page_list); | |
1101 | continue; | |
1102 | } | |
1103 | } | |
1104 | ||
1105 | if (!force_reclaim) | |
1106 | references = page_check_references(page, sc); | |
1107 | ||
1108 | switch (references) { | |
1109 | case PAGEREF_ACTIVATE: | |
1110 | goto activate_locked; | |
1111 | case PAGEREF_KEEP: | |
1112 | nr_ref_keep++; | |
1113 | goto keep_locked; | |
1114 | case PAGEREF_RECLAIM: | |
1115 | case PAGEREF_RECLAIM_CLEAN: | |
1116 | ; /* try to reclaim the page below */ | |
1117 | } | |
1118 | ||
1119 | /* | |
1120 | * Anonymous process memory has backing store? | |
1121 | * Try to allocate it some swap space here. | |
1122 | * Lazyfree page could be freed directly | |
1123 | */ | |
1124 | if (PageAnon(page) && PageSwapBacked(page) && | |
1125 | !PageSwapCache(page)) { | |
1126 | if (!(sc->gfp_mask & __GFP_IO)) | |
1127 | goto keep_locked; | |
1128 | if (PageTransHuge(page)) { | |
1129 | /* cannot split THP, skip it */ | |
1130 | if (!can_split_huge_page(page, NULL)) | |
1131 | goto activate_locked; | |
1132 | /* | |
1133 | * Split pages without a PMD map right | |
1134 | * away. Chances are some or all of the | |
1135 | * tail pages can be freed without IO. | |
1136 | */ | |
1137 | if (!compound_mapcount(page) && | |
1138 | split_huge_page_to_list(page, page_list)) | |
1139 | goto activate_locked; | |
1140 | } | |
1141 | if (!add_to_swap(page)) { | |
1142 | if (!PageTransHuge(page)) | |
1143 | goto activate_locked; | |
1144 | /* Split THP and swap individual base pages */ | |
1145 | if (split_huge_page_to_list(page, page_list)) | |
1146 | goto activate_locked; | |
1147 | if (!add_to_swap(page)) | |
1148 | goto activate_locked; | |
1149 | } | |
1150 | ||
1151 | /* XXX: We don't support THP writes */ | |
1152 | if (PageTransHuge(page) && | |
1153 | split_huge_page_to_list(page, page_list)) { | |
1154 | delete_from_swap_cache(page); | |
1155 | goto activate_locked; | |
1156 | } | |
1157 | ||
1158 | may_enter_fs = 1; | |
1159 | ||
1160 | /* Adding to swap updated mapping */ | |
1161 | mapping = page_mapping(page); | |
1162 | } else if (unlikely(PageTransHuge(page))) { | |
1163 | /* Split file THP */ | |
1164 | if (split_huge_page_to_list(page, page_list)) | |
1165 | goto keep_locked; | |
1166 | } | |
1167 | ||
1168 | VM_BUG_ON_PAGE(PageTransHuge(page), page); | |
1169 | ||
1170 | /* | |
1171 | * The page is mapped into the page tables of one or more | |
1172 | * processes. Try to unmap it here. | |
1173 | */ | |
1174 | if (page_mapped(page)) { | |
1175 | if (!try_to_unmap(page, ttu_flags | TTU_BATCH_FLUSH)) { | |
1176 | nr_unmap_fail++; | |
1177 | goto activate_locked; | |
1178 | } | |
1179 | } | |
1180 | ||
1181 | if (PageDirty(page)) { | |
1182 | /* | |
1183 | * Only kswapd can writeback filesystem pages | |
1184 | * to avoid risk of stack overflow. But avoid | |
1185 | * injecting inefficient single-page IO into | |
1186 | * flusher writeback as much as possible: only | |
1187 | * write pages when we've encountered many | |
1188 | * dirty pages, and when we've already scanned | |
1189 | * the rest of the LRU for clean pages and see | |
1190 | * the same dirty pages again (PageReclaim). | |
1191 | */ | |
1192 | if (page_is_file_cache(page) && | |
1193 | (!current_is_kswapd() || !PageReclaim(page) || | |
1194 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
1195 | /* | |
1196 | * Immediately reclaim when written back. | |
1197 | * Similar in principal to deactivate_page() | |
1198 | * except we already have the page isolated | |
1199 | * and know it's dirty | |
1200 | */ | |
1201 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); | |
1202 | SetPageReclaim(page); | |
1203 | ||
1204 | goto activate_locked; | |
1205 | } | |
1206 | ||
1207 | if (references == PAGEREF_RECLAIM_CLEAN) | |
1208 | goto keep_locked; | |
1209 | if (!may_enter_fs) | |
1210 | goto keep_locked; | |
1211 | if (!sc->may_writepage) | |
1212 | goto keep_locked; | |
1213 | ||
1214 | /* | |
1215 | * Page is dirty. Flush the TLB if a writable entry | |
1216 | * potentially exists to avoid CPU writes after IO | |
1217 | * starts and then write it out here. | |
1218 | */ | |
1219 | try_to_unmap_flush_dirty(); | |
1220 | switch (pageout(page, mapping, sc)) { | |
1221 | case PAGE_KEEP: | |
1222 | goto keep_locked; | |
1223 | case PAGE_ACTIVATE: | |
1224 | goto activate_locked; | |
1225 | case PAGE_SUCCESS: | |
1226 | if (PageWriteback(page)) | |
1227 | goto keep; | |
1228 | if (PageDirty(page)) | |
1229 | goto keep; | |
1230 | ||
1231 | /* | |
1232 | * A synchronous write - probably a ramdisk. Go | |
1233 | * ahead and try to reclaim the page. | |
1234 | */ | |
1235 | if (!trylock_page(page)) | |
1236 | goto keep; | |
1237 | if (PageDirty(page) || PageWriteback(page)) | |
1238 | goto keep_locked; | |
1239 | mapping = page_mapping(page); | |
1240 | case PAGE_CLEAN: | |
1241 | ; /* try to free the page below */ | |
1242 | } | |
1243 | } | |
1244 | ||
1245 | /* | |
1246 | * If the page has buffers, try to free the buffer mappings | |
1247 | * associated with this page. If we succeed we try to free | |
1248 | * the page as well. | |
1249 | * | |
1250 | * We do this even if the page is PageDirty(). | |
1251 | * try_to_release_page() does not perform I/O, but it is | |
1252 | * possible for a page to have PageDirty set, but it is actually | |
1253 | * clean (all its buffers are clean). This happens if the | |
1254 | * buffers were written out directly, with submit_bh(). ext3 | |
1255 | * will do this, as well as the blockdev mapping. | |
1256 | * try_to_release_page() will discover that cleanness and will | |
1257 | * drop the buffers and mark the page clean - it can be freed. | |
1258 | * | |
1259 | * Rarely, pages can have buffers and no ->mapping. These are | |
1260 | * the pages which were not successfully invalidated in | |
1261 | * truncate_complete_page(). We try to drop those buffers here | |
1262 | * and if that worked, and the page is no longer mapped into | |
1263 | * process address space (page_count == 1) it can be freed. | |
1264 | * Otherwise, leave the page on the LRU so it is swappable. | |
1265 | */ | |
1266 | if (page_has_private(page)) { | |
1267 | if (!try_to_release_page(page, sc->gfp_mask)) | |
1268 | goto activate_locked; | |
1269 | if (!mapping && page_count(page) == 1) { | |
1270 | unlock_page(page); | |
1271 | if (put_page_testzero(page)) | |
1272 | goto free_it; | |
1273 | else { | |
1274 | /* | |
1275 | * rare race with speculative reference. | |
1276 | * the speculative reference will free | |
1277 | * this page shortly, so we may | |
1278 | * increment nr_reclaimed here (and | |
1279 | * leave it off the LRU). | |
1280 | */ | |
1281 | nr_reclaimed++; | |
1282 | continue; | |
1283 | } | |
1284 | } | |
1285 | } | |
1286 | ||
1287 | if (PageAnon(page) && !PageSwapBacked(page)) { | |
1288 | /* follow __remove_mapping for reference */ | |
1289 | if (!page_ref_freeze(page, 1)) | |
1290 | goto keep_locked; | |
1291 | if (PageDirty(page)) { | |
1292 | page_ref_unfreeze(page, 1); | |
1293 | goto keep_locked; | |
1294 | } | |
1295 | ||
1296 | count_vm_event(PGLAZYFREED); | |
1297 | count_memcg_page_event(page, PGLAZYFREED); | |
1298 | } else if (!mapping || !__remove_mapping(mapping, page, true)) | |
1299 | goto keep_locked; | |
1300 | /* | |
1301 | * At this point, we have no other references and there is | |
1302 | * no way to pick any more up (removed from LRU, removed | |
1303 | * from pagecache). Can use non-atomic bitops now (and | |
1304 | * we obviously don't have to worry about waking up a process | |
1305 | * waiting on the page lock, because there are no references. | |
1306 | */ | |
1307 | __ClearPageLocked(page); | |
1308 | free_it: | |
1309 | nr_reclaimed++; | |
1310 | ||
1311 | /* | |
1312 | * Is there need to periodically free_page_list? It would | |
1313 | * appear not as the counts should be low | |
1314 | */ | |
1315 | list_add(&page->lru, &free_pages); | |
1316 | continue; | |
1317 | ||
1318 | activate_locked: | |
1319 | /* Not a candidate for swapping, so reclaim swap space. */ | |
1320 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || | |
1321 | PageMlocked(page))) | |
1322 | try_to_free_swap(page); | |
1323 | VM_BUG_ON_PAGE(PageActive(page), page); | |
1324 | if (!PageMlocked(page)) { | |
1325 | SetPageActive(page); | |
1326 | pgactivate++; | |
1327 | count_memcg_page_event(page, PGACTIVATE); | |
1328 | } | |
1329 | keep_locked: | |
1330 | unlock_page(page); | |
1331 | keep: | |
1332 | list_add(&page->lru, &ret_pages); | |
1333 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); | |
1334 | } | |
1335 | ||
1336 | mem_cgroup_uncharge_list(&free_pages); | |
1337 | try_to_unmap_flush(); | |
1338 | free_hot_cold_page_list(&free_pages, true); | |
1339 | ||
1340 | list_splice(&ret_pages, page_list); | |
1341 | count_vm_events(PGACTIVATE, pgactivate); | |
1342 | ||
1343 | if (stat) { | |
1344 | stat->nr_dirty = nr_dirty; | |
1345 | stat->nr_congested = nr_congested; | |
1346 | stat->nr_unqueued_dirty = nr_unqueued_dirty; | |
1347 | stat->nr_writeback = nr_writeback; | |
1348 | stat->nr_immediate = nr_immediate; | |
1349 | stat->nr_activate = pgactivate; | |
1350 | stat->nr_ref_keep = nr_ref_keep; | |
1351 | stat->nr_unmap_fail = nr_unmap_fail; | |
1352 | } | |
1353 | return nr_reclaimed; | |
1354 | } | |
1355 | ||
1356 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, | |
1357 | struct list_head *page_list) | |
1358 | { | |
1359 | struct scan_control sc = { | |
1360 | .gfp_mask = GFP_KERNEL, | |
1361 | .priority = DEF_PRIORITY, | |
1362 | .may_unmap = 1, | |
1363 | }; | |
1364 | unsigned long ret; | |
1365 | struct page *page, *next; | |
1366 | LIST_HEAD(clean_pages); | |
1367 | ||
1368 | list_for_each_entry_safe(page, next, page_list, lru) { | |
1369 | if (page_is_file_cache(page) && !PageDirty(page) && | |
1370 | !__PageMovable(page)) { | |
1371 | ClearPageActive(page); | |
1372 | list_move(&page->lru, &clean_pages); | |
1373 | } | |
1374 | } | |
1375 | ||
1376 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, | |
1377 | TTU_IGNORE_ACCESS, NULL, true); | |
1378 | list_splice(&clean_pages, page_list); | |
1379 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); | |
1380 | return ret; | |
1381 | } | |
1382 | ||
1383 | /* | |
1384 | * Attempt to remove the specified page from its LRU. Only take this page | |
1385 | * if it is of the appropriate PageActive status. Pages which are being | |
1386 | * freed elsewhere are also ignored. | |
1387 | * | |
1388 | * page: page to consider | |
1389 | * mode: one of the LRU isolation modes defined above | |
1390 | * | |
1391 | * returns 0 on success, -ve errno on failure. | |
1392 | */ | |
1393 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) | |
1394 | { | |
1395 | int ret = -EINVAL; | |
1396 | ||
1397 | /* Only take pages on the LRU. */ | |
1398 | if (!PageLRU(page)) | |
1399 | return ret; | |
1400 | ||
1401 | /* Compaction should not handle unevictable pages but CMA can do so */ | |
1402 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
1403 | return ret; | |
1404 | ||
1405 | ret = -EBUSY; | |
1406 | ||
1407 | /* | |
1408 | * To minimise LRU disruption, the caller can indicate that it only | |
1409 | * wants to isolate pages it will be able to operate on without | |
1410 | * blocking - clean pages for the most part. | |
1411 | * | |
1412 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1413 | * that it is possible to migrate without blocking | |
1414 | */ | |
1415 | if (mode & ISOLATE_ASYNC_MIGRATE) { | |
1416 | /* All the caller can do on PageWriteback is block */ | |
1417 | if (PageWriteback(page)) | |
1418 | return ret; | |
1419 | ||
1420 | if (PageDirty(page)) { | |
1421 | struct address_space *mapping; | |
1422 | ||
1423 | /* | |
1424 | * Only pages without mappings or that have a | |
1425 | * ->migratepage callback are possible to migrate | |
1426 | * without blocking | |
1427 | */ | |
1428 | mapping = page_mapping(page); | |
1429 | if (mapping && !mapping->a_ops->migratepage) | |
1430 | return ret; | |
1431 | } | |
1432 | } | |
1433 | ||
1434 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) | |
1435 | return ret; | |
1436 | ||
1437 | if (likely(get_page_unless_zero(page))) { | |
1438 | /* | |
1439 | * Be careful not to clear PageLRU until after we're | |
1440 | * sure the page is not being freed elsewhere -- the | |
1441 | * page release code relies on it. | |
1442 | */ | |
1443 | ClearPageLRU(page); | |
1444 | ret = 0; | |
1445 | } | |
1446 | ||
1447 | return ret; | |
1448 | } | |
1449 | ||
1450 | ||
1451 | /* | |
1452 | * Update LRU sizes after isolating pages. The LRU size updates must | |
1453 | * be complete before mem_cgroup_update_lru_size due to a santity check. | |
1454 | */ | |
1455 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
1456 | enum lru_list lru, unsigned long *nr_zone_taken) | |
1457 | { | |
1458 | int zid; | |
1459 | ||
1460 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
1461 | if (!nr_zone_taken[zid]) | |
1462 | continue; | |
1463 | ||
1464 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
1465 | #ifdef CONFIG_MEMCG | |
1466 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
1467 | #endif | |
1468 | } | |
1469 | ||
1470 | } | |
1471 | ||
1472 | /* | |
1473 | * zone_lru_lock is heavily contended. Some of the functions that | |
1474 | * shrink the lists perform better by taking out a batch of pages | |
1475 | * and working on them outside the LRU lock. | |
1476 | * | |
1477 | * For pagecache intensive workloads, this function is the hottest | |
1478 | * spot in the kernel (apart from copy_*_user functions). | |
