blob: 33b88cc838dad3e284e73c0b42d86e3cfe66228a
1 | /* |
2 | * linux/mm/vmscan.c |
3 | * |
4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds |
5 | * |
6 | * Swap reorganised 29.12.95, Stephen Tweedie. |
7 | * kswapd added: 7.1.96 sct |
8 | * Removed kswapd_ctl limits, and swap out as many pages as needed |
9 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. |
10 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). |
11 | * Multiqueue VM started 5.8.00, Rik van Riel. |
12 | */ |
13 | |
14 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
15 | |
16 | #include <linux/mm.h> |
17 | #include <linux/module.h> |
18 | #include <linux/gfp.h> |
19 | #include <linux/kernel_stat.h> |
20 | #include <linux/swap.h> |
21 | #include <linux/pagemap.h> |
22 | #include <linux/init.h> |
23 | #include <linux/highmem.h> |
24 | #include <linux/vmpressure.h> |
25 | #include <linux/vmstat.h> |
26 | #include <linux/file.h> |
27 | #include <linux/writeback.h> |
28 | #include <linux/blkdev.h> |
29 | #include <linux/buffer_head.h> /* for try_to_release_page(), |
30 | buffer_heads_over_limit */ |
31 | #include <linux/mm_inline.h> |
32 | #include <linux/backing-dev.h> |
33 | #include <linux/rmap.h> |
34 | #include <linux/topology.h> |
35 | #include <linux/cpu.h> |
36 | #include <linux/cpuset.h> |
37 | #include <linux/compaction.h> |
38 | #include <linux/notifier.h> |
39 | #include <linux/rwsem.h> |
40 | #include <linux/delay.h> |
41 | #include <linux/kthread.h> |
42 | #include <linux/freezer.h> |
43 | #include <linux/memcontrol.h> |
44 | #include <linux/delayacct.h> |
45 | #include <linux/sysctl.h> |
46 | #include <linux/oom.h> |
47 | #include <linux/prefetch.h> |
48 | #include <linux/printk.h> |
49 | #include <linux/dax.h> |
50 | #include <linux/psi.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 | unsigned int may_writepage:1; |
92 | |
93 | /* Can mapped pages be reclaimed? */ |
94 | unsigned int may_unmap:1; |
95 | |
96 | /* Can pages be swapped as part of reclaim? */ |
97 | unsigned int may_swap:1; |
98 | |
99 | /* Can cgroups be reclaimed below their normal consumption range? */ |
100 | unsigned int may_thrash:1; |
101 | |
102 | unsigned int hibernation_mode:1; |
103 | |
104 | /* One of the zones is ready for compaction */ |
105 | unsigned int compaction_ready:1; |
106 | |
107 | /* Incremented by the number of inactive pages that were scanned */ |
108 | unsigned long nr_scanned; |
109 | |
110 | /* Number of pages freed so far during a call to shrink_zones() */ |
111 | unsigned long nr_reclaimed; |
112 | }; |
113 | |
114 | #ifdef ARCH_HAS_PREFETCH |
115 | #define prefetch_prev_lru_page(_page, _base, _field) \ |
116 | do { \ |
117 | if ((_page)->lru.prev != _base) { \ |
118 | struct page *prev; \ |
119 | \ |
120 | prev = lru_to_page(&(_page->lru)); \ |
121 | prefetch(&prev->_field); \ |
122 | } \ |
123 | } while (0) |
124 | #else |
125 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) |
126 | #endif |
127 | |
128 | #ifdef ARCH_HAS_PREFETCHW |
129 | #define prefetchw_prev_lru_page(_page, _base, _field) \ |
130 | do { \ |
131 | if ((_page)->lru.prev != _base) { \ |
132 | struct page *prev; \ |
133 | \ |
134 | prev = lru_to_page(&(_page->lru)); \ |
135 | prefetchw(&prev->_field); \ |
136 | } \ |
137 | } while (0) |
138 | #else |
139 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) |
140 | #endif |
141 | |
142 | /* |
143 | * From 0 .. 100. Higher means more swappy. |
144 | */ |
145 | int vm_swappiness = 60; |
146 | /* |
147 | * The total number of pages which are beyond the high watermark within all |
148 | * zones. |
149 | */ |
150 | unsigned long vm_total_pages; |
151 | |
152 | static LIST_HEAD(shrinker_list); |
153 | static DECLARE_RWSEM(shrinker_rwsem); |
154 | |
155 | #ifdef CONFIG_MEMCG |
156 | static bool global_reclaim(struct scan_control *sc) |
157 | { |
158 | return !sc->target_mem_cgroup; |
159 | } |
160 | |
161 | /** |
162 | * sane_reclaim - is the usual dirty throttling mechanism operational? |
163 | * @sc: scan_control in question |
164 | * |
165 | * The normal page dirty throttling mechanism in balance_dirty_pages() is |
166 | * completely broken with the legacy memcg and direct stalling in |
167 | * shrink_page_list() is used for throttling instead, which lacks all the |
168 | * niceties such as fairness, adaptive pausing, bandwidth proportional |
169 | * allocation and configurability. |
170 | * |
171 | * This function tests whether the vmscan currently in progress can assume |
172 | * that the normal dirty throttling mechanism is operational. |
173 | */ |
174 | static bool sane_reclaim(struct scan_control *sc) |
175 | { |
176 | struct mem_cgroup *memcg = sc->target_mem_cgroup; |
177 | |
178 | if (!memcg) |
179 | return true; |
180 | #ifdef CONFIG_CGROUP_WRITEBACK |
181 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
182 | return true; |
183 | #endif |
184 | return false; |
185 | } |
186 | #else |
187 | static bool global_reclaim(struct scan_control *sc) |
188 | { |
189 | return true; |
190 | } |
191 | |
192 | static bool sane_reclaim(struct scan_control *sc) |
193 | { |
194 | return true; |
195 | } |
196 | #endif |
197 | |
198 | /* |
199 | * This misses isolated pages which are not accounted for to save counters. |
200 | * As the data only determines if reclaim or compaction continues, it is |
201 | * not expected that isolated pages will be a dominating factor. |
202 | */ |
203 | unsigned long zone_reclaimable_pages(struct zone *zone) |
204 | { |
205 | unsigned long nr; |
206 | |
207 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + |
208 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); |
209 | if (get_nr_swap_pages() > 0) |
210 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + |
211 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); |
212 | |
213 | return nr; |
214 | } |
215 | |
216 | unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat) |
217 | { |
218 | unsigned long nr; |
219 | |
220 | nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) + |
221 | node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) + |
222 | node_page_state_snapshot(pgdat, NR_ISOLATED_FILE); |
223 | |
224 | if (get_nr_swap_pages() > 0) |
225 | nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) + |
226 | node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) + |
227 | node_page_state_snapshot(pgdat, NR_ISOLATED_ANON); |
228 | |
229 | return nr; |
230 | } |
231 | |
232 | bool pgdat_reclaimable(struct pglist_data *pgdat) |
233 | { |
234 | return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) < |
235 | pgdat_reclaimable_pages(pgdat) * 6; |
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 | if (!shrinker->nr_deferred) |
300 | return; |
301 | down_write(&shrinker_rwsem); |
302 | list_del(&shrinker->list); |
303 | up_write(&shrinker_rwsem); |
304 | kfree(shrinker->nr_deferred); |
305 | shrinker->nr_deferred = NULL; |
306 | } |
307 | EXPORT_SYMBOL(unregister_shrinker); |
308 | |
309 | #define SHRINK_BATCH 128 |
310 | |
311 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
312 | struct shrinker *shrinker, |
313 | unsigned long nr_scanned, |
314 | unsigned long nr_eligible) |
315 | { |
316 | unsigned long freed = 0; |
317 | unsigned long long delta; |
318 | long total_scan; |
319 | long freeable; |
320 | long nr; |
321 | long new_nr; |
322 | int nid = shrinkctl->nid; |
323 | long batch_size = shrinker->batch ? shrinker->batch |
324 | : SHRINK_BATCH; |
325 | long scanned = 0, next_deferred; |
326 | |
327 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
328 | if (freeable == 0) |
329 | return 0; |
330 | |
331 | /* |
332 | * copy the current shrinker scan count into a local variable |
333 | * and zero it so that other concurrent shrinker invocations |
334 | * don't also do this scanning work. |
335 | */ |
336 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); |
337 | |
338 | total_scan = nr; |
339 | delta = (4 * nr_scanned) / shrinker->seeks; |
340 | delta *= freeable; |
341 | do_div(delta, nr_eligible + 1); |
342 | total_scan += delta; |
343 | if (total_scan < 0) { |
344 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
345 | shrinker->scan_objects, total_scan); |
346 | total_scan = freeable; |
347 | next_deferred = nr; |
348 | } else |
349 | next_deferred = total_scan; |
350 | |
351 | /* |
352 | * We need to avoid excessive windup on filesystem shrinkers |
353 | * due to large numbers of GFP_NOFS allocations causing the |
354 | * shrinkers to return -1 all the time. This results in a large |
355 | * nr being built up so when a shrink that can do some work |
356 | * comes along it empties the entire cache due to nr >>> |
357 | * freeable. This is bad for sustaining a working set in |
358 | * memory. |
359 | * |
360 | * Hence only allow the shrinker to scan the entire cache when |
361 | * a large delta change is calculated directly. |
362 | */ |
363 | if (delta < freeable / 4) |
364 | total_scan = min(total_scan, freeable / 2); |
365 | |
366 | /* |
367 | * Avoid risking looping forever due to too large nr value: |
368 | * never try to free more than twice the estimate number of |
369 | * freeable entries. |
370 | */ |
371 | if (total_scan > freeable * 2) |
372 | total_scan = freeable * 2; |
373 | |
374 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, |
375 | nr_scanned, nr_eligible, |
376 | freeable, delta, total_scan); |
377 | |
378 | /* |
379 | * Normally, we should not scan less than batch_size objects in one |
380 | * pass to avoid too frequent shrinker calls, but if the slab has less |
381 | * than batch_size objects in total and we are really tight on memory, |
382 | * we will try to reclaim all available objects, otherwise we can end |
383 | * up failing allocations although there are plenty of reclaimable |
384 | * objects spread over several slabs with usage less than the |
385 | * batch_size. |
386 | * |
387 | * We detect the "tight on memory" situations by looking at the total |
388 | * number of objects we want to scan (total_scan). If it is greater |
389 | * than the total number of objects on slab (freeable), we must be |
390 | * scanning at high prio and therefore should try to reclaim as much as |
391 | * possible. |
392 | */ |
393 | while (total_scan >= batch_size || |
394 | total_scan >= freeable) { |
395 | unsigned long ret; |
396 | unsigned long nr_to_scan = min(batch_size, total_scan); |
397 | |
398 | shrinkctl->nr_to_scan = nr_to_scan; |
399 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
400 | if (ret == SHRINK_STOP) |
401 | break; |
402 | freed += ret; |
403 | |
404 | count_vm_events(SLABS_SCANNED, nr_to_scan); |
405 | total_scan -= nr_to_scan; |
406 | scanned += nr_to_scan; |
407 | |
408 | cond_resched(); |
409 | } |
410 | |
411 | if (next_deferred >= scanned) |
412 | next_deferred -= scanned; |
413 | else |
414 | next_deferred = 0; |
415 | /* |
416 | * move the unused scan count back into the shrinker in a |
417 | * manner that handles concurrent updates. If we exhausted the |
418 | * scan, there is no need to do an update. |
419 | */ |
420 | if (next_deferred > 0) |
421 | new_nr = atomic_long_add_return(next_deferred, |
422 | &shrinker->nr_deferred[nid]); |
423 | else |
424 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); |
425 | |
426 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
427 | return freed; |
428 | } |
429 | |
430 | /** |
431 | * shrink_slab - shrink slab caches |
432 | * @gfp_mask: allocation context |
433 | * @nid: node whose slab caches to target |
434 | * @memcg: memory cgroup whose slab caches to target |
435 | * @nr_scanned: pressure numerator |
436 | * @nr_eligible: pressure denominator |
437 | * |
438 | * Call the shrink functions to age shrinkable caches. |
439 | * |
440 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
441 | * unaware shrinkers will receive a node id of 0 instead. |
442 | * |
443 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
444 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan |
445 | * objects from the memory cgroup specified. Otherwise, only unaware |
446 | * shrinkers are called. |
447 | * |
448 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of |
449 | * the available objects should be scanned. Page reclaim for example |
450 | * passes the number of pages scanned and the number of pages on the |
451 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned |
452 | * when it encountered mapped pages. The ratio is further biased by |
453 | * the ->seeks setting of the shrink function, which indicates the |
454 | * cost to recreate an object relative to that of an LRU page. |
455 | * |
456 | * Returns the number of reclaimed slab objects. |
457 | */ |
458 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
459 | struct mem_cgroup *memcg, |
460 | unsigned long nr_scanned, |
461 | unsigned long nr_eligible) |
462 | { |
463 | struct shrinker *shrinker; |
464 | unsigned long freed = 0; |
465 | |
466 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) |
467 | return 0; |
468 | |
469 | if (nr_scanned == 0) |
470 | nr_scanned = SWAP_CLUSTER_MAX; |
471 | |
472 | if (!down_read_trylock(&shrinker_rwsem)) { |
473 | /* |
474 | * If we would return 0, our callers would understand that we |
475 | * have nothing else to shrink and give up trying. By returning |
476 | * 1 we keep it going and assume we'll be able to shrink next |
477 | * time. |
478 | */ |
479 | freed = 1; |
480 | goto out; |
481 | } |
482 | |
483 | list_for_each_entry(shrinker, &shrinker_list, list) { |
484 | struct shrink_control sc = { |
485 | .gfp_mask = gfp_mask, |
486 | .nid = nid, |
487 | .memcg = memcg, |
488 | }; |
489 | |
490 | /* |
491 | * If kernel memory accounting is disabled, we ignore |
492 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers |
493 | * passing NULL for memcg. |
494 | */ |
495 | if (memcg_kmem_enabled() && |
496 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) |
497 | continue; |
498 | |
499 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
500 | sc.nid = 0; |
501 | |
502 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); |
503 | } |
504 | |
505 | up_read(&shrinker_rwsem); |
506 | out: |
507 | cond_resched(); |
508 | return freed; |
509 | } |
510 | |
511 | void drop_slab_node(int nid) |
512 | { |
513 | unsigned long freed; |
514 | |
515 | do { |
516 | struct mem_cgroup *memcg = NULL; |
517 | |
518 | freed = 0; |
519 | do { |
520 | freed += shrink_slab(GFP_KERNEL, nid, memcg, |
521 | 1000, 1000); |
522 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
523 | } while (freed > 10); |
524 | } |
525 | |
526 | void drop_slab(void) |
527 | { |
528 | int nid; |
529 | |
530 | for_each_online_node(nid) |
531 | drop_slab_node(nid); |
532 | } |
533 | |
534 | static inline int is_page_cache_freeable(struct page *page) |
535 | { |
536 | /* |
537 | * A freeable page cache page is referenced only by the caller |
538 | * that isolated the page, the page cache radix tree and |
539 | * optional buffer heads at page->private. |
540 | */ |
541 | return page_count(page) - page_has_private(page) == 2; |
542 | } |
543 | |
544 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
545 | { |
546 | if (current->flags & PF_SWAPWRITE) |
547 | return 1; |
548 | if (!inode_write_congested(inode)) |
549 | return 1; |
550 | if (inode_to_bdi(inode) == current->backing_dev_info) |
551 | return 1; |
552 | return 0; |
553 | } |
554 | |
555 | /* |
556 | * We detected a synchronous write error writing a page out. Probably |
557 | * -ENOSPC. We need to propagate that into the address_space for a subsequent |
558 | * fsync(), msync() or close(). |
559 | * |
560 | * The tricky part is that after writepage we cannot touch the mapping: nothing |
561 | * prevents it from being freed up. But we have a ref on the page and once |
562 | * that page is locked, the mapping is pinned. |
563 | * |
564 | * We're allowed to run sleeping lock_page() here because we know the caller has |
565 | * __GFP_FS. |
566 | */ |
567 | static void handle_write_error(struct address_space *mapping, |
568 | struct page *page, int error) |
569 | { |
570 | lock_page(page); |
571 | if (page_mapping(page) == mapping) |
572 | mapping_set_error(mapping, error); |
573 | unlock_page(page); |
574 | } |
575 | |
576 | /* possible outcome of pageout() */ |
577 | typedef enum { |
578 | /* failed to write page out, page is locked */ |
579 | PAGE_KEEP, |
580 | /* move page to the active list, page is locked */ |
581 | PAGE_ACTIVATE, |
582 | /* page has been sent to the disk successfully, page is unlocked */ |
583 | PAGE_SUCCESS, |
584 | /* page is clean and locked */ |
585 | PAGE_CLEAN, |
586 | } pageout_t; |
587 | |
588 | /* |
589 | * pageout is called by shrink_page_list() for each dirty page. |
590 | * Calls ->writepage(). |
591 | */ |
592 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
593 | struct scan_control *sc) |
594 | { |
595 | /* |
596 | * If the page is dirty, only perform writeback if that write |
597 | * will be non-blocking. To prevent this allocation from being |
598 | * stalled by pagecache activity. But note that there may be |
599 | * stalls if we need to run get_block(). We could test |
600 | * PagePrivate for that. |
601 | * |
602 | * If this process is currently in __generic_file_write_iter() against |
603 | * this page's queue, we can perform writeback even if that |
604 | * will block. |
605 | * |
