]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/mm/vmscan.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
5 | * | |
6 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
7 | * kswapd added: 7.1.96 sct | |
8 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
9 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
10 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
11 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
12 | */ | |
13 | ||
14 | #include <linux/mm.h> | |
15 | #include <linux/module.h> | |
16 | #include <linux/gfp.h> | |
17 | #include <linux/kernel_stat.h> | |
18 | #include <linux/swap.h> | |
19 | #include <linux/pagemap.h> | |
20 | #include <linux/init.h> | |
21 | #include <linux/highmem.h> | |
22 | #include <linux/vmstat.h> | |
23 | #include <linux/file.h> | |
24 | #include <linux/writeback.h> | |
25 | #include <linux/blkdev.h> | |
26 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
27 | buffer_heads_over_limit */ | |
28 | #include <linux/mm_inline.h> | |
29 | #include <linux/pagevec.h> | |
30 | #include <linux/backing-dev.h> | |
31 | #include <linux/rmap.h> | |
32 | #include <linux/topology.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/cpuset.h> | |
35 | #include <linux/compaction.h> | |
36 | #include <linux/notifier.h> | |
37 | #include <linux/rwsem.h> | |
38 | #include <linux/delay.h> | |
39 | #include <linux/kthread.h> | |
40 | #include <linux/freezer.h> | |
41 | #include <linux/memcontrol.h> | |
42 | #include <linux/delayacct.h> | |
43 | #include <linux/sysctl.h> | |
44 | #include <linux/oom.h> | |
45 | #include <linux/prefetch.h> | |
46 | ||
47 | #include <asm/tlbflush.h> | |
48 | #include <asm/div64.h> | |
49 | ||
50 | #include <linux/swapops.h> | |
51 | ||
52 | #include "internal.h" | |
53 | ||
54 | #define CREATE_TRACE_POINTS | |
55 | #include <trace/events/vmscan.h> | |
56 | ||
57 | /* | |
58 | * reclaim_mode determines how the inactive list is shrunk | |
59 | * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages | |
60 | * RECLAIM_MODE_ASYNC: Do not block | |
61 | * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback | |
62 | * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference | |
63 | * page from the LRU and reclaim all pages within a | |
64 | * naturally aligned range | |
65 | * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of | |
66 | * order-0 pages and then compact the zone | |
67 | */ | |
68 | typedef unsigned __bitwise__ reclaim_mode_t; | |
69 | #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u) | |
70 | #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u) | |
71 | #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u) | |
72 | #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u) | |
73 | #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u) | |
74 | ||
75 | struct scan_control { | |
76 | /* Incremented by the number of inactive pages that were scanned */ | |
77 | unsigned long nr_scanned; | |
78 | ||
79 | /* Number of pages freed so far during a call to shrink_zones() */ | |
80 | unsigned long nr_reclaimed; | |
81 | ||
82 | /* How many pages shrink_list() should reclaim */ | |
83 | unsigned long nr_to_reclaim; | |
84 | ||
85 | unsigned long hibernation_mode; | |
86 | ||
87 | /* This context's GFP mask */ | |
88 | gfp_t gfp_mask; | |
89 | ||
90 | int may_writepage; | |
91 | ||
92 | /* Can mapped pages be reclaimed? */ | |
93 | int may_unmap; | |
94 | ||
95 | /* Can pages be swapped as part of reclaim? */ | |
96 | int may_swap; | |
97 | ||
98 | int order; | |
99 | ||
100 | /* | |
101 | * Intend to reclaim enough continuous memory rather than reclaim | |
102 | * enough amount of memory. i.e, mode for high order allocation. | |
103 | */ | |
104 | reclaim_mode_t reclaim_mode; | |
105 | ||
106 | /* Which cgroup do we reclaim from */ | |
107 | struct mem_cgroup *mem_cgroup; | |
108 | ||
109 | /* | |
110 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
111 | * are scanned. | |
112 | */ | |
113 | nodemask_t *nodemask; | |
114 | }; | |
115 | ||
116 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) | |
117 | ||
118 | #ifdef ARCH_HAS_PREFETCH | |
119 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
120 | do { \ | |
121 | if ((_page)->lru.prev != _base) { \ | |
122 | struct page *prev; \ | |
123 | \ | |
124 | prev = lru_to_page(&(_page->lru)); \ | |
125 | prefetch(&prev->_field); \ | |
126 | } \ | |
127 | } while (0) | |
128 | #else | |
129 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
130 | #endif | |
131 | ||
132 | #ifdef ARCH_HAS_PREFETCHW | |
133 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
134 | do { \ | |
135 | if ((_page)->lru.prev != _base) { \ | |
136 | struct page *prev; \ | |
137 | \ | |
138 | prev = lru_to_page(&(_page->lru)); \ | |
139 | prefetchw(&prev->_field); \ | |
140 | } \ | |
141 | } while (0) | |
142 | #else | |
143 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
144 | #endif | |
145 | ||
146 | /* | |
147 | * From 0 .. 100. Higher means more swappy. | |
148 | */ | |
149 | int vm_swappiness = 60; | |
150 | long vm_total_pages; /* The total number of pages which the VM controls */ | |
151 | ||
152 | static LIST_HEAD(shrinker_list); | |
153 | static DECLARE_RWSEM(shrinker_rwsem); | |
154 | ||
155 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR | |
156 | #define scanning_global_lru(sc) (!(sc)->mem_cgroup) | |
157 | #else | |
158 | #define scanning_global_lru(sc) (1) | |
159 | #endif | |
160 | ||
161 | static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone, | |
162 | struct scan_control *sc) | |
163 | { | |
164 | if (!scanning_global_lru(sc)) | |
165 | return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone); | |
166 | ||
167 | return &zone->reclaim_stat; | |
168 | } | |
169 | ||
170 | static unsigned long zone_nr_lru_pages(struct zone *zone, | |
171 | struct scan_control *sc, enum lru_list lru) | |
172 | { | |
173 | if (!scanning_global_lru(sc)) | |
174 | return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, | |
175 | zone_to_nid(zone), zone_idx(zone), BIT(lru)); | |
176 | ||
177 | return zone_page_state(zone, NR_LRU_BASE + lru); | |
178 | } | |
179 | ||
180 | ||
181 | /* | |
182 | * Add a shrinker callback to be called from the vm | |
183 | */ | |
184 | void register_shrinker(struct shrinker *shrinker) | |
185 | { | |
186 | shrinker->nr = 0; | |
187 | down_write(&shrinker_rwsem); | |
188 | list_add_tail(&shrinker->list, &shrinker_list); | |
189 | up_write(&shrinker_rwsem); | |
190 | } | |
191 | EXPORT_SYMBOL(register_shrinker); | |
192 | ||
193 | /* | |
194 | * Remove one | |
195 | */ | |
196 | void unregister_shrinker(struct shrinker *shrinker) | |
197 | { | |
198 | down_write(&shrinker_rwsem); | |
199 | list_del(&shrinker->list); | |
200 | up_write(&shrinker_rwsem); | |
201 | } | |
202 | EXPORT_SYMBOL(unregister_shrinker); | |
203 | ||
204 | static inline int do_shrinker_shrink(struct shrinker *shrinker, | |
205 | struct shrink_control *sc, | |
206 | unsigned long nr_to_scan) | |
207 | { | |
208 | sc->nr_to_scan = nr_to_scan; | |
209 | return (*shrinker->shrink)(shrinker, sc); | |
210 | } | |
211 | ||
212 | #define SHRINK_BATCH 128 | |
213 | /* | |
214 | * Call the shrink functions to age shrinkable caches | |
215 | * | |
216 | * Here we assume it costs one seek to replace a lru page and that it also | |
217 | * takes a seek to recreate a cache object. With this in mind we age equal | |
218 | * percentages of the lru and ageable caches. This should balance the seeks | |
219 | * generated by these structures. | |
220 | * | |
221 | * If the vm encountered mapped pages on the LRU it increase the pressure on | |
222 | * slab to avoid swapping. | |
223 | * | |