1479 | * | |
1480 | * Appropriate locks must be held before calling this function. | |
1481 | * | |
1482 | * @nr_to_scan: The number of eligible pages to look through on the list. | |
1483 | * @lruvec: The LRU vector to pull pages from. | |
1484 | * @dst: The temp list to put pages on to. | |
1485 | * @nr_scanned: The number of pages that were scanned. | |
1486 | * @sc: The scan_control struct for this reclaim session | |
1487 | * @mode: One of the LRU isolation modes | |
1488 | * @lru: LRU list id for isolating | |
1489 | * | |
1490 | * returns how many pages were moved onto *@dst. | |
1491 | */ | |
1492 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, | |
1493 | struct lruvec *lruvec, struct list_head *dst, | |
1494 | unsigned long *nr_scanned, struct scan_control *sc, | |
1495 | isolate_mode_t mode, enum lru_list lru) | |
1496 | { | |
1497 | struct list_head *src = &lruvec->lists[lru]; | |
1498 | unsigned long nr_taken = 0; | |
1499 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; | |
1500 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; | |
1501 | unsigned long skipped = 0; | |
1502 | unsigned long scan, total_scan, nr_pages; | |
1503 | LIST_HEAD(pages_skipped); | |
1504 | ||
1505 | scan = 0; | |
1506 | for (total_scan = 0; | |
1507 | scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src); | |
1508 | total_scan++) { | |
1509 | struct page *page; | |
1510 | ||
1511 | page = lru_to_page(src); | |
1512 | prefetchw_prev_lru_page(page, src, flags); | |
1513 | ||
1514 | VM_BUG_ON_PAGE(!PageLRU(page), page); | |
1515 | ||
1516 | if (page_zonenum(page) > sc->reclaim_idx) { | |
1517 | list_move(&page->lru, &pages_skipped); | |
1518 | nr_skipped[page_zonenum(page)]++; | |
1519 | continue; | |
1520 | } | |
1521 | ||
1522 | /* | |
1523 | * Do not count skipped pages because that makes the function | |
1524 | * return with no isolated pages if the LRU mostly contains | |
1525 | * ineligible pages. This causes the VM to not reclaim any | |
1526 | * pages, triggering a premature OOM. | |
1527 | */ | |
1528 | scan++; | |
1529 | switch (__isolate_lru_page(page, mode)) { | |
1530 | case 0: | |
1531 | nr_pages = hpage_nr_pages(page); | |
1532 | nr_taken += nr_pages; | |
1533 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
1534 | list_move(&page->lru, dst); | |
1535 | break; | |
1536 | ||
1537 | case -EBUSY: | |
1538 | /* else it is being freed elsewhere */ | |
1539 | list_move(&page->lru, src); | |
1540 | continue; | |
1541 | ||
1542 | default: | |
1543 | BUG(); | |
1544 | } | |
1545 | } | |
1546 | ||
1547 | /* | |
1548 | * Splice any skipped pages to the start of the LRU list. Note that | |
1549 | * this disrupts the LRU order when reclaiming for lower zones but | |
1550 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1551 | * scanning would soon rescan the same pages to skip and put the | |
1552 | * system at risk of premature OOM. | |
1553 | */ | |
1554 | if (!list_empty(&pages_skipped)) { | |
1555 | int zid; | |
1556 | ||
1557 | list_splice(&pages_skipped, src); | |
1558 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
1559 | if (!nr_skipped[zid]) | |
1560 | continue; | |
1561 | ||
1562 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1563 | skipped += nr_skipped[zid]; | |
1564 | } | |
1565 | } | |
1566 | *nr_scanned = total_scan; | |
1567 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, | |
1568 | total_scan, skipped, nr_taken, mode, lru); | |
1569 | update_lru_sizes(lruvec, lru, nr_zone_taken); | |
1570 | return nr_taken; | |
1571 | } | |
1572 | ||
1573 | /** | |
1574 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1575 | * @page: page to isolate from its LRU list | |
1576 | * | |
1577 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1578 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1579 | * | |
1580 | * Returns 0 if the page was removed from an LRU list. | |
1581 | * Returns -EBUSY if the page was not on an LRU list. | |
1582 | * | |
1583 | * The returned page will have PageLRU() cleared. If it was found on | |
1584 | * the active list, it will have PageActive set. If it was found on | |
1585 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1586 | * may need to be cleared by the caller before letting the page go. | |
1587 | * | |
1588 | * The vmstat statistic corresponding to the list on which the page was | |
1589 | * found will be decremented. | |
1590 | * | |
1591 | * Restrictions: | |
1592 | * (1) Must be called with an elevated refcount on the page. This is a | |
1593 | * fundamentnal difference from isolate_lru_pages (which is called | |
1594 | * without a stable reference). | |
1595 | * (2) the lru_lock must not be held. | |
1596 | * (3) interrupts must be enabled. | |
1597 | */ | |
1598 | int isolate_lru_page(struct page *page) | |
1599 | { | |
1600 | int ret = -EBUSY; | |
1601 | ||
1602 | VM_BUG_ON_PAGE(!page_count(page), page); | |
1603 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); | |
1604 | ||
1605 | if (PageLRU(page)) { | |
1606 | struct zone *zone = page_zone(page); | |
1607 | struct lruvec *lruvec; | |
1608 | ||
1609 | spin_lock_irq(zone_lru_lock(zone)); | |
1610 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); | |
1611 | if (PageLRU(page)) { | |
1612 | int lru = page_lru(page); | |
1613 | get_page(page); | |
1614 | ClearPageLRU(page); | |
1615 | del_page_from_lru_list(page, lruvec, lru); | |
1616 | ret = 0; | |
1617 | } | |
1618 | spin_unlock_irq(zone_lru_lock(zone)); | |
1619 | } | |
1620 | return ret; | |
1621 | } | |
1622 | ||
1623 | /* | |
1624 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and | |
1625 | * then get resheduled. When there are massive number of tasks doing page | |
1626 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1627 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1628 | * unnecessary swapping, thrashing and OOM. | |
1629 | */ | |
1630 | static int too_many_isolated(struct pglist_data *pgdat, int file, | |
1631 | struct scan_control *sc) | |
1632 | { | |
1633 | unsigned long inactive, isolated; | |
1634 | ||
1635 | if (current_is_kswapd()) | |
1636 | return 0; | |
1637 | ||
1638 | if (!sane_reclaim(sc)) | |
1639 | return 0; | |
1640 | ||
1641 | if (file) { | |
1642 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); | |
1643 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
1644 | } else { | |
1645 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); | |
1646 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
1647 | } | |
1648 | ||
1649 | /* | |
1650 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1651 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1652 | * deadlock. | |
1653 | */ | |
1654 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) | |
1655 | inactive >>= 3; | |
1656 | ||
1657 | return isolated > inactive; | |
1658 | } | |
1659 | ||
1660 | static noinline_for_stack void | |
1661 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) | |
1662 | { | |
1663 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1664 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
1665 | LIST_HEAD(pages_to_free); | |
1666 | ||
1667 | /* | |
1668 | * Put back any unfreeable pages. | |
1669 | */ | |
1670 | while (!list_empty(page_list)) { | |
1671 | struct page *page = lru_to_page(page_list); | |
1672 | int lru; | |
1673 | ||
1674 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
1675 | list_del(&page->lru); | |
1676 | if (unlikely(!page_evictable(page))) { | |
1677 | spin_unlock_irq(&pgdat->lru_lock); | |
1678 | putback_lru_page(page); | |
1679 | spin_lock_irq(&pgdat->lru_lock); | |
1680 | continue; | |
1681 | } | |
1682 | ||
1683 | lruvec = mem_cgroup_page_lruvec(page, pgdat); | |
1684 | ||
1685 | SetPageLRU(page); | |
1686 | lru = page_lru(page); | |
1687 | add_page_to_lru_list(page, lruvec, lru); | |
1688 | ||
1689 | if (is_active_lru(lru)) { | |
1690 | int file = is_file_lru(lru); | |
1691 | int numpages = hpage_nr_pages(page); | |
1692 | reclaim_stat->recent_rotated[file] += numpages; | |
1693 | } | |
1694 | if (put_page_testzero(page)) { | |
1695 | __ClearPageLRU(page); | |
1696 | __ClearPageActive(page); | |
1697 | del_page_from_lru_list(page, lruvec, lru); | |
1698 | ||
1699 | if (unlikely(PageCompound(page))) { | |
1700 | spin_unlock_irq(&pgdat->lru_lock); | |
1701 | mem_cgroup_uncharge(page); | |
1702 | (*get_compound_page_dtor(page))(page); | |
1703 | spin_lock_irq(&pgdat->lru_lock); | |
1704 | } else | |
1705 | list_add(&page->lru, &pages_to_free); | |
1706 | } | |
1707 | } | |
1708 | ||
1709 | /* | |
1710 | * To save our caller's stack, now use input list for pages to free. | |
1711 | */ | |
1712 | list_splice(&pages_to_free, page_list); | |
1713 | } | |
1714 | ||
1715 | /* | |
1716 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1717 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1718 | * In that case we should only throttle if the backing device it is | |
1719 | * writing to is congested. In other cases it is safe to throttle. | |
1720 | */ | |
1721 | static int current_may_throttle(void) | |
1722 | { | |
1723 | return !(current->flags & PF_LESS_THROTTLE) || | |
1724 | current->backing_dev_info == NULL || | |
1725 | bdi_write_congested(current->backing_dev_info); | |
1726 | } | |
1727 | ||
1728 | /* | |
1729 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number | |
1730 | * of reclaimed pages | |
1731 | */ | |
1732 | static noinline_for_stack unsigned long | |
1733 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, | |
1734 | struct scan_control *sc, enum lru_list lru) | |
1735 | { | |
1736 | LIST_HEAD(page_list); | |
1737 | unsigned long nr_scanned; | |
1738 | unsigned long nr_reclaimed = 0; | |
1739 | unsigned long nr_taken; | |
1740 | struct reclaim_stat stat = {}; | |
1741 | isolate_mode_t isolate_mode = 0; | |
1742 | int file = is_file_lru(lru); | |
1743 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
1744 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1745 | ||
1746 | while (unlikely(too_many_isolated(pgdat, file, sc))) { | |
1747 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
1748 | ||
1749 | /* We are about to die and free our memory. Return now. */ | |
1750 | if (fatal_signal_pending(current)) | |
1751 | return SWAP_CLUSTER_MAX; | |
1752 | } | |
1753 | ||
1754 | lru_add_drain(); | |
1755 | ||
1756 | if (!sc->may_unmap) | |
1757 | isolate_mode |= ISOLATE_UNMAPPED; | |
1758 | ||
1759 | spin_lock_irq(&pgdat->lru_lock); | |
1760 | ||
1761 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, | |
1762 | &nr_scanned, sc, isolate_mode, lru); | |
1763 | ||
1764 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); | |
1765 | reclaim_stat->recent_scanned[file] += nr_taken; | |
1766 | ||
1767 | if (current_is_kswapd()) { | |
1768 | if (global_reclaim(sc)) | |
1769 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); | |
1770 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD, | |
1771 | nr_scanned); | |
1772 | } else { | |
1773 | if (global_reclaim(sc)) | |
1774 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); | |
1775 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT, | |
1776 | nr_scanned); | |
1777 | } | |
1778 | spin_unlock_irq(&pgdat->lru_lock); | |
1779 | ||
1780 | if (nr_taken == 0) | |
1781 | return 0; | |
1782 | ||
1783 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, | |
1784 | &stat, false); | |
1785 | ||
1786 | spin_lock_irq(&pgdat->lru_lock); | |
1787 | ||
1788 | if (current_is_kswapd()) { | |
1789 | if (global_reclaim(sc)) | |
1790 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); | |
1791 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD, | |
1792 | nr_reclaimed); | |
1793 | } else { | |
1794 | if (global_reclaim(sc)) | |
1795 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); | |
1796 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT, | |
1797 | nr_reclaimed); | |
1798 | } | |
1799 | ||
1800 | putback_inactive_pages(lruvec, &page_list); | |
1801 | ||
1802 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); | |
1803 | ||
1804 | spin_unlock_irq(&pgdat->lru_lock); | |
1805 | ||
1806 | mem_cgroup_uncharge_list(&page_list); | |
1807 | free_hot_cold_page_list(&page_list, true); | |
1808 | ||
1809 | /* | |
1810 | * If reclaim is isolating dirty pages under writeback, it implies | |
1811 | * that the long-lived page allocation rate is exceeding the page | |
1812 | * laundering rate. Either the global limits are not being effective | |
1813 | * at throttling processes due to the page distribution throughout | |
1814 | * zones or there is heavy usage of a slow backing device. The | |
1815 | * only option is to throttle from reclaim context which is not ideal | |
1816 | * as there is no guarantee the dirtying process is throttled in the | |
1817 | * same way balance_dirty_pages() manages. | |
1818 | * | |
1819 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number | |
1820 | * of pages under pages flagged for immediate reclaim and stall if any | |
1821 | * are encountered in the nr_immediate check below. | |
1822 | */ | |
1823 | if (stat.nr_writeback && stat.nr_writeback == nr_taken) | |
1824 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
1825 | ||
1826 | /* | |
1827 | * Legacy memcg will stall in page writeback so avoid forcibly | |
1828 | * stalling here. | |
1829 | */ | |
1830 | if (sane_reclaim(sc)) { | |
1831 | /* | |
1832 | * Tag a zone as congested if all the dirty pages scanned were | |
1833 | * backed by a congested BDI and wait_iff_congested will stall. | |
1834 | */ | |
1835 | if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested) | |