606 | * If the page is swapcache, write it back even if that would |
607 | * block, for some throttling. This happens by accident, because |
608 | * swap_backing_dev_info is bust: it doesn't reflect the |
609 | * congestion state of the swapdevs. Easy to fix, if needed. |
610 | */ |
611 | if (!is_page_cache_freeable(page)) |
612 | return PAGE_KEEP; |
613 | if (!mapping) { |
614 | /* |
615 | * Some data journaling orphaned pages can have |
616 | * page->mapping == NULL while being dirty with clean buffers. |
617 | */ |
618 | if (page_has_private(page)) { |
619 | if (try_to_free_buffers(page)) { |
620 | ClearPageDirty(page); |
621 | pr_info("%s: orphaned page\n", __func__); |
622 | return PAGE_CLEAN; |
623 | } |
624 | } |
625 | return PAGE_KEEP; |
626 | } |
627 | if (mapping->a_ops->writepage == NULL) |
628 | return PAGE_ACTIVATE; |
629 | if (!may_write_to_inode(mapping->host, sc)) |
630 | return PAGE_KEEP; |
631 | |
632 | if (clear_page_dirty_for_io(page)) { |
633 | int res; |
634 | struct writeback_control wbc = { |
635 | .sync_mode = WB_SYNC_NONE, |
636 | .nr_to_write = SWAP_CLUSTER_MAX, |
637 | .range_start = 0, |
638 | .range_end = LLONG_MAX, |
639 | .for_reclaim = 1, |
640 | }; |
641 | |
642 | SetPageReclaim(page); |
643 | res = mapping->a_ops->writepage(page, &wbc); |
644 | if (res < 0) |
645 | handle_write_error(mapping, page, res); |
646 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
647 | ClearPageReclaim(page); |
648 | return PAGE_ACTIVATE; |
649 | } |
650 | |
651 | if (!PageWriteback(page)) { |
652 | /* synchronous write or broken a_ops? */ |
653 | ClearPageReclaim(page); |
654 | } |
655 | trace_mm_vmscan_writepage(page); |
656 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
657 | return PAGE_SUCCESS; |
658 | } |
659 | |
660 | return PAGE_CLEAN; |
661 | } |
662 | |
663 | /* |
664 | * Same as remove_mapping, but if the page is removed from the mapping, it |
665 | * gets returned with a refcount of 0. |
666 | */ |
667 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
668 | bool reclaimed) |
669 | { |
670 | unsigned long flags; |
671 | |
672 | BUG_ON(!PageLocked(page)); |
673 | BUG_ON(mapping != page_mapping(page)); |
674 | |
675 | spin_lock_irqsave(&mapping->tree_lock, flags); |
676 | /* |
677 | * The non racy check for a busy page. |
678 | * |
679 | * Must be careful with the order of the tests. When someone has |
680 | * a ref to the page, it may be possible that they dirty it then |
681 | * drop the reference. So if PageDirty is tested before page_count |
682 | * here, then the following race may occur: |
683 | * |
684 | * get_user_pages(&page); |
685 | * [user mapping goes away] |
686 | * write_to(page); |
687 | * !PageDirty(page) [good] |
688 | * SetPageDirty(page); |
689 | * put_page(page); |
690 | * !page_count(page) [good, discard it] |
691 | * |
692 | * [oops, our write_to data is lost] |
693 | * |
694 | * Reversing the order of the tests ensures such a situation cannot |
695 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags |
696 | * load is not satisfied before that of page->_refcount. |
697 | * |
698 | * Note that if SetPageDirty is always performed via set_page_dirty, |
699 | * and thus under tree_lock, then this ordering is not required. |
700 | */ |
701 | if (!page_ref_freeze(page, 2)) |
702 | goto cannot_free; |
703 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
704 | if (unlikely(PageDirty(page))) { |
705 | page_ref_unfreeze(page, 2); |
706 | goto cannot_free; |
707 | } |
708 | |
709 | if (PageSwapCache(page)) { |
710 | swp_entry_t swap = { .val = page_private(page) }; |
711 | mem_cgroup_swapout(page, swap); |
712 | __delete_from_swap_cache(page); |
713 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
714 | swapcache_free(swap); |
715 | } else { |
716 | void (*freepage)(struct page *); |
717 | void *shadow = NULL; |
718 | |
719 | freepage = mapping->a_ops->freepage; |
720 | /* |
721 | * Remember a shadow entry for reclaimed file cache in |
722 | * order to detect refaults, thus thrashing, later on. |
723 | * |
724 | * But don't store shadows in an address space that is |
725 | * already exiting. This is not just an optizimation, |
726 | * inode reclaim needs to empty out the radix tree or |
727 | * the nodes are lost. Don't plant shadows behind its |
728 | * back. |
729 | * |
730 | * We also don't store shadows for DAX mappings because the |
731 | * only page cache pages found in these are zero pages |
732 | * covering holes, and because we don't want to mix DAX |
733 | * exceptional entries and shadow exceptional entries in the |
734 | * same page_tree. |
735 | */ |
736 | if (reclaimed && page_is_file_cache(page) && |
737 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
738 | shadow = workingset_eviction(mapping, page); |
739 | __delete_from_page_cache(page, shadow); |
740 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
741 | |
742 | if (freepage != NULL) |
743 | freepage(page); |
744 | } |
745 | |
746 | return 1; |
747 | |
748 | cannot_free: |
749 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
750 | return 0; |
751 | } |
752 | |
753 | /* |
754 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if |
755 | * someone else has a ref on the page, abort and return 0. If it was |
756 | * successfully detached, return 1. Assumes the caller has a single ref on |
757 | * this page. |
758 | */ |
759 | int remove_mapping(struct address_space *mapping, struct page *page) |
760 | { |
761 | if (__remove_mapping(mapping, page, false)) { |
762 | /* |
763 | * Unfreezing the refcount with 1 rather than 2 effectively |
764 | * drops the pagecache ref for us without requiring another |
765 | * atomic operation. |
766 | */ |
767 | page_ref_unfreeze(page, 1); |
768 | return 1; |
769 | } |
770 | return 0; |
771 | } |
772 | |
773 | /** |
774 | * putback_lru_page - put previously isolated page onto appropriate LRU list |
775 | * @page: page to be put back to appropriate lru list |
776 | * |
777 | * Add previously isolated @page to appropriate LRU list. |
778 | * Page may still be unevictable for other reasons. |
779 | * |
780 | * lru_lock must not be held, interrupts must be enabled. |
781 | */ |
782 | void putback_lru_page(struct page *page) |
783 | { |
784 | bool is_unevictable; |
785 | int was_unevictable = PageUnevictable(page); |
786 | |
787 | VM_BUG_ON_PAGE(PageLRU(page), page); |
788 | |
789 | redo: |
790 | ClearPageUnevictable(page); |
791 | |
792 | if (page_evictable(page)) { |
793 | /* |
794 | * For evictable pages, we can use the cache. |
795 | * In event of a race, worst case is we end up with an |
796 | * unevictable page on [in]active list. |
797 | * We know how to handle that. |
798 | */ |
799 | is_unevictable = false; |
800 | lru_cache_add(page); |
801 | } else { |
802 | /* |
803 | * Put unevictable pages directly on zone's unevictable |
804 | * list. |
805 | */ |
806 | is_unevictable = true; |
807 | add_page_to_unevictable_list(page); |
808 | /* |
809 | * When racing with an mlock or AS_UNEVICTABLE clearing |
810 | * (page is unlocked) make sure that if the other thread |
811 | * does not observe our setting of PG_lru and fails |
812 | * isolation/check_move_unevictable_pages, |
813 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
814 | * the page back to the evictable list. |
815 | * |
816 | * The other side is TestClearPageMlocked() or shmem_lock(). |
817 | */ |
818 | smp_mb(); |
819 | } |
820 | |
821 | /* |
822 | * page's status can change while we move it among lru. If an evictable |
823 | * page is on unevictable list, it never be freed. To avoid that, |
824 | * check after we added it to the list, again. |
825 | */ |
826 | if (is_unevictable && page_evictable(page)) { |
827 | if (!isolate_lru_page(page)) { |
828 | put_page(page); |
829 | goto redo; |
830 | } |
831 | /* This means someone else dropped this page from LRU |
832 | * So, it will be freed or putback to LRU again. There is |
833 | * nothing to do here. |
834 | */ |
835 | } |
836 | |
837 | if (was_unevictable && !is_unevictable) |
838 | count_vm_event(UNEVICTABLE_PGRESCUED); |
839 | else if (!was_unevictable && is_unevictable) |
840 | count_vm_event(UNEVICTABLE_PGCULLED); |
841 | |
842 | put_page(page); /* drop ref from isolate */ |
843 | } |
844 | |
845 | enum page_references { |
846 | PAGEREF_RECLAIM, |
847 | PAGEREF_RECLAIM_CLEAN, |
848 | PAGEREF_KEEP, |
849 | PAGEREF_ACTIVATE, |
850 | }; |
851 | |
852 | static enum page_references page_check_references(struct page *page, |
853 | struct scan_control *sc) |
854 | { |
855 | int referenced_ptes, referenced_page; |
856 | unsigned long vm_flags; |
857 | |
858 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
859 | &vm_flags); |
860 | referenced_page = TestClearPageReferenced(page); |
861 | |
862 | /* |
863 | * Mlock lost the isolation race with us. Let try_to_unmap() |
864 | * move the page to the unevictable list. |
865 | */ |
866 | if (vm_flags & VM_LOCKED) |
867 | return PAGEREF_RECLAIM; |
868 | |
869 | if (referenced_ptes) { |
870 | if (PageSwapBacked(page)) |
871 | return PAGEREF_ACTIVATE; |
872 | /* |
873 | * All mapped pages start out with page table |
874 | * references from the instantiating fault, so we need |
875 | * to look twice if a mapped file page is used more |
876 | * than once. |
877 | * |
878 | * Mark it and spare it for another trip around the |
879 | * inactive list. Another page table reference will |
880 | * lead to its activation. |
881 | * |
882 | * Note: the mark is set for activated pages as well |
883 | * so that recently deactivated but used pages are |
884 | * quickly recovered. |
885 | */ |
886 | SetPageReferenced(page); |
887 | |
888 | if (referenced_page || referenced_ptes > 1) |
889 | return PAGEREF_ACTIVATE; |
890 | |
891 | /* |
892 | * Activate file-backed executable pages after first usage. |
893 | */ |
894 | if (vm_flags & VM_EXEC) |
895 | return PAGEREF_ACTIVATE; |
896 | |
897 | return PAGEREF_KEEP; |
898 | } |
899 | |
900 | /* Reclaim if clean, defer dirty pages to writeback */ |
901 | if (referenced_page && !PageSwapBacked(page)) |
902 | return PAGEREF_RECLAIM_CLEAN; |
903 | |
904 | return PAGEREF_RECLAIM; |
905 | } |
906 | |
907 | /* Check if a page is dirty or under writeback */ |
908 | static void page_check_dirty_writeback(struct page *page, |
909 | bool *dirty, bool *writeback) |
910 | { |
911 | struct address_space *mapping; |
912 | |
913 | /* |
914 | * Anonymous pages are not handled by flushers and must be written |
915 | * from reclaim context. Do not stall reclaim based on them |
916 | */ |
917 | if (!page_is_file_cache(page)) { |
918 | *dirty = false; |
919 | *writeback = false; |
920 | return; |
921 | } |
922 | |
923 | /* By default assume that the page flags are accurate */ |
924 | *dirty = PageDirty(page); |
925 | *writeback = PageWriteback(page); |
926 | |
927 | /* Verify dirty/writeback state if the filesystem supports it */ |
928 | if (!page_has_private(page)) |
929 | return; |
930 | |
931 | mapping = page_mapping(page); |
932 | if (mapping && mapping->a_ops->is_dirty_writeback) |
933 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); |
934 | } |
935 | |
936 | /* |
937 | * shrink_page_list() returns the number of reclaimed pages |
938 | */ |
939 | static unsigned long shrink_page_list(struct list_head *page_list, |
940 | struct pglist_data *pgdat, |
941 | struct scan_control *sc, |
942 | enum ttu_flags ttu_flags, |
943 | unsigned long *ret_nr_dirty, |
944 | unsigned long *ret_nr_unqueued_dirty, |
945 | unsigned long *ret_nr_congested, |
946 | unsigned long *ret_nr_writeback, |
947 | unsigned long *ret_nr_immediate, |
948 | bool force_reclaim) |
949 | { |
950 | LIST_HEAD(ret_pages); |
951 | LIST_HEAD(free_pages); |
952 | int pgactivate = 0; |
953 | unsigned long nr_unqueued_dirty = 0; |
954 | unsigned long nr_dirty = 0; |
955 | unsigned long nr_congested = 0; |
956 | unsigned long nr_reclaimed = 0; |
957 | unsigned long nr_writeback = 0; |
958 | unsigned long nr_immediate = 0; |
959 | |
960 | cond_resched(); |
961 | |
962 | while (!list_empty(page_list)) { |
963 | struct address_space *mapping; |
964 | struct page *page; |
965 | int may_enter_fs; |
966 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
967 | bool dirty, writeback; |
968 | bool lazyfree = false; |
969 | int ret = SWAP_SUCCESS; |
970 | |
971 | cond_resched(); |
972 | |
973 | page = lru_to_page(page_list); |
974 | list_del(&page->lru); |
975 | |
976 | if (!trylock_page(page)) |
977 | goto keep; |
978 | |
979 | VM_BUG_ON_PAGE(PageActive(page), page); |
980 | |
981 | sc->nr_scanned++; |
982 | |
983 | if (unlikely(!page_evictable(page))) |
984 | goto cull_mlocked; |
985 | |
986 | if (!sc->may_unmap && page_mapped(page)) |
987 | goto keep_locked; |
988 | |
989 | /* Double the slab pressure for mapped and swapcache pages */ |
990 | if (page_mapped(page) || PageSwapCache(page)) |
991 | sc->nr_scanned++; |
992 | |
993 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
994 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); |
995 | |
996 | /* |
997 | * The number of dirty pages determines if a zone is marked |
998 | * reclaim_congested which affects wait_iff_congested. kswapd |
999 | * will stall and start writing pages if the tail of the LRU |
1000 | * is all dirty unqueued pages. |
1001 | */ |
1002 | page_check_dirty_writeback(page, &dirty, &writeback); |
1003 | if (dirty || writeback) |
1004 | nr_dirty++; |
1005 | |
1006 | if (dirty && !writeback) |
1007 | nr_unqueued_dirty++; |
1008 | |
1009 | /* |
1010 | * Treat this page as congested if the underlying BDI is or if |
1011 | * pages are cycling through the LRU so quickly that the |
1012 | * pages marked for immediate reclaim are making it to the |
1013 | * end of the LRU a second time. |
1014 | */ |
1015 | mapping = page_mapping(page); |
1016 | if (((dirty || writeback) && mapping && |
1017 | inode_write_congested(mapping->host)) || |
1018 | (writeback && PageReclaim(page))) |
1019 | nr_congested++; |
1020 | |
1021 | /* |
1022 | * If a page at the tail of the LRU is under writeback, there |
1023 | * are three cases to consider. |
1024 | * |
1025 | * 1) If reclaim is encountering an excessive number of pages |
1026 | * under writeback and this page is both under writeback and |
1027 | * PageReclaim then it indicates that pages are being queued |
1028 | * for IO but are being recycled through the LRU before the |
1029 | * IO can complete. Waiting on the page itself risks an |
1030 | * indefinite stall if it is impossible to writeback the |
1031 | * page due to IO error or disconnected storage so instead |
1032 | * note that the LRU is being scanned too quickly and the |
1033 | * caller can stall after page list has been processed. |
1034 | * |
1035 | * 2) Global or new memcg reclaim encounters a page that is |
1036 | * not marked for immediate reclaim, or the caller does not |
1037 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, |
1038 | * not to fs). In this case mark the page for immediate |
1039 | * reclaim and continue scanning. |
1040 | * |
1041 | * Require may_enter_fs because we would wait on fs, which |
1042 | * may not have submitted IO yet. And the loop driver might |
1043 | * enter reclaim, and deadlock if it waits on a page for |
1044 | * which it is needed to do the write (loop masks off |
1045 | * __GFP_IO|__GFP_FS for this reason); but more thought |
1046 | * would probably show more reasons. |
1047 | * |
1048 | * 3) Legacy memcg encounters a page that is already marked |
1049 | * PageReclaim. memcg does not have any dirty pages |
1050 | * throttling so we could easily OOM just because too many |
1051 | * pages are in writeback and there is nothing else to |
1052 | * reclaim. Wait for the writeback to complete. |
1053 | */ |
1054 | if (PageWriteback(page)) { |
1055 | /* Case 1 above */ |
1056 | if (current_is_kswapd() && |
1057 | PageReclaim(page) && |
1058 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
1059 | nr_immediate++; |
1060 | goto keep_locked; |
1061 | |
1062 | /* Case 2 above */ |
1063 | } else if (sane_reclaim(sc) || |
1064 | !PageReclaim(page) || !may_enter_fs) { |
1065 | /* |
1066 | * This is slightly racy - end_page_writeback() |
1067 | * might have just cleared PageReclaim, then |
1068 | * setting PageReclaim here end up interpreted |
1069 | * as PageReadahead - but that does not matter |
1070 | * enough to care. What we do want is for this |
1071 | * page to have PageReclaim set next time memcg |
1072 | * reclaim reaches the tests above, so it will |
1073 | * then wait_on_page_writeback() to avoid OOM; |
1074 | * and it's also appropriate in global reclaim. |
1075 | */ |
1076 | SetPageReclaim(page); |
1077 | nr_writeback++; |
1078 | goto keep_locked; |
1079 | |
1080 | /* Case 3 above */ |
1081 | } else { |
1082 | unlock_page(page); |
1083 | wait_on_page_writeback(page); |
1084 | /* then go back and try same page again */ |
1085 | list_add_tail(&page->lru, page_list); |
1086 | continue; |
1087 | } |
1088 | } |
1089 | |
1090 | if (!force_reclaim) |
1091 | references = page_check_references(page, sc); |
1092 | |
1093 | switch (references) { |
1094 | case PAGEREF_ACTIVATE: |
1095 | goto activate_locked; |
1096 | case PAGEREF_KEEP: |
1097 | goto keep_locked; |
1098 | case PAGEREF_RECLAIM: |
1099 | case PAGEREF_RECLAIM_CLEAN: |
1100 | ; /* try to reclaim the page below */ |
1101 | } |
1102 | |
1103 | /* |