224 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
225 | * | |
226 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
227 | * are eligible for the caller's allocation attempt. It is used for balancing | |
228 | * slab reclaim versus page reclaim. | |
229 | * | |
230 | * Returns the number of slab objects which we shrunk. | |
231 | */ | |
232 | unsigned long shrink_slab(struct shrink_control *shrink, | |
233 | unsigned long nr_pages_scanned, | |
234 | unsigned long lru_pages) | |
235 | { | |
236 | struct shrinker *shrinker; | |
237 | unsigned long ret = 0; | |
238 | ||
239 | if (nr_pages_scanned == 0) | |
240 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
241 | ||
242 | if (!down_read_trylock(&shrinker_rwsem)) { | |
243 | /* Assume we'll be able to shrink next time */ | |
244 | ret = 1; | |
245 | goto out; | |
246 | } | |
247 | ||
248 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
249 | unsigned long long delta; | |
250 | unsigned long total_scan; | |
251 | unsigned long max_pass; | |
252 | int shrink_ret = 0; | |
253 | long nr; | |
254 | long new_nr; | |
255 | long batch_size = shrinker->batch ? shrinker->batch | |
256 | : SHRINK_BATCH; | |
257 | ||
258 | /* | |
259 | * copy the current shrinker scan count into a local variable | |
260 | * and zero it so that other concurrent shrinker invocations | |
261 | * don't also do this scanning work. | |
262 | */ | |
263 | do { | |
264 | nr = shrinker->nr; | |
265 | } while (cmpxchg(&shrinker->nr, nr, 0) != nr); | |
266 | ||
267 | total_scan = nr; | |
268 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); | |
269 | delta = (4 * nr_pages_scanned) / shrinker->seeks; | |
270 | delta *= max_pass; | |
271 | do_div(delta, lru_pages + 1); | |
272 | total_scan += delta; | |
273 | if (total_scan < 0) { | |
274 | printk(KERN_ERR "shrink_slab: %pF negative objects to " | |
275 | "delete nr=%ld\n", | |
276 | shrinker->shrink, total_scan); | |
277 | total_scan = max_pass; | |
278 | } | |
279 | ||
280 | /* | |
281 | * We need to avoid excessive windup on filesystem shrinkers | |
282 | * due to large numbers of GFP_NOFS allocations causing the | |
283 | * shrinkers to return -1 all the time. This results in a large | |
284 | * nr being built up so when a shrink that can do some work | |
285 | * comes along it empties the entire cache due to nr >>> | |
286 | * max_pass. This is bad for sustaining a working set in | |
287 | * memory. | |
288 | * | |
289 | * Hence only allow the shrinker to scan the entire cache when | |
290 | * a large delta change is calculated directly. | |
291 | */ | |
292 | if (delta < max_pass / 4) | |
293 | total_scan = min(total_scan, max_pass / 2); | |
294 | ||
295 | /* | |
296 | * Avoid risking looping forever due to too large nr value: | |
297 | * never try to free more than twice the estimate number of | |
298 | * freeable entries. | |
299 | */ | |
300 | if (total_scan > max_pass * 2) | |
301 | total_scan = max_pass * 2; | |
302 | ||
303 | trace_mm_shrink_slab_start(shrinker, shrink, nr, | |
304 | nr_pages_scanned, lru_pages, | |
305 | max_pass, delta, total_scan); | |
306 | ||
307 | while (total_scan >= batch_size) { | |
308 | int nr_before; | |
309 | ||
310 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); | |
311 | shrink_ret = do_shrinker_shrink(shrinker, shrink, | |
312 | batch_size); | |
313 | if (shrink_ret == -1) | |
314 | break; | |
315 | if (shrink_ret < nr_before) | |
316 | ret += nr_before - shrink_ret; | |
317 | count_vm_events(SLABS_SCANNED, batch_size); | |
318 | total_scan -= batch_size; | |
319 | ||
320 | cond_resched(); | |
321 | } | |
322 | ||
323 | /* | |
324 | * move the unused scan count back into the shrinker in a | |
325 | * manner that handles concurrent updates. If we exhausted the | |
326 | * scan, there is no need to do an update. | |
327 | */ | |
328 | do { | |
329 | nr = shrinker->nr; | |
330 | new_nr = total_scan + nr; | |
331 | if (total_scan <= 0) | |
332 | break; | |
333 | } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr); | |
334 | ||
335 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); | |
336 | } | |
337 | up_read(&shrinker_rwsem); | |
338 | out: | |
339 | cond_resched(); | |
340 | return ret; | |
341 | } | |
342 | ||
343 | static void set_reclaim_mode(int priority, struct scan_control *sc, | |
344 | bool sync) | |
345 | { | |
346 | reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC; | |
347 | ||
348 | /* | |
349 | * Initially assume we are entering either lumpy reclaim or | |
350 | * reclaim/compaction.Depending on the order, we will either set the | |
351 | * sync mode or just reclaim order-0 pages later. | |
352 | */ | |
353 | if (COMPACTION_BUILD) | |
354 | sc->reclaim_mode = RECLAIM_MODE_COMPACTION; | |
355 | else | |
356 | sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM; | |
357 | ||
358 | /* | |
359 | * Avoid using lumpy reclaim or reclaim/compaction if possible by | |
360 | * restricting when its set to either costly allocations or when | |
361 | * under memory pressure | |
362 | */ | |
363 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) | |
364 | sc->reclaim_mode |= syncmode; | |
365 | else if (sc->order && priority < DEF_PRIORITY - 2) | |
366 | sc->reclaim_mode |= syncmode; | |
367 | else | |
368 | sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; | |
369 | } | |
370 | ||
371 | static void reset_reclaim_mode(struct scan_control *sc) | |
372 | { | |
373 | sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; | |
374 | } | |
375 | ||
376 | static inline int is_page_cache_freeable(struct page *page) | |
377 | { | |
378 | /* | |
379 | * A freeable page cache page is referenced only by the caller | |
380 | * that isolated the page, the page cache radix tree and | |
381 | * optional buffer heads at page->private. | |
382 | */ | |
383 | return page_count(page) - page_has_private(page) == 2; | |
384 | } | |
385 | ||
386 | static int may_write_to_queue(struct backing_dev_info *bdi, | |
387 | struct scan_control *sc) | |
388 | { | |
389 | if (current->flags & PF_SWAPWRITE) | |
390 | return 1; | |
391 | if (!bdi_write_congested(bdi)) | |
392 | return 1; | |
393 | if (bdi == current->backing_dev_info) | |
394 | return 1; | |
395 | ||
396 | /* lumpy reclaim for hugepage often need a lot of write */ | |
397 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) | |
398 | return 1; | |
399 | return 0; | |
400 | } | |
401 | ||
402 | /* | |
403 | * We detected a synchronous write error writing a page out. Probably | |
404 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
405 | * fsync(), msync() or close(). | |
406 | * | |
407 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
408 | * prevents it from being freed up. But we have a ref on the page and once | |
409 | * that page is locked, the mapping is pinned. | |
410 | * | |
411 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
412 | * __GFP_FS. | |
413 | */ | |
414 | static void handle_write_error(struct address_space *mapping, | |
415 | struct page *page, int error) | |
416 | { | |
417 | lock_page(page); | |
418 | if (page_mapping(page) == mapping) | |
419 | mapping_set_error(mapping, error); | |
420 | unlock_page(page); | |
421 | } | |
422 | ||
423 | /* possible outcome of pageout() */ | |
424 | typedef enum { | |
425 | /* failed to write page out, page is locked */ | |
426 | PAGE_KEEP, | |
427 | /* move page to the active list, page is locked */ | |
428 | PAGE_ACTIVATE, | |
429 | /* page has been sent to the disk successfully, page is unlocked */ | |
430 | PAGE_SUCCESS, | |
431 | /* page is clean and locked */ | |
432 | PAGE_CLEAN, | |
433 | } pageout_t; | |
434 | ||
435 | /* | |
436 | * pageout is called by shrink_page_list() for each dirty page. | |