1836 | set_bit(PGDAT_CONGESTED, &pgdat->flags); | |
1837 | ||
1838 | /* | |
1839 | * If dirty pages are scanned that are not queued for IO, it | |
1840 | * implies that flushers are not doing their job. This can | |
1841 | * happen when memory pressure pushes dirty pages to the end of | |
1842 | * the LRU before the dirty limits are breached and the dirty | |
1843 | * data has expired. It can also happen when the proportion of | |
1844 | * dirty pages grows not through writes but through memory | |
1845 | * pressure reclaiming all the clean cache. And in some cases, | |
1846 | * the flushers simply cannot keep up with the allocation | |
1847 | * rate. Nudge the flusher threads in case they are asleep, but | |
1848 | * also allow kswapd to start writing pages during reclaim. | |
1849 | */ | |
1850 | if (stat.nr_unqueued_dirty == nr_taken) { | |
1851 | wakeup_flusher_threads(0, WB_REASON_VMSCAN); | |
1852 | set_bit(PGDAT_DIRTY, &pgdat->flags); | |
1853 | } | |
1854 | ||
1855 | /* | |
1856 | * If kswapd scans pages marked marked for immediate | |
1857 | * reclaim and under writeback (nr_immediate), it implies | |
1858 | * that pages are cycling through the LRU faster than | |
1859 | * they are written so also forcibly stall. | |
1860 | */ | |
1861 | if (stat.nr_immediate && current_may_throttle()) | |
1862 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
1863 | } | |
1864 | ||
1865 | /* | |
1866 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1867 | * is congested. Allow kswapd to continue until it starts encountering | |
1868 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1869 | */ | |
1870 | if (!sc->hibernation_mode && !current_is_kswapd() && | |
1871 | current_may_throttle()) | |
1872 | wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); | |
1873 | ||
1874 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, | |
1875 | nr_scanned, nr_reclaimed, | |
1876 | stat.nr_dirty, stat.nr_writeback, | |
1877 | stat.nr_congested, stat.nr_immediate, | |
1878 | stat.nr_activate, stat.nr_ref_keep, | |
1879 | stat.nr_unmap_fail, | |
1880 | sc->priority, file); | |
1881 | return nr_reclaimed; | |
1882 | } | |
1883 | ||
1884 | /* | |
1885 | * This moves pages from the active list to the inactive list. | |
1886 | * | |
1887 | * We move them the other way if the page is referenced by one or more | |
1888 | * processes, from rmap. | |
1889 | * | |
1890 | * If the pages are mostly unmapped, the processing is fast and it is | |
1891 | * appropriate to hold zone_lru_lock across the whole operation. But if | |
1892 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1893 | * should drop zone_lru_lock around each page. It's impossible to balance | |
1894 | * this, so instead we remove the pages from the LRU while processing them. | |
1895 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1896 | * nobody will play with that bit on a non-LRU page. | |
1897 | * | |
1898 | * The downside is that we have to touch page->_refcount against each page. | |
1899 | * But we had to alter page->flags anyway. | |
1900 | * | |
1901 | * Returns the number of pages moved to the given lru. | |
1902 | */ | |
1903 | ||
1904 | static unsigned move_active_pages_to_lru(struct lruvec *lruvec, | |
1905 | struct list_head *list, | |
1906 | struct list_head *pages_to_free, | |
1907 | enum lru_list lru) | |
1908 | { | |
1909 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
1910 | struct page *page; | |
1911 | int nr_pages; | |
1912 | int nr_moved = 0; | |
1913 | ||
1914 | while (!list_empty(list)) { | |
1915 | page = lru_to_page(list); | |
1916 | lruvec = mem_cgroup_page_lruvec(page, pgdat); | |
1917 | ||
1918 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
1919 | SetPageLRU(page); | |
1920 | ||
1921 | nr_pages = hpage_nr_pages(page); | |
1922 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); | |
1923 | list_move(&page->lru, &lruvec->lists[lru]); | |
1924 | ||
1925 | if (put_page_testzero(page)) { | |
1926 | __ClearPageLRU(page); | |
1927 | __ClearPageActive(page); | |
1928 | del_page_from_lru_list(page, lruvec, lru); | |
1929 | ||
1930 | if (unlikely(PageCompound(page))) { | |
1931 | spin_unlock_irq(&pgdat->lru_lock); | |
1932 | mem_cgroup_uncharge(page); | |
1933 | (*get_compound_page_dtor(page))(page); | |
1934 | spin_lock_irq(&pgdat->lru_lock); | |
1935 | } else | |
1936 | list_add(&page->lru, pages_to_free); | |
1937 | } else { | |
1938 | nr_moved += nr_pages; | |
1939 | } | |
1940 | } | |
1941 | ||
1942 | if (!is_active_lru(lru)) { | |
1943 | __count_vm_events(PGDEACTIVATE, nr_moved); | |
1944 | count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, | |
1945 | nr_moved); | |
1946 | } | |
1947 | ||
1948 | return nr_moved; | |
1949 | } | |
1950 | ||
1951 | static void shrink_active_list(unsigned long nr_to_scan, | |
1952 | struct lruvec *lruvec, | |
1953 | struct scan_control *sc, | |
1954 | enum lru_list lru) | |
1955 | { | |
1956 | unsigned long nr_taken; | |
1957 | unsigned long nr_scanned; | |
1958 | unsigned long vm_flags; | |
1959 | LIST_HEAD(l_hold); /* The pages which were snipped off */ | |
1960 | LIST_HEAD(l_active); | |
1961 | LIST_HEAD(l_inactive); | |
1962 | struct page *page; | |
1963 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1964 | unsigned nr_deactivate, nr_activate; | |
1965 | unsigned nr_rotated = 0; | |
1966 | isolate_mode_t isolate_mode = 0; | |
1967 | int file = is_file_lru(lru); | |
1968 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
1969 | ||
1970 | lru_add_drain(); | |
1971 | ||
1972 | if (!sc->may_unmap) | |
1973 | isolate_mode |= ISOLATE_UNMAPPED; | |
1974 | ||
1975 | spin_lock_irq(&pgdat->lru_lock); | |
1976 | ||
1977 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, | |
1978 | &nr_scanned, sc, isolate_mode, lru); | |
1979 | ||
1980 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); | |
1981 | reclaim_stat->recent_scanned[file] += nr_taken; | |
1982 | ||
1983 | __count_vm_events(PGREFILL, nr_scanned); | |
1984 | count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); | |
1985 | ||
1986 | spin_unlock_irq(&pgdat->lru_lock); | |
1987 | ||
1988 | while (!list_empty(&l_hold)) { | |
1989 | cond_resched(); | |
1990 | page = lru_to_page(&l_hold); | |
1991 | list_del(&page->lru); | |
1992 | ||
1993 | if (unlikely(!page_evictable(page))) { | |
1994 | putback_lru_page(page); | |
1995 | continue; | |
1996 | } | |
1997 | ||
1998 | if (unlikely(buffer_heads_over_limit)) { | |
1999 | if (page_has_private(page) && trylock_page(page)) { | |
2000 | if (page_has_private(page)) | |
2001 | try_to_release_page(page, 0); | |
2002 | unlock_page(page); | |
2003 | } | |
2004 | } | |
2005 | ||
2006 | if (page_referenced(page, 0, sc->target_mem_cgroup, | |
2007 | &vm_flags)) { | |
2008 | nr_rotated += hpage_nr_pages(page); | |
2009 | /* | |
2010 | * Identify referenced, file-backed active pages and | |
2011 | * give them one more trip around the active list. So | |
2012 | * that executable code get better chances to stay in | |
2013 | * memory under moderate memory pressure. Anon pages | |
2014 | * are not likely to be evicted by use-once streaming | |
2015 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
2016 | * so we ignore them here. | |
2017 | */ | |
2018 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { | |
2019 | list_add(&page->lru, &l_active); | |
2020 | continue; | |
2021 | } | |
2022 | } | |
2023 | ||
2024 | ClearPageActive(page); /* we are de-activating */ | |
2025 | list_add(&page->lru, &l_inactive); | |
2026 | } | |
2027 | ||
2028 | /* | |
2029 | * Move pages back to the lru list. | |
2030 | */ | |
2031 | spin_lock_irq(&pgdat->lru_lock); | |
2032 | /* | |
2033 | * Count referenced pages from currently used mappings as rotated, | |
2034 | * even though only some of them are actually re-activated. This | |
2035 | * helps balance scan pressure between file and anonymous pages in | |
2036 | * get_scan_count. | |
2037 | */ | |
2038 | reclaim_stat->recent_rotated[file] += nr_rotated; | |
2039 | ||
2040 | nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); | |
2041 | nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
2042 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); | |
2043 | spin_unlock_irq(&pgdat->lru_lock); | |
2044 | ||
2045 | mem_cgroup_uncharge_list(&l_hold); | |
2046 | free_hot_cold_page_list(&l_hold, true); | |
2047 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, | |
2048 | nr_deactivate, nr_rotated, sc->priority, file); | |
2049 | } | |
2050 | ||
2051 | /* | |
2052 | * The inactive anon list should be small enough that the VM never has | |
2053 | * to do too much work. | |
2054 | * | |
2055 | * The inactive file list should be small enough to leave most memory | |
2056 | * to the established workingset on the scan-resistant active list, | |
2057 | * but large enough to avoid thrashing the aggregate readahead window. | |
2058 | * | |
2059 | * Both inactive lists should also be large enough that each inactive | |
2060 | * page has a chance to be referenced again before it is reclaimed. | |
2061 | * | |
2062 | * If that fails and refaulting is observed, the inactive list grows. | |
2063 | * | |
2064 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages | |
2065 | * on this LRU, maintained by the pageout code. A zone->inactive_ratio | |
2066 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. | |
2067 | * | |
2068 | * total target max | |
2069 | * memory ratio inactive | |
2070 | * ------------------------------------- | |
2071 | * 10MB 1 5MB | |
2072 | * 100MB 1 50MB | |
2073 | * 1GB 3 250MB | |
2074 | * 10GB 10 0.9GB | |
2075 | * 100GB 31 3GB | |
2076 | * 1TB 101 10GB | |
2077 | * 10TB 320 32GB | |
2078 | */ | |
2079 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, | |
2080 | struct mem_cgroup *memcg, | |
2081 | struct scan_control *sc, bool actual_reclaim) | |
2082 | { | |
2083 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; | |
2084 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
2085 | enum lru_list inactive_lru = file * LRU_FILE; | |
2086 | unsigned long inactive, active; | |
2087 | unsigned long inactive_ratio; | |
2088 | unsigned long refaults; | |
2089 | unsigned long gb; | |
2090 | ||
2091 | /* | |
2092 | * If we don't have swap space, anonymous page deactivation | |
2093 | * is pointless. | |
2094 | */ | |
2095 | if (!file && !total_swap_pages) | |
2096 | return false; | |
2097 | ||
2098 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); | |
2099 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); | |
2100 | ||
2101 | if (memcg) | |
2102 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); | |
2103 | else | |
2104 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); | |
2105 | ||
2106 | /* | |
2107 | * When refaults are being observed, it means a new workingset | |
2108 | * is being established. Disable active list protection to get | |
2109 | * rid of the stale workingset quickly. | |
2110 | */ | |
2111 | if (file && actual_reclaim && lruvec->refaults != refaults) { | |
2112 | inactive_ratio = 0; | |
2113 | } else { | |
2114 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
2115 | if (gb) | |
2116 | inactive_ratio = int_sqrt(10 * gb); | |
2117 | else | |
2118 | inactive_ratio = 1; | |
2119 | } | |
2120 | ||
2121 | if (actual_reclaim) | |
2122 | trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, | |
2123 | lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, | |
2124 | lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, | |
2125 | inactive_ratio, file); | |
2126 | ||
2127 | return inactive * inactive_ratio < active; | |
2128 | } | |
2129 | ||
2130 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, | |
2131 | struct lruvec *lruvec, struct mem_cgroup *memcg, | |
2132 | struct scan_control *sc) | |
2133 | { | |
2134 | if (is_active_lru(lru)) { | |
2135 | if (inactive_list_is_low(lruvec, is_file_lru(lru), | |
2136 | memcg, sc, true)) | |
2137 | shrink_active_list(nr_to_scan, lruvec, sc, lru); | |
2138 | return 0; | |
2139 | } | |
2140 | ||
2141 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); | |
2142 | } | |
2143 | ||
2144 | enum scan_balance { | |
2145 | SCAN_EQUAL, | |
2146 | SCAN_FRACT, | |
2147 | SCAN_ANON, | |
2148 | SCAN_FILE, | |
2149 | }; | |
2150 | ||
2151 | /* | |
2152 | * Determine how aggressively the anon and file LRU lists should be | |
2153 | * scanned. The relative value of each set of LRU lists is determined | |
2154 | * by looking at the fraction of the pages scanned we did rotate back | |
2155 | * onto the active list instead of evict. | |
2156 | * | |
2157 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan | |
2158 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
2159 | */ | |
2160 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, | |
2161 | struct scan_control *sc, unsigned long *nr, | |
2162 | unsigned long *lru_pages) | |
2163 | { | |
2164 | int swappiness = mem_cgroup_swappiness(memcg); | |
2165 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
2166 | u64 fraction[2]; | |
2167 | u64 denominator = 0; /* gcc */ | |
2168 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
2169 | unsigned long anon_prio, file_prio; | |
2170 | enum scan_balance scan_balance; | |
2171 | unsigned long anon, file; | |
2172 | unsigned long ap, fp; | |
2173 | enum lru_list lru; | |
2174 | ||
2175 | /* If we have no swap space, do not bother scanning anon pages. */ | |
2176 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { | |
2177 | scan_balance = SCAN_FILE; | |
2178 | goto out; | |
2179 | } | |
2180 | ||
2181 | /* | |
2182 | * Global reclaim will swap to prevent OOM even with no | |