1104 | * Anonymous process memory has backing store? |
1105 | * Try to allocate it some swap space here. |
1106 | */ |
1107 | if (PageAnon(page) && !PageSwapCache(page)) { |
1108 | if (!(sc->gfp_mask & __GFP_IO)) |
1109 | goto keep_locked; |
1110 | if (!add_to_swap(page, page_list)) |
1111 | goto activate_locked; |
1112 | lazyfree = true; |
1113 | may_enter_fs = 1; |
1114 | |
1115 | /* Adding to swap updated mapping */ |
1116 | mapping = page_mapping(page); |
1117 | } else if (unlikely(PageTransHuge(page))) { |
1118 | /* Split file THP */ |
1119 | if (split_huge_page_to_list(page, page_list)) |
1120 | goto keep_locked; |
1121 | } |
1122 | |
1123 | VM_BUG_ON_PAGE(PageTransHuge(page), page); |
1124 | |
1125 | /* |
1126 | * The page is mapped into the page tables of one or more |
1127 | * processes. Try to unmap it here. |
1128 | */ |
1129 | if (page_mapped(page) && mapping) { |
1130 | switch (ret = try_to_unmap(page, lazyfree ? |
1131 | (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) : |
1132 | (ttu_flags | TTU_BATCH_FLUSH))) { |
1133 | case SWAP_FAIL: |
1134 | goto activate_locked; |
1135 | case SWAP_AGAIN: |
1136 | goto keep_locked; |
1137 | case SWAP_MLOCK: |
1138 | goto cull_mlocked; |
1139 | case SWAP_LZFREE: |
1140 | goto lazyfree; |
1141 | case SWAP_SUCCESS: |
1142 | ; /* try to free the page below */ |
1143 | } |
1144 | } |
1145 | |
1146 | if (PageDirty(page)) { |
1147 | /* |
1148 | * Only kswapd can writeback filesystem pages to |
1149 | * avoid risk of stack overflow but only writeback |
1150 | * if many dirty pages have been encountered. |
1151 | */ |
1152 | if (page_is_file_cache(page) && |
1153 | (!current_is_kswapd() || |
1154 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { |
1155 | /* |
1156 | * Immediately reclaim when written back. |
1157 | * Similar in principal to deactivate_page() |
1158 | * except we already have the page isolated |
1159 | * and know it's dirty |
1160 | */ |
1161 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
1162 | SetPageReclaim(page); |
1163 | |
1164 | goto keep_locked; |
1165 | } |
1166 | |
1167 | if (references == PAGEREF_RECLAIM_CLEAN) |
1168 | goto keep_locked; |
1169 | if (!may_enter_fs) |
1170 | goto keep_locked; |
1171 | if (!sc->may_writepage) |
1172 | goto keep_locked; |
1173 | |
1174 | /* |
1175 | * Page is dirty. Flush the TLB if a writable entry |
1176 | * potentially exists to avoid CPU writes after IO |
1177 | * starts and then write it out here. |
1178 | */ |
1179 | try_to_unmap_flush_dirty(); |
1180 | switch (pageout(page, mapping, sc)) { |
1181 | case PAGE_KEEP: |
1182 | goto keep_locked; |
1183 | case PAGE_ACTIVATE: |
1184 | goto activate_locked; |
1185 | case PAGE_SUCCESS: |
1186 | if (PageWriteback(page)) |
1187 | goto keep; |
1188 | if (PageDirty(page)) |
1189 | goto keep; |
1190 | |
1191 | /* |
1192 | * A synchronous write - probably a ramdisk. Go |
1193 | * ahead and try to reclaim the page. |
1194 | */ |
1195 | if (!trylock_page(page)) |
1196 | goto keep; |
1197 | if (PageDirty(page) || PageWriteback(page)) |
1198 | goto keep_locked; |
1199 | mapping = page_mapping(page); |
1200 | case PAGE_CLEAN: |
1201 | ; /* try to free the page below */ |
1202 | } |
1203 | } |
1204 | |
1205 | /* |
1206 | * If the page has buffers, try to free the buffer mappings |
1207 | * associated with this page. If we succeed we try to free |
1208 | * the page as well. |
1209 | * |
1210 | * We do this even if the page is PageDirty(). |
1211 | * try_to_release_page() does not perform I/O, but it is |
1212 | * possible for a page to have PageDirty set, but it is actually |
1213 | * clean (all its buffers are clean). This happens if the |
1214 | * buffers were written out directly, with submit_bh(). ext3 |
1215 | * will do this, as well as the blockdev mapping. |
1216 | * try_to_release_page() will discover that cleanness and will |
1217 | * drop the buffers and mark the page clean - it can be freed. |
1218 | * |
1219 | * Rarely, pages can have buffers and no ->mapping. These are |
1220 | * the pages which were not successfully invalidated in |
1221 | * truncate_complete_page(). We try to drop those buffers here |
1222 | * and if that worked, and the page is no longer mapped into |
1223 | * process address space (page_count == 1) it can be freed. |
1224 | * Otherwise, leave the page on the LRU so it is swappable. |
1225 | */ |
1226 | if (page_has_private(page)) { |
1227 | if (!try_to_release_page(page, sc->gfp_mask)) |
1228 | goto activate_locked; |
1229 | if (!mapping && page_count(page) == 1) { |
1230 | unlock_page(page); |
1231 | if (put_page_testzero(page)) |
1232 | goto free_it; |
1233 | else { |
1234 | /* |
1235 | * rare race with speculative reference. |
1236 | * the speculative reference will free |
1237 | * this page shortly, so we may |
1238 | * increment nr_reclaimed here (and |
1239 | * leave it off the LRU). |
1240 | */ |
1241 | nr_reclaimed++; |
1242 | continue; |
1243 | } |
1244 | } |
1245 | } |
1246 | |
1247 | lazyfree: |
1248 | if (!mapping || !__remove_mapping(mapping, page, true)) |
1249 | goto keep_locked; |
1250 | |
1251 | /* |
1252 | * At this point, we have no other references and there is |
1253 | * no way to pick any more up (removed from LRU, removed |
1254 | * from pagecache). Can use non-atomic bitops now (and |
1255 | * we obviously don't have to worry about waking up a process |
1256 | * waiting on the page lock, because there are no references. |
1257 | */ |
1258 | __ClearPageLocked(page); |
1259 | free_it: |
1260 | if (ret == SWAP_LZFREE) |
1261 | count_vm_event(PGLAZYFREED); |
1262 | |
1263 | nr_reclaimed++; |
1264 | |
1265 | /* |
1266 | * Is there need to periodically free_page_list? It would |
1267 | * appear not as the counts should be low |
1268 | */ |
1269 | list_add(&page->lru, &free_pages); |
1270 | #ifdef CONFIG_AMLOGIC_CMA |
1271 | if (ttu_flags & TTU_IGNORE_ACCESS) |
1272 | ClearPageCmaAllocating(page); |
1273 | #endif |
1274 | continue; |
1275 | |
1276 | cull_mlocked: |
1277 | if (PageSwapCache(page)) |
1278 | try_to_free_swap(page); |
1279 | unlock_page(page); |
1280 | list_add(&page->lru, &ret_pages); |
1281 | continue; |
1282 | |
1283 | activate_locked: |
1284 | /* Not a candidate for swapping, so reclaim swap space. */ |
1285 | if (PageSwapCache(page) && mem_cgroup_swap_full(page)) |
1286 | try_to_free_swap(page); |
1287 | VM_BUG_ON_PAGE(PageActive(page), page); |
1288 | SetPageActive(page); |
1289 | pgactivate++; |
1290 | keep_locked: |
1291 | unlock_page(page); |
1292 | keep: |
1293 | list_add(&page->lru, &ret_pages); |
1294 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1295 | } |
1296 | |
1297 | mem_cgroup_uncharge_list(&free_pages); |
1298 | try_to_unmap_flush(); |
1299 | free_hot_cold_page_list(&free_pages, true); |
1300 | |
1301 | list_splice(&ret_pages, page_list); |
1302 | count_vm_events(PGACTIVATE, pgactivate); |
1303 | |
1304 | *ret_nr_dirty += nr_dirty; |
1305 | *ret_nr_congested += nr_congested; |
1306 | *ret_nr_unqueued_dirty += nr_unqueued_dirty; |
1307 | *ret_nr_writeback += nr_writeback; |
1308 | *ret_nr_immediate += nr_immediate; |
1309 | return nr_reclaimed; |
1310 | } |
1311 | |
1312 | #ifdef CONFIG_AMLOGIC_CMA |
1313 | #define ACTIVE_MIGRATE 3 |
1314 | #define INACTIVE_MIGRATE (ACTIVE_MIGRATE * 4) |
1315 | static int filecache_need_migrate(struct page *page) |
1316 | { |
1317 | if (PageActive(page) && page_mapcount(page) >= ACTIVE_MIGRATE) |
1318 | return 1; |
1319 | |
1320 | if (!PageActive(page) && page_mapcount(page) >= INACTIVE_MIGRATE) |
1321 | return 1; |
1322 | |
1323 | if (PageUnevictable(page)) |
1324 | return 0; |
1325 | |
1326 | return 0; |
1327 | } |
1328 | #endif |
1329 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1330 | struct list_head *page_list) |
1331 | { |
1332 | struct scan_control sc = { |
1333 | .gfp_mask = GFP_KERNEL, |
1334 | .priority = DEF_PRIORITY, |
1335 | .may_unmap = 1, |
1336 | }; |
1337 | unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; |
1338 | struct page *page, *next; |
1339 | LIST_HEAD(clean_pages); |
1340 | #ifdef CONFIG_AMLOGIC_CMA |
1341 | LIST_HEAD(high_active_pages); |
1342 | unsigned long nr_high_active = 0; |
1343 | unsigned long nr_normal_pages = 0; |
1344 | #endif |
1345 | |
1346 | list_for_each_entry_safe(page, next, page_list, lru) { |
1347 | if (page_is_file_cache(page) && !PageDirty(page) && |
1348 | !__PageMovable(page)) { |
1349 | #ifdef CONFIG_AMLOGIC_CMA |
1350 | if (filecache_need_migrate(page)) { |
1351 | /* |
1352 | * leaving pages with high map count to migrate |
1353 | * instead of reclaimed. This can help to avoid |
1354 | * file cache jolt if reclaim large cma size |
1355 | */ |
1356 | list_move(&page->lru, &high_active_pages); |
1357 | nr_high_active++; |
1358 | } else { |
1359 | ClearPageActive(page); |
1360 | list_move(&page->lru, &clean_pages); |
1361 | nr_normal_pages++; |
1362 | } |
1363 | #else |
1364 | ClearPageActive(page); |
1365 | list_move(&page->lru, &clean_pages); |
1366 | #endif |
1367 | } |
1368 | } |
1369 | |
1370 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
1371 | TTU_UNMAP|TTU_IGNORE_ACCESS, |
1372 | &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); |
1373 | list_splice(&clean_pages, page_list); |
1374 | #ifdef CONFIG_AMLOGIC_CMA |
1375 | list_splice(&high_active_pages, page_list); |
1376 | pr_debug("high_active:%4ld, normal:%4ld, reclaimed:%4ld\n", |
1377 | nr_high_active, nr_normal_pages, ret); |
1378 | #endif |
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_CLEAN means that only clean pages should be isolated. This |
1413 | * is used by reclaim when it is cannot write to backing storage |
1414 | * |
1415 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1416 | * that it is possible to migrate without blocking |
1417 | */ |
1418 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { |
1419 | /* All the caller can do on PageWriteback is block */ |
1420 | if (PageWriteback(page)) |
1421 | return ret; |
1422 | |
1423 | if (PageDirty(page)) { |
1424 | struct address_space *mapping; |
1425 | bool migrate_dirty; |
1426 | |
1427 | /* ISOLATE_CLEAN means only clean pages */ |
1428 | if (mode & ISOLATE_CLEAN) |
1429 | return ret; |
1430 | |
1431 | /* |
1432 | * Only pages without mappings or that have a |
1433 | * ->migratepage callback are possible to migrate |
1434 | * without blocking. However, we can be racing with |
1435 | * truncation so it's necessary to lock the page |
1436 | * to stabilise the mapping as truncation holds |
1437 | * the page lock until after the page is removed |
1438 | * from the page cache. |
1439 | */ |
1440 | if (!trylock_page(page)) |
1441 | return ret; |
1442 | |
1443 | mapping = page_mapping(page); |
1444 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
1445 | unlock_page(page); |
1446 | if (!migrate_dirty) |
1447 | return ret; |
1448 | } |
1449 | } |
1450 | |
1451 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1452 | return ret; |
1453 | |
1454 | if (likely(get_page_unless_zero(page))) { |
1455 | /* |
1456 | * Be careful not to clear PageLRU until after we're |
1457 | * sure the page is not being freed elsewhere -- the |
1458 | * page release code relies on it. |
1459 | */ |
1460 | ClearPageLRU(page); |
1461 | ret = 0; |
1462 | } |
1463 | |
1464 | return ret; |
1465 | } |
1466 | |
1467 | |
1468 | /* |
1469 | * Update LRU sizes after isolating pages. The LRU size updates must |
1470 | * be complete before mem_cgroup_update_lru_size due to a santity check. |
1471 | */ |
1472 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, |
1473 | enum lru_list lru, unsigned long *nr_zone_taken) |
1474 | { |
1475 | int zid; |
1476 | |
1477 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1478 | if (!nr_zone_taken[zid]) |
1479 | continue; |
1480 | |
1481 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
1482 | #ifdef CONFIG_MEMCG |
1483 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
1484 | #endif |
1485 | } |
1486 | |
1487 | } |
1488 | |
1489 | /* |
1490 | * zone_lru_lock is heavily contended. Some of the functions that |
1491 | * shrink the lists perform better by taking out a batch of pages |
1492 | * and working on them outside the LRU lock. |
1493 | * |
1494 | * For pagecache intensive workloads, this function is the hottest |
1495 | * spot in the kernel (apart from copy_*_user functions). |
1496 | * |
1497 | * Appropriate locks must be held before calling this function. |
1498 | * |
1499 | * @nr_to_scan: The number of pages to look through on the list. |
1500 | * @lruvec: The LRU vector to pull pages from. |
1501 | * @dst: The temp list to put pages on to. |
1502 | * @nr_scanned: The number of pages that were scanned. |
1503 | * @sc: The scan_control struct for this reclaim session |
1504 | * @mode: One of the LRU isolation modes |
1505 | * @lru: LRU list id for isolating |
1506 | * |
1507 | * returns how many pages were moved onto *@dst. |
1508 | */ |
1509 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
1510 | struct lruvec *lruvec, struct list_head *dst, |
1511 | unsigned long *nr_scanned, struct scan_control *sc, |
1512 | isolate_mode_t mode, enum lru_list lru) |
1513 | { |
1514 | struct list_head *src = &lruvec->lists[lru]; |
1515 | unsigned long nr_taken = 0; |
1516 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
1517 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
1518 | unsigned long scan, nr_pages; |
1519 | LIST_HEAD(pages_skipped); |
1520 | #ifdef CONFIG_AMLOGIC_MODIFY |
1521 | int num = NR_INACTIVE_ANON_CMA - NR_INACTIVE_ANON; |
1522 | int migrate_type = 0; |
1523 | #endif /* CONFIG_AMLOGIC_MODIFY */ |
1524 | |
1525 | for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan && |
1526 | !list_empty(src);) { |
1527 | struct page *page; |
1528 | |
1529 | page = lru_to_page(src); |
1530 | prefetchw_prev_lru_page(page, src, flags); |
1531 | |
1532 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
1533 | |
1534 | if (page_zonenum(page) > sc->reclaim_idx) { |
1535 | list_move(&page->lru, &pages_skipped); |
1536 | nr_skipped[page_zonenum(page)]++; |
1537 | continue; |
1538 | } |
1539 | |
1540 | /* |
1541 | * Account for scanned and skipped separetly to avoid the pgdat |
1542 | * being prematurely marked unreclaimable by pgdat_reclaimable. |
1543 | */ |
1544 | scan++; |
1545 | |
1546 | switch (__isolate_lru_page(page, mode)) { |
1547 | case 0: |
1548 | nr_pages = hpage_nr_pages(page); |
1549 | nr_taken += nr_pages; |
1550 | nr_zone_taken[page_zonenum(page)] += nr_pages; |
1551 | list_move(&page->lru, dst); |
1552 | #ifdef CONFIG_AMLOGIC_MODIFY |
1553 | migrate_type = get_pageblock_migratetype(page); |
1554 | if (is_migrate_cma(migrate_type) || |
1555 | is_migrate_isolate(migrate_type)) |
1556 | __mod_zone_page_state(page_zone(page), |
1557 | NR_LRU_BASE + lru + num, |
1558 | -nr_pages); |
1559 | #endif /* CONFIG_AMLOGIC_MODIFY */ |
1560 | break; |
1561 | |
1562 | case -EBUSY: |
1563 | /* else it is being freed elsewhere */ |
1564 | list_move(&page->lru, src); |
1565 | continue; |
1566 | |
1567 | default: |
1568 | BUG(); |
1569 | } |
1570 | } |
1571 | |
1572 | /* |
1573 | * Splice any skipped pages to the start of the LRU list. Note that |
1574 | * this disrupts the LRU order when reclaiming for lower zones but |
1575 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX |
1576 | * scanning would soon rescan the same pages to skip and put the |
1577 | * system at risk of premature OOM. |
1578 | */ |
1579 | if (!list_empty(&pages_skipped)) { |
1580 | int zid; |
1581 | unsigned long total_skipped = 0; |
1582 | |
1583 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1584 | if (!nr_skipped[zid]) |
1585 | continue; |
1586 | |
1587 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); |
1588 | total_skipped += nr_skipped[zid]; |
1589 | } |
1590 | |
1591 | /* |
1592 | * Account skipped pages as a partial scan as the pgdat may be |
1593 | * close to unreclaimable. If the LRU list is empty, account |
1594 | * skipped pages as a full scan. |
1595 | */ |
1596 | scan += list_empty(src) ? total_skipped : total_skipped >> 2; |
1597 | |
1598 | list_splice(&pages_skipped, src); |
1599 | } |
1600 | *nr_scanned = scan; |
1601 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan, |
1602 | nr_taken, mode, is_file_lru(lru)); |
1603 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1604 | return nr_taken; |
1605 | } |
1606 | |
1607 | /** |
1608 | * isolate_lru_page - tries to isolate a page from its LRU list |
1609 | * @page: page to isolate from its LRU list |
1610 | * |
1611 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the |
1612 | * vmstat statistic corresponding to whatever LRU list the page was on. |
1613 | * |
1614 | * Returns 0 if the page was removed from an LRU list. |
1615 | * Returns -EBUSY if the page was not on an LRU list. |
1616 | * |
1617 | * The returned page will have PageLRU() cleared. If it was found on |
1618 | * the active list, it will have PageActive set. If it was found on |
1619 | * the unevictable list, it will have the PageUnevictable bit set. That flag |
1620 | * may need to be cleared by the caller before letting the page go. |
1621 | * |
1622 | * The vmstat statistic corresponding to the list on which the page was |
1623 | * found will be decremented. |
1624 | * |
1625 | * Restrictions: |
1626 | * (1) Must be called with an elevated refcount on the page. This is a |
1627 | * fundamentnal difference from isolate_lru_pages (which is called |