437 | * Calls ->writepage(). | |
438 | */ | |
439 | static pageout_t pageout(struct page *page, struct address_space *mapping, | |
440 | struct scan_control *sc) | |
441 | { | |
442 | /* | |
443 | * If the page is dirty, only perform writeback if that write | |
444 | * will be non-blocking. To prevent this allocation from being | |
445 | * stalled by pagecache activity. But note that there may be | |
446 | * stalls if we need to run get_block(). We could test | |
447 | * PagePrivate for that. | |
448 | * | |
449 | * If this process is currently in __generic_file_aio_write() against | |
450 | * this page's queue, we can perform writeback even if that | |
451 | * will block. | |
452 | * | |
453 | * If the page is swapcache, write it back even if that would | |
454 | * block, for some throttling. This happens by accident, because | |
455 | * swap_backing_dev_info is bust: it doesn't reflect the | |
456 | * congestion state of the swapdevs. Easy to fix, if needed. | |
457 | */ | |
458 | if (!is_page_cache_freeable(page)) | |
459 | return PAGE_KEEP; | |
460 | if (!mapping) { | |
461 | /* | |
462 | * Some data journaling orphaned pages can have | |
463 | * page->mapping == NULL while being dirty with clean buffers. | |
464 | */ | |
465 | if (page_has_private(page)) { | |
466 | if (try_to_free_buffers(page)) { | |
467 | ClearPageDirty(page); | |
468 | printk("%s: orphaned page\n", __func__); | |
469 | return PAGE_CLEAN; | |
470 | } | |
471 | } | |
472 | return PAGE_KEEP; | |
473 | } | |
474 | if (mapping->a_ops->writepage == NULL) | |
475 | return PAGE_ACTIVATE; | |
476 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) | |
477 | return PAGE_KEEP; | |
478 | ||
479 | if (clear_page_dirty_for_io(page)) { | |
480 | int res; | |
481 | struct writeback_control wbc = { | |
482 | .sync_mode = WB_SYNC_NONE, | |
483 | .nr_to_write = SWAP_CLUSTER_MAX, | |
484 | .range_start = 0, | |
485 | .range_end = LLONG_MAX, | |
486 | .for_reclaim = 1, | |
487 | }; | |
488 | ||
489 | SetPageReclaim(page); | |
490 | res = mapping->a_ops->writepage(page, &wbc); | |
491 | if (res < 0) | |
492 | handle_write_error(mapping, page, res); | |
493 | if (res == AOP_WRITEPAGE_ACTIVATE) { | |
494 | ClearPageReclaim(page); | |
495 | return PAGE_ACTIVATE; | |
496 | } | |
497 | ||
498 | if (!PageWriteback(page)) { | |
499 | /* synchronous write or broken a_ops? */ | |
500 | ClearPageReclaim(page); | |
501 | } | |
502 | trace_mm_vmscan_writepage(page, | |
503 | trace_reclaim_flags(page, sc->reclaim_mode)); | |
504 | inc_zone_page_state(page, NR_VMSCAN_WRITE); | |
505 | return PAGE_SUCCESS; | |
506 | } | |
507 | ||
508 | return PAGE_CLEAN; | |
509 | } | |
510 | ||
511 | /* | |
512 | * Same as remove_mapping, but if the page is removed from the mapping, it | |
513 | * gets returned with a refcount of 0. | |
514 | */ | |
515 | static int __remove_mapping(struct address_space *mapping, struct page *page) | |
516 | { | |
517 | BUG_ON(!PageLocked(page)); | |
518 | BUG_ON(mapping != page_mapping(page)); | |
519 | ||
520 | spin_lock_irq(&mapping->tree_lock); | |
521 | /* | |
522 | * The non racy check for a busy page. | |
523 | * | |
524 | * Must be careful with the order of the tests. When someone has | |
525 | * a ref to the page, it may be possible that they dirty it then | |
526 | * drop the reference. So if PageDirty is tested before page_count | |
527 | * here, then the following race may occur: | |
528 | * | |
529 | * get_user_pages(&page); | |
530 | * [user mapping goes away] | |
531 | * write_to(page); | |
532 | * !PageDirty(page) [good] | |
533 | * SetPageDirty(page); | |
534 | * put_page(page); | |
535 | * !page_count(page) [good, discard it] | |
536 | * | |
537 | * [oops, our write_to data is lost] | |
538 | * | |
539 | * Reversing the order of the tests ensures such a situation cannot | |
540 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
541 | * load is not satisfied before that of page->_count. | |
542 | * | |
543 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
544 | * and thus under tree_lock, then this ordering is not required. | |
545 | */ | |
546 | if (!page_freeze_refs(page, 2)) | |
547 | goto cannot_free; | |
548 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ | |
549 | if (unlikely(PageDirty(page))) { | |
550 | page_unfreeze_refs(page, 2); | |
551 | goto cannot_free; | |
552 | } | |
553 | ||
554 | if (PageSwapCache(page)) { | |
555 | swp_entry_t swap = { .val = page_private(page) }; | |
556 | __delete_from_swap_cache(page); | |
557 | spin_unlock_irq(&mapping->tree_lock); | |
558 | swapcache_free(swap, page); | |
559 | } else { | |
560 | void (*freepage)(struct page *); | |
561 | ||
562 | freepage = mapping->a_ops->freepage; | |
563 | ||
564 | __delete_from_page_cache(page); | |
565 | spin_unlock_irq(&mapping->tree_lock); | |
566 | mem_cgroup_uncharge_cache_page(page); | |
567 | ||
568 | if (freepage != NULL) | |
569 | freepage(page); | |
570 | } | |
571 | ||
572 | return 1; | |
573 | ||
574 | cannot_free: | |
575 | spin_unlock_irq(&mapping->tree_lock); | |
576 | return 0; | |
577 | } | |
578 | ||
579 | /* | |
580 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
581 | * someone else has a ref on the page, abort and return 0. If it was | |
582 | * successfully detached, return 1. Assumes the caller has a single ref on | |
583 | * this page. | |
584 | */ | |
585 | int remove_mapping(struct address_space *mapping, struct page *page) | |
586 | { | |
587 | if (__remove_mapping(mapping, page)) { | |
588 | /* | |
589 | * Unfreezing the refcount with 1 rather than 2 effectively | |
590 | * drops the pagecache ref for us without requiring another | |
591 | * atomic operation. | |
592 | */ | |
593 | page_unfreeze_refs(page, 1); | |
594 | return 1; | |
595 | } | |
596 | return 0; | |
597 | } | |
598 | ||
599 | /** | |
600 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
601 | * @page: page to be put back to appropriate lru list | |
602 | * | |
603 | * Add previously isolated @page to appropriate LRU list. | |
604 | * Page may still be unevictable for other reasons. | |
605 | * | |
606 | * lru_lock must not be held, interrupts must be enabled. | |
607 | */ | |
608 | void putback_lru_page(struct page *page) | |
609 | { | |
610 | int lru; | |
611 | int active = !!TestClearPageActive(page); | |
612 | int was_unevictable = PageUnevictable(page); | |
613 | ||
614 | VM_BUG_ON(PageLRU(page)); | |
615 | ||
616 | redo: | |
617 | ClearPageUnevictable(page); | |
618 | ||
619 | if (page_evictable(page, NULL)) { | |
620 | /* | |
621 | * For evictable pages, we can use the cache. | |
622 | * In event of a race, worst case is we end up with an | |
623 | * unevictable page on [in]active list. | |
624 | * We know how to handle that. | |
625 | */ | |
626 | lru = active + page_lru_base_type(page); | |
627 | lru_cache_add_lru(page, lru); | |
628 | } else { | |
629 | /* | |
630 | * Put unevictable pages directly on zone's unevictable | |
631 | * list. | |
632 | */ | |
633 | lru = LRU_UNEVICTABLE; | |
634 | add_page_to_unevictable_list(page); | |
635 | /* | |
636 | * When racing with an mlock clearing (page is | |
637 | * unlocked), make sure that if the other thread does | |
638 | * not observe our setting of PG_lru and fails | |
639 | * isolation, we see PG_mlocked cleared below and move | |
640 | * the page back to the evictable list. | |
641 | * | |
642 | * The other side is TestClearPageMlocked(). | |
643 | */ | |
644 | smp_mb(); | |
645 | } | |
646 | ||
647 | /* | |
648 | * page's status can change while we move it among lru. If an evictable | |
649 | * page is on unevictable list, it never be freed. To avoid that, | |
650 | * check after we added it to the list, again. | |