2183 | * swappiness, but memcg users want to use this knob to | |
2184 | * disable swapping for individual groups completely when | |
2185 | * using the memory controller's swap limit feature would be | |
2186 | * too expensive. | |
2187 | */ | |
2188 | if (!global_reclaim(sc) && !swappiness) { | |
2189 | scan_balance = SCAN_FILE; | |
2190 | goto out; | |
2191 | } | |
2192 | ||
2193 | /* | |
2194 | * Do not apply any pressure balancing cleverness when the | |
2195 | * system is close to OOM, scan both anon and file equally | |
2196 | * (unless the swappiness setting disagrees with swapping). | |
2197 | */ | |
2198 | if (!sc->priority && swappiness) { | |
2199 | scan_balance = SCAN_EQUAL; | |
2200 | goto out; | |
2201 | } | |
2202 | ||
2203 | /* | |
2204 | * Prevent the reclaimer from falling into the cache trap: as | |
2205 | * cache pages start out inactive, every cache fault will tip | |
2206 | * the scan balance towards the file LRU. And as the file LRU | |
2207 | * shrinks, so does the window for rotation from references. | |
2208 | * This means we have a runaway feedback loop where a tiny | |
2209 | * thrashing file LRU becomes infinitely more attractive than | |
2210 | * anon pages. Try to detect this based on file LRU size. | |
2211 | */ | |
2212 | if (global_reclaim(sc)) { | |
2213 | unsigned long pgdatfile; | |
2214 | unsigned long pgdatfree; | |
2215 | int z; | |
2216 | unsigned long total_high_wmark = 0; | |
2217 | ||
2218 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); | |
2219 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2220 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2221 | ||
2222 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2223 | struct zone *zone = &pgdat->node_zones[z]; | |
2224 | if (!managed_zone(zone)) | |
2225 | continue; | |
2226 | ||
2227 | total_high_wmark += high_wmark_pages(zone); | |
2228 | } | |
2229 | ||
2230 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { | |
2231 | /* | |
2232 | * Force SCAN_ANON if there are enough inactive | |
2233 | * anonymous pages on the LRU in eligible zones. | |
2234 | * Otherwise, the small LRU gets thrashed. | |
2235 | */ | |
2236 | if (!inactive_list_is_low(lruvec, false, memcg, sc, false) && | |
2237 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx) | |
2238 | >> sc->priority) { | |
2239 | scan_balance = SCAN_ANON; | |
2240 | goto out; | |
2241 | } | |
2242 | } | |
2243 | } | |
2244 | ||
2245 | /* | |
2246 | * If there is enough inactive page cache, i.e. if the size of the | |
2247 | * inactive list is greater than that of the active list *and* the | |
2248 | * inactive list actually has some pages to scan on this priority, we | |
2249 | * do not reclaim anything from the anonymous working set right now. | |
2250 | * Without the second condition we could end up never scanning an | |
2251 | * lruvec even if it has plenty of old anonymous pages unless the | |
2252 | * system is under heavy pressure. | |
2253 | */ | |
2254 | if (!inactive_list_is_low(lruvec, true, memcg, sc, false) && | |
2255 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { | |
2256 | scan_balance = SCAN_FILE; | |
2257 | goto out; | |
2258 | } | |
2259 | ||
2260 | scan_balance = SCAN_FRACT; | |
2261 | ||
2262 | /* | |
2263 | * With swappiness at 100, anonymous and file have the same priority. | |
2264 | * This scanning priority is essentially the inverse of IO cost. | |
2265 | */ | |
2266 | anon_prio = swappiness; | |
2267 | file_prio = 200 - anon_prio; | |
2268 | ||
2269 | /* | |
2270 | * OK, so we have swap space and a fair amount of page cache | |
2271 | * pages. We use the recently rotated / recently scanned | |
2272 | * ratios to determine how valuable each cache is. | |
2273 | * | |
2274 | * Because workloads change over time (and to avoid overflow) | |
2275 | * we keep these statistics as a floating average, which ends | |
2276 | * up weighing recent references more than old ones. | |
2277 | * | |
2278 | * anon in [0], file in [1] | |
2279 | */ | |
2280 | ||
2281 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + | |
2282 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); | |
2283 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + | |
2284 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); | |
2285 | ||
2286 | spin_lock_irq(&pgdat->lru_lock); | |
2287 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { | |
2288 | reclaim_stat->recent_scanned[0] /= 2; | |
2289 | reclaim_stat->recent_rotated[0] /= 2; | |
2290 | } | |
2291 | ||
2292 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { | |
2293 | reclaim_stat->recent_scanned[1] /= 2; | |
2294 | reclaim_stat->recent_rotated[1] /= 2; | |
2295 | } | |
2296 | ||
2297 | /* | |
2298 | * The amount of pressure on anon vs file pages is inversely | |
2299 | * proportional to the fraction of recently scanned pages on | |
2300 | * each list that were recently referenced and in active use. | |
2301 | */ | |
2302 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); | |
2303 | ap /= reclaim_stat->recent_rotated[0] + 1; | |
2304 | ||
2305 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); | |
2306 | fp /= reclaim_stat->recent_rotated[1] + 1; | |
2307 | spin_unlock_irq(&pgdat->lru_lock); | |
2308 | ||
2309 | fraction[0] = ap; | |
2310 | fraction[1] = fp; | |
2311 | denominator = ap + fp + 1; | |
2312 | out: | |
2313 | *lru_pages = 0; | |
2314 | for_each_evictable_lru(lru) { | |
2315 | int file = is_file_lru(lru); | |
2316 | unsigned long size; | |
2317 | unsigned long scan; | |
2318 | ||
2319 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); | |
2320 | scan = size >> sc->priority; | |
2321 | /* | |
2322 | * If the cgroup's already been deleted, make sure to | |
2323 | * scrape out the remaining cache. | |
2324 | */ | |
2325 | if (!scan && !mem_cgroup_online(memcg)) | |
2326 | scan = min(size, SWAP_CLUSTER_MAX); | |
2327 | ||
2328 | switch (scan_balance) { | |
2329 | case SCAN_EQUAL: | |
2330 | /* Scan lists relative to size */ | |
2331 | break; | |
2332 | case SCAN_FRACT: | |
2333 | /* | |
2334 | * Scan types proportional to swappiness and | |
2335 | * their relative recent reclaim efficiency. | |
2336 | */ | |
2337 | scan = div64_u64(scan * fraction[file], | |
2338 | denominator); | |
2339 | break; | |
2340 | case SCAN_FILE: | |
2341 | case SCAN_ANON: | |
2342 | /* Scan one type exclusively */ | |
2343 | if ((scan_balance == SCAN_FILE) != file) { | |
2344 | size = 0; | |
2345 | scan = 0; | |
2346 | } | |
2347 | break; | |
2348 | default: | |
2349 | /* Look ma, no brain */ | |
2350 | BUG(); | |
2351 | } | |
2352 | ||
2353 | *lru_pages += size; | |
2354 | nr[lru] = scan; | |
2355 | } | |
2356 | } | |
2357 | ||
2358 | /* | |
2359 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. | |
2360 | */ | |
2361 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, | |
2362 | struct scan_control *sc, unsigned long *lru_pages) | |
2363 | { | |
2364 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); | |
2365 | unsigned long nr[NR_LRU_LISTS]; | |
2366 | unsigned long targets[NR_LRU_LISTS]; | |
2367 | unsigned long nr_to_scan; | |
2368 | enum lru_list lru; | |
2369 | unsigned long nr_reclaimed = 0; | |
2370 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2371 | struct blk_plug plug; | |
2372 | bool scan_adjusted; | |
2373 | ||
2374 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); | |
2375 | ||
2376 | /* Record the original scan target for proportional adjustments later */ | |
2377 | memcpy(targets, nr, sizeof(nr)); | |
2378 | ||
2379 | /* | |
2380 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2381 | * event that can occur when there is little memory pressure e.g. | |
2382 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2383 | * when the requested number of pages are reclaimed when scanning at | |
2384 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2385 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2386 | * do a batch of work at once. For memcg reclaim one check is made to | |
2387 | * abort proportional reclaim if either the file or anon lru has already | |
2388 | * dropped to zero at the first pass. | |
2389 | */ | |
2390 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2391 | sc->priority == DEF_PRIORITY); | |
2392 | ||
2393 | blk_start_plug(&plug); | |
2394 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2395 | nr[LRU_INACTIVE_FILE]) { | |
2396 | unsigned long nr_anon, nr_file, percentage; | |
2397 | unsigned long nr_scanned; | |
2398 | ||
2399 | for_each_evictable_lru(lru) { | |
2400 | if (nr[lru]) { | |
2401 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2402 | nr[lru] -= nr_to_scan; | |
2403 | ||
2404 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2405 | lruvec, memcg, sc); | |
2406 | } | |
2407 | } | |
2408 | ||
2409 | cond_resched(); | |
2410 | ||
2411 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) | |
2412 | continue; | |
2413 | ||
2414 | /* | |
2415 | * For kswapd and memcg, reclaim at least the number of pages | |
2416 | * requested. Ensure that the anon and file LRUs are scanned | |
2417 | * proportionally what was requested by get_scan_count(). We | |
2418 | * stop reclaiming one LRU and reduce the amount scanning | |
2419 | * proportional to the original scan target. | |
2420 | */ | |
2421 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2422 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2423 | ||
2424 | /* | |
2425 | * It's just vindictive to attack the larger once the smaller | |
2426 | * has gone to zero. And given the way we stop scanning the | |
2427 | * smaller below, this makes sure that we only make one nudge | |
2428 | * towards proportionality once we've got nr_to_reclaim. | |
2429 | */ | |
2430 | if (!nr_file || !nr_anon) | |
2431 | break; | |
2432 | ||
2433 | if (nr_file > nr_anon) { | |
2434 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2435 | targets[LRU_ACTIVE_ANON] + 1; | |
2436 | lru = LRU_BASE; | |
2437 | percentage = nr_anon * 100 / scan_target; | |
2438 | } else { | |
2439 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2440 | targets[LRU_ACTIVE_FILE] + 1; | |
2441 | lru = LRU_FILE; | |
2442 | percentage = nr_file * 100 / scan_target; | |
2443 | } | |
2444 | ||
2445 | /* Stop scanning the smaller of the LRU */ | |
2446 | nr[lru] = 0; | |
2447 | nr[lru + LRU_ACTIVE] = 0; | |
2448 | ||
2449 | /* | |
2450 | * Recalculate the other LRU scan count based on its original | |
2451 | * scan target and the percentage scanning already complete | |
2452 | */ | |
2453 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2454 | nr_scanned = targets[lru] - nr[lru]; | |
2455 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2456 | nr[lru] -= min(nr[lru], nr_scanned); | |
2457 | ||
2458 | lru += LRU_ACTIVE; | |
2459 | nr_scanned = targets[lru] - nr[lru]; | |
2460 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2461 | nr[lru] -= min(nr[lru], nr_scanned); | |
2462 | ||
2463 | scan_adjusted = true; | |
2464 | } | |
2465 | blk_finish_plug(&plug); | |
2466 | sc->nr_reclaimed += nr_reclaimed; | |
2467 | ||
2468 | /* | |
2469 | * Even if we did not try to evict anon pages at all, we want to | |
2470 | * rebalance the anon lru active/inactive ratio. | |
2471 | */ | |
2472 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) | |
2473 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
2474 | sc, LRU_ACTIVE_ANON); | |
2475 | } | |
2476 | ||
2477 | /* Use reclaim/compaction for costly allocs or under memory pressure */ | |
2478 | static bool in_reclaim_compaction(struct scan_control *sc) | |
2479 | { | |
2480 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && | |
2481 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || | |
2482 | sc->priority < DEF_PRIORITY - 2)) | |
2483 | return true; | |
2484 | ||
2485 | return false; | |
2486 | } | |
2487 | ||
2488 | /* | |
2489 | * Reclaim/compaction is used for high-order allocation requests. It reclaims | |
2490 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2491 | * true if more pages should be reclaimed such that when the page allocator | |
2492 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2493 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
2494 | */ | |
2495 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, | |
2496 | unsigned long nr_reclaimed, | |
2497 | unsigned long nr_scanned, | |
2498 | struct scan_control *sc) | |
2499 | { | |
2500 | unsigned long pages_for_compaction; | |
2501 | unsigned long inactive_lru_pages; | |
2502 | int z; | |
2503 | ||
2504 | /* If not in reclaim/compaction mode, stop */ | |
2505 | if (!in_reclaim_compaction(sc)) | |
2506 | return false; | |
2507 | ||
2508 | /* Consider stopping depending on scan and reclaim activity */ | |
2509 | if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) { | |
2510 | /* | |
2511 | * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the | |
2512 | * full LRU list has been scanned and we are still failing | |
2513 | * to reclaim pages. This full LRU scan is potentially | |
2514 | * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed | |
2515 | */ | |
2516 | if (!nr_reclaimed && !nr_scanned) | |
2517 | return false; | |
2518 | } else { | |
2519 | /* | |
2520 | * For non-__GFP_RETRY_MAYFAIL allocations which can presumably | |
2521 | * fail without consequence, stop if we failed to reclaim | |
2522 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2523 | * pages that were scanned. This will return to the | |
2524 | * caller faster at the risk reclaim/compaction and | |
2525 | * the resulting allocation attempt fails | |
2526 | */ | |