1628 | * without a stable reference). |
1629 | * (2) the lru_lock must not be held. |
1630 | * (3) interrupts must be enabled. |
1631 | */ |
1632 | int isolate_lru_page(struct page *page) |
1633 | { |
1634 | int ret = -EBUSY; |
1635 | |
1636 | VM_BUG_ON_PAGE(!page_count(page), page); |
1637 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
1638 | |
1639 | if (PageLRU(page)) { |
1640 | struct zone *zone = page_zone(page); |
1641 | struct lruvec *lruvec; |
1642 | |
1643 | spin_lock_irq(zone_lru_lock(zone)); |
1644 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
1645 | if (PageLRU(page)) { |
1646 | int lru = page_lru(page); |
1647 | get_page(page); |
1648 | ClearPageLRU(page); |
1649 | del_page_from_lru_list(page, lruvec, lru); |
1650 | ret = 0; |
1651 | } |
1652 | spin_unlock_irq(zone_lru_lock(zone)); |
1653 | } |
1654 | return ret; |
1655 | } |
1656 | |
1657 | /* |
1658 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1659 | * then get resheduled. When there are massive number of tasks doing page |
1660 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, |
1661 | * the LRU list will go small and be scanned faster than necessary, leading to |
1662 | * unnecessary swapping, thrashing and OOM. |
1663 | */ |
1664 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
1665 | struct scan_control *sc) |
1666 | { |
1667 | #ifdef CONFIG_AMLOGIC_CMA |
1668 | signed long inactive, isolated; |
1669 | #else |
1670 | unsigned long inactive, isolated; |
1671 | #endif /* CONFIG_AMLOGIC_CMA */ |
1672 | |
1673 | if (current_is_kswapd()) |
1674 | return 0; |
1675 | |
1676 | if (!sane_reclaim(sc)) |
1677 | return 0; |
1678 | |
1679 | if (file) { |
1680 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1681 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); |
1682 | } else { |
1683 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1684 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); |
1685 | } |
1686 | |
1687 | #ifdef CONFIG_AMLOGIC_CMA |
1688 | isolated -= node_page_state(pgdat, NR_CMA_ISOLATED); |
1689 | #endif /* CONFIG_AMLOGIC_CMA */ |
1690 | /* |
1691 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they |
1692 | * won't get blocked by normal direct-reclaimers, forming a circular |
1693 | * deadlock. |
1694 | */ |
1695 | #ifndef CONFIG_AMLOGIC_MODIFY |
1696 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
1697 | inactive >>= 3; |
1698 | #endif |
1699 | #ifdef CONFIG_AMLOGIC_CMA |
1700 | WARN_ONCE(isolated > inactive, |
1701 | "isolated:%ld, cma:%ld, inactive:%ld, mask:%x, file:%d\n", |
1702 | isolated, node_page_state(pgdat, NR_CMA_ISOLATED), |
1703 | inactive, sc->gfp_mask, file); |
1704 | #endif /* CONFIG_AMLOGIC_CMA */ |
1705 | return isolated > inactive; |
1706 | } |
1707 | |
1708 | static noinline_for_stack void |
1709 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
1710 | { |
1711 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1712 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1713 | LIST_HEAD(pages_to_free); |
1714 | |
1715 | /* |
1716 | * Put back any unfreeable pages. |
1717 | */ |
1718 | while (!list_empty(page_list)) { |
1719 | struct page *page = lru_to_page(page_list); |
1720 | int lru; |
1721 | |
1722 | VM_BUG_ON_PAGE(PageLRU(page), page); |
1723 | list_del(&page->lru); |
1724 | if (unlikely(!page_evictable(page))) { |
1725 | spin_unlock_irq(&pgdat->lru_lock); |
1726 | putback_lru_page(page); |
1727 | spin_lock_irq(&pgdat->lru_lock); |
1728 | continue; |
1729 | } |
1730 | |
1731 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
1732 | |
1733 | SetPageLRU(page); |
1734 | lru = page_lru(page); |
1735 | add_page_to_lru_list(page, lruvec, lru); |
1736 | |
1737 | if (is_active_lru(lru)) { |
1738 | int file = is_file_lru(lru); |
1739 | int numpages = hpage_nr_pages(page); |
1740 | reclaim_stat->recent_rotated[file] += numpages; |
1741 | } |
1742 | if (put_page_testzero(page)) { |
1743 | __ClearPageLRU(page); |
1744 | __ClearPageActive(page); |
1745 | del_page_from_lru_list(page, lruvec, lru); |
1746 | |
1747 | if (unlikely(PageCompound(page))) { |
1748 | spin_unlock_irq(&pgdat->lru_lock); |
1749 | mem_cgroup_uncharge(page); |
1750 | (*get_compound_page_dtor(page))(page); |
1751 | spin_lock_irq(&pgdat->lru_lock); |
1752 | } else |
1753 | list_add(&page->lru, &pages_to_free); |
1754 | } |
1755 | } |
1756 | |
1757 | /* |
1758 | * To save our caller's stack, now use input list for pages to free. |
1759 | */ |
1760 | list_splice(&pages_to_free, page_list); |
1761 | } |
1762 | |
1763 | /* |
1764 | * If a kernel thread (such as nfsd for loop-back mounts) services |
1765 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. |
1766 | * In that case we should only throttle if the backing device it is |
1767 | * writing to is congested. In other cases it is safe to throttle. |
1768 | */ |
1769 | static int current_may_throttle(void) |
1770 | { |
1771 | return !(current->flags & PF_LESS_THROTTLE) || |
1772 | current->backing_dev_info == NULL || |
1773 | bdi_write_congested(current->backing_dev_info); |
1774 | } |
1775 | |
1776 | static bool inactive_reclaimable_pages(struct lruvec *lruvec, |
1777 | struct scan_control *sc, enum lru_list lru) |
1778 | { |
1779 | int zid; |
1780 | struct zone *zone; |
1781 | int file = is_file_lru(lru); |
1782 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1783 | |
1784 | if (!global_reclaim(sc)) |
1785 | return true; |
1786 | |
1787 | for (zid = sc->reclaim_idx; zid >= 0; zid--) { |
1788 | zone = &pgdat->node_zones[zid]; |
1789 | if (!managed_zone(zone)) |
1790 | continue; |
1791 | |
1792 | if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE + |
1793 | LRU_FILE * file) >= SWAP_CLUSTER_MAX) |
1794 | return true; |
1795 | } |
1796 | |
1797 | return false; |
1798 | } |
1799 | |
1800 | /* |
1801 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1802 | * of reclaimed pages |
1803 | */ |
1804 | static noinline_for_stack unsigned long |
1805 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
1806 | struct scan_control *sc, enum lru_list lru) |
1807 | { |
1808 | LIST_HEAD(page_list); |
1809 | unsigned long nr_scanned; |
1810 | unsigned long nr_reclaimed = 0; |
1811 | unsigned long nr_taken; |
1812 | unsigned long nr_dirty = 0; |
1813 | unsigned long nr_congested = 0; |
1814 | unsigned long nr_unqueued_dirty = 0; |
1815 | unsigned long nr_writeback = 0; |
1816 | unsigned long nr_immediate = 0; |
1817 | isolate_mode_t isolate_mode = 0; |
1818 | int file = is_file_lru(lru); |
1819 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1820 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1821 | |
1822 | if (!inactive_reclaimable_pages(lruvec, sc, lru)) |
1823 | return 0; |
1824 | |
1825 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
1826 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
1827 | |
1828 | /* We are about to die and free our memory. Return now. */ |
1829 | if (fatal_signal_pending(current)) |
1830 | return SWAP_CLUSTER_MAX; |
1831 | } |
1832 | |
1833 | lru_add_drain(); |
1834 | |
1835 | if (!sc->may_unmap) |
1836 | isolate_mode |= ISOLATE_UNMAPPED; |
1837 | if (!sc->may_writepage) |
1838 | isolate_mode |= ISOLATE_CLEAN; |
1839 | |
1840 | spin_lock_irq(&pgdat->lru_lock); |
1841 | |
1842 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1843 | &nr_scanned, sc, isolate_mode, lru); |
1844 | |
1845 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
1846 | reclaim_stat->recent_scanned[file] += nr_taken; |
1847 | |
1848 | if (global_reclaim(sc)) { |
1849 | __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned); |
1850 | if (current_is_kswapd()) |
1851 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
1852 | else |
1853 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
1854 | } |
1855 | spin_unlock_irq(&pgdat->lru_lock); |
1856 | |
1857 | if (nr_taken == 0) |
1858 | return 0; |
1859 | |
1860 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP, |
1861 | &nr_dirty, &nr_unqueued_dirty, &nr_congested, |
1862 | &nr_writeback, &nr_immediate, |
1863 | false); |
1864 | |
1865 | spin_lock_irq(&pgdat->lru_lock); |
1866 | |
1867 | if (global_reclaim(sc)) { |
1868 | if (current_is_kswapd()) |
1869 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
1870 | else |
1871 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
1872 | } |
1873 | |
1874 | putback_inactive_pages(lruvec, &page_list); |
1875 | |
1876 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
1877 | |
1878 | spin_unlock_irq(&pgdat->lru_lock); |
1879 | |
1880 | mem_cgroup_uncharge_list(&page_list); |
1881 | free_hot_cold_page_list(&page_list, true); |
1882 | |
1883 | /* |
1884 | * If reclaim is isolating dirty pages under writeback, it implies |
1885 | * that the long-lived page allocation rate is exceeding the page |
1886 | * laundering rate. Either the global limits are not being effective |
1887 | * at throttling processes due to the page distribution throughout |
1888 | * zones or there is heavy usage of a slow backing device. The |
1889 | * only option is to throttle from reclaim context which is not ideal |
1890 | * as there is no guarantee the dirtying process is throttled in the |
1891 | * same way balance_dirty_pages() manages. |
1892 | * |
1893 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1894 | * of pages under pages flagged for immediate reclaim and stall if any |
1895 | * are encountered in the nr_immediate check below. |
1896 | */ |
1897 | if (nr_writeback && nr_writeback == nr_taken) |
1898 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); |
1899 | |
1900 | /* |
1901 | * Legacy memcg will stall in page writeback so avoid forcibly |
1902 | * stalling here. |
1903 | */ |
1904 | if (sane_reclaim(sc)) { |
1905 | /* |
1906 | * Tag a zone as congested if all the dirty pages scanned were |
1907 | * backed by a congested BDI and wait_iff_congested will stall. |
1908 | */ |
1909 | if (nr_dirty && nr_dirty == nr_congested) |
1910 | set_bit(PGDAT_CONGESTED, &pgdat->flags); |
1911 | |
1912 | /* |
1913 | * If dirty pages are scanned that are not queued for IO, it |
1914 | * implies that flushers are not keeping up. In this case, flag |
1915 | * the pgdat PGDAT_DIRTY and kswapd will start writing pages from |
1916 | * reclaim context. |
1917 | */ |
1918 | if (nr_unqueued_dirty == nr_taken) |
1919 | set_bit(PGDAT_DIRTY, &pgdat->flags); |
1920 | |
1921 | /* |
1922 | * If kswapd scans pages marked marked for immediate |
1923 | * reclaim and under writeback (nr_immediate), it implies |
1924 | * that pages are cycling through the LRU faster than |
1925 | * they are written so also forcibly stall. |
1926 | */ |
1927 | if (nr_immediate && current_may_throttle()) |
1928 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
1929 | } |
1930 | |
1931 | /* |
1932 | * Stall direct reclaim for IO completions if underlying BDIs or zone |
1933 | * is congested. Allow kswapd to continue until it starts encountering |
1934 | * unqueued dirty pages or cycling through the LRU too quickly. |
1935 | */ |
1936 | if (!sc->hibernation_mode && !current_is_kswapd() && |
1937 | current_may_throttle()) |
1938 | wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); |
1939 | |
1940 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
1941 | nr_scanned, nr_reclaimed, |
1942 | sc->priority, file); |
1943 | return nr_reclaimed; |
1944 | } |
1945 | |
1946 | /* |
1947 | * This moves pages from the active list to the inactive list. |
1948 | * |
1949 | * We move them the other way if the page is referenced by one or more |
1950 | * processes, from rmap. |
1951 | * |
1952 | * If the pages are mostly unmapped, the processing is fast and it is |
1953 | * appropriate to hold zone_lru_lock across the whole operation. But if |
1954 | * the pages are mapped, the processing is slow (page_referenced()) so we |
1955 | * should drop zone_lru_lock around each page. It's impossible to balance |
1956 | * this, so instead we remove the pages from the LRU while processing them. |
1957 | * It is safe to rely on PG_active against the non-LRU pages in here because |
1958 | * nobody will play with that bit on a non-LRU page. |
1959 | * |
1960 | * The downside is that we have to touch page->_refcount against each page. |
1961 | * But we had to alter page->flags anyway. |
1962 | */ |
1963 | |
1964 | static void move_active_pages_to_lru(struct lruvec *lruvec, |
1965 | struct list_head *list, |
1966 | struct list_head *pages_to_free, |
1967 | enum lru_list lru) |
1968 | { |
1969 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1970 | unsigned long pgmoved = 0; |
1971 | struct page *page; |
1972 | int nr_pages; |
1973 | #ifdef CONFIG_AMLOGIC_MODIFY |
1974 | int num = NR_INACTIVE_ANON_CMA - NR_INACTIVE_ANON; |
1975 | int migrate_type = 0; |
1976 | #endif /* CONFIG_AMLOGIC_MODIFY */ |
1977 | |
1978 | while (!list_empty(list)) { |
1979 | page = lru_to_page(list); |
1980 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
1981 | |
1982 | VM_BUG_ON_PAGE(PageLRU(page), page); |
1983 | SetPageLRU(page); |
1984 | |
1985 | nr_pages = hpage_nr_pages(page); |
1986 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
1987 | list_move(&page->lru, &lruvec->lists[lru]); |
1988 | pgmoved += nr_pages; |
1989 | #ifdef CONFIG_AMLOGIC_MODIFY |
1990 | migrate_type = get_pageblock_migratetype(page); |
1991 | if (is_migrate_cma(migrate_type) || |
1992 | is_migrate_isolate(migrate_type)) |
1993 | __mod_zone_page_state(page_zone(page), |
1994 | NR_LRU_BASE + lru + num, |
1995 | nr_pages); |
1996 | #endif /* CONFIG_AMLOGIC_MODIFY */ |
1997 | |
1998 | if (put_page_testzero(page)) { |
1999 | __ClearPageLRU(page); |
2000 | __ClearPageActive(page); |
2001 | del_page_from_lru_list(page, lruvec, lru); |
2002 | |
2003 | if (unlikely(PageCompound(page))) { |
2004 | spin_unlock_irq(&pgdat->lru_lock); |
2005 | mem_cgroup_uncharge(page); |
2006 | (*get_compound_page_dtor(page))(page); |
2007 | spin_lock_irq(&pgdat->lru_lock); |
2008 | } else |
2009 | list_add(&page->lru, pages_to_free); |
2010 | } |
2011 | } |
2012 | |
2013 | if (!is_active_lru(lru)) |
2014 | __count_vm_events(PGDEACTIVATE, pgmoved); |
2015 | } |
2016 | |
2017 | static void shrink_active_list(unsigned long nr_to_scan, |
2018 | struct lruvec *lruvec, |
2019 | struct scan_control *sc, |
2020 | enum lru_list lru) |
2021 | { |
2022 | unsigned long nr_taken; |
2023 | unsigned long nr_scanned; |
2024 | unsigned long vm_flags; |
2025 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
2026 | LIST_HEAD(l_active); |
2027 | LIST_HEAD(l_inactive); |
2028 | struct page *page; |
2029 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2030 | unsigned long nr_rotated = 0; |
2031 | isolate_mode_t isolate_mode = 0; |
2032 | int file = is_file_lru(lru); |
2033 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
2034 | |
2035 | lru_add_drain(); |
2036 | |
2037 | if (!sc->may_unmap) |
2038 | isolate_mode |= ISOLATE_UNMAPPED; |
2039 | if (!sc->may_writepage) |
2040 | isolate_mode |= ISOLATE_CLEAN; |
2041 | |
2042 | spin_lock_irq(&pgdat->lru_lock); |
2043 | |
2044 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
2045 | &nr_scanned, sc, isolate_mode, lru); |
2046 | |
2047 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
2048 | reclaim_stat->recent_scanned[file] += nr_taken; |
2049 | |
2050 | if (global_reclaim(sc)) |
2051 | __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned); |
2052 | __count_vm_events(PGREFILL, nr_scanned); |
2053 | |
2054 | spin_unlock_irq(&pgdat->lru_lock); |
2055 | |
2056 | while (!list_empty(&l_hold)) { |
2057 | cond_resched(); |
2058 | page = lru_to_page(&l_hold); |
2059 | list_del(&page->lru); |
2060 | |
2061 | if (unlikely(!page_evictable(page))) { |
2062 | putback_lru_page(page); |
2063 | continue; |
2064 | } |
2065 | |
2066 | if (unlikely(buffer_heads_over_limit)) { |
2067 | if (page_has_private(page) && trylock_page(page)) { |
2068 | if (page_has_private(page)) |
2069 | try_to_release_page(page, 0); |
2070 | unlock_page(page); |
2071 | } |
2072 | } |
2073 | |
2074 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
2075 | &vm_flags)) { |
2076 | nr_rotated += hpage_nr_pages(page); |
2077 | /* |
2078 | * Identify referenced, file-backed active pages and |
2079 | * give them one more trip around the active list. So |
2080 | * that executable code get better chances to stay in |
2081 | * memory under moderate memory pressure. Anon pages |
2082 | * are not likely to be evicted by use-once streaming |
2083 | * IO, plus JVM can create lots of anon VM_EXEC pages, |
2084 | * so we ignore them here. |
2085 | */ |
2086 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
2087 | list_add(&page->lru, &l_active); |
2088 | continue; |
2089 | } |
2090 | } |
2091 | |
2092 | ClearPageActive(page); /* we are de-activating */ |
2093 | SetPageWorkingset(page); |
2094 | list_add(&page->lru, &l_inactive); |
2095 | } |
2096 | |
2097 | /* |
2098 | * Move pages back to the lru list. |
2099 | */ |
2100 | spin_lock_irq(&pgdat->lru_lock); |
2101 | /* |
2102 | * Count referenced pages from currently used mappings as rotated, |
2103 | * even though only some of them are actually re-activated. This |
2104 | * helps balance scan pressure between file and anonymous pages in |
2105 | * get_scan_count. |
2106 | */ |
2107 | reclaim_stat->recent_rotated[file] += nr_rotated; |
2108 | |
2109 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