651 | */ | |
652 | if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { | |
653 | if (!isolate_lru_page(page)) { | |
654 | put_page(page); | |
655 | goto redo; | |
656 | } | |
657 | /* This means someone else dropped this page from LRU | |
658 | * So, it will be freed or putback to LRU again. There is | |
659 | * nothing to do here. | |
660 | */ | |
661 | } | |
662 | ||
663 | if (was_unevictable && lru != LRU_UNEVICTABLE) | |
664 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
665 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) | |
666 | count_vm_event(UNEVICTABLE_PGCULLED); | |
667 | ||
668 | put_page(page); /* drop ref from isolate */ | |
669 | } | |
670 | ||
671 | enum page_references { | |
672 | PAGEREF_RECLAIM, | |
673 | PAGEREF_RECLAIM_CLEAN, | |
674 | PAGEREF_KEEP, | |
675 | PAGEREF_ACTIVATE, | |
676 | }; | |
677 | ||
678 | static enum page_references page_check_references(struct page *page, | |
679 | struct scan_control *sc) | |
680 | { | |
681 | int referenced_ptes, referenced_page; | |
682 | unsigned long vm_flags; | |
683 | ||
684 | referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags); | |
685 | referenced_page = TestClearPageReferenced(page); | |
686 | ||
687 | /* Lumpy reclaim - ignore references */ | |
688 | if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM) | |
689 | return PAGEREF_RECLAIM; | |
690 | ||
691 | /* | |
692 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
693 | * move the page to the unevictable list. | |
694 | */ | |
695 | if (vm_flags & VM_LOCKED) | |
696 | return PAGEREF_RECLAIM; | |
697 | ||
698 | if (referenced_ptes) { | |
699 | if (PageAnon(page)) | |
700 | return PAGEREF_ACTIVATE; | |
701 | /* | |
702 | * All mapped pages start out with page table | |
703 | * references from the instantiating fault, so we need | |
704 | * to look twice if a mapped file page is used more | |
705 | * than once. | |
706 | * | |
707 | * Mark it and spare it for another trip around the | |
708 | * inactive list. Another page table reference will | |
709 | * lead to its activation. | |
710 | * | |
711 | * Note: the mark is set for activated pages as well | |
712 | * so that recently deactivated but used pages are | |
713 | * quickly recovered. | |
714 | */ | |
715 | SetPageReferenced(page); | |
716 | ||
717 | if (referenced_page) | |
718 | return PAGEREF_ACTIVATE; | |
719 | ||
720 | return PAGEREF_KEEP; | |
721 | } | |
722 | ||
723 | /* Reclaim if clean, defer dirty pages to writeback */ | |
724 | if (referenced_page && !PageSwapBacked(page)) | |
725 | return PAGEREF_RECLAIM_CLEAN; | |
726 | ||
727 | return PAGEREF_RECLAIM; | |
728 | } | |
729 | ||
730 | static noinline_for_stack void free_page_list(struct list_head *free_pages) | |
731 | { | |
732 | struct pagevec freed_pvec; | |
733 | struct page *page, *tmp; | |
734 | ||
735 | pagevec_init(&freed_pvec, 1); | |
736 | ||
737 | list_for_each_entry_safe(page, tmp, free_pages, lru) { | |
738 | list_del(&page->lru); | |
739 | if (!pagevec_add(&freed_pvec, page)) { | |
740 | __pagevec_free(&freed_pvec); | |
741 | pagevec_reinit(&freed_pvec); | |
742 | } | |
743 | } | |
744 | ||
745 | pagevec_free(&freed_pvec); | |
746 | } | |
747 | ||
748 | /* | |
749 | * shrink_page_list() returns the number of reclaimed pages | |
750 | */ | |
751 | static unsigned long shrink_page_list(struct list_head *page_list, | |
752 | struct zone *zone, | |
753 | struct scan_control *sc, | |
754 | int priority) | |
755 | { | |
756 | LIST_HEAD(ret_pages); | |
757 | LIST_HEAD(free_pages); | |
758 | int pgactivate = 0; | |
759 | unsigned long nr_dirty = 0; | |
760 | unsigned long nr_congested = 0; | |
761 | unsigned long nr_reclaimed = 0; | |
762 | ||
763 | cond_resched(); | |
764 | ||
765 | while (!list_empty(page_list)) { | |
766 | enum page_references references; | |
767 | struct address_space *mapping; | |
768 | struct page *page; | |
769 | int may_enter_fs; | |
770 | ||
771 | cond_resched(); | |
772 | ||
773 | page = lru_to_page(page_list); | |
774 | list_del(&page->lru); | |
775 | ||
776 | if (!trylock_page(page)) | |
777 | goto keep; | |
778 | ||
779 | VM_BUG_ON(PageActive(page)); | |
780 | VM_BUG_ON(page_zone(page) != zone); | |
781 | ||
782 | sc->nr_scanned++; | |
783 | ||
784 | if (unlikely(!page_evictable(page, NULL))) | |
785 | goto cull_mlocked; | |
786 | ||
787 | if (!sc->may_unmap && page_mapped(page)) | |
788 | goto keep_locked; | |
789 | ||
790 | /* Double the slab pressure for mapped and swapcache pages */ | |
791 | if (page_mapped(page) || PageSwapCache(page)) | |
792 | sc->nr_scanned++; | |
793 | ||
794 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || | |
795 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
796 | ||
797 | if (PageWriteback(page)) { | |
798 | /* | |
799 | * Synchronous reclaim cannot queue pages for | |
800 | * writeback due to the possibility of stack overflow | |
801 | * but if it encounters a page under writeback, wait | |
802 | * for the IO to complete. | |
803 | */ | |
804 | if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) && | |
805 | may_enter_fs) | |
806 | wait_on_page_writeback(page); | |
807 | else { | |
808 | unlock_page(page); | |
809 | goto keep_lumpy; | |
810 | } | |
811 | } | |
812 | ||
813 | references = page_check_references(page, sc); | |
814 | switch (references) { | |
815 | case PAGEREF_ACTIVATE: | |
816 | goto activate_locked; | |
817 | case PAGEREF_KEEP: | |
818 | goto keep_locked; | |
819 | case PAGEREF_RECLAIM: | |
820 | case PAGEREF_RECLAIM_CLEAN: | |
821 | ; /* try to reclaim the page below */ | |
822 | } | |
823 | ||
824 | /* | |
825 | * Anonymous process memory has backing store? | |
826 | * Try to allocate it some swap space here. | |
827 | */ | |
828 | if (PageAnon(page) && !PageSwapCache(page)) { | |
829 | if (!(sc->gfp_mask & __GFP_IO)) | |
830 | goto keep_locked; | |
831 | if (!add_to_swap(page)) | |
832 | goto activate_locked; | |
833 | may_enter_fs = 1; | |
834 | } | |
835 | ||
836 | mapping = page_mapping(page); | |
837 | ||
838 | /* | |
839 | * The page is mapped into the page tables of one or more | |
840 | * processes. Try to unmap it here. | |
841 | */ | |
842 | if (page_mapped(page) && mapping) { | |
843 | switch (try_to_unmap(page, TTU_UNMAP)) { | |
844 | case SWAP_FAIL: | |
845 | goto activate_locked; | |
846 | case SWAP_AGAIN: | |
847 | goto keep_locked; | |
848 | case SWAP_MLOCK: | |
849 | goto cull_mlocked; | |
850 | case SWAP_SUCCESS: | |
851 | ; /* try to free the page below */ | |
852 | } | |
853 | } | |
854 | ||
855 | if (PageDirty(page)) { | |
856 | nr_dirty++; | |
857 | ||
858 | /* | |
859 | * Only kswapd can writeback filesystem pages to | |
860 | * avoid risk of stack overflow but do not writeback | |
861 | * unless under significant pressure. | |
862 | */ | |
863 | if (page_is_file_cache(page) && | |
864 | (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) { | |
865 | inc_zone_page_state(page, NR_VMSCAN_WRITE_SKIP); | |
866 | goto keep_locked; | |
867 | } | |
868 | ||
869 | if (references == PAGEREF_RECLAIM_CLEAN) | |
870 | goto keep_locked; | |
871 | if (!may_enter_fs) | |
872 | goto keep_locked; | |
873 | if (!sc->may_writepage) | |
874 | goto keep_locked; | |
875 | ||
876 | /* Page is dirty, try to write it out here */ | |
877 | switch (pageout(page, mapping, sc)) { | |
878 | case PAGE_KEEP: | |
879 | nr_congested++; | |
880 | goto keep_locked; | |
881 | case PAGE_ACTIVATE: | |
882 | goto activate_locked; | |
883 | case PAGE_SUCCESS: | |
884 | if (PageWriteback(page)) | |
885 | goto keep_lumpy; | |
886 | if (PageDirty(page)) | |
887 | goto keep; | |
888 | ||
889 | /* | |
890 | * A synchronous write - probably a ramdisk. Go | |
891 | * ahead and try to reclaim the page. | |