2527 | if (!nr_reclaimed) | |
2528 | return false; | |
2529 | } | |
2530 | ||
2531 | /* | |
2532 | * If we have not reclaimed enough pages for compaction and the | |
2533 | * inactive lists are large enough, continue reclaiming | |
2534 | */ | |
2535 | pages_for_compaction = compact_gap(sc->order); | |
2536 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); | |
2537 | if (get_nr_swap_pages() > 0) | |
2538 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); | |
2539 | if (sc->nr_reclaimed < pages_for_compaction && | |
2540 | inactive_lru_pages > pages_for_compaction) | |
2541 | return true; | |
2542 | ||
2543 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
2544 | for (z = 0; z <= sc->reclaim_idx; z++) { | |
2545 | struct zone *zone = &pgdat->node_zones[z]; | |
2546 | if (!managed_zone(zone)) | |
2547 | continue; | |
2548 | ||
2549 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
2550 | case COMPACT_SUCCESS: | |
2551 | case COMPACT_CONTINUE: | |
2552 | return false; | |
2553 | default: | |
2554 | /* check next zone */ | |
2555 | ; | |
2556 | } | |
2557 | } | |
2558 | return true; | |
2559 | } | |
2560 | ||
2561 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) | |
2562 | { | |
2563 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
2564 | unsigned long nr_reclaimed, nr_scanned; | |
2565 | bool reclaimable = false; | |
2566 | ||
2567 | do { | |
2568 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2569 | struct mem_cgroup_reclaim_cookie reclaim = { | |
2570 | .pgdat = pgdat, | |
2571 | .priority = sc->priority, | |
2572 | }; | |
2573 | unsigned long node_lru_pages = 0; | |
2574 | struct mem_cgroup *memcg; | |
2575 | ||
2576 | nr_reclaimed = sc->nr_reclaimed; | |
2577 | nr_scanned = sc->nr_scanned; | |
2578 | ||
2579 | memcg = mem_cgroup_iter(root, NULL, &reclaim); | |
2580 | do { | |
2581 | unsigned long lru_pages; | |
2582 | unsigned long reclaimed; | |
2583 | unsigned long scanned; | |
2584 | ||
2585 | if (mem_cgroup_low(root, memcg)) { | |
2586 | if (!sc->memcg_low_reclaim) { | |
2587 | sc->memcg_low_skipped = 1; | |
2588 | continue; | |
2589 | } | |
2590 | mem_cgroup_event(memcg, MEMCG_LOW); | |
2591 | } | |
2592 | ||
2593 | reclaimed = sc->nr_reclaimed; | |
2594 | scanned = sc->nr_scanned; | |
2595 | ||
2596 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); | |
2597 | node_lru_pages += lru_pages; | |
2598 | ||
2599 | if (memcg) | |
2600 | shrink_slab(sc->gfp_mask, pgdat->node_id, | |
2601 | memcg, sc->nr_scanned - scanned, | |
2602 | lru_pages); | |
2603 | ||
2604 | /* Record the group's reclaim efficiency */ | |
2605 | vmpressure(sc->gfp_mask, memcg, false, | |
2606 | sc->nr_scanned - scanned, | |
2607 | sc->nr_reclaimed - reclaimed); | |
2608 | ||
2609 | /* | |
2610 | * Direct reclaim and kswapd have to scan all memory | |
2611 | * cgroups to fulfill the overall scan target for the | |
2612 | * node. | |
2613 | * | |
2614 | * Limit reclaim, on the other hand, only cares about | |
2615 | * nr_to_reclaim pages to be reclaimed and it will | |
2616 | * retry with decreasing priority if one round over the | |
2617 | * whole hierarchy is not sufficient. | |
2618 | */ | |
2619 | if (!global_reclaim(sc) && | |
2620 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
2621 | mem_cgroup_iter_break(root, memcg); | |
2622 | break; | |
2623 | } | |
2624 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); | |
2625 | ||
2626 | /* | |
2627 | * Shrink the slab caches in the same proportion that | |
2628 | * the eligible LRU pages were scanned. | |
2629 | */ | |
2630 | if (global_reclaim(sc)) | |
2631 | shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, | |
2632 | sc->nr_scanned - nr_scanned, | |
2633 | node_lru_pages); | |
2634 | ||
2635 | if (reclaim_state) { | |
2636 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2637 | reclaim_state->reclaimed_slab = 0; | |
2638 | } | |
2639 | ||
2640 | /* Record the subtree's reclaim efficiency */ | |
2641 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
2642 | sc->nr_scanned - nr_scanned, | |
2643 | sc->nr_reclaimed - nr_reclaimed); | |
2644 | ||
2645 | if (sc->nr_reclaimed - nr_reclaimed) | |
2646 | reclaimable = true; | |
2647 | ||
2648 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, | |
2649 | sc->nr_scanned - nr_scanned, sc)); | |
2650 | ||
2651 | /* | |
2652 | * Kswapd gives up on balancing particular nodes after too | |
2653 | * many failures to reclaim anything from them and goes to | |
2654 | * sleep. On reclaim progress, reset the failure counter. A | |
2655 | * successful direct reclaim run will revive a dormant kswapd. | |
2656 | */ | |
2657 | if (reclaimable) | |
2658 | pgdat->kswapd_failures = 0; | |
2659 | ||
2660 | return reclaimable; | |
2661 | } | |
2662 | ||
2663 | /* | |
2664 | * Returns true if compaction should go ahead for a costly-order request, or | |
2665 | * the allocation would already succeed without compaction. Return false if we | |
2666 | * should reclaim first. | |
2667 | */ | |
2668 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
2669 | { | |
2670 | unsigned long watermark; | |
2671 | enum compact_result suitable; | |
2672 | ||
2673 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); | |
2674 | if (suitable == COMPACT_SUCCESS) | |
2675 | /* Allocation should succeed already. Don't reclaim. */ | |
2676 | return true; | |
2677 | if (suitable == COMPACT_SKIPPED) | |
2678 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2679 | return false; | |
2680 | ||
2681 | /* | |
2682 | * Compaction is already possible, but it takes time to run and there | |
2683 | * are potentially other callers using the pages just freed. So proceed | |
2684 | * with reclaim to make a buffer of free pages available to give | |
2685 | * compaction a reasonable chance of completing and allocating the page. | |
2686 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2687 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2688 | * we are already above the high+gap watermark, don't reclaim at all. | |
2689 | */ | |
2690 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); | |
2691 | ||
2692 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); | |
2693 | } | |
2694 | ||
2695 | /* | |
2696 | * This is the direct reclaim path, for page-allocating processes. We only | |
2697 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2698 | * request. | |
2699 | * | |
2700 | * If a zone is deemed to be full of pinned pages then just give it a light | |
2701 | * scan then give up on it. | |
2702 | */ | |
2703 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) | |
2704 | { | |
2705 | struct zoneref *z; | |
2706 | struct zone *zone; | |
2707 | unsigned long nr_soft_reclaimed; | |
2708 | unsigned long nr_soft_scanned; | |
2709 | gfp_t orig_mask; | |
2710 | pg_data_t *last_pgdat = NULL; | |
2711 | ||
2712 | /* | |
2713 | * If the number of buffer_heads in the machine exceeds the maximum | |
2714 | * allowed level, force direct reclaim to scan the highmem zone as | |
2715 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2716 | */ | |
2717 | orig_mask = sc->gfp_mask; | |
2718 | if (buffer_heads_over_limit) { | |
2719 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2720 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); | |
2721 | } | |
2722 | ||
2723 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2724 | sc->reclaim_idx, sc->nodemask) { | |
2725 | /* | |
2726 | * Take care memory controller reclaiming has small influence | |
2727 | * to global LRU. | |
2728 | */ | |
2729 | if (global_reclaim(sc)) { | |
2730 | if (!cpuset_zone_allowed(zone, | |
2731 | GFP_KERNEL | __GFP_HARDWALL)) | |
2732 | continue; | |
2733 | ||
2734 | /* | |
2735 | * If we already have plenty of memory free for | |
2736 | * compaction in this zone, don't free any more. | |
2737 | * Even though compaction is invoked for any | |
2738 | * non-zero order, only frequent costly order | |
2739 | * reclamation is disruptive enough to become a | |
2740 | * noticeable problem, like transparent huge | |
2741 | * page allocations. | |
2742 | */ | |
2743 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2744 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
2745 | compaction_ready(zone, sc)) { | |
2746 | sc->compaction_ready = true; | |
2747 | continue; | |
2748 | } | |
2749 | ||
2750 | /* | |
2751 | * Shrink each node in the zonelist once. If the | |
2752 | * zonelist is ordered by zone (not the default) then a | |
2753 | * node may be shrunk multiple times but in that case | |
2754 | * the user prefers lower zones being preserved. | |
2755 | */ | |
2756 | if (zone->zone_pgdat == last_pgdat) | |
2757 | continue; | |
2758 | ||
2759 | /* | |
2760 | * This steals pages from memory cgroups over softlimit | |
2761 | * and returns the number of reclaimed pages and | |
2762 | * scanned pages. This works for global memory pressure | |
2763 | * and balancing, not for a memcg's limit. | |
2764 | */ | |
2765 | nr_soft_scanned = 0; | |
2766 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, | |
2767 | sc->order, sc->gfp_mask, | |
2768 | &nr_soft_scanned); | |
2769 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2770 | sc->nr_scanned += nr_soft_scanned; | |
2771 | /* need some check for avoid more shrink_zone() */ | |
2772 | } | |
2773 | ||
2774 | /* See comment about same check for global reclaim above */ | |
2775 | if (zone->zone_pgdat == last_pgdat) | |
2776 | continue; | |
2777 | last_pgdat = zone->zone_pgdat; | |
2778 | shrink_node(zone->zone_pgdat, sc); | |
2779 | } | |
2780 | ||
2781 | /* | |
2782 | * Restore to original mask to avoid the impact on the caller if we | |
2783 | * promoted it to __GFP_HIGHMEM. | |
2784 | */ | |
2785 | sc->gfp_mask = orig_mask; | |
2786 | } | |
2787 | ||
2788 | static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) | |
2789 | { | |
2790 | struct mem_cgroup *memcg; | |
2791 | ||
2792 | memcg = mem_cgroup_iter(root_memcg, NULL, NULL); | |
2793 | do { | |
2794 | unsigned long refaults; | |
2795 | struct lruvec *lruvec; | |
2796 | ||
2797 | if (memcg) | |
2798 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); | |
2799 | else | |
2800 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); | |
2801 | ||
2802 | lruvec = mem_cgroup_lruvec(pgdat, memcg); | |
2803 | lruvec->refaults = refaults; | |
2804 | } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); | |
2805 | } | |
2806 | ||
2807 | /* | |
2808 | * This is the main entry point to direct page reclaim. | |
2809 | * | |
2810 | * If a full scan of the inactive list fails to free enough memory then we | |
2811 | * are "out of memory" and something needs to be killed. | |
2812 | * | |
2813 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2814 | * high - the zone may be full of dirty or under-writeback pages, which this | |
2815 | * caller can't do much about. We kick the writeback threads and take explicit | |
2816 | * naps in the hope that some of these pages can be written. But if the | |
2817 | * allocating task holds filesystem locks which prevent writeout this might not | |
2818 | * work, and the allocation attempt will fail. | |
2819 | * | |
2820 | * returns: 0, if no pages reclaimed | |
2821 | * else, the number of pages reclaimed | |
2822 | */ | |
2823 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, | |
2824 | struct scan_control *sc) | |
2825 | { | |
2826 | int initial_priority = sc->priority; | |
2827 | pg_data_t *last_pgdat; | |
2828 | struct zoneref *z; | |
2829 | struct zone *zone; | |
2830 | retry: | |
2831 | delayacct_freepages_start(); | |
2832 | ||
2833 | if (global_reclaim(sc)) | |
2834 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); | |
2835 | ||
2836 | do { | |
2837 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, | |
2838 | sc->priority); | |
2839 | sc->nr_scanned = 0; | |
2840 | shrink_zones(zonelist, sc); | |
2841 | ||
2842 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) | |
2843 | break; | |
2844 | ||
2845 | if (sc->compaction_ready) | |
2846 | break; | |
2847 | ||
2848 | /* | |
2849 | * If we're getting trouble reclaiming, start doing | |
2850 | * writepage even in laptop mode. | |
2851 | */ | |
2852 | if (sc->priority < DEF_PRIORITY - 2) | |
2853 | sc->may_writepage = 1; | |
2854 | } while (--sc->priority >= 0); | |
2855 | ||
2856 | last_pgdat = NULL; | |
2857 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
2858 | sc->nodemask) { | |
2859 | if (zone->zone_pgdat == last_pgdat) | |
2860 | continue; | |
2861 | last_pgdat = zone->zone_pgdat; | |
2862 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); | |
2863 | } | |
2864 | ||
2865 | delayacct_freepages_end(); | |
2866 | ||
2867 | if (sc->nr_reclaimed) | |
2868 | return sc->nr_reclaimed; | |
2869 | ||
2870 | /* Aborted reclaim to try compaction? don't OOM, then */ | |
2871 | if (sc->compaction_ready) | |
2872 | return 1; | |
2873 | ||
2874 | /* Untapped cgroup reserves? Don't OOM, retry. */ | |
2875 | if (sc->memcg_low_skipped) { | |
2876 | sc->priority = initial_priority; | |
2877 | sc->memcg_low_reclaim = 1; | |
2878 | sc->memcg_low_skipped = 0; | |
2879 | goto retry; | |
2880 | } | |
2881 | ||
2882 | return 0; | |
2883 | } | |
2884 | ||
2885 | static bool allow_direct_reclaim(pg_data_t *pgdat) | |
2886 | { | |
2887 | struct zone *zone; | |
2888 | unsigned long pfmemalloc_reserve = 0; | |
2889 | unsigned long free_pages = 0; | |
2890 | int i; | |
2891 | bool wmark_ok; | |
2892 | ||
2893 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
2894 | return true; | |
2895 | ||
2896 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2897 | zone = &pgdat->node_zones[i]; | |