2110 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); |
2111 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
2112 | spin_unlock_irq(&pgdat->lru_lock); |
2113 | |
2114 | mem_cgroup_uncharge_list(&l_hold); |
2115 | free_hot_cold_page_list(&l_hold, true); |
2116 | } |
2117 | |
2118 | /* |
2119 | * The inactive anon list should be small enough that the VM never has |
2120 | * to do too much work. |
2121 | * |
2122 | * The inactive file list should be small enough to leave most memory |
2123 | * to the established workingset on the scan-resistant active list, |
2124 | * but large enough to avoid thrashing the aggregate readahead window. |
2125 | * |
2126 | * Both inactive lists should also be large enough that each inactive |
2127 | * page has a chance to be referenced again before it is reclaimed. |
2128 | * |
2129 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
2130 | * on this LRU, maintained by the pageout code. A zone->inactive_ratio |
2131 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
2132 | * |
2133 | * total target max |
2134 | * memory ratio inactive |
2135 | * ------------------------------------- |
2136 | * 10MB 1 5MB |
2137 | * 100MB 1 50MB |
2138 | * 1GB 3 250MB |
2139 | * 10GB 10 0.9GB |
2140 | * 100GB 31 3GB |
2141 | * 1TB 101 10GB |
2142 | * 10TB 320 32GB |
2143 | */ |
2144 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
2145 | struct scan_control *sc) |
2146 | { |
2147 | unsigned long inactive_ratio; |
2148 | unsigned long inactive, active; |
2149 | enum lru_list inactive_lru = file * LRU_FILE; |
2150 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; |
2151 | unsigned long gb; |
2152 | |
2153 | /* |
2154 | * If we don't have swap space, anonymous page deactivation |
2155 | * is pointless. |
2156 | */ |
2157 | if (!file && !total_swap_pages) |
2158 | return false; |
2159 | |
2160 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
2161 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); |
2162 | |
2163 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
2164 | if (gb) |
2165 | inactive_ratio = int_sqrt(10 * gb); |
2166 | #ifdef CONFIG_AMLOGIC_MEMORY_EXTEND |
2167 | else if (!file && (totalram_pages >> (20 - PAGE_SHIFT)) >= 512) |
2168 | inactive_ratio = 2; |
2169 | #endif /* CONFIG_AMLOGIC_MEMORY_EXTEND */ |
2170 | else |
2171 | inactive_ratio = 1; |
2172 | |
2173 | return inactive * inactive_ratio < active; |
2174 | } |
2175 | |
2176 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
2177 | struct lruvec *lruvec, struct scan_control *sc) |
2178 | { |
2179 | if (is_active_lru(lru)) { |
2180 | if (inactive_list_is_low(lruvec, is_file_lru(lru), sc)) |
2181 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
2182 | return 0; |
2183 | } |
2184 | |
2185 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
2186 | } |
2187 | |
2188 | enum scan_balance { |
2189 | SCAN_EQUAL, |
2190 | SCAN_FRACT, |
2191 | SCAN_ANON, |
2192 | SCAN_FILE, |
2193 | }; |
2194 | |
2195 | /* |
2196 | * Determine how aggressively the anon and file LRU lists should be |
2197 | * scanned. The relative value of each set of LRU lists is determined |
2198 | * by looking at the fraction of the pages scanned we did rotate back |
2199 | * onto the active list instead of evict. |
2200 | * |
2201 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2202 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan |
2203 | */ |
2204 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
2205 | struct scan_control *sc, unsigned long *nr, |
2206 | unsigned long *lru_pages) |
2207 | { |
2208 | int swappiness = mem_cgroup_swappiness(memcg); |
2209 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2210 | u64 fraction[2]; |
2211 | u64 denominator = 0; /* gcc */ |
2212 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
2213 | unsigned long anon_prio, file_prio; |
2214 | enum scan_balance scan_balance; |
2215 | unsigned long anon, file; |
2216 | bool force_scan = false; |
2217 | unsigned long ap, fp; |
2218 | enum lru_list lru; |
2219 | bool some_scanned; |
2220 | int pass; |
2221 | |
2222 | /* |
2223 | * If the zone or memcg is small, nr[l] can be 0. This |
2224 | * results in no scanning on this priority and a potential |
2225 | * priority drop. Global direct reclaim can go to the next |
2226 | * zone and tends to have no problems. Global kswapd is for |
2227 | * zone balancing and it needs to scan a minimum amount. When |
2228 | * reclaiming for a memcg, a priority drop can cause high |
2229 | * latencies, so it's better to scan a minimum amount there as |
2230 | * well. |
2231 | */ |
2232 | if (current_is_kswapd()) { |
2233 | if (!pgdat_reclaimable(pgdat)) |
2234 | force_scan = true; |
2235 | if (!mem_cgroup_online(memcg)) |
2236 | force_scan = true; |
2237 | } |
2238 | if (!global_reclaim(sc)) |
2239 | force_scan = true; |
2240 | |
2241 | /* If we have no swap space, do not bother scanning anon pages. */ |
2242 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
2243 | scan_balance = SCAN_FILE; |
2244 | goto out; |
2245 | } |
2246 | |
2247 | /* |
2248 | * Global reclaim will swap to prevent OOM even with no |
2249 | * swappiness, but memcg users want to use this knob to |
2250 | * disable swapping for individual groups completely when |
2251 | * using the memory controller's swap limit feature would be |
2252 | * too expensive. |
2253 | */ |
2254 | if (!global_reclaim(sc) && !swappiness) { |
2255 | scan_balance = SCAN_FILE; |
2256 | goto out; |
2257 | } |
2258 | |
2259 | /* |
2260 | * Do not apply any pressure balancing cleverness when the |
2261 | * system is close to OOM, scan both anon and file equally |
2262 | * (unless the swappiness setting disagrees with swapping). |
2263 | */ |
2264 | if (!sc->priority && swappiness) { |
2265 | scan_balance = SCAN_EQUAL; |
2266 | goto out; |
2267 | } |
2268 | |
2269 | /* |
2270 | * Prevent the reclaimer from falling into the cache trap: as |
2271 | * cache pages start out inactive, every cache fault will tip |
2272 | * the scan balance towards the file LRU. And as the file LRU |
2273 | * shrinks, so does the window for rotation from references. |
2274 | * This means we have a runaway feedback loop where a tiny |
2275 | * thrashing file LRU becomes infinitely more attractive than |
2276 | * anon pages. Try to detect this based on file LRU size. |
2277 | */ |
2278 | if (global_reclaim(sc)) { |
2279 | unsigned long pgdatfile; |
2280 | unsigned long pgdatfree; |
2281 | int z; |
2282 | unsigned long total_high_wmark = 0; |
2283 | |
2284 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2285 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + |
2286 | node_page_state(pgdat, NR_INACTIVE_FILE); |
2287 | |
2288 | for (z = 0; z < MAX_NR_ZONES; z++) { |
2289 | struct zone *zone = &pgdat->node_zones[z]; |
2290 | if (!managed_zone(zone)) |
2291 | continue; |
2292 | |
2293 | total_high_wmark += high_wmark_pages(zone); |
2294 | } |
2295 | |
2296 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
2297 | scan_balance = SCAN_ANON; |
2298 | goto out; |
2299 | } |
2300 | } |
2301 | |
2302 | /* |
2303 | * If there is enough inactive page cache, i.e. if the size of the |
2304 | * inactive list is greater than that of the active list *and* the |
2305 | * inactive list actually has some pages to scan on this priority, we |
2306 | * do not reclaim anything from the anonymous working set right now. |
2307 | * Without the second condition we could end up never scanning an |
2308 | * lruvec even if it has plenty of old anonymous pages unless the |
2309 | * system is under heavy pressure. |
2310 | */ |
2311 | if (!inactive_list_is_low(lruvec, true, sc) && |
2312 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
2313 | scan_balance = SCAN_FILE; |
2314 | goto out; |
2315 | } |
2316 | |
2317 | scan_balance = SCAN_FRACT; |
2318 | |
2319 | /* |
2320 | * With swappiness at 100, anonymous and file have the same priority. |
2321 | * This scanning priority is essentially the inverse of IO cost. |
2322 | */ |
2323 | anon_prio = swappiness; |
2324 | #ifdef CONFIG_AMLOGIC_MEMORY_EXTEND |
2325 | if (get_nr_swap_pages() * 3 < total_swap_pages) |
2326 | anon_prio >>= 1; |
2327 | #endif /* CONFIG_AMLOGIC_MEMORY_EXTEND */ |
2328 | file_prio = 200 - anon_prio; |
2329 | |
2330 | /* |
2331 | * OK, so we have swap space and a fair amount of page cache |
2332 | * pages. We use the recently rotated / recently scanned |
2333 | * ratios to determine how valuable each cache is. |
2334 | * |
2335 | * Because workloads change over time (and to avoid overflow) |
2336 | * we keep these statistics as a floating average, which ends |
2337 | * up weighing recent references more than old ones. |
2338 | * |
2339 | * anon in [0], file in [1] |
2340 | */ |
2341 | |
2342 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
2343 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); |
2344 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + |
2345 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); |
2346 | |
2347 | spin_lock_irq(&pgdat->lru_lock); |
2348 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
2349 | reclaim_stat->recent_scanned[0] /= 2; |
2350 | reclaim_stat->recent_rotated[0] /= 2; |
2351 | } |
2352 | |
2353 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
2354 | reclaim_stat->recent_scanned[1] /= 2; |
2355 | reclaim_stat->recent_rotated[1] /= 2; |
2356 | } |
2357 | |
2358 | /* |
2359 | * The amount of pressure on anon vs file pages is inversely |
2360 | * proportional to the fraction of recently scanned pages on |
2361 | * each list that were recently referenced and in active use. |
2362 | */ |
2363 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
2364 | ap /= reclaim_stat->recent_rotated[0] + 1; |
2365 | |
2366 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
2367 | fp /= reclaim_stat->recent_rotated[1] + 1; |
2368 | spin_unlock_irq(&pgdat->lru_lock); |
2369 | |
2370 | fraction[0] = ap; |
2371 | fraction[1] = fp; |
2372 | denominator = ap + fp + 1; |
2373 | out: |
2374 | some_scanned = false; |
2375 | /* Only use force_scan on second pass. */ |
2376 | for (pass = 0; !some_scanned && pass < 2; pass++) { |
2377 | *lru_pages = 0; |
2378 | for_each_evictable_lru(lru) { |
2379 | int file = is_file_lru(lru); |
2380 | unsigned long size; |
2381 | unsigned long scan; |
2382 | |
2383 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); |
2384 | scan = size >> sc->priority; |
2385 | |
2386 | if (!scan && pass && force_scan) |
2387 | scan = min(size, SWAP_CLUSTER_MAX); |
2388 | |
2389 | switch (scan_balance) { |
2390 | case SCAN_EQUAL: |
2391 | /* Scan lists relative to size */ |
2392 | break; |
2393 | case SCAN_FRACT: |
2394 | /* |
2395 | * Scan types proportional to swappiness and |
2396 | * their relative recent reclaim efficiency. |
2397 | */ |
2398 | scan = div64_u64(scan * fraction[file], |
2399 | denominator); |
2400 | break; |
2401 | case SCAN_FILE: |
2402 | case SCAN_ANON: |
2403 | /* Scan one type exclusively */ |
2404 | if ((scan_balance == SCAN_FILE) != file) { |
2405 | size = 0; |
2406 | scan = 0; |
2407 | } |
2408 | break; |
2409 | default: |
2410 | /* Look ma, no brain */ |
2411 | BUG(); |
2412 | } |
2413 | |
2414 | *lru_pages += size; |
2415 | nr[lru] = scan; |
2416 | |
2417 | /* |
2418 | * Skip the second pass and don't force_scan, |
2419 | * if we found something to scan. |
2420 | */ |
2421 | some_scanned |= !!scan; |
2422 | } |
2423 | } |
2424 | } |
2425 | |
2426 | /* |
2427 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. |
2428 | */ |
2429 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, |
2430 | struct scan_control *sc, unsigned long *lru_pages) |
2431 | { |
2432 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
2433 | unsigned long nr[NR_LRU_LISTS]; |
2434 | unsigned long targets[NR_LRU_LISTS]; |
2435 | unsigned long nr_to_scan; |
2436 | enum lru_list lru; |
2437 | unsigned long nr_reclaimed = 0; |
2438 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; |
2439 | struct blk_plug plug; |
2440 | bool scan_adjusted; |
2441 | |
2442 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
2443 | |
2444 | /* Record the original scan target for proportional adjustments later */ |
2445 | memcpy(targets, nr, sizeof(nr)); |
2446 | |
2447 | /* |
2448 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal |
2449 | * event that can occur when there is little memory pressure e.g. |
2450 | * multiple streaming readers/writers. Hence, we do not abort scanning |
2451 | * when the requested number of pages are reclaimed when scanning at |
2452 | * DEF_PRIORITY on the assumption that the fact we are direct |
2453 | * reclaiming implies that kswapd is not keeping up and it is best to |
2454 | * do a batch of work at once. For memcg reclaim one check is made to |
2455 | * abort proportional reclaim if either the file or anon lru has already |
2456 | * dropped to zero at the first pass. |
2457 | */ |
2458 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && |
2459 | sc->priority == DEF_PRIORITY); |
2460 | |
2461 | blk_start_plug(&plug); |
2462 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || |
2463 | nr[LRU_INACTIVE_FILE]) { |
2464 | unsigned long nr_anon, nr_file, percentage; |
2465 | unsigned long nr_scanned; |
2466 | |
2467 | for_each_evictable_lru(lru) { |
2468 | if (nr[lru]) { |
2469 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); |
2470 | nr[lru] -= nr_to_scan; |
2471 | |
2472 | nr_reclaimed += shrink_list(lru, nr_to_scan, |
2473 | lruvec, sc); |
2474 | } |
2475 | } |
2476 | |
2477 | cond_resched(); |
2478 | |
2479 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2480 | continue; |
2481 | |
2482 | /* |
2483 | * For kswapd and memcg, reclaim at least the number of pages |
2484 | * requested. Ensure that the anon and file LRUs are scanned |
2485 | * proportionally what was requested by get_scan_count(). We |
2486 | * stop reclaiming one LRU and reduce the amount scanning |
2487 | * proportional to the original scan target. |
2488 | */ |
2489 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; |
2490 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; |
2491 | |
2492 | /* |
2493 | * It's just vindictive to attack the larger once the smaller |
2494 | * has gone to zero. And given the way we stop scanning the |
2495 | * smaller below, this makes sure that we only make one nudge |
2496 | * towards proportionality once we've got nr_to_reclaim. |
2497 | */ |
2498 | if (!nr_file || !nr_anon) |
2499 | break; |
2500 | |
2501 | if (nr_file > nr_anon) { |
2502 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + |
2503 | targets[LRU_ACTIVE_ANON] + 1; |
2504 | lru = LRU_BASE; |
2505 | percentage = nr_anon * 100 / scan_target; |
2506 | } else { |
2507 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + |
2508 | targets[LRU_ACTIVE_FILE] + 1; |
2509 | lru = LRU_FILE; |
2510 | percentage = nr_file * 100 / scan_target; |
2511 | } |
2512 | |
2513 | /* Stop scanning the smaller of the LRU */ |
2514 | nr[lru] = 0; |
2515 | nr[lru + LRU_ACTIVE] = 0; |
2516 | |
2517 | /* |
2518 | * Recalculate the other LRU scan count based on its original |
2519 | * scan target and the percentage scanning already complete |
2520 | */ |
2521 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; |
2522 | nr_scanned = targets[lru] - nr[lru]; |
2523 | nr[lru] = targets[lru] * (100 - percentage) / 100; |
2524 | nr[lru] -= min(nr[lru], nr_scanned); |
2525 | |
2526 | lru += LRU_ACTIVE; |
2527 | nr_scanned = targets[lru] - nr[lru]; |
2528 | nr[lru] = targets[lru] * (100 - percentage) / 100; |
2529 | nr[lru] -= min(nr[lru], nr_scanned); |
2530 | |
2531 | scan_adjusted = true; |
2532 | } |
2533 | blk_finish_plug(&plug); |
2534 | sc->nr_reclaimed += nr_reclaimed; |
2535 | |
2536 | /* |
2537 | * Even if we did not try to evict anon pages at all, we want to |
2538 | * rebalance the anon lru active/inactive ratio. |
2539 | */ |
2540 | if (inactive_list_is_low(lruvec, false, sc)) |
2541 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2542 | sc, LRU_ACTIVE_ANON); |
2543 | } |
2544 | |
2545 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
2546 | static bool in_reclaim_compaction(struct scan_control *sc) |
2547 | { |
2548 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
2549 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
2550 | sc->priority < DEF_PRIORITY - 2)) |
2551 | return true; |
2552 | |
2553 | return false; |
2554 | } |
2555 | |
2556 | /* |
2557 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2558 | * order-0 pages before compacting the zone. should_continue_reclaim() returns |
2559 | * true if more pages should be reclaimed such that when the page allocator |
2560 | * calls try_to_compact_zone() that it will have enough free pages to succeed. |
2561 | * It will give up earlier than that if there is difficulty reclaiming pages. |
2562 | */ |
2563 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
2564 | unsigned long nr_reclaimed, |
2565 | unsigned long nr_scanned, |
2566 | struct scan_control *sc) |
2567 | { |
2568 | unsigned long pages_for_compaction; |
2569 | unsigned long inactive_lru_pages; |
2570 | int z; |
2571 | |
2572 | /* If not in reclaim/compaction mode, stop */ |
2573 | if (!in_reclaim_compaction(sc)) |
2574 | return false; |
2575 | |
2576 | /* Consider stopping depending on scan and reclaim activity */ |
2577 | if (sc->gfp_mask & __GFP_REPEAT) { |