892 | */ | |
893 | if (!trylock_page(page)) | |
894 | goto keep; | |
895 | if (PageDirty(page) || PageWriteback(page)) | |
896 | goto keep_locked; | |
897 | mapping = page_mapping(page); | |
898 | case PAGE_CLEAN: | |
899 | ; /* try to free the page below */ | |
900 | } | |
901 | } | |
902 | ||
903 | /* | |
904 | * If the page has buffers, try to free the buffer mappings | |
905 | * associated with this page. If we succeed we try to free | |
906 | * the page as well. | |
907 | * | |
908 | * We do this even if the page is PageDirty(). | |
909 | * try_to_release_page() does not perform I/O, but it is | |
910 | * possible for a page to have PageDirty set, but it is actually | |
911 | * clean (all its buffers are clean). This happens if the | |
912 | * buffers were written out directly, with submit_bh(). ext3 | |
913 | * will do this, as well as the blockdev mapping. | |
914 | * try_to_release_page() will discover that cleanness and will | |
915 | * drop the buffers and mark the page clean - it can be freed. | |
916 | * | |
917 | * Rarely, pages can have buffers and no ->mapping. These are | |
918 | * the pages which were not successfully invalidated in | |
919 | * truncate_complete_page(). We try to drop those buffers here | |
920 | * and if that worked, and the page is no longer mapped into | |
921 | * process address space (page_count == 1) it can be freed. | |
922 | * Otherwise, leave the page on the LRU so it is swappable. | |
923 | */ | |
924 | if (page_has_private(page)) { | |
925 | if (!try_to_release_page(page, sc->gfp_mask)) | |
926 | goto activate_locked; | |
927 | if (!mapping && page_count(page) == 1) { | |
928 | unlock_page(page); | |
929 | if (put_page_testzero(page)) | |
930 | goto free_it; | |
931 | else { | |
932 | /* | |
933 | * rare race with speculative reference. | |
934 | * the speculative reference will free | |
935 | * this page shortly, so we may | |
936 | * increment nr_reclaimed here (and | |
937 | * leave it off the LRU). | |
938 | */ | |
939 | nr_reclaimed++; | |
940 | continue; | |
941 | } | |
942 | } | |
943 | } | |
944 | ||
945 | if (!mapping || !__remove_mapping(mapping, page)) | |
946 | goto keep_locked; | |
947 | ||
948 | /* | |
949 | * At this point, we have no other references and there is | |
950 | * no way to pick any more up (removed from LRU, removed | |
951 | * from pagecache). Can use non-atomic bitops now (and | |
952 | * we obviously don't have to worry about waking up a process | |
953 | * waiting on the page lock, because there are no references. | |
954 | */ | |
955 | __clear_page_locked(page); | |
956 | free_it: | |
957 | nr_reclaimed++; | |
958 | ||
959 | /* | |
960 | * Is there need to periodically free_page_list? It would | |
961 | * appear not as the counts should be low | |
962 | */ | |
963 | list_add(&page->lru, &free_pages); | |
964 | continue; | |
965 | ||
966 | cull_mlocked: | |
967 | if (PageSwapCache(page)) | |
968 | try_to_free_swap(page); | |
969 | unlock_page(page); | |
970 | putback_lru_page(page); | |
971 | reset_reclaim_mode(sc); | |
972 | continue; | |
973 | ||
974 | activate_locked: | |
975 | /* Not a candidate for swapping, so reclaim swap space. */ | |
976 | if (PageSwapCache(page) && vm_swap_full()) | |
977 | try_to_free_swap(page); | |
978 | VM_BUG_ON(PageActive(page)); | |
979 | SetPageActive(page); | |
980 | pgactivate++; | |
981 | keep_locked: | |
982 | unlock_page(page); | |
983 | keep: | |
984 | reset_reclaim_mode(sc); | |
985 | keep_lumpy: | |
986 | list_add(&page->lru, &ret_pages); | |
987 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); | |
988 | } | |
989 | ||
990 | /* | |
991 | * Tag a zone as congested if all the dirty pages encountered were | |
992 | * backed by a congested BDI. In this case, reclaimers should just | |
993 | * back off and wait for congestion to clear because further reclaim | |
994 | * will encounter the same problem | |
995 | */ | |
996 | if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc)) | |
997 | zone_set_flag(zone, ZONE_CONGESTED); | |
998 | ||
999 | free_page_list(&free_pages); | |
1000 | ||
1001 | list_splice(&ret_pages, page_list); | |
1002 | count_vm_events(PGACTIVATE, pgactivate); | |
1003 | return nr_reclaimed; | |
1004 | } | |
1005 | ||
1006 | /* | |
1007 | * Attempt to remove the specified page from its LRU. Only take this page | |
1008 | * if it is of the appropriate PageActive status. Pages which are being | |
1009 | * freed elsewhere are also ignored. | |
1010 | * | |
1011 | * page: page to consider | |
1012 | * mode: one of the LRU isolation modes defined above | |
1013 | * | |
1014 | * returns 0 on success, -ve errno on failure. | |
1015 | */ | |
1016 | int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file) | |
1017 | { | |
1018 | bool all_lru_mode; | |
1019 | int ret = -EINVAL; | |
1020 | ||
1021 | /* Only take pages on the LRU. */ | |
1022 | if (!PageLRU(page)) | |
1023 | return ret; | |
1024 | ||
1025 | all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) == | |
1026 | (ISOLATE_ACTIVE|ISOLATE_INACTIVE); | |
1027 | ||
1028 | /* | |
1029 | * When checking the active state, we need to be sure we are | |
1030 | * dealing with comparible boolean values. Take the logical not | |
1031 | * of each. | |
1032 | */ | |
1033 | if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE)) | |
1034 | return ret; | |
1035 | ||
1036 | if (!all_lru_mode && !!page_is_file_cache(page) != file) | |
1037 | return ret; | |
1038 | ||
1039 | /* | |
1040 | * When this function is being called for lumpy reclaim, we | |
1041 | * initially look into all LRU pages, active, inactive and | |
1042 | * unevictable; only give shrink_page_list evictable pages. | |
1043 | */ | |
1044 | if (PageUnevictable(page)) | |
1045 | return ret; | |
1046 | ||
1047 | ret = -EBUSY; | |
1048 | ||
1049 | if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page))) | |
1050 | return ret; | |
1051 | ||
1052 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) | |
1053 | return ret; | |
1054 | ||
1055 | if (likely(get_page_unless_zero(page))) { | |
1056 | /* | |
1057 | * Be careful not to clear PageLRU until after we're | |
1058 | * sure the page is not being freed elsewhere -- the | |
1059 | * page release code relies on it. | |
1060 | */ | |
1061 | ClearPageLRU(page); | |
1062 | ret = 0; | |
1063 | } | |
1064 | ||
1065 | return ret; | |
1066 | } | |
1067 | ||
1068 | /* | |
1069 | * zone->lru_lock is heavily contended. Some of the functions that | |
1070 | * shrink the lists perform better by taking out a batch of pages | |
1071 | * and working on them outside the LRU lock. | |
1072 | * | |
1073 | * For pagecache intensive workloads, this function is the hottest | |
1074 | * spot in the kernel (apart from copy_*_user functions). | |
1075 | * | |
1076 | * Appropriate locks must be held before calling this function. | |
1077 | * | |
1078 | * @nr_to_scan: The number of pages to look through on the list. | |
1079 | * @src: The LRU list to pull pages off. | |
1080 | * @dst: The temp list to put pages on to. | |
1081 | * @scanned: The number of pages that were scanned. | |
1082 | * @order: The caller's attempted allocation order | |
1083 | * @mode: One of the LRU isolation modes | |
1084 | * @file: True [1] if isolating file [!anon] pages | |
1085 | * | |
1086 | * returns how many pages were moved onto *@dst. | |
1087 | */ | |
1088 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, | |
1089 | struct list_head *src, struct list_head *dst, | |
1090 | unsigned long *scanned, int order, isolate_mode_t mode, | |
1091 | int file) | |
1092 | { | |
1093 | unsigned long nr_taken = 0; | |
1094 | unsigned long nr_lumpy_taken = 0; | |
1095 | unsigned long nr_lumpy_dirty = 0; | |