2898 | if (!managed_zone(zone)) | |
2899 | continue; | |
2900 | ||
2901 | if (!zone_reclaimable_pages(zone)) | |
2902 | continue; | |
2903 | ||
2904 | pfmemalloc_reserve += min_wmark_pages(zone); | |
2905 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2906 | } | |
2907 | ||
2908 | /* If there are no reserves (unexpected config) then do not throttle */ | |
2909 | if (!pfmemalloc_reserve) | |
2910 | return true; | |
2911 | ||
2912 | wmark_ok = free_pages > pfmemalloc_reserve / 2; | |
2913 | ||
2914 | /* kswapd must be awake if processes are being throttled */ | |
2915 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2916 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, | |
2917 | (enum zone_type)ZONE_NORMAL); | |
2918 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2919 | } | |
2920 | ||
2921 | return wmark_ok; | |
2922 | } | |
2923 | ||
2924 | /* | |
2925 | * Throttle direct reclaimers if backing storage is backed by the network | |
2926 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2927 | * depleted. kswapd will continue to make progress and wake the processes | |
2928 | * when the low watermark is reached. | |
2929 | * | |
2930 | * Returns true if a fatal signal was delivered during throttling. If this | |
2931 | * happens, the page allocator should not consider triggering the OOM killer. | |
2932 | */ | |
2933 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, | |
2934 | nodemask_t *nodemask) | |
2935 | { | |
2936 | struct zoneref *z; | |
2937 | struct zone *zone; | |
2938 | pg_data_t *pgdat = NULL; | |
2939 | ||
2940 | /* | |
2941 | * Kernel threads should not be throttled as they may be indirectly | |
2942 | * responsible for cleaning pages necessary for reclaim to make forward | |
2943 | * progress. kjournald for example may enter direct reclaim while | |
2944 | * committing a transaction where throttling it could forcing other | |
2945 | * processes to block on log_wait_commit(). | |
2946 | */ | |
2947 | if (current->flags & PF_KTHREAD) | |
2948 | goto out; | |
2949 | ||
2950 | /* | |
2951 | * If a fatal signal is pending, this process should not throttle. | |
2952 | * It should return quickly so it can exit and free its memory | |
2953 | */ | |
2954 | if (fatal_signal_pending(current)) | |
2955 | goto out; | |
2956 | ||
2957 | /* | |
2958 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2959 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2960 | * GFP_KERNEL will be required for allocating network buffers when | |
2961 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2962 | * | |
2963 | * Throttling is based on the first usable node and throttled processes | |
2964 | * wait on a queue until kswapd makes progress and wakes them. There | |
2965 | * is an affinity then between processes waking up and where reclaim | |
2966 | * progress has been made assuming the process wakes on the same node. | |
2967 | * More importantly, processes running on remote nodes will not compete | |
2968 | * for remote pfmemalloc reserves and processes on different nodes | |
2969 | * should make reasonable progress. | |
2970 | */ | |
2971 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2972 | gfp_zone(gfp_mask), nodemask) { | |
2973 | if (zone_idx(zone) > ZONE_NORMAL) | |
2974 | continue; | |
2975 | ||
2976 | /* Throttle based on the first usable node */ | |
2977 | pgdat = zone->zone_pgdat; | |
2978 | if (allow_direct_reclaim(pgdat)) | |
2979 | goto out; | |
2980 | break; | |
2981 | } | |
2982 | ||
2983 | /* If no zone was usable by the allocation flags then do not throttle */ | |
2984 | if (!pgdat) | |
2985 | goto out; | |
2986 | ||
2987 | /* Account for the throttling */ | |
2988 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2989 | ||
2990 | /* | |
2991 | * If the caller cannot enter the filesystem, it's possible that it | |
2992 | * is due to the caller holding an FS lock or performing a journal | |
2993 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2994 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2995 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2996 | * second before continuing. | |
2997 | */ | |
2998 | if (!(gfp_mask & __GFP_FS)) { | |
2999 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
3000 | allow_direct_reclaim(pgdat), HZ); | |
3001 | ||
3002 | goto check_pending; | |
3003 | } | |
3004 | ||
3005 | /* Throttle until kswapd wakes the process */ | |
3006 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
3007 | allow_direct_reclaim(pgdat)); | |
3008 | ||
3009 | check_pending: | |
3010 | if (fatal_signal_pending(current)) | |
3011 | return true; | |
3012 | ||
3013 | out: | |
3014 | return false; | |
3015 | } | |
3016 | ||
3017 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, | |
3018 | gfp_t gfp_mask, nodemask_t *nodemask) | |
3019 | { | |
3020 | unsigned long nr_reclaimed; | |
3021 | struct scan_control sc = { | |
3022 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
3023 | .gfp_mask = current_gfp_context(gfp_mask), | |
3024 | .reclaim_idx = gfp_zone(gfp_mask), | |
3025 | .order = order, | |
3026 | .nodemask = nodemask, | |
3027 | .priority = DEF_PRIORITY, | |
3028 | .may_writepage = !laptop_mode, | |
3029 | .may_unmap = 1, | |
3030 | .may_swap = 1, | |
3031 | }; | |
3032 | ||
3033 | /* | |
3034 | * Do not enter reclaim if fatal signal was delivered while throttled. | |
3035 | * 1 is returned so that the page allocator does not OOM kill at this | |
3036 | * point. | |
3037 | */ | |
3038 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) | |
3039 | return 1; | |
3040 | ||
3041 | trace_mm_vmscan_direct_reclaim_begin(order, | |
3042 | sc.may_writepage, | |
3043 | sc.gfp_mask, | |
3044 | sc.reclaim_idx); | |
3045 | ||
3046 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); | |
3047 | ||
3048 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
3049 | ||
3050 | return nr_reclaimed; | |
3051 | } | |
3052 | ||
3053 | #ifdef CONFIG_MEMCG | |
3054 | ||
3055 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, | |
3056 | gfp_t gfp_mask, bool noswap, | |
3057 | pg_data_t *pgdat, | |
3058 | unsigned long *nr_scanned) | |
3059 | { | |
3060 | struct scan_control sc = { | |
3061 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
3062 | .target_mem_cgroup = memcg, | |
3063 | .may_writepage = !laptop_mode, | |
3064 | .may_unmap = 1, | |
3065 | .reclaim_idx = MAX_NR_ZONES - 1, | |
3066 | .may_swap = !noswap, | |
3067 | }; | |
3068 | unsigned long lru_pages; | |
3069 | ||
3070 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
3071 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
3072 | ||
3073 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, | |
3074 | sc.may_writepage, | |
3075 | sc.gfp_mask, | |
3076 | sc.reclaim_idx); | |
3077 | ||
3078 | /* | |
3079 | * NOTE: Although we can get the priority field, using it | |
3080 | * here is not a good idea, since it limits the pages we can scan. | |
3081 | * if we don't reclaim here, the shrink_node from balance_pgdat | |
3082 | * will pick up pages from other mem cgroup's as well. We hack | |
3083 | * the priority and make it zero. | |
3084 | */ | |
3085 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); | |
3086 | ||
3087 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3088 | ||
3089 | *nr_scanned = sc.nr_scanned; | |
3090 | return sc.nr_reclaimed; | |
3091 | } | |
3092 | ||
3093 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, | |
3094 | unsigned long nr_pages, | |
3095 | gfp_t gfp_mask, | |
3096 | bool may_swap) | |
3097 | { | |
3098 | struct zonelist *zonelist; | |
3099 | unsigned long nr_reclaimed; | |
3100 | int nid; | |
3101 | unsigned int noreclaim_flag; | |
3102 | struct scan_control sc = { | |
3103 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), | |
3104 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | | |
3105 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
3106 | .reclaim_idx = MAX_NR_ZONES - 1, | |
3107 | .target_mem_cgroup = memcg, | |
3108 | .priority = DEF_PRIORITY, | |
3109 | .may_writepage = !laptop_mode, | |
3110 | .may_unmap = 1, | |
3111 | .may_swap = may_swap, | |
3112 | }; | |
3113 | ||
3114 | /* | |
3115 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
3116 | * take care of from where we get pages. So the node where we start the | |
3117 | * scan does not need to be the current node. | |
3118 | */ | |
3119 | nid = mem_cgroup_select_victim_node(memcg); | |
3120 | ||
3121 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; | |
3122 | ||
3123 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
3124 | sc.may_writepage, | |
3125 | sc.gfp_mask, | |
3126 | sc.reclaim_idx); | |
3127 | ||
3128 | noreclaim_flag = memalloc_noreclaim_save(); | |
3129 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); | |
3130 | memalloc_noreclaim_restore(noreclaim_flag); | |
3131 | ||
3132 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
3133 | ||
3134 | return nr_reclaimed; | |
3135 | } | |
3136 | #endif | |
3137 | ||
3138 | static void age_active_anon(struct pglist_data *pgdat, | |
3139 | struct scan_control *sc) | |
3140 | { | |
3141 | struct mem_cgroup *memcg; | |
3142 | ||
3143 | if (!total_swap_pages) | |
3144 | return; | |
3145 | ||
3146 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
3147 | do { | |
3148 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); | |
3149 | ||
3150 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) | |
3151 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
3152 | sc, LRU_ACTIVE_ANON); | |
3153 | ||
3154 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
3155 | } while (memcg); | |
3156 | } | |
3157 | ||
3158 | /* | |
3159 | * Returns true if there is an eligible zone balanced for the request order | |
3160 | * and classzone_idx | |
3161 | */ | |
3162 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
3163 | { | |
3164 | int i; | |
3165 | unsigned long mark = -1; | |
3166 | struct zone *zone; | |
3167 | ||
3168 | for (i = 0; i <= classzone_idx; i++) { | |
3169 | zone = pgdat->node_zones + i; | |
3170 | ||
3171 | if (!managed_zone(zone)) | |
3172 | continue; | |
3173 | ||
3174 | mark = high_wmark_pages(zone); | |
3175 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) | |
3176 | return true; | |
3177 | } | |
3178 | ||
3179 | /* | |
3180 | * If a node has no populated zone within classzone_idx, it does not | |
3181 | * need balancing by definition. This can happen if a zone-restricted | |
3182 | * allocation tries to wake a remote kswapd. | |
3183 | */ | |
3184 | if (mark == -1) | |
3185 | return true; | |
3186 | ||
3187 | return false; | |
3188 | } | |
3189 | ||
3190 | /* Clear pgdat state for congested, dirty or under writeback. */ | |
3191 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
3192 | { | |
3193 | clear_bit(PGDAT_CONGESTED, &pgdat->flags); | |
3194 | clear_bit(PGDAT_DIRTY, &pgdat->flags); | |
3195 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
3196 | } | |
3197 | ||
3198 | /* | |
3199 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3200 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3201 | * | |
3202 | * Returns true if kswapd is ready to sleep | |
3203 | */ | |
3204 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) | |
3205 | { | |
3206 | /* | |
3207 | * The throttled processes are normally woken up in balance_pgdat() as | |
3208 | * soon as allow_direct_reclaim() is true. But there is a potential | |
3209 | * race between when kswapd checks the watermarks and a process gets | |
3210 | * throttled. There is also a potential race if processes get | |
3211 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3212 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3213 | * the wake up checks. If kswapd is going to sleep, no process should | |
3214 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3215 | * the wake up is premature, processes will wake kswapd and get | |
3216 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3217 | * that here we are under prepare_to_wait(). | |
3218 | */ | |
3219 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) | |
3220 | wake_up_all(&pgdat->pfmemalloc_wait); | |
3221 | ||
3222 | /* Hopeless node, leave it to direct reclaim */ | |
3223 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3224 | return true; | |
3225 | ||
3226 | if (pgdat_balanced(pgdat, order, classzone_idx)) { | |
3227 | clear_pgdat_congested(pgdat); | |
3228 | return true; | |
3229 | } | |
3230 | ||
3231 | return false; | |
3232 | } | |
3233 | ||
3234 | /* | |
3235 | * kswapd shrinks a node of pages that are at or below the highest usable | |
3236 | * zone that is currently unbalanced. | |
3237 | * | |
3238 | * Returns true if kswapd scanned at least the requested number of pages to | |
3239 | * reclaim or if the lack of progress was due to pages under writeback. | |
3240 | * This is used to determine if the scanning priority needs to be raised. | |
3241 | */ | |
3242 | static bool kswapd_shrink_node(pg_data_t *pgdat, | |
3243 | struct scan_control *sc) | |
3244 | { | |
3245 | struct zone *zone; | |
3246 | int z; | |
3247 | ||
3248 | /* Reclaim a number of pages proportional to the number of zones */ | |
3249 | sc->nr_to_reclaim = 0; | |
3250 | for (z = 0; z <= sc->reclaim_idx; z++) { | |
3251 | zone = pgdat->node_zones + z; | |
3252 | if (!managed_zone(zone)) | |
3253 | continue; | |
3254 | ||
3255 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); | |
3256 | } | |
3257 | ||
3258 | /* | |
3259 | * Historically care was taken to put equal pressure on all zones but | |
3260 | * now pressure is applied based on node LRU order. | |