2578 | /* |
2579 | * For __GFP_REPEAT allocations, stop reclaiming if the |
2580 | * full LRU list has been scanned and we are still failing |
2581 | * to reclaim pages. This full LRU scan is potentially |
2582 | * expensive but a __GFP_REPEAT caller really wants to succeed |
2583 | */ |
2584 | if (!nr_reclaimed && !nr_scanned) |
2585 | return false; |
2586 | } else { |
2587 | /* |
2588 | * For non-__GFP_REPEAT allocations which can presumably |
2589 | * fail without consequence, stop if we failed to reclaim |
2590 | * any pages from the last SWAP_CLUSTER_MAX number of |
2591 | * pages that were scanned. This will return to the |
2592 | * caller faster at the risk reclaim/compaction and |
2593 | * the resulting allocation attempt fails |
2594 | */ |
2595 | if (!nr_reclaimed) |
2596 | return false; |
2597 | } |
2598 | |
2599 | /* |
2600 | * If we have not reclaimed enough pages for compaction and the |
2601 | * inactive lists are large enough, continue reclaiming |
2602 | */ |
2603 | pages_for_compaction = compact_gap(sc->order); |
2604 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
2605 | if (get_nr_swap_pages() > 0) |
2606 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
2607 | if (sc->nr_reclaimed < pages_for_compaction && |
2608 | inactive_lru_pages > pages_for_compaction) |
2609 | return true; |
2610 | |
2611 | /* If compaction would go ahead or the allocation would succeed, stop */ |
2612 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2613 | struct zone *zone = &pgdat->node_zones[z]; |
2614 | if (!managed_zone(zone)) |
2615 | continue; |
2616 | |
2617 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { |
2618 | case COMPACT_SUCCESS: |
2619 | case COMPACT_CONTINUE: |
2620 | return false; |
2621 | default: |
2622 | /* check next zone */ |
2623 | ; |
2624 | } |
2625 | } |
2626 | return true; |
2627 | } |
2628 | |
2629 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
2630 | { |
2631 | struct reclaim_state *reclaim_state = current->reclaim_state; |
2632 | unsigned long nr_reclaimed, nr_scanned; |
2633 | bool reclaimable = false; |
2634 | |
2635 | do { |
2636 | struct mem_cgroup *root = sc->target_mem_cgroup; |
2637 | struct mem_cgroup_reclaim_cookie reclaim = { |
2638 | .pgdat = pgdat, |
2639 | .priority = sc->priority, |
2640 | }; |
2641 | unsigned long node_lru_pages = 0; |
2642 | struct mem_cgroup *memcg; |
2643 | |
2644 | nr_reclaimed = sc->nr_reclaimed; |
2645 | nr_scanned = sc->nr_scanned; |
2646 | |
2647 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2648 | do { |
2649 | unsigned long lru_pages; |
2650 | unsigned long reclaimed; |
2651 | unsigned long scanned; |
2652 | |
2653 | if (mem_cgroup_low(root, memcg)) { |
2654 | if (!sc->may_thrash) |
2655 | continue; |
2656 | mem_cgroup_events(memcg, MEMCG_LOW, 1); |
2657 | } |
2658 | |
2659 | reclaimed = sc->nr_reclaimed; |
2660 | scanned = sc->nr_scanned; |
2661 | |
2662 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
2663 | node_lru_pages += lru_pages; |
2664 | |
2665 | if (memcg) |
2666 | shrink_slab(sc->gfp_mask, pgdat->node_id, |
2667 | memcg, sc->nr_scanned - scanned, |
2668 | lru_pages); |
2669 | |
2670 | /* Record the group's reclaim efficiency */ |
2671 | vmpressure(sc->gfp_mask, memcg, false, |
2672 | sc->nr_scanned - scanned, |
2673 | sc->nr_reclaimed - reclaimed); |
2674 | |
2675 | /* |
2676 | * Direct reclaim and kswapd have to scan all memory |
2677 | * cgroups to fulfill the overall scan target for the |
2678 | * node. |
2679 | * |
2680 | * Limit reclaim, on the other hand, only cares about |
2681 | * nr_to_reclaim pages to be reclaimed and it will |
2682 | * retry with decreasing priority if one round over the |
2683 | * whole hierarchy is not sufficient. |
2684 | */ |
2685 | if (!global_reclaim(sc) && |
2686 | sc->nr_reclaimed >= sc->nr_to_reclaim) { |
2687 | mem_cgroup_iter_break(root, memcg); |
2688 | break; |
2689 | } |
2690 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
2691 | |
2692 | /* |
2693 | * Shrink the slab caches in the same proportion that |
2694 | * the eligible LRU pages were scanned. |
2695 | */ |
2696 | if (global_reclaim(sc)) |
2697 | shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, |
2698 | sc->nr_scanned - nr_scanned, |
2699 | node_lru_pages); |
2700 | |
2701 | if (reclaim_state) { |
2702 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
2703 | reclaim_state->reclaimed_slab = 0; |
2704 | } |
2705 | |
2706 | /* Record the subtree's reclaim efficiency */ |
2707 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, |
2708 | sc->nr_scanned - nr_scanned, |
2709 | sc->nr_reclaimed - nr_reclaimed); |
2710 | |
2711 | if (sc->nr_reclaimed - nr_reclaimed) |
2712 | reclaimable = true; |
2713 | |
2714 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
2715 | sc->nr_scanned - nr_scanned, sc)); |
2716 | |
2717 | /* |
2718 | * Kswapd gives up on balancing particular nodes after too |
2719 | * many failures to reclaim anything from them and goes to |
2720 | * sleep. On reclaim progress, reset the failure counter. A |
2721 | * successful direct reclaim run will revive a dormant kswapd. |
2722 | */ |
2723 | if (reclaimable) |
2724 | pgdat->kswapd_failures = 0; |
2725 | |
2726 | return reclaimable; |
2727 | } |
2728 | |
2729 | /* |
2730 | * Returns true if compaction should go ahead for a costly-order request, or |
2731 | * the allocation would already succeed without compaction. Return false if we |
2732 | * should reclaim first. |
2733 | */ |
2734 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
2735 | { |
2736 | unsigned long watermark; |
2737 | enum compact_result suitable; |
2738 | |
2739 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2740 | if (suitable == COMPACT_SUCCESS) |
2741 | /* Allocation should succeed already. Don't reclaim. */ |
2742 | return true; |
2743 | if (suitable == COMPACT_SKIPPED) |
2744 | /* Compaction cannot yet proceed. Do reclaim. */ |
2745 | return false; |
2746 | |
2747 | /* |
2748 | * Compaction is already possible, but it takes time to run and there |
2749 | * are potentially other callers using the pages just freed. So proceed |
2750 | * with reclaim to make a buffer of free pages available to give |
2751 | * compaction a reasonable chance of completing and allocating the page. |
2752 | * Note that we won't actually reclaim the whole buffer in one attempt |
2753 | * as the target watermark in should_continue_reclaim() is lower. But if |
2754 | * we are already above the high+gap watermark, don't reclaim at all. |
2755 | */ |
2756 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
2757 | |
2758 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
2759 | } |
2760 | |
2761 | /* |
2762 | * This is the direct reclaim path, for page-allocating processes. We only |
2763 | * try to reclaim pages from zones which will satisfy the caller's allocation |
2764 | * request. |
2765 | * |
2766 | * If a zone is deemed to be full of pinned pages then just give it a light |
2767 | * scan then give up on it. |
2768 | */ |
2769 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
2770 | { |
2771 | struct zoneref *z; |
2772 | struct zone *zone; |
2773 | unsigned long nr_soft_reclaimed; |
2774 | unsigned long nr_soft_scanned; |
2775 | gfp_t orig_mask; |
2776 | pg_data_t *last_pgdat = NULL; |
2777 | |
2778 | /* |
2779 | * If the number of buffer_heads in the machine exceeds the maximum |
2780 | * allowed level, force direct reclaim to scan the highmem zone as |
2781 | * highmem pages could be pinning lowmem pages storing buffer_heads |
2782 | */ |
2783 | orig_mask = sc->gfp_mask; |
2784 | if (buffer_heads_over_limit) { |
2785 | sc->gfp_mask |= __GFP_HIGHMEM; |
2786 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
2787 | } |
2788 | |
2789 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
2790 | sc->reclaim_idx, sc->nodemask) { |
2791 | /* |
2792 | * Take care memory controller reclaiming has small influence |
2793 | * to global LRU. |
2794 | */ |
2795 | if (global_reclaim(sc)) { |
2796 | if (!cpuset_zone_allowed(zone, |
2797 | GFP_KERNEL | __GFP_HARDWALL)) |
2798 | continue; |
2799 | |
2800 | /* |
2801 | * If we already have plenty of memory free for |
2802 | * compaction in this zone, don't free any more. |
2803 | * Even though compaction is invoked for any |
2804 | * non-zero order, only frequent costly order |
2805 | * reclamation is disruptive enough to become a |
2806 | * noticeable problem, like transparent huge |
2807 | * page allocations. |
2808 | */ |
2809 | if (IS_ENABLED(CONFIG_COMPACTION) && |
2810 | sc->order > PAGE_ALLOC_COSTLY_ORDER && |
2811 | compaction_ready(zone, sc)) { |
2812 | sc->compaction_ready = true; |
2813 | continue; |
2814 | } |
2815 | |
2816 | /* |
2817 | * Shrink each node in the zonelist once. If the |
2818 | * zonelist is ordered by zone (not the default) then a |
2819 | * node may be shrunk multiple times but in that case |
2820 | * the user prefers lower zones being preserved. |
2821 | */ |
2822 | if (zone->zone_pgdat == last_pgdat) |
2823 | continue; |
2824 | |
2825 | /* |
2826 | * This steals pages from memory cgroups over softlimit |
2827 | * and returns the number of reclaimed pages and |
2828 | * scanned pages. This works for global memory pressure |
2829 | * and balancing, not for a memcg's limit. |
2830 | */ |
2831 | nr_soft_scanned = 0; |
2832 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
2833 | sc->order, sc->gfp_mask, |
2834 | &nr_soft_scanned); |
2835 | sc->nr_reclaimed += nr_soft_reclaimed; |
2836 | sc->nr_scanned += nr_soft_scanned; |
2837 | /* need some check for avoid more shrink_zone() */ |
2838 | } |
2839 | |
2840 | /* See comment about same check for global reclaim above */ |
2841 | if (zone->zone_pgdat == last_pgdat) |
2842 | continue; |
2843 | last_pgdat = zone->zone_pgdat; |
2844 | shrink_node(zone->zone_pgdat, sc); |
2845 | } |
2846 | |
2847 | /* |
2848 | * Restore to original mask to avoid the impact on the caller if we |
2849 | * promoted it to __GFP_HIGHMEM. |
2850 | */ |
2851 | sc->gfp_mask = orig_mask; |
2852 | } |
2853 | |
2854 | /* |
2855 | * This is the main entry point to direct page reclaim. |
2856 | * |
2857 | * If a full scan of the inactive list fails to free enough memory then we |
2858 | * are "out of memory" and something needs to be killed. |
2859 | * |
2860 | * If the caller is !__GFP_FS then the probability of a failure is reasonably |
2861 | * high - the zone may be full of dirty or under-writeback pages, which this |
2862 | * caller can't do much about. We kick the writeback threads and take explicit |
2863 | * naps in the hope that some of these pages can be written. But if the |
2864 | * allocating task holds filesystem locks which prevent writeout this might not |
2865 | * work, and the allocation attempt will fail. |
2866 | * |
2867 | * returns: 0, if no pages reclaimed |
2868 | * else, the number of pages reclaimed |
2869 | */ |
2870 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
2871 | struct scan_control *sc) |
2872 | { |
2873 | int initial_priority = sc->priority; |
2874 | unsigned long total_scanned = 0; |
2875 | unsigned long writeback_threshold; |
2876 | retry: |
2877 | delayacct_freepages_start(); |
2878 | |
2879 | if (global_reclaim(sc)) |
2880 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
2881 | |
2882 | do { |
2883 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2884 | sc->priority); |
2885 | sc->nr_scanned = 0; |
2886 | shrink_zones(zonelist, sc); |
2887 | |
2888 | total_scanned += sc->nr_scanned; |
2889 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
2890 | break; |
2891 | |
2892 | if (sc->compaction_ready) |
2893 | break; |
2894 | |
2895 | /* |
2896 | * If we're getting trouble reclaiming, start doing |
2897 | * writepage even in laptop mode. |
2898 | */ |
2899 | if (sc->priority < DEF_PRIORITY - 2) |
2900 | sc->may_writepage = 1; |
2901 | |
2902 | /* |
2903 | * Try to write back as many pages as we just scanned. This |
2904 | * tends to cause slow streaming writers to write data to the |
2905 | * disk smoothly, at the dirtying rate, which is nice. But |
2906 | * that's undesirable in laptop mode, where we *want* lumpy |
2907 | * writeout. So in laptop mode, write out the whole world. |
2908 | */ |
2909 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2910 | if (total_scanned > writeback_threshold) { |
2911 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2912 | WB_REASON_TRY_TO_FREE_PAGES); |
2913 | sc->may_writepage = 1; |
2914 | } |
2915 | } while (--sc->priority >= 0); |
2916 | |
2917 | delayacct_freepages_end(); |
2918 | |
2919 | if (sc->nr_reclaimed) |
2920 | return sc->nr_reclaimed; |
2921 | |
2922 | /* Aborted reclaim to try compaction? don't OOM, then */ |
2923 | if (sc->compaction_ready) |
2924 | return 1; |
2925 | |
2926 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
2927 | if (!sc->may_thrash) { |
2928 | sc->priority = initial_priority; |
2929 | sc->may_thrash = 1; |
2930 | goto retry; |
2931 | } |
2932 | |
2933 | return 0; |
2934 | } |
2935 | |
2936 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
2937 | { |
2938 | struct zone *zone; |
2939 | unsigned long pfmemalloc_reserve = 0; |
2940 | unsigned long free_pages = 0; |
2941 | int i; |
2942 | bool wmark_ok; |
2943 | |
2944 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
2945 | return true; |
2946 | |
2947 | for (i = 0; i <= ZONE_NORMAL; i++) { |
2948 | zone = &pgdat->node_zones[i]; |
2949 | if (!managed_zone(zone)) |
2950 | continue; |
2951 | |
2952 | if (!zone_reclaimable_pages(zone)) |
2953 | continue; |
2954 | |
2955 | pfmemalloc_reserve += min_wmark_pages(zone); |
2956 | free_pages += zone_page_state(zone, NR_FREE_PAGES); |
2957 | } |
2958 | |
2959 | /* If there are no reserves (unexpected config) then do not throttle */ |
2960 | if (!pfmemalloc_reserve) |
2961 | return true; |
2962 | |
2963 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
2964 | |
2965 | /* kswapd must be awake if processes are being throttled */ |
2966 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { |
2967 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
2968 | (enum zone_type)ZONE_NORMAL); |
2969 | wake_up_interruptible(&pgdat->kswapd_wait); |
2970 | } |
2971 | |
2972 | return wmark_ok; |
2973 | } |
2974 | |
2975 | /* |
2976 | * Throttle direct reclaimers if backing storage is backed by the network |
2977 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously |
2978 | * depleted. kswapd will continue to make progress and wake the processes |
2979 | * when the low watermark is reached. |
2980 | * |
2981 | * Returns true if a fatal signal was delivered during throttling. If this |
2982 | * happens, the page allocator should not consider triggering the OOM killer. |
2983 | */ |
2984 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
2985 | nodemask_t *nodemask) |
2986 | { |
2987 | struct zoneref *z; |
2988 | struct zone *zone; |
2989 | pg_data_t *pgdat = NULL; |
2990 | |
2991 | /* |
2992 | * Kernel threads should not be throttled as they may be indirectly |
2993 | * responsible for cleaning pages necessary for reclaim to make forward |
2994 | * progress. kjournald for example may enter direct reclaim while |
2995 | * committing a transaction where throttling it could forcing other |
2996 | * processes to block on log_wait_commit(). |
2997 | */ |
2998 | if (current->flags & PF_KTHREAD) |
2999 | goto out; |
3000 | |
3001 | /* |
3002 | * If a fatal signal is pending, this process should not throttle. |
3003 | * It should return quickly so it can exit and free its memory |
3004 | */ |
3005 | if (fatal_signal_pending(current)) |
3006 | goto out; |
3007 | |
3008 | /* |
3009 | * Check if the pfmemalloc reserves are ok by finding the first node |
3010 | * with a usable ZONE_NORMAL or lower zone. The expectation is that |
3011 | * GFP_KERNEL will be required for allocating network buffers when |
3012 | * swapping over the network so ZONE_HIGHMEM is unusable. |
3013 | * |
3014 | * Throttling is based on the first usable node and throttled processes |
3015 | * wait on a queue until kswapd makes progress and wakes them. There |
3016 | * is an affinity then between processes waking up and where reclaim |
3017 | * progress has been made assuming the process wakes on the same node. |
3018 | * More importantly, processes running on remote nodes will not compete |
3019 | * for remote pfmemalloc reserves and processes on different nodes |
3020 | * should make reasonable progress. |
3021 | */ |
3022 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
3023 | gfp_zone(gfp_mask), nodemask) { |
3024 | if (zone_idx(zone) > ZONE_NORMAL) |
3025 | continue; |
3026 | |
3027 | /* Throttle based on the first usable node */ |
3028 | pgdat = zone->zone_pgdat; |
3029 | if (allow_direct_reclaim(pgdat)) |
3030 | goto out; |
3031 | break; |
3032 | } |
3033 | |
3034 | /* If no zone was usable by the allocation flags then do not throttle */ |
3035 | if (!pgdat) |
3036 | goto out; |
3037 | |
3038 | /* Account for the throttling */ |
3039 | count_vm_event(PGSCAN_DIRECT_THROTTLE); |
3040 | |
3041 | /* |
3042 | * If the caller cannot enter the filesystem, it's possible that it |
3043 | * is due to the caller holding an FS lock or performing a journal |
3044 | * transaction in the case of a filesystem like ext[3|4]. In this case, |
3045 | * it is not safe to block on pfmemalloc_wait as kswapd could be |