1096 | unsigned long nr_lumpy_failed = 0; | |
1097 | unsigned long scan; | |
1098 | ||
1099 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { | |
1100 | struct page *page; | |
1101 | unsigned long pfn; | |
1102 | unsigned long end_pfn; | |
1103 | unsigned long page_pfn; | |
1104 | int zone_id; | |
1105 | ||
1106 | page = lru_to_page(src); | |
1107 | prefetchw_prev_lru_page(page, src, flags); | |
1108 | ||
1109 | VM_BUG_ON(!PageLRU(page)); | |
1110 | ||
1111 | switch (__isolate_lru_page(page, mode, file)) { | |
1112 | case 0: | |
1113 | list_move(&page->lru, dst); | |
1114 | mem_cgroup_del_lru(page); | |
1115 | nr_taken += hpage_nr_pages(page); | |
1116 | break; | |
1117 | ||
1118 | case -EBUSY: | |
1119 | /* else it is being freed elsewhere */ | |
1120 | list_move(&page->lru, src); | |
1121 | mem_cgroup_rotate_lru_list(page, page_lru(page)); | |
1122 | continue; | |
1123 | ||
1124 | default: | |
1125 | BUG(); | |
1126 | } | |
1127 | ||
1128 | if (!order) | |
1129 | continue; | |
1130 | ||
1131 | /* | |
1132 | * Attempt to take all pages in the order aligned region | |
1133 | * surrounding the tag page. Only take those pages of | |
1134 | * the same active state as that tag page. We may safely | |
1135 | * round the target page pfn down to the requested order | |
1136 | * as the mem_map is guaranteed valid out to MAX_ORDER, | |
1137 | * where that page is in a different zone we will detect | |
1138 | * it from its zone id and abort this block scan. | |
1139 | */ | |
1140 | zone_id = page_zone_id(page); | |
1141 | page_pfn = page_to_pfn(page); | |
1142 | pfn = page_pfn & ~((1 << order) - 1); | |
1143 | end_pfn = pfn + (1 << order); | |
1144 | for (; pfn < end_pfn; pfn++) { | |
1145 | struct page *cursor_page; | |
1146 | ||
1147 | /* The target page is in the block, ignore it. */ | |
1148 | if (unlikely(pfn == page_pfn)) | |
1149 | continue; | |
1150 | ||
1151 | /* Avoid holes within the zone. */ | |
1152 | if (unlikely(!pfn_valid_within(pfn))) | |
1153 | break; | |
1154 | ||
1155 | cursor_page = pfn_to_page(pfn); | |
1156 | ||
1157 | /* Check that we have not crossed a zone boundary. */ | |
1158 | if (unlikely(page_zone_id(cursor_page) != zone_id)) | |
1159 | break; | |
1160 | ||
1161 | /* | |
1162 | * If we don't have enough swap space, reclaiming of | |
1163 | * anon page which don't already have a swap slot is | |
1164 | * pointless. | |
1165 | */ | |
1166 | if (nr_swap_pages <= 0 && PageAnon(cursor_page) && | |
1167 | !PageSwapCache(cursor_page)) | |
1168 | break; | |
1169 | ||
1170 | if (__isolate_lru_page(cursor_page, mode, file) == 0) { | |
1171 | list_move(&cursor_page->lru, dst); | |
1172 | mem_cgroup_del_lru(cursor_page); | |
1173 | nr_taken += hpage_nr_pages(page); | |
1174 | nr_lumpy_taken++; | |
1175 | if (PageDirty(cursor_page)) | |
1176 | nr_lumpy_dirty++; | |
1177 | scan++; | |
1178 | } else { | |
1179 | /* | |
1180 | * Check if the page is freed already. | |
1181 | * | |
1182 | * We can't use page_count() as that | |
1183 | * requires compound_head and we don't | |
1184 | * have a pin on the page here. If a | |
1185 | * page is tail, we may or may not | |
1186 | * have isolated the head, so assume | |
1187 | * it's not free, it'd be tricky to | |
1188 | * track the head status without a | |
1189 | * page pin. | |
1190 | */ | |
1191 | if (!PageTail(cursor_page) && | |
1192 | !atomic_read(&cursor_page->_count)) | |
1193 | continue; | |
1194 | break; | |
1195 | } | |
1196 | } | |
1197 | ||
1198 | /* If we break out of the loop above, lumpy reclaim failed */ | |
1199 | if (pfn < end_pfn) | |
1200 | nr_lumpy_failed++; | |
1201 | } | |
1202 | ||
1203 | *scanned = scan; | |
1204 | ||
1205 | trace_mm_vmscan_lru_isolate(order, | |
1206 | nr_to_scan, scan, | |
1207 | nr_taken, | |
1208 | nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed, | |
1209 | mode); | |
1210 | return nr_taken; | |
1211 | } | |
1212 | ||
1213 | static unsigned long isolate_pages_global(unsigned long nr, | |
1214 | struct list_head *dst, | |
1215 | unsigned long *scanned, int order, | |
1216 | isolate_mode_t mode, | |
1217 | struct zone *z, int active, int file) | |
1218 | { | |
1219 | int lru = LRU_BASE; | |
1220 | if (active) | |
1221 | lru += LRU_ACTIVE; | |
1222 | if (file) | |
1223 | lru += LRU_FILE; | |
1224 | return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order, | |
1225 | mode, file); | |
1226 | } | |
1227 | ||
1228 | /* | |
1229 | * clear_active_flags() is a helper for shrink_active_list(), clearing | |
1230 | * any active bits from the pages in the list. | |
1231 | */ | |
1232 | static unsigned long clear_active_flags(struct list_head *page_list, | |
1233 | unsigned int *count) | |
1234 | { | |
1235 | int nr_active = 0; | |
1236 | int lru; | |
1237 | struct page *page; | |
1238 | ||
1239 | list_for_each_entry(page, page_list, lru) { | |
1240 | int numpages = hpage_nr_pages(page); | |
1241 | lru = page_lru_base_type(page); | |
1242 | if (PageActive(page)) { | |
1243 | lru += LRU_ACTIVE; | |
1244 | ClearPageActive(page); | |
1245 | nr_active += numpages; | |
1246 | } | |
1247 | if (count) | |
1248 | count[lru] += numpages; | |
1249 | } | |
1250 | ||
1251 | return nr_active; | |
1252 | } | |
1253 | ||
1254 | /** | |
1255 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1256 | * @page: page to isolate from its LRU list | |
1257 | * | |
1258 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1259 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1260 | * | |
1261 | * Returns 0 if the page was removed from an LRU list. | |
1262 | * Returns -EBUSY if the page was not on an LRU list. | |
1263 | * | |
1264 | * The returned page will have PageLRU() cleared. If it was found on | |
1265 | * the active list, it will have PageActive set. If it was found on | |
1266 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1267 | * may need to be cleared by the caller before letting the page go. | |
1268 | * | |
1269 | * The vmstat statistic corresponding to the list on which the page was | |
1270 | * found will be decremented. | |
1271 | * | |
1272 | * Restrictions: | |
1273 | * (1) Must be called with an elevated refcount on the page. This is a | |
1274 | * fundamentnal difference from isolate_lru_pages (which is called | |
1275 | * without a stable reference). | |
1276 | * (2) the lru_lock must not be held. | |
1277 | * (3) interrupts must be enabled. | |
1278 | */ | |
1279 | int isolate_lru_page(struct page *page) | |
1280 | { | |
1281 | int ret = -EBUSY; | |
1282 | ||
1283 | VM_BUG_ON(!page_count(page)); | |
1284 | ||
1285 | if (PageLRU(page)) { | |
1286 | struct zone *zone = page_zone(page); | |
1287 | ||
1288 | spin_lock_irq(&zone->lru_lock); | |
1289 | if (PageLRU(page)) { | |
1290 | int lru = page_lru(page); | |
1291 | ret = 0; | |
1292 | get_page(page); | |
1293 | ClearPageLRU(page); | |
1294 | ||
1295 | del_page_from_lru_list(zone, page, lru); | |
1296 | } | |
1297 | spin_unlock_irq(&zone->lru_lock); | |
1298 | } | |
1299 | return ret; | |
1300 | } | |
1301 | ||
1302 | /* | |
1303 | * Are there way too many processes in the direct reclaim path already? | |
1304 | */ | |
1305 | static int too_many_isolated(struct zone *zone, int file, | |
1306 | struct scan_control *sc) | |
1307 | { | |
1308 | unsigned long inactive, isolated; | |
1309 | ||
1310 | if (current_is_kswapd()) | |
1311 | return 0; | |
1312 | ||
1313 | if (!scanning_global_lru(sc)) | |
1314 | return 0; | |
1315 | ||
1316 | if (file) { | |
1317 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1318 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1319 | } else { | |
1320 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1321 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1322 | } | |