3261 | */ | |
3262 | shrink_node(pgdat, sc); | |
3263 | ||
3264 | /* | |
3265 | * Fragmentation may mean that the system cannot be rebalanced for | |
3266 | * high-order allocations. If twice the allocation size has been | |
3267 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3268 | * excessive reclaim. Assume that a process requested a high-order | |
3269 | * can direct reclaim/compact. | |
3270 | */ | |
3271 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) | |
3272 | sc->order = 0; | |
3273 | ||
3274 | return sc->nr_scanned >= sc->nr_to_reclaim; | |
3275 | } | |
3276 | ||
3277 | /* | |
3278 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones | |
3279 | * that are eligible for use by the caller until at least one zone is | |
3280 | * balanced. | |
3281 | * | |
3282 | * Returns the order kswapd finished reclaiming at. | |
3283 | * | |
3284 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
3285 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is | |
3286 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone | |
3287 | * or lower is eligible for reclaim until at least one usable zone is | |
3288 | * balanced. | |
3289 | */ | |
3290 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) | |
3291 | { | |
3292 | int i; | |
3293 | unsigned long nr_soft_reclaimed; | |
3294 | unsigned long nr_soft_scanned; | |
3295 | struct zone *zone; | |
3296 | struct scan_control sc = { | |
3297 | .gfp_mask = GFP_KERNEL, | |
3298 | .order = order, | |
3299 | .priority = DEF_PRIORITY, | |
3300 | .may_writepage = !laptop_mode, | |
3301 | .may_unmap = 1, | |
3302 | .may_swap = 1, | |
3303 | }; | |
3304 | count_vm_event(PAGEOUTRUN); | |
3305 | ||
3306 | do { | |
3307 | unsigned long nr_reclaimed = sc.nr_reclaimed; | |
3308 | bool raise_priority = true; | |
3309 | ||
3310 | sc.reclaim_idx = classzone_idx; | |
3311 | ||
3312 | /* | |
3313 | * If the number of buffer_heads exceeds the maximum allowed | |
3314 | * then consider reclaiming from all zones. This has a dual | |
3315 | * purpose -- on 64-bit systems it is expected that | |
3316 | * buffer_heads are stripped during active rotation. On 32-bit | |
3317 | * systems, highmem pages can pin lowmem memory and shrinking | |
3318 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3319 | * go ahead if all eligible zones for the original allocation | |
3320 | * request are balanced to avoid excessive reclaim from kswapd. | |
3321 | */ | |
3322 | if (buffer_heads_over_limit) { | |
3323 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3324 | zone = pgdat->node_zones + i; | |
3325 | if (!managed_zone(zone)) | |
3326 | continue; | |
3327 | ||
3328 | sc.reclaim_idx = i; | |
3329 | break; | |
3330 | } | |
3331 | } | |
3332 | ||
3333 | /* | |
3334 | * Only reclaim if there are no eligible zones. Note that | |
3335 | * sc.reclaim_idx is not used as buffer_heads_over_limit may | |
3336 | * have adjusted it. | |
3337 | */ | |
3338 | if (pgdat_balanced(pgdat, sc.order, classzone_idx)) | |
3339 | goto out; | |
3340 | ||
3341 | /* | |
3342 | * Do some background aging of the anon list, to give | |
3343 | * pages a chance to be referenced before reclaiming. All | |
3344 | * pages are rotated regardless of classzone as this is | |
3345 | * about consistent aging. | |
3346 | */ | |
3347 | age_active_anon(pgdat, &sc); | |
3348 | ||
3349 | /* | |
3350 | * If we're getting trouble reclaiming, start doing writepage | |
3351 | * even in laptop mode. | |
3352 | */ | |
3353 | if (sc.priority < DEF_PRIORITY - 2) | |
3354 | sc.may_writepage = 1; | |
3355 | ||
3356 | /* Call soft limit reclaim before calling shrink_node. */ | |
3357 | sc.nr_scanned = 0; | |
3358 | nr_soft_scanned = 0; | |
3359 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, | |
3360 | sc.gfp_mask, &nr_soft_scanned); | |
3361 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3362 | ||
3363 | /* | |
3364 | * There should be no need to raise the scanning priority if | |
3365 | * enough pages are already being scanned that that high | |
3366 | * watermark would be met at 100% efficiency. | |
3367 | */ | |
3368 | if (kswapd_shrink_node(pgdat, &sc)) | |
3369 | raise_priority = false; | |
3370 | ||
3371 | /* | |
3372 | * If the low watermark is met there is no need for processes | |
3373 | * to be throttled on pfmemalloc_wait as they should not be | |
3374 | * able to safely make forward progress. Wake them | |
3375 | */ | |
3376 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
3377 | allow_direct_reclaim(pgdat)) | |
3378 | wake_up_all(&pgdat->pfmemalloc_wait); | |
3379 | ||
3380 | /* Check if kswapd should be suspending */ | |
3381 | if (try_to_freeze() || kthread_should_stop()) | |
3382 | break; | |
3383 | ||
3384 | /* | |
3385 | * Raise priority if scanning rate is too low or there was no | |
3386 | * progress in reclaiming pages | |
3387 | */ | |
3388 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; | |
3389 | if (raise_priority || !nr_reclaimed) | |
3390 | sc.priority--; | |
3391 | } while (sc.priority >= 1); | |
3392 | ||
3393 | if (!sc.nr_reclaimed) | |
3394 | pgdat->kswapd_failures++; | |
3395 | ||
3396 | out: | |
3397 | snapshot_refaults(NULL, pgdat); | |
3398 | /* | |
3399 | * Return the order kswapd stopped reclaiming at as | |
3400 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3401 | * entered the allocator slow path while kswapd was awake, order will | |
3402 | * remain at the higher level. | |
3403 | */ | |
3404 | return sc.order; | |
3405 | } | |
3406 | ||
3407 | /* | |
3408 | * pgdat->kswapd_classzone_idx is the highest zone index that a recent | |
3409 | * allocation request woke kswapd for. When kswapd has not woken recently, | |
3410 | * the value is MAX_NR_ZONES which is not a valid index. This compares a | |
3411 | * given classzone and returns it or the highest classzone index kswapd | |
3412 | * was recently woke for. | |
3413 | */ | |
3414 | static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, | |
3415 | enum zone_type classzone_idx) | |
3416 | { | |
3417 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
3418 | return classzone_idx; | |
3419 | ||
3420 | return max(pgdat->kswapd_classzone_idx, classzone_idx); | |
3421 | } | |
3422 | ||
3423 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, | |
3424 | unsigned int classzone_idx) | |
3425 | { | |
3426 | long remaining = 0; | |
3427 | DEFINE_WAIT(wait); | |
3428 | ||
3429 | if (freezing(current) || kthread_should_stop()) | |
3430 | return; | |
3431 | ||
3432 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3433 | ||
3434 | /* | |
3435 | * Try to sleep for a short interval. Note that kcompactd will only be | |
3436 | * woken if it is possible to sleep for a short interval. This is | |
3437 | * deliberate on the assumption that if reclaim cannot keep an | |
3438 | * eligible zone balanced that it's also unlikely that compaction will | |
3439 | * succeed. | |
3440 | */ | |
3441 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
3442 | /* | |
3443 | * Compaction records what page blocks it recently failed to | |
3444 | * isolate pages from and skips them in the future scanning. | |
3445 | * When kswapd is going to sleep, it is reasonable to assume | |
3446 | * that pages and compaction may succeed so reset the cache. | |
3447 | */ | |
3448 | reset_isolation_suitable(pgdat); | |
3449 | ||
3450 | /* | |
3451 | * We have freed the memory, now we should compact it to make | |
3452 | * allocation of the requested order possible. | |
3453 | */ | |
3454 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); | |
3455 | ||
3456 | remaining = schedule_timeout(HZ/10); | |
3457 | ||
3458 | /* | |
3459 | * If woken prematurely then reset kswapd_classzone_idx and | |
3460 | * order. The values will either be from a wakeup request or | |
3461 | * the previous request that slept prematurely. | |
3462 | */ | |
3463 | if (remaining) { | |
3464 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); | |
3465 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); | |
3466 | } | |
3467 | ||
3468 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3469 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3470 | } | |
3471 | ||
3472 | /* | |
3473 | * After a short sleep, check if it was a premature sleep. If not, then | |
3474 | * go fully to sleep until explicitly woken up. | |
3475 | */ | |
3476 | if (!remaining && | |
3477 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
3478 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); | |
3479 | ||
3480 | /* | |
3481 | * vmstat counters are not perfectly accurate and the estimated | |
3482 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3483 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3484 | * watermarks being breached while under pressure, we reduce the | |
3485 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3486 | * them before going back to sleep. | |
3487 | */ | |
3488 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
3489 | ||
3490 | if (!kthread_should_stop()) | |
3491 | schedule(); | |
3492 | ||
3493 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); | |
3494 | } else { | |
3495 | if (remaining) | |
3496 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3497 | else | |
3498 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3499 | } | |
3500 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3501 | } | |
3502 | ||
3503 | /* | |
3504 | * The background pageout daemon, started as a kernel thread | |
3505 | * from the init process. | |
3506 | * | |
3507 | * This basically trickles out pages so that we have _some_ | |
3508 | * free memory available even if there is no other activity | |
3509 | * that frees anything up. This is needed for things like routing | |
3510 | * etc, where we otherwise might have all activity going on in | |
3511 | * asynchronous contexts that cannot page things out. | |
3512 | * | |
3513 | * If there are applications that are active memory-allocators | |
3514 | * (most normal use), this basically shouldn't matter. | |
3515 | */ | |
3516 | static int kswapd(void *p) | |
3517 | { | |
3518 | unsigned int alloc_order, reclaim_order; | |
3519 | unsigned int classzone_idx = MAX_NR_ZONES - 1; | |
3520 | pg_data_t *pgdat = (pg_data_t*)p; | |
3521 | struct task_struct *tsk = current; | |
3522 | ||
3523 | struct reclaim_state reclaim_state = { | |
3524 | .reclaimed_slab = 0, | |
3525 | }; | |
3526 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
3527 | ||
3528 | lockdep_set_current_reclaim_state(GFP_KERNEL); | |
3529 | ||
3530 | if (!cpumask_empty(cpumask)) | |
3531 | set_cpus_allowed_ptr(tsk, cpumask); | |
3532 | current->reclaim_state = &reclaim_state; | |
3533 | ||
3534 | /* | |
3535 | * Tell the memory management that we're a "memory allocator", | |
3536 | * and that if we need more memory we should get access to it | |
3537 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3538 | * never get caught in the normal page freeing logic. | |
3539 | * | |
3540 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3541 | * you need a small amount of memory in order to be able to | |
3542 | * page out something else, and this flag essentially protects | |
3543 | * us from recursively trying to free more memory as we're | |
3544 | * trying to free the first piece of memory in the first place). | |
3545 | */ | |
3546 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; | |
3547 | set_freezable(); | |
3548 | ||
3549 | pgdat->kswapd_order = 0; | |
3550 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; | |
3551 | for ( ; ; ) { | |
3552 | bool ret; | |
3553 | ||
3554 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
3555 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); | |
3556 | ||
3557 | kswapd_try_sleep: | |
3558 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3559 | classzone_idx); | |
3560 | ||
3561 | /* Read the new order and classzone_idx */ | |
3562 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
3563 | classzone_idx = kswapd_classzone_idx(pgdat, 0); | |
3564 | pgdat->kswapd_order = 0; | |
3565 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; | |
3566 | ||
3567 | ret = try_to_freeze(); | |
3568 | if (kthread_should_stop()) | |
3569 | break; | |
3570 | ||
3571 | /* | |
3572 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3573 | * after returning from the refrigerator | |
3574 | */ | |
3575 | if (ret) | |
3576 | continue; | |
3577 | ||
3578 | /* | |
3579 | * Reclaim begins at the requested order but if a high-order | |
3580 | * reclaim fails then kswapd falls back to reclaiming for | |
3581 | * order-0. If that happens, kswapd will consider sleeping | |
3582 | * for the order it finished reclaiming at (reclaim_order) | |
3583 | * but kcompactd is woken to compact for the original | |
3584 | * request (alloc_order). | |
3585 | */ | |
3586 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, | |
3587 | alloc_order); | |
3588 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); | |
3589 | if (reclaim_order < alloc_order) | |
3590 | goto kswapd_try_sleep; | |
3591 | } | |
3592 | ||
3593 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); | |
3594 | current->reclaim_state = NULL; | |
3595 | lockdep_clear_current_reclaim_state(); | |
3596 | ||
3597 | return 0; | |
3598 | } | |
3599 | ||
3600 | /* | |
3601 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3602 | */ | |
3603 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) | |
3604 | { | |
3605 | pg_data_t *pgdat; | |
3606 | ||
3607 | if (!managed_zone(zone)) | |
3608 | return; | |
3609 | ||