3046 | * blocked waiting on the same lock. Instead, throttle for up to a |
3047 | * second before continuing. |
3048 | */ |
3049 | if (!(gfp_mask & __GFP_FS)) { |
3050 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, |
3051 | allow_direct_reclaim(pgdat), HZ); |
3052 | |
3053 | goto check_pending; |
3054 | } |
3055 | |
3056 | /* Throttle until kswapd wakes the process */ |
3057 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, |
3058 | allow_direct_reclaim(pgdat)); |
3059 | |
3060 | check_pending: |
3061 | if (fatal_signal_pending(current)) |
3062 | return true; |
3063 | |
3064 | out: |
3065 | return false; |
3066 | } |
3067 | |
3068 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
3069 | gfp_t gfp_mask, nodemask_t *nodemask) |
3070 | { |
3071 | unsigned long nr_reclaimed; |
3072 | struct scan_control sc = { |
3073 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
3074 | .gfp_mask = memalloc_noio_flags(gfp_mask), |
3075 | .reclaim_idx = gfp_zone(gfp_mask), |
3076 | .order = order, |
3077 | .nodemask = nodemask, |
3078 | .priority = DEF_PRIORITY, |
3079 | .may_writepage = !laptop_mode, |
3080 | .may_unmap = 1, |
3081 | .may_swap = 1, |
3082 | }; |
3083 | |
3084 | /* |
3085 | * Do not enter reclaim if fatal signal was delivered while throttled. |
3086 | * 1 is returned so that the page allocator does not OOM kill at this |
3087 | * point. |
3088 | */ |
3089 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
3090 | return 1; |
3091 | |
3092 | trace_mm_vmscan_direct_reclaim_begin(order, |
3093 | sc.may_writepage, |
3094 | sc.gfp_mask, |
3095 | sc.reclaim_idx); |
3096 | |
3097 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
3098 | |
3099 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); |
3100 | |
3101 | return nr_reclaimed; |
3102 | } |
3103 | |
3104 | #ifdef CONFIG_MEMCG |
3105 | |
3106 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
3107 | gfp_t gfp_mask, bool noswap, |
3108 | pg_data_t *pgdat, |
3109 | unsigned long *nr_scanned) |
3110 | { |
3111 | struct scan_control sc = { |
3112 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
3113 | .target_mem_cgroup = memcg, |
3114 | .may_writepage = !laptop_mode, |
3115 | .may_unmap = 1, |
3116 | .reclaim_idx = MAX_NR_ZONES - 1, |
3117 | .may_swap = !noswap, |
3118 | }; |
3119 | unsigned long lru_pages; |
3120 | |
3121 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3122 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); |
3123 | |
3124 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
3125 | sc.may_writepage, |
3126 | sc.gfp_mask, |
3127 | sc.reclaim_idx); |
3128 | |
3129 | /* |
3130 | * NOTE: Although we can get the priority field, using it |
3131 | * here is not a good idea, since it limits the pages we can scan. |
3132 | * if we don't reclaim here, the shrink_node from balance_pgdat |
3133 | * will pick up pages from other mem cgroup's as well. We hack |
3134 | * the priority and make it zero. |
3135 | */ |
3136 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
3137 | |
3138 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); |
3139 | |
3140 | *nr_scanned = sc.nr_scanned; |
3141 | return sc.nr_reclaimed; |
3142 | } |
3143 | |
3144 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
3145 | unsigned long nr_pages, |
3146 | gfp_t gfp_mask, |
3147 | bool may_swap) |
3148 | { |
3149 | struct zonelist *zonelist; |
3150 | unsigned long nr_reclaimed; |
3151 | unsigned long pflags; |
3152 | int nid; |
3153 | struct scan_control sc = { |
3154 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
3155 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3156 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
3157 | .reclaim_idx = MAX_NR_ZONES - 1, |
3158 | .target_mem_cgroup = memcg, |
3159 | .priority = DEF_PRIORITY, |
3160 | .may_writepage = !laptop_mode, |
3161 | .may_unmap = 1, |
3162 | .may_swap = may_swap, |
3163 | }; |
3164 | |
3165 | /* |
3166 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't |
3167 | * take care of from where we get pages. So the node where we start the |
3168 | * scan does not need to be the current node. |
3169 | */ |
3170 | nid = mem_cgroup_select_victim_node(memcg); |
3171 | |
3172 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; |
3173 | |
3174 | trace_mm_vmscan_memcg_reclaim_begin(0, |
3175 | sc.may_writepage, |
3176 | sc.gfp_mask, |
3177 | sc.reclaim_idx); |
3178 | |
3179 | psi_memstall_enter(&pflags); |
3180 | current->flags |= PF_MEMALLOC; |
3181 | |
3182 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
3183 | |
3184 | current->flags &= ~PF_MEMALLOC; |
3185 | psi_memstall_leave(&pflags); |
3186 | |
3187 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); |
3188 | |
3189 | return nr_reclaimed; |
3190 | } |
3191 | #endif |
3192 | |
3193 | static void age_active_anon(struct pglist_data *pgdat, |
3194 | struct scan_control *sc) |
3195 | { |
3196 | struct mem_cgroup *memcg; |
3197 | |
3198 | if (!total_swap_pages) |
3199 | return; |
3200 | |
3201 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
3202 | do { |
3203 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
3204 | |
3205 | if (inactive_list_is_low(lruvec, false, sc)) |
3206 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
3207 | sc, LRU_ACTIVE_ANON); |
3208 | |
3209 | memcg = mem_cgroup_iter(NULL, memcg, NULL); |
3210 | } while (memcg); |
3211 | } |
3212 | |
3213 | static bool zone_balanced(struct zone *zone, int order, int classzone_idx) |
3214 | { |
3215 | unsigned long mark = high_wmark_pages(zone); |
3216 | |
3217 | if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx)) |
3218 | return false; |
3219 | |
3220 | /* |
3221 | * If any eligible zone is balanced then the node is not considered |
3222 | * to be congested or dirty |
3223 | */ |
3224 | clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags); |
3225 | clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags); |
3226 | clear_bit(PGDAT_WRITEBACK, &zone->zone_pgdat->flags); |
3227 | |
3228 | return true; |
3229 | } |
3230 | |
3231 | /* |
3232 | * Prepare kswapd for sleeping. This verifies that there are no processes |
3233 | * waiting in throttle_direct_reclaim() and that watermarks have been met. |
3234 | * |
3235 | * Returns true if kswapd is ready to sleep |
3236 | */ |
3237 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
3238 | { |
3239 | int i; |
3240 | |
3241 | /* |
3242 | * The throttled processes are normally woken up in balance_pgdat() as |
3243 | * soon as allow_direct_reclaim() is true. But there is a potential |
3244 | * race between when kswapd checks the watermarks and a process gets |
3245 | * throttled. There is also a potential race if processes get |
3246 | * throttled, kswapd wakes, a large process exits thereby balancing the |
3247 | * zones, which causes kswapd to exit balance_pgdat() before reaching |
3248 | * the wake up checks. If kswapd is going to sleep, no process should |
3249 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If |
3250 | * the wake up is premature, processes will wake kswapd and get |
3251 | * throttled again. The difference from wake ups in balance_pgdat() is |
3252 | * that here we are under prepare_to_wait(). |
3253 | */ |
3254 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3255 | wake_up_all(&pgdat->pfmemalloc_wait); |
3256 | |
3257 | /* Hopeless node, leave it to direct reclaim */ |
3258 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
3259 | return true; |
3260 | |
3261 | for (i = 0; i <= classzone_idx; i++) { |
3262 | struct zone *zone = pgdat->node_zones + i; |
3263 | |
3264 | if (!managed_zone(zone)) |
3265 | continue; |
3266 | |
3267 | if (!zone_balanced(zone, order, classzone_idx)) |
3268 | return false; |
3269 | } |
3270 | |
3271 | return true; |
3272 | } |
3273 | |
3274 | /* |
3275 | * kswapd shrinks a node of pages that are at or below the highest usable |
3276 | * zone that is currently unbalanced. |
3277 | * |
3278 | * Returns true if kswapd scanned at least the requested number of pages to |
3279 | * reclaim or if the lack of progress was due to pages under writeback. |
3280 | * This is used to determine if the scanning priority needs to be raised. |
3281 | */ |
3282 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
3283 | struct scan_control *sc) |
3284 | { |
3285 | struct zone *zone; |
3286 | int z; |
3287 | |
3288 | /* Reclaim a number of pages proportional to the number of zones */ |
3289 | sc->nr_to_reclaim = 0; |
3290 | for (z = 0; z <= sc->reclaim_idx; z++) { |
3291 | zone = pgdat->node_zones + z; |
3292 | if (!managed_zone(zone)) |
3293 | continue; |
3294 | |
3295 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3296 | } |
3297 | |
3298 | /* |
3299 | * Historically care was taken to put equal pressure on all zones but |
3300 | * now pressure is applied based on node LRU order. |
3301 | */ |
3302 | shrink_node(pgdat, sc); |
3303 | |
3304 | /* |
3305 | * Fragmentation may mean that the system cannot be rebalanced for |
3306 | * high-order allocations. If twice the allocation size has been |
3307 | * reclaimed then recheck watermarks only at order-0 to prevent |
3308 | * excessive reclaim. Assume that a process requested a high-order |
3309 | * can direct reclaim/compact. |
3310 | */ |
3311 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
3312 | sc->order = 0; |
3313 | |
3314 | return sc->nr_scanned >= sc->nr_to_reclaim; |
3315 | } |
3316 | |
3317 | /* |
3318 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3319 | * that are eligible for use by the caller until at least one zone is |
3320 | * balanced. |
3321 | * |
3322 | * Returns the order kswapd finished reclaiming at. |
3323 | * |
3324 | * kswapd scans the zones in the highmem->normal->dma direction. It skips |
3325 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
3326 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone |
3327 | * or lower is eligible for reclaim until at least one usable zone is |
3328 | * balanced. |
3329 | */ |
3330 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
3331 | { |
3332 | int i; |
3333 | unsigned long nr_soft_reclaimed; |
3334 | unsigned long nr_soft_scanned; |
3335 | unsigned long pflags; |
3336 | struct zone *zone; |
3337 | struct scan_control sc = { |
3338 | .gfp_mask = GFP_KERNEL, |
3339 | .order = order, |
3340 | .priority = DEF_PRIORITY, |
3341 | .may_writepage = !laptop_mode, |
3342 | .may_unmap = 1, |
3343 | .may_swap = 1, |
3344 | }; |
3345 | |
3346 | psi_memstall_enter(&pflags); |
3347 | count_vm_event(PAGEOUTRUN); |
3348 | |
3349 | do { |
3350 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
3351 | bool raise_priority = true; |
3352 | |
3353 | sc.reclaim_idx = classzone_idx; |
3354 | |
3355 | /* |
3356 | * If the number of buffer_heads exceeds the maximum allowed |
3357 | * then consider reclaiming from all zones. This has a dual |
3358 | * purpose -- on 64-bit systems it is expected that |
3359 | * buffer_heads are stripped during active rotation. On 32-bit |
3360 | * systems, highmem pages can pin lowmem memory and shrinking |
3361 | * buffers can relieve lowmem pressure. Reclaim may still not |
3362 | * go ahead if all eligible zones for the original allocation |
3363 | * request are balanced to avoid excessive reclaim from kswapd. |
3364 | */ |
3365 | if (buffer_heads_over_limit) { |
3366 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { |
3367 | zone = pgdat->node_zones + i; |
3368 | if (!managed_zone(zone)) |
3369 | continue; |
3370 | |
3371 | sc.reclaim_idx = i; |
3372 | break; |
3373 | } |
3374 | } |
3375 | |
3376 | /* |
3377 | * Only reclaim if there are no eligible zones. Check from |
3378 | * high to low zone as allocations prefer higher zones. |
3379 | * Scanning from low to high zone would allow congestion to be |
3380 | * cleared during a very small window when a small low |
3381 | * zone was balanced even under extreme pressure when the |
3382 | * overall node may be congested. Note that sc.reclaim_idx |
3383 | * is not used as buffer_heads_over_limit may have adjusted |
3384 | * it. |
3385 | */ |
3386 | for (i = classzone_idx; i >= 0; i--) { |
3387 | zone = pgdat->node_zones + i; |
3388 | if (!managed_zone(zone)) |
3389 | continue; |
3390 | |
3391 | if (zone_balanced(zone, sc.order, classzone_idx)) |
3392 | goto out; |
3393 | } |
3394 | |
3395 | /* |
3396 | * Do some background aging of the anon list, to give |
3397 | * pages a chance to be referenced before reclaiming. All |
3398 | * pages are rotated regardless of classzone as this is |
3399 | * about consistent aging. |
3400 | */ |
3401 | age_active_anon(pgdat, &sc); |
3402 | |
3403 | /* |
3404 | * If we're getting trouble reclaiming, start doing writepage |
3405 | * even in laptop mode. |
3406 | */ |
3407 | if (sc.priority < DEF_PRIORITY - 2) |
3408 | sc.may_writepage = 1; |
3409 | |
3410 | /* Call soft limit reclaim before calling shrink_node. */ |
3411 | sc.nr_scanned = 0; |
3412 | nr_soft_scanned = 0; |
3413 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
3414 | sc.gfp_mask, &nr_soft_scanned); |
3415 | sc.nr_reclaimed += nr_soft_reclaimed; |
3416 | |
3417 | /* |
3418 | * There should be no need to raise the scanning priority if |
3419 | * enough pages are already being scanned that that high |
3420 | * watermark would be met at 100% efficiency. |
3421 | */ |
3422 | if (kswapd_shrink_node(pgdat, &sc)) |
3423 | raise_priority = false; |
3424 | |
3425 | /* |
3426 | * If the low watermark is met there is no need for processes |
3427 | * to be throttled on pfmemalloc_wait as they should not be |
3428 | * able to safely make forward progress. Wake them |
3429 | */ |
3430 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && |
3431 | allow_direct_reclaim(pgdat)) |
3432 | wake_up_all(&pgdat->pfmemalloc_wait); |
3433 | |
3434 | /* Check if kswapd should be suspending */ |
3435 | if (try_to_freeze() || kthread_should_stop()) |
3436 | break; |
3437 | |
3438 | /* |
3439 | * Raise priority if scanning rate is too low or there was no |
3440 | * progress in reclaiming pages |
3441 | */ |
3442 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
3443 | if (raise_priority || !nr_reclaimed) |
3444 | sc.priority--; |
3445 | } while (sc.priority >= 1); |
3446 | |
3447 | if (!sc.nr_reclaimed) |
3448 | pgdat->kswapd_failures++; |
3449 | |
3450 | out: |
3451 | psi_memstall_leave(&pflags); |
3452 | /* |
3453 | * Return the order kswapd stopped reclaiming at as |
3454 | * prepare_kswapd_sleep() takes it into account. If another caller |
3455 | * entered the allocator slow path while kswapd was awake, order will |
3456 | * remain at the higher level. |
3457 | */ |
3458 | return sc.order; |
3459 | } |
3460 | |
3461 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3462 | unsigned int classzone_idx) |
3463 | { |
3464 | long remaining = 0; |
3465 | DEFINE_WAIT(wait); |
3466 | |
3467 | if (freezing(current) || kthread_should_stop()) |
3468 | return; |
3469 | |
3470 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); |
3471 | |
3472 | /* Try to sleep for a short interval */ |
3473 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
3474 | /* |
3475 | * Compaction records what page blocks it recently failed to |
3476 | * isolate pages from and skips them in the future scanning. |
3477 | * When kswapd is going to sleep, it is reasonable to assume |
3478 | * that pages and compaction may succeed so reset the cache. |
3479 | */ |
3480 | reset_isolation_suitable(pgdat); |
3481 | |
3482 | /* |
3483 | * We have freed the memory, now we should compact it to make |
3484 | * allocation of the requested order possible. |
3485 | */ |
3486 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
3487 | |
3488 | remaining = schedule_timeout(HZ/10); |
3489 | |
3490 | /* |
3491 | * If woken prematurely then reset kswapd_classzone_idx and |
3492 | * order. The values will either be from a wakeup request or |
3493 | * the previous request that slept prematurely. |
3494 | */ |
3495 | if (remaining) { |
3496 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx); |
3497 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); |
3498 | } |
3499 | |
3500 | finish_wait(&pgdat->kswapd_wait, &wait); |
3501 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); |
3502 | } |
3503 | |
3504 | /* |
3505 | * After a short sleep, check if it was a premature sleep. If not, then |
3506 | * go fully to sleep until explicitly woken up. |
3507 | */ |
3508 | if (!remaining && |
3509 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
3510 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3511 | |
3512 | /* |
3513 | * vmstat counters are not perfectly accurate and the estimated |
3514 | * value for counters such as NR_FREE_PAGES can deviate from the |
3515 | * true value by nr_online_cpus * threshold. To avoid the zone |
3516 | * watermarks being breached while under pressure, we reduce the |
3517 | * per-cpu vmstat threshold while kswapd is awake and restore |
3518 | * them before going back to sleep. |
3519 | */ |
3520 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); |
3521 | |
3522 | if (!kthread_should_stop()) |
3523 | schedule(); |
3524 | |
3525 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3526 | } else { |
3527 | if (remaining) |
3528 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); |
3529 | else |
3530 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); |