1323 | ||
1324 | return isolated > inactive; | |
1325 | } | |
1326 | ||
1327 | /* | |
1328 | * TODO: Try merging with migrations version of putback_lru_pages | |
1329 | */ | |
1330 | static noinline_for_stack void | |
1331 | putback_lru_pages(struct zone *zone, struct scan_control *sc, | |
1332 | unsigned long nr_anon, unsigned long nr_file, | |
1333 | struct list_head *page_list) | |
1334 | { | |
1335 | struct page *page; | |
1336 | struct pagevec pvec; | |
1337 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); | |
1338 | ||
1339 | pagevec_init(&pvec, 1); | |
1340 | ||
1341 | /* | |
1342 | * Put back any unfreeable pages. | |
1343 | */ | |
1344 | spin_lock(&zone->lru_lock); | |
1345 | while (!list_empty(page_list)) { | |
1346 | int lru; | |
1347 | page = lru_to_page(page_list); | |
1348 | VM_BUG_ON(PageLRU(page)); | |
1349 | list_del(&page->lru); | |
1350 | if (unlikely(!page_evictable(page, NULL))) { | |
1351 | spin_unlock_irq(&zone->lru_lock); | |
1352 | putback_lru_page(page); | |
1353 | spin_lock_irq(&zone->lru_lock); | |
1354 | continue; | |
1355 | } | |
1356 | SetPageLRU(page); | |
1357 | lru = page_lru(page); | |
1358 | add_page_to_lru_list(zone, page, lru); | |
1359 | if (is_active_lru(lru)) { | |
1360 | int file = is_file_lru(lru); | |
1361 | int numpages = hpage_nr_pages(page); | |
1362 | reclaim_stat->recent_rotated[file] += numpages; | |
1363 | } | |
1364 | if (!pagevec_add(&pvec, page)) { | |
1365 | spin_unlock_irq(&zone->lru_lock); | |
1366 | __pagevec_release(&pvec); | |
1367 | spin_lock_irq(&zone->lru_lock); | |
1368 | } | |
1369 | } | |
1370 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon); | |
1371 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file); | |
1372 | ||
1373 | spin_unlock_irq(&zone->lru_lock); | |
1374 | pagevec_release(&pvec); | |
1375 | } | |
1376 | ||
1377 | static noinline_for_stack void update_isolated_counts(struct zone *zone, | |
1378 | struct scan_control *sc, | |
1379 | unsigned long *nr_anon, | |
1380 | unsigned long *nr_file, | |
1381 | struct list_head *isolated_list) | |
1382 | { | |
1383 | unsigned long nr_active; | |
1384 | unsigned int count[NR_LRU_LISTS] = { 0, }; | |
1385 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); | |
1386 | ||
1387 | nr_active = clear_active_flags(isolated_list, count); | |
1388 | __count_vm_events(PGDEACTIVATE, nr_active); | |
1389 | ||
1390 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, | |
1391 | -count[LRU_ACTIVE_FILE]); | |
1392 | __mod_zone_page_state(zone, NR_INACTIVE_FILE, | |
1393 | -count[LRU_INACTIVE_FILE]); | |
1394 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, | |
1395 | -count[LRU_ACTIVE_ANON]); | |
1396 | __mod_zone_page_state(zone, NR_INACTIVE_ANON, | |
1397 | -count[LRU_INACTIVE_ANON]); | |
1398 | ||
1399 | *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON]; | |
1400 | *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE]; | |
1401 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon); | |
1402 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file); | |
1403 | ||
1404 | reclaim_stat->recent_scanned[0] += *nr_anon; | |
1405 | reclaim_stat->recent_scanned[1] += *nr_file; | |
1406 | } | |
1407 | ||
1408 | /* | |
1409 | * Returns true if a direct reclaim should wait on pages under writeback. | |
1410 | * | |
1411 | * If we are direct reclaiming for contiguous pages and we do not reclaim | |
1412 | * everything in the list, try again and wait for writeback IO to complete. | |
1413 | * This will stall high-order allocations noticeably. Only do that when really | |
1414 | * need to free the pages under high memory pressure. | |
1415 | */ | |
1416 | static inline bool should_reclaim_stall(unsigned long nr_taken, | |
1417 | unsigned long nr_freed, | |
1418 | int priority, | |
1419 | struct scan_control *sc) | |
1420 | { | |
1421 | int lumpy_stall_priority; | |
1422 | ||
1423 | /* kswapd should not stall on sync IO */ | |
1424 | if (current_is_kswapd()) | |
1425 | return false; | |
1426 | ||
1427 | /* Only stall on lumpy reclaim */ | |
1428 | if (sc->reclaim_mode & RECLAIM_MODE_SINGLE) | |
1429 | return false; | |
1430 | ||
1431 | /* If we have reclaimed everything on the isolated list, no stall */ | |
1432 | if (nr_freed == nr_taken) | |
1433 | return false; | |
1434 | ||
1435 | /* | |
1436 | * For high-order allocations, there are two stall thresholds. | |
1437 | * High-cost allocations stall immediately where as lower | |
1438 | * order allocations such as stacks require the scanning | |
1439 | * priority to be much higher before stalling. | |
1440 | */ | |
1441 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) | |
1442 | lumpy_stall_priority = DEF_PRIORITY; | |
1443 | else | |
1444 | lumpy_stall_priority = DEF_PRIORITY / 3; | |
1445 | ||
1446 | return priority <= lumpy_stall_priority; | |
1447 | } | |
1448 | ||
1449 | /* | |
1450 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number | |
1451 | * of reclaimed pages | |
1452 | */ | |
1453 | static noinline_for_stack unsigned long | |
1454 | shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone, | |
1455 | struct scan_control *sc, int priority, int file) | |
1456 | { | |
1457 | LIST_HEAD(page_list); | |
1458 | unsigned long nr_scanned; | |
1459 | unsigned long nr_reclaimed = 0; | |
1460 | unsigned long nr_taken; | |
1461 | unsigned long nr_anon; | |
1462 | unsigned long nr_file; | |
1463 | isolate_mode_t reclaim_mode = ISOLATE_INACTIVE; | |
1464 | ||
1465 | while (unlikely(too_many_isolated(zone, file, sc))) { | |
1466 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
1467 | ||
1468 | /* We are about to die and free our memory. Return now. */ | |
1469 | if (fatal_signal_pending(current)) | |
1470 | return SWAP_CLUSTER_MAX; | |
1471 | } | |
1472 | ||
1473 | set_reclaim_mode(priority, sc, false); | |
1474 | if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM) | |
1475 | reclaim_mode |= ISOLATE_ACTIVE; | |
1476 | ||
1477 | lru_add_drain(); | |
1478 | ||
1479 | if (!sc->may_unmap) | |
1480 | reclaim_mode |= ISOLATE_UNMAPPED; | |
1481 | if (!sc->may_writepage) | |
1482 | reclaim_mode |= ISOLATE_CLEAN; | |
1483 | ||
1484 | spin_lock_irq(&zone->lru_lock); | |
1485 | ||
1486 | if (scanning_global_lru(sc)) { | |
1487 | nr_taken = isolate_pages_global(nr_to_scan, &page_list, | |
1488 | &nr_scanned, sc->order, reclaim_mode, zone, 0, file); | |
1489 | zone->pages_scanned += nr_scanned; | |
1490 | if (current_is_kswapd()) | |
1491 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, | |
1492 | nr_scanned); | |
1493 | else | |
1494 | __count_zone_vm_events(PGSCAN_DIRECT, zone, | |
1495 | nr_scanned); | |
1496 | } else { | |
1497 | nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list, | |
1498 | &nr_scanned, sc->order, reclaim_mode, zone, | |
1499 | sc->mem_cgroup, 0, file); | |
1500 | /* | |
1501 | * mem_cgroup_isolate_pages() keeps track of | |
1502 | * scanned pages on its own. | |
1503 | */ | |
1504 | } | |
1505 | ||
1506 | if (nr_taken == 0) { | |
1507 | spin_unlock_irq(&zone->lru_lock); | |
1508 | return 0; | |
1509 | } | |
1510 | ||
1511 | update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list); | |
1512 | ||
1513 | spin_unlock_irq(&zone->lru_lock); | |
1514 | ||
1515 | nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority); | |
1516 | ||
1517 | /* Check if we should syncronously wait for writeback */ | |
1518 | if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) { | |
1519 | set_reclaim_mode(priority, sc, true); | |
1520 | nr_reclaimed += shrink_page_list(&page_list, zone, sc, priority); | |
1521 | } | |
1522 | ||
1523 | local_irq_disable(); | |
1524 | if (current_is_kswapd()) | |