3610 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) | |
3611 | return; | |
3612 | pgdat = zone->zone_pgdat; | |
3613 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, | |
3614 | classzone_idx); | |
3615 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); | |
3616 | if (!waitqueue_active(&pgdat->kswapd_wait)) | |
3617 | return; | |
3618 | ||
3619 | /* Hopeless node, leave it to direct reclaim */ | |
3620 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3621 | return; | |
3622 | ||
3623 | if (pgdat_balanced(pgdat, order, classzone_idx)) | |
3624 | return; | |
3625 | ||
3626 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order); | |
3627 | wake_up_interruptible(&pgdat->kswapd_wait); | |
3628 | } | |
3629 | ||
3630 | #ifdef CONFIG_HIBERNATION | |
3631 | /* | |
3632 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of | |
3633 | * freed pages. | |
3634 | * | |
3635 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3636 | * LRU order by reclaiming preferentially | |
3637 | * inactive > active > active referenced > active mapped | |
3638 | */ | |
3639 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) | |
3640 | { | |
3641 | struct reclaim_state reclaim_state; | |
3642 | struct scan_control sc = { | |
3643 | .nr_to_reclaim = nr_to_reclaim, | |
3644 | .gfp_mask = GFP_HIGHUSER_MOVABLE, | |
3645 | .reclaim_idx = MAX_NR_ZONES - 1, | |
3646 | .priority = DEF_PRIORITY, | |
3647 | .may_writepage = 1, | |
3648 | .may_unmap = 1, | |
3649 | .may_swap = 1, | |
3650 | .hibernation_mode = 1, | |
3651 | }; | |
3652 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
3653 | struct task_struct *p = current; | |
3654 | unsigned long nr_reclaimed; | |
3655 | unsigned int noreclaim_flag; | |
3656 | ||
3657 | noreclaim_flag = memalloc_noreclaim_save(); | |
3658 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3659 | reclaim_state.reclaimed_slab = 0; | |
3660 | p->reclaim_state = &reclaim_state; | |
3661 | ||
3662 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); | |
3663 | ||
3664 | p->reclaim_state = NULL; | |
3665 | lockdep_clear_current_reclaim_state(); | |
3666 | memalloc_noreclaim_restore(noreclaim_flag); | |
3667 | ||
3668 | return nr_reclaimed; | |
3669 | } | |
3670 | #endif /* CONFIG_HIBERNATION */ | |
3671 | ||
3672 | /* It's optimal to keep kswapds on the same CPUs as their memory, but | |
3673 | not required for correctness. So if the last cpu in a node goes | |
3674 | away, we get changed to run anywhere: as the first one comes back, | |
3675 | restore their cpu bindings. */ | |
3676 | static int kswapd_cpu_online(unsigned int cpu) | |
3677 | { | |
3678 | int nid; | |
3679 | ||
3680 | for_each_node_state(nid, N_MEMORY) { | |
3681 | pg_data_t *pgdat = NODE_DATA(nid); | |
3682 | const struct cpumask *mask; | |
3683 | ||
3684 | mask = cpumask_of_node(pgdat->node_id); | |
3685 | ||
3686 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) | |
3687 | /* One of our CPUs online: restore mask */ | |
3688 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
3689 | } | |
3690 | return 0; | |
3691 | } | |
3692 | ||
3693 | /* | |
3694 | * This kswapd start function will be called by init and node-hot-add. | |
3695 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3696 | */ | |
3697 | int kswapd_run(int nid) | |
3698 | { | |
3699 | pg_data_t *pgdat = NODE_DATA(nid); | |
3700 | int ret = 0; | |
3701 | ||
3702 | if (pgdat->kswapd) | |
3703 | return 0; | |
3704 | ||
3705 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3706 | if (IS_ERR(pgdat->kswapd)) { | |
3707 | /* failure at boot is fatal */ | |
3708 | BUG_ON(system_state < SYSTEM_RUNNING); | |
3709 | pr_err("Failed to start kswapd on node %d\n", nid); | |
3710 | ret = PTR_ERR(pgdat->kswapd); | |
3711 | pgdat->kswapd = NULL; | |
3712 | } | |
3713 | return ret; | |
3714 | } | |
3715 | ||
3716 | /* | |
3717 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
3718 | * hold mem_hotplug_begin/end(). | |
3719 | */ | |
3720 | void kswapd_stop(int nid) | |
3721 | { | |
3722 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3723 | ||
3724 | if (kswapd) { | |
3725 | kthread_stop(kswapd); | |
3726 | NODE_DATA(nid)->kswapd = NULL; | |
3727 | } | |
3728 | } | |
3729 | ||
3730 | static int __init kswapd_init(void) | |
3731 | { | |
3732 | int nid, ret; | |
3733 | ||
3734 | swap_setup(); | |
3735 | for_each_node_state(nid, N_MEMORY) | |
3736 | kswapd_run(nid); | |
3737 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, | |
3738 | "mm/vmscan:online", kswapd_cpu_online, | |
3739 | NULL); | |
3740 | WARN_ON(ret < 0); | |
3741 | return 0; | |
3742 | } | |
3743 | ||
3744 | module_init(kswapd_init) | |
3745 | ||
3746 | #ifdef CONFIG_NUMA | |
3747 | /* | |
3748 | * Node reclaim mode | |
3749 | * | |
3750 | * If non-zero call node_reclaim when the number of free pages falls below | |
3751 | * the watermarks. | |
3752 | */ | |
3753 | int node_reclaim_mode __read_mostly; | |
3754 | ||
3755 | #define RECLAIM_OFF 0 | |
3756 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ | |
3757 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ | |
3758 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ | |
3759 | ||
3760 | /* | |
3761 | * Priority for NODE_RECLAIM. This determines the fraction of pages | |
3762 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3763 | * a zone. | |
3764 | */ | |
3765 | #define NODE_RECLAIM_PRIORITY 4 | |
3766 | ||
3767 | /* | |
3768 | * Percentage of pages in a zone that must be unmapped for node_reclaim to | |
3769 | * occur. | |
3770 | */ | |
3771 | int sysctl_min_unmapped_ratio = 1; | |
3772 | ||
3773 | /* | |
3774 | * If the number of slab pages in a zone grows beyond this percentage then | |
3775 | * slab reclaim needs to occur. | |
3776 | */ | |
3777 | int sysctl_min_slab_ratio = 5; | |
3778 | ||
3779 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) | |
3780 | { | |
3781 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); | |
3782 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
3783 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
3784 | ||
3785 | /* | |
3786 | * It's possible for there to be more file mapped pages than | |
3787 | * accounted for by the pages on the file LRU lists because | |
3788 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3789 | */ | |
3790 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3791 | } | |
3792 | ||
3793 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3794 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) | |
3795 | { | |
3796 | unsigned long nr_pagecache_reclaimable; | |
3797 | unsigned long delta = 0; | |
3798 | ||
3799 | /* | |
3800 | * If RECLAIM_UNMAP is set, then all file pages are considered | |
3801 | * potentially reclaimable. Otherwise, we have to worry about | |
3802 | * pages like swapcache and node_unmapped_file_pages() provides | |
3803 | * a better estimate | |
3804 | */ | |
3805 | if (node_reclaim_mode & RECLAIM_UNMAP) | |
3806 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
3807 | else | |
3808 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); | |
3809 | ||
3810 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3811 | if (!(node_reclaim_mode & RECLAIM_WRITE)) | |
3812 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
3813 | ||
3814 | /* Watch for any possible underflows due to delta */ | |
3815 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3816 | delta = nr_pagecache_reclaimable; | |
3817 | ||
3818 | return nr_pagecache_reclaimable - delta; | |
3819 | } | |
3820 | ||
3821 | /* | |
3822 | * Try to free up some pages from this node through reclaim. | |
3823 | */ | |
3824 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) | |
3825 | { | |
3826 | /* Minimum pages needed in order to stay on node */ | |
3827 | const unsigned long nr_pages = 1 << order; | |
3828 | struct task_struct *p = current; | |
3829 | struct reclaim_state reclaim_state; | |
3830 | unsigned int noreclaim_flag; | |
3831 | struct scan_control sc = { | |
3832 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), | |
3833 | .gfp_mask = current_gfp_context(gfp_mask), | |
3834 | .order = order, | |
3835 | .priority = NODE_RECLAIM_PRIORITY, | |
3836 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
3837 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
3838 | .may_swap = 1, | |
3839 | .reclaim_idx = gfp_zone(gfp_mask), | |
3840 | }; | |
3841 | ||
3842 | cond_resched(); | |
3843 | /* | |
3844 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP | |
3845 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3846 | * and RECLAIM_UNMAP. | |
3847 | */ | |
3848 | noreclaim_flag = memalloc_noreclaim_save(); | |
3849 | p->flags |= PF_SWAPWRITE; | |
3850 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3851 | reclaim_state.reclaimed_slab = 0; | |
3852 | p->reclaim_state = &reclaim_state; | |
3853 | ||
3854 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { | |
3855 | /* | |
3856 | * Free memory by calling shrink zone with increasing | |
3857 | * priorities until we have enough memory freed. | |
3858 | */ | |
3859 | do { | |
3860 | shrink_node(pgdat, &sc); | |
3861 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); | |
3862 | } | |
3863 | ||
3864 | p->reclaim_state = NULL; | |
3865 | current->flags &= ~PF_SWAPWRITE; | |
3866 | memalloc_noreclaim_restore(noreclaim_flag); | |
3867 | lockdep_clear_current_reclaim_state(); | |
3868 | return sc.nr_reclaimed >= nr_pages; | |
3869 | } | |
3870 | ||
3871 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) | |
3872 | { | |
3873 | int ret; | |
3874 | ||
3875 | /* | |
3876 | * Node reclaim reclaims unmapped file backed pages and | |
3877 | * slab pages if we are over the defined limits. | |
3878 | * | |
3879 | * A small portion of unmapped file backed pages is needed for | |
3880 | * file I/O otherwise pages read by file I/O will be immediately | |
3881 | * thrown out if the node is overallocated. So we do not reclaim | |
3882 | * if less than a specified percentage of the node is used by | |
3883 | * unmapped file backed pages. | |
3884 | */ | |
3885 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && | |
3886 | node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) | |
3887 | return NODE_RECLAIM_FULL; | |
3888 | ||
3889 | /* | |
3890 | * Do not scan if the allocation should not be delayed. | |
3891 | */ | |
3892 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) | |
3893 | return NODE_RECLAIM_NOSCAN; | |
3894 | ||
3895 | /* | |
3896 | * Only run node reclaim on the local node or on nodes that do not | |
3897 | * have associated processors. This will favor the local processor | |
3898 | * over remote processors and spread off node memory allocations | |
3899 | * as wide as possible. | |
3900 | */ | |
3901 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) | |
3902 | return NODE_RECLAIM_NOSCAN; | |
3903 | ||
3904 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) | |
3905 | return NODE_RECLAIM_NOSCAN; | |
3906 | ||
3907 | ret = __node_reclaim(pgdat, gfp_mask, order); | |
3908 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
3909 | ||
3910 | if (!ret) | |
3911 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3912 | ||
3913 | return ret; | |
3914 | } | |
3915 | #endif | |
3916 | ||
3917 | /* | |
3918 | * page_evictable - test whether a page is evictable | |
3919 | * @page: the page to test | |
3920 | * | |
3921 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
3922 | * lists vs unevictable list. | |
3923 | * | |
3924 | * Reasons page might not be evictable: | |
3925 | * (1) page's mapping marked unevictable | |
3926 | * (2) page is part of an mlocked VMA | |
3927 | * | |
3928 | */ | |
3929 | int page_evictable(struct page *page) | |
3930 | { | |
3931 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); | |
3932 | } | |
3933 | ||
3934 | #ifdef CONFIG_SHMEM | |
3935 | /** | |
3936 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list | |
3937 | * @pages: array of pages to check | |
3938 | * @nr_pages: number of pages to check | |
3939 | * | |
3940 | * Checks pages for evictability and moves them to the appropriate lru list. | |
3941 | * | |
3942 | * This function is only used for SysV IPC SHM_UNLOCK. | |
3943 | */ | |
3944 | void check_move_unevictable_pages(struct page **pages, int nr_pages) | |
3945 | { | |
3946 | struct lruvec *lruvec; | |
3947 | struct pglist_data *pgdat = NULL; | |
3948 | int pgscanned = 0; | |
3949 | int pgrescued = 0; | |
3950 | int i; | |
3951 | ||
3952 | for (i = 0; i < nr_pages; i++) { | |
3953 | struct page *page = pages[i]; | |
3954 | struct pglist_data *pagepgdat = page_pgdat(page); | |
3955 | ||
3956 | pgscanned++; | |
3957 | if (pagepgdat != pgdat) { | |
3958 | if (pgdat) | |
3959 | spin_unlock_irq(&pgdat->lru_lock); | |
3960 | pgdat = pagepgdat; | |
3961 | spin_lock_irq(&pgdat->lru_lock); | |
3962 | } | |
3963 | lruvec = mem_cgroup_page_lruvec(page, pgdat); | |
3964 | ||
3965 | if (!PageLRU(page) || !PageUnevictable(page)) | |
3966 | continue; | |
3967 | ||
3968 | if (page_evictable(page)) { | |
3969 | enum lru_list lru = page_lru_base_type(page); | |
3970 | ||
3971 | VM_BUG_ON_PAGE(PageActive(page), page); | |
3972 | ClearPageUnevictable(page); | |
3973 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); | |
3974 | add_page_to_lru_list(page, lruvec, lru); | |
3975 | pgrescued++; | |
3976 | } | |
3977 | } | |
3978 | ||
3979 | if (pgdat) { | |
3980 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3981 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3982 | spin_unlock_irq(&pgdat->lru_lock); | |
3983 | } | |
3984 | } | |
3985 | #endif /* CONFIG_SHMEM */ |