3531 | } |
3532 | finish_wait(&pgdat->kswapd_wait, &wait); |
3533 | } |
3534 | |
3535 | /* |
3536 | * The background pageout daemon, started as a kernel thread |
3537 | * from the init process. |
3538 | * |
3539 | * This basically trickles out pages so that we have _some_ |
3540 | * free memory available even if there is no other activity |
3541 | * that frees anything up. This is needed for things like routing |
3542 | * etc, where we otherwise might have all activity going on in |
3543 | * asynchronous contexts that cannot page things out. |
3544 | * |
3545 | * If there are applications that are active memory-allocators |
3546 | * (most normal use), this basically shouldn't matter. |
3547 | */ |
3548 | static int kswapd(void *p) |
3549 | { |
3550 | unsigned int alloc_order, reclaim_order, classzone_idx; |
3551 | pg_data_t *pgdat = (pg_data_t*)p; |
3552 | struct task_struct *tsk = current; |
3553 | |
3554 | struct reclaim_state reclaim_state = { |
3555 | .reclaimed_slab = 0, |
3556 | }; |
3557 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
3558 | |
3559 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3560 | |
3561 | if (!cpumask_empty(cpumask)) |
3562 | set_cpus_allowed_ptr(tsk, cpumask); |
3563 | current->reclaim_state = &reclaim_state; |
3564 | |
3565 | /* |
3566 | * Tell the memory management that we're a "memory allocator", |
3567 | * and that if we need more memory we should get access to it |
3568 | * regardless (see "__alloc_pages()"). "kswapd" should |
3569 | * never get caught in the normal page freeing logic. |
3570 | * |
3571 | * (Kswapd normally doesn't need memory anyway, but sometimes |
3572 | * you need a small amount of memory in order to be able to |
3573 | * page out something else, and this flag essentially protects |
3574 | * us from recursively trying to free more memory as we're |
3575 | * trying to free the first piece of memory in the first place). |
3576 | */ |
3577 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
3578 | set_freezable(); |
3579 | |
3580 | pgdat->kswapd_order = alloc_order = reclaim_order = 0; |
3581 | pgdat->kswapd_classzone_idx = classzone_idx = 0; |
3582 | #ifdef CONFIG_AMLOGIC_CMA |
3583 | set_user_nice(current, -5); |
3584 | #endif /* CONFIG_AMLOGIC_CMA */ |
3585 | for ( ; ; ) { |
3586 | bool ret; |
3587 | |
3588 | kswapd_try_sleep: |
3589 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, |
3590 | classzone_idx); |
3591 | |
3592 | /* Read the new order and classzone_idx */ |
3593 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3594 | classzone_idx = pgdat->kswapd_classzone_idx; |
3595 | pgdat->kswapd_order = 0; |
3596 | pgdat->kswapd_classzone_idx = 0; |
3597 | |
3598 | ret = try_to_freeze(); |
3599 | if (kthread_should_stop()) |
3600 | break; |
3601 | |
3602 | /* |
3603 | * We can speed up thawing tasks if we don't call balance_pgdat |
3604 | * after returning from the refrigerator |
3605 | */ |
3606 | if (ret) |
3607 | continue; |
3608 | |
3609 | /* |
3610 | * Reclaim begins at the requested order but if a high-order |
3611 | * reclaim fails then kswapd falls back to reclaiming for |
3612 | * order-0. If that happens, kswapd will consider sleeping |
3613 | * for the order it finished reclaiming at (reclaim_order) |
3614 | * but kcompactd is woken to compact for the original |
3615 | * request (alloc_order). |
3616 | */ |
3617 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
3618 | alloc_order); |
3619 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
3620 | if (reclaim_order < alloc_order) |
3621 | goto kswapd_try_sleep; |
3622 | |
3623 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3624 | classzone_idx = pgdat->kswapd_classzone_idx; |
3625 | } |
3626 | |
3627 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
3628 | current->reclaim_state = NULL; |
3629 | lockdep_clear_current_reclaim_state(); |
3630 | |
3631 | return 0; |
3632 | } |
3633 | |
3634 | /* |
3635 | * A zone is low on free memory, so wake its kswapd task to service it. |
3636 | */ |
3637 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
3638 | { |
3639 | pg_data_t *pgdat; |
3640 | int z; |
3641 | |
3642 | if (!managed_zone(zone)) |
3643 | return; |
3644 | |
3645 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) |
3646 | return; |
3647 | pgdat = zone->zone_pgdat; |
3648 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx); |
3649 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
3650 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
3651 | return; |
3652 | |
3653 | /* Hopeless node, leave it to direct reclaim */ |
3654 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
3655 | return; |
3656 | |
3657 | /* Only wake kswapd if all zones are unbalanced */ |
3658 | for (z = 0; z <= classzone_idx; z++) { |
3659 | zone = pgdat->node_zones + z; |
3660 | if (!managed_zone(zone)) |
3661 | continue; |
3662 | |
3663 | if (zone_balanced(zone, order, classzone_idx)) |
3664 | return; |
3665 | } |
3666 | |
3667 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); |
3668 | wake_up_interruptible(&pgdat->kswapd_wait); |
3669 | } |
3670 | |
3671 | #ifdef CONFIG_HIBERNATION |
3672 | /* |
3673 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
3674 | * freed pages. |
3675 | * |
3676 | * Rather than trying to age LRUs the aim is to preserve the overall |
3677 | * LRU order by reclaiming preferentially |
3678 | * inactive > active > active referenced > active mapped |
3679 | */ |
3680 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
3681 | { |
3682 | struct reclaim_state reclaim_state; |
3683 | struct scan_control sc = { |
3684 | .nr_to_reclaim = nr_to_reclaim, |
3685 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
3686 | .reclaim_idx = MAX_NR_ZONES - 1, |
3687 | .priority = DEF_PRIORITY, |
3688 | .may_writepage = 1, |
3689 | .may_unmap = 1, |
3690 | .may_swap = 1, |
3691 | .hibernation_mode = 1, |
3692 | }; |
3693 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
3694 | struct task_struct *p = current; |
3695 | unsigned long nr_reclaimed; |
3696 | |
3697 | p->flags |= PF_MEMALLOC; |
3698 | lockdep_set_current_reclaim_state(sc.gfp_mask); |
3699 | reclaim_state.reclaimed_slab = 0; |
3700 | p->reclaim_state = &reclaim_state; |
3701 | |
3702 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
3703 | |
3704 | p->reclaim_state = NULL; |
3705 | lockdep_clear_current_reclaim_state(); |
3706 | p->flags &= ~PF_MEMALLOC; |
3707 | |
3708 | return nr_reclaimed; |
3709 | } |
3710 | #endif /* CONFIG_HIBERNATION */ |
3711 | |
3712 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3713 | not required for correctness. So if the last cpu in a node goes |
3714 | away, we get changed to run anywhere: as the first one comes back, |
3715 | restore their cpu bindings. */ |
3716 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
3717 | void *hcpu) |
3718 | { |
3719 | int nid; |
3720 | |
3721 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
3722 | for_each_node_state(nid, N_MEMORY) { |
3723 | pg_data_t *pgdat = NODE_DATA(nid); |
3724 | const struct cpumask *mask; |
3725 | |
3726 | mask = cpumask_of_node(pgdat->node_id); |
3727 | |
3728 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3729 | /* One of our CPUs online: restore mask */ |
3730 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
3731 | } |
3732 | } |
3733 | return NOTIFY_OK; |
3734 | } |
3735 | |
3736 | /* |
3737 | * This kswapd start function will be called by init and node-hot-add. |
3738 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. |
3739 | */ |
3740 | int kswapd_run(int nid) |
3741 | { |
3742 | pg_data_t *pgdat = NODE_DATA(nid); |
3743 | int ret = 0; |
3744 | |
3745 | if (pgdat->kswapd) |
3746 | return 0; |
3747 | |
3748 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); |
3749 | if (IS_ERR(pgdat->kswapd)) { |
3750 | /* failure at boot is fatal */ |
3751 | BUG_ON(system_state == SYSTEM_BOOTING); |
3752 | pr_err("Failed to start kswapd on node %d\n", nid); |
3753 | ret = PTR_ERR(pgdat->kswapd); |
3754 | pgdat->kswapd = NULL; |
3755 | } |
3756 | return ret; |
3757 | } |
3758 | |
3759 | /* |
3760 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
3761 | * hold mem_hotplug_begin/end(). |
3762 | */ |
3763 | void kswapd_stop(int nid) |
3764 | { |
3765 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; |
3766 | |
3767 | if (kswapd) { |
3768 | kthread_stop(kswapd); |
3769 | NODE_DATA(nid)->kswapd = NULL; |
3770 | } |
3771 | } |
3772 | |
3773 | static int __init kswapd_init(void) |
3774 | { |
3775 | int nid; |
3776 | |
3777 | swap_setup(); |
3778 | for_each_node_state(nid, N_MEMORY) |
3779 | kswapd_run(nid); |
3780 | hotcpu_notifier(cpu_callback, 0); |
3781 | return 0; |
3782 | } |
3783 | |
3784 | module_init(kswapd_init) |
3785 | |
3786 | #ifdef CONFIG_NUMA |
3787 | /* |
3788 | * Node reclaim mode |
3789 | * |
3790 | * If non-zero call node_reclaim when the number of free pages falls below |
3791 | * the watermarks. |
3792 | */ |
3793 | int node_reclaim_mode __read_mostly; |
3794 | |
3795 | #define RECLAIM_OFF 0 |
3796 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
3797 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
3798 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
3799 | |
3800 | /* |
3801 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
3802 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
3803 | * a zone. |
3804 | */ |
3805 | #define NODE_RECLAIM_PRIORITY 4 |
3806 | |
3807 | /* |
3808 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
3809 | * occur. |
3810 | */ |
3811 | int sysctl_min_unmapped_ratio = 1; |
3812 | |
3813 | /* |
3814 | * If the number of slab pages in a zone grows beyond this percentage then |
3815 | * slab reclaim needs to occur. |
3816 | */ |
3817 | int sysctl_min_slab_ratio = 5; |
3818 | |
3819 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
3820 | { |
3821 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
3822 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + |
3823 | node_page_state(pgdat, NR_ACTIVE_FILE); |
3824 | |
3825 | /* |
3826 | * It's possible for there to be more file mapped pages than |
3827 | * accounted for by the pages on the file LRU lists because |
3828 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED |
3829 | */ |
3830 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; |
3831 | } |
3832 | |
3833 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ |
3834 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
3835 | { |
3836 | unsigned long nr_pagecache_reclaimable; |
3837 | unsigned long delta = 0; |
3838 | |
3839 | /* |
3840 | * If RECLAIM_UNMAP is set, then all file pages are considered |
3841 | * potentially reclaimable. Otherwise, we have to worry about |
3842 | * pages like swapcache and node_unmapped_file_pages() provides |
3843 | * a better estimate |
3844 | */ |
3845 | if (node_reclaim_mode & RECLAIM_UNMAP) |
3846 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); |
3847 | else |
3848 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
3849 | |
3850 | /* If we can't clean pages, remove dirty pages from consideration */ |
3851 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
3852 | delta += node_page_state(pgdat, NR_FILE_DIRTY); |
3853 | |
3854 | /* Watch for any possible underflows due to delta */ |
3855 | if (unlikely(delta > nr_pagecache_reclaimable)) |
3856 | delta = nr_pagecache_reclaimable; |
3857 | |
3858 | return nr_pagecache_reclaimable - delta; |
3859 | } |
3860 | |
3861 | /* |
3862 | * Try to free up some pages from this node through reclaim. |
3863 | */ |
3864 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
3865 | { |
3866 | /* Minimum pages needed in order to stay on node */ |
3867 | const unsigned long nr_pages = 1 << order; |
3868 | struct task_struct *p = current; |
3869 | struct reclaim_state reclaim_state; |
3870 | struct scan_control sc = { |
3871 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
3872 | .gfp_mask = memalloc_noio_flags(gfp_mask), |
3873 | .order = order, |
3874 | .priority = NODE_RECLAIM_PRIORITY, |
3875 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), |
3876 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), |
3877 | .may_swap = 1, |
3878 | .reclaim_idx = gfp_zone(gfp_mask), |
3879 | }; |
3880 | |
3881 | cond_resched(); |
3882 | /* |
3883 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
3884 | * and we also need to be able to write out pages for RECLAIM_WRITE |
3885 | * and RECLAIM_UNMAP. |
3886 | */ |
3887 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; |
3888 | lockdep_set_current_reclaim_state(sc.gfp_mask); |
3889 | reclaim_state.reclaimed_slab = 0; |
3890 | p->reclaim_state = &reclaim_state; |
3891 | |
3892 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
3893 | /* |
3894 | * Free memory by calling shrink zone with increasing |
3895 | * priorities until we have enough memory freed. |
3896 | */ |
3897 | do { |
3898 | shrink_node(pgdat, &sc); |
3899 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
3900 | } |
3901 | |
3902 | p->reclaim_state = NULL; |
3903 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
3904 | lockdep_clear_current_reclaim_state(); |
3905 | return sc.nr_reclaimed >= nr_pages; |
3906 | } |
3907 | |
3908 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
3909 | { |
3910 | int ret; |
3911 | |
3912 | /* |
3913 | * Node reclaim reclaims unmapped file backed pages and |
3914 | * slab pages if we are over the defined limits. |
3915 | * |
3916 | * A small portion of unmapped file backed pages is needed for |
3917 | * file I/O otherwise pages read by file I/O will be immediately |
3918 | * thrown out if the node is overallocated. So we do not reclaim |
3919 | * if less than a specified percentage of the node is used by |
3920 | * unmapped file backed pages. |
3921 | */ |
3922 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
3923 | sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) |
3924 | return NODE_RECLAIM_FULL; |
3925 | |
3926 | /* |
3927 | * Do not scan if the allocation should not be delayed. |
3928 | */ |
3929 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
3930 | return NODE_RECLAIM_NOSCAN; |
3931 | |
3932 | /* |
3933 | * Only run node reclaim on the local node or on nodes that do not |
3934 | * have associated processors. This will favor the local processor |
3935 | * over remote processors and spread off node memory allocations |
3936 | * as wide as possible. |
3937 | */ |
3938 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
3939 | return NODE_RECLAIM_NOSCAN; |
3940 | |
3941 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
3942 | return NODE_RECLAIM_NOSCAN; |
3943 | |
3944 | ret = __node_reclaim(pgdat, gfp_mask, order); |
3945 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); |
3946 | |
3947 | if (!ret) |
3948 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); |
3949 | |
3950 | return ret; |
3951 | } |
3952 | #endif |
3953 | |
3954 | /* |
3955 | * page_evictable - test whether a page is evictable |
3956 | * @page: the page to test |
3957 | * |
3958 | * Test whether page is evictable--i.e., should be placed on active/inactive |
3959 | * lists vs unevictable list. |
3960 | * |
3961 | * Reasons page might not be evictable: |
3962 | * (1) page's mapping marked unevictable |
3963 | * (2) page is part of an mlocked VMA |
3964 | * |
3965 | */ |
3966 | int page_evictable(struct page *page) |
3967 | { |
3968 | int ret; |
3969 | |
3970 | /* Prevent address_space of inode and swap cache from being freed */ |
3971 | rcu_read_lock(); |
3972 | ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
3973 | rcu_read_unlock(); |
3974 | return ret; |
3975 | } |
3976 | |
3977 | #ifdef CONFIG_SHMEM |
3978 | /** |
3979 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3980 | * @pages: array of pages to check |
3981 | * @nr_pages: number of pages to check |
3982 | * |
3983 | * Checks pages for evictability and moves them to the appropriate lru list. |
3984 | * |
3985 | * This function is only used for SysV IPC SHM_UNLOCK. |
3986 | */ |
3987 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
3988 | { |
3989 | struct lruvec *lruvec; |
3990 | struct pglist_data *pgdat = NULL; |
3991 | int pgscanned = 0; |
3992 | int pgrescued = 0; |
3993 | int i; |
3994 | |
3995 | for (i = 0; i < nr_pages; i++) { |
3996 | struct page *page = pages[i]; |
3997 | struct pglist_data *pagepgdat = page_pgdat(page); |
3998 | |
3999 | pgscanned++; |
4000 | if (pagepgdat != pgdat) { |
4001 | if (pgdat) |
4002 | spin_unlock_irq(&pgdat->lru_lock); |
4003 | pgdat = pagepgdat; |
4004 | spin_lock_irq(&pgdat->lru_lock); |
4005 | } |
4006 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
4007 | |
4008 | if (!PageLRU(page) || !PageUnevictable(page)) |
4009 | continue; |
4010 | |
4011 | if (page_evictable(page)) { |
4012 | enum lru_list lru = page_lru_base_type(page); |
4013 | |
4014 | VM_BUG_ON_PAGE(PageActive(page), page); |
4015 | ClearPageUnevictable(page); |
4016 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
4017 | add_page_to_lru_list(page, lruvec, lru); |
4018 | pgrescued++; |
4019 | } |
4020 | } |
4021 | |
4022 | if (pgdat) { |
4023 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
4024 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); |
4025 | spin_unlock_irq(&pgdat->lru_lock); |
4026 | } |
4027 | } |
4028 | #endif /* CONFIG_SHMEM */ |
4029 |