1525 | __count_vm_events(KSWAPD_STEAL, nr_reclaimed); | |
1526 | __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed); | |
1527 | ||
1528 | putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list); | |
1529 | ||
1530 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, | |
1531 | zone_idx(zone), | |
1532 | nr_scanned, nr_reclaimed, | |
1533 | priority, | |
1534 | trace_shrink_flags(file, sc->reclaim_mode)); | |
1535 | return nr_reclaimed; | |
1536 | } | |
1537 | ||
1538 | /* | |
1539 | * This moves pages from the active list to the inactive list. | |
1540 | * | |
1541 | * We move them the other way if the page is referenced by one or more | |
1542 | * processes, from rmap. | |
1543 | * | |
1544 | * If the pages are mostly unmapped, the processing is fast and it is | |
1545 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1546 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1547 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1548 | * this, so instead we remove the pages from the LRU while processing them. | |
1549 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1550 | * nobody will play with that bit on a non-LRU page. | |
1551 | * | |
1552 | * The downside is that we have to touch page->_count against each page. | |
1553 | * But we had to alter page->flags anyway. | |
1554 | */ | |
1555 | ||
1556 | static void move_active_pages_to_lru(struct zone *zone, | |
1557 | struct list_head *list, | |
1558 | enum lru_list lru) | |
1559 | { | |
1560 | unsigned long pgmoved = 0; | |
1561 | struct pagevec pvec; | |
1562 | struct page *page; | |
1563 | ||
1564 | pagevec_init(&pvec, 1); | |
1565 | ||
1566 | while (!list_empty(list)) { | |
1567 | page = lru_to_page(list); | |
1568 | ||
1569 | VM_BUG_ON(PageLRU(page)); | |
1570 | SetPageLRU(page); | |
1571 | ||
1572 | list_move(&page->lru, &zone->lru[lru].list); | |
1573 | mem_cgroup_add_lru_list(page, lru); | |
1574 | pgmoved += hpage_nr_pages(page); | |
1575 | ||
1576 | if (!pagevec_add(&pvec, page) || list_empty(list)) { | |
1577 | spin_unlock_irq(&zone->lru_lock); | |
1578 | if (buffer_heads_over_limit) | |
1579 | pagevec_strip(&pvec); | |
1580 | __pagevec_release(&pvec); | |
1581 | spin_lock_irq(&zone->lru_lock); | |
1582 | } | |
1583 | } | |
1584 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1585 | if (!is_active_lru(lru)) | |
1586 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1587 | } | |
1588 | ||
1589 | static void shrink_active_list(unsigned long nr_pages, struct zone *zone, | |
1590 | struct scan_control *sc, int priority, int file) | |
1591 | { | |
1592 | unsigned long nr_taken; | |
1593 | unsigned long pgscanned; | |
1594 | unsigned long vm_flags; | |
1595 | LIST_HEAD(l_hold); /* The pages which were snipped off */ | |
1596 | LIST_HEAD(l_active); | |
1597 | LIST_HEAD(l_inactive); | |
1598 | struct page *page; | |
1599 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); | |
1600 | unsigned long nr_rotated = 0; | |
1601 | isolate_mode_t reclaim_mode = ISOLATE_ACTIVE; | |
1602 | ||
1603 | lru_add_drain(); | |
1604 | ||
1605 | if (!sc->may_unmap) | |
1606 | reclaim_mode |= ISOLATE_UNMAPPED; | |
1607 | if (!sc->may_writepage) | |
1608 | reclaim_mode |= ISOLATE_CLEAN; | |
1609 | ||
1610 | spin_lock_irq(&zone->lru_lock); | |
1611 | if (scanning_global_lru(sc)) { | |
1612 | nr_taken = isolate_pages_global(nr_pages, &l_hold, | |
1613 | &pgscanned, sc->order, | |
1614 | reclaim_mode, zone, | |
1615 | 1, file); | |
1616 | zone->pages_scanned += pgscanned; | |
1617 | } else { | |
1618 | nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold, | |
1619 | &pgscanned, sc->order, | |
1620 | reclaim_mode, zone, | |
1621 | sc->mem_cgroup, 1, file); | |
1622 | /* | |
1623 | * mem_cgroup_isolate_pages() keeps track of | |
1624 | * scanned pages on its own. | |
1625 | */ | |
1626 | } | |
1627 | ||
1628 | reclaim_stat->recent_scanned[file] += nr_taken; | |
1629 | ||
1630 | __count_zone_vm_events(PGREFILL, zone, pgscanned); | |
1631 | if (file) | |
1632 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken); | |
1633 | else | |
1634 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken); | |
1635 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1636 | spin_unlock_irq(&zone->lru_lock); | |
1637 | ||
1638 | while (!list_empty(&l_hold)) { | |
1639 | cond_resched(); | |
1640 | page = lru_to_page(&l_hold); | |
1641 | list_del(&page->lru); | |
1642 | ||
1643 | if (unlikely(!page_evictable(page, NULL))) { | |
1644 | putback_lru_page(page); | |
1645 | continue; | |
1646 | } | |
1647 | ||
1648 | if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) { | |
1649 | nr_rotated += hpage_nr_pages(page); | |
1650 | /* | |
1651 | * Identify referenced, file-backed active pages and | |
1652 | * give them one more trip around the active list. So | |
1653 | * that executable code get better chances to stay in | |
1654 | * memory under moderate memory pressure. Anon pages | |
1655 | * are not likely to be evicted by use-once streaming | |
1656 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1657 | * so we ignore them here. | |
1658 | */ | |
1659 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { | |
1660 | list_add(&page->lru, &l_active); | |
1661 | continue; | |
1662 | } | |
1663 | } | |
1664 | ||
1665 | ClearPageActive(page); /* we are de-activating */ | |
1666 | list_add(&page->lru, &l_inactive); | |
1667 | } | |
1668 | ||
1669 | /* | |
1670 | * Move pages back to the lru list. | |
1671 | */ | |
1672 | spin_lock_irq(&zone->lru_lock); | |
1673 | /* | |
1674 | * Count referenced pages from currently used mappings as rotated, | |
1675 | * even though only some of them are actually re-activated. This | |
1676 | * helps balance scan pressure between file and anonymous pages in | |
1677 | * get_scan_ratio. | |
1678 | */ | |
1679 | reclaim_stat->recent_rotated[file] += nr_rotated; | |
1680 | ||
1681 | move_active_pages_to_lru(zone, &l_active, | |
1682 | LRU_ACTIVE + file * LRU_FILE); | |
1683 | move_active_pages_to_lru(zone, &l_inactive, | |
1684 | LRU_BASE + file * LRU_FILE); | |
1685 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); | |
1686 | spin_unlock_irq(&zone->lru_lock); | |
1687 | } | |
1688 | ||
1689 | #ifdef CONFIG_SWAP | |
1690 | static int inactive_anon_is_low_global(struct zone *zone) | |
1691 | { | |
1692 | unsigned long active, inactive; | |
1693 | ||
1694 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1695 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1696 | ||
1697 | if (inactive * zone->inactive_ratio < active) | |
1698 | return 1; | |
1699 | ||
1700 | return 0; | |
1701 | } | |
1702 | ||
1703 | /** | |
1704 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
1705 | * @zone: zone to check | |
1706 | * @sc: scan control of this context | |
1707 | * | |
1708 | * Returns true if the zone does not have enough inactive anon pages, | |
1709 | * meaning some active anon pages need to be deactivated. | |
1710 | */ | |
1711 | static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc) | |
1712 | { | |
1713 | int low; | |
1714 | ||
1715 | /* | |
1716 | * If we don't have swap space, anonymous page deactivation | |
1717 | * is pointless. | |
1718 | */ | |
1719 | if (!total_swap_pages) | |
1720 | return 0; | |
1721 | ||
1722 | if (scanning_global_lru(sc)) | |
1723 | low = inactive_anon_is_low_global(zone); | |
1724 | else | |
1725 | low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup); | |
1726 | return low; | |
1727 | } | |
1728 | #else | |
1729 | static inline int inactive_anon_is_low(struct zone *zone, | |
1730 | struct scan_control *sc) | |
1731 | { | |
1732 | return 0; | |
1733 | } | |
1734 | #endif | |
1735 | ||
1736 | static int inactive_file_is_low_global(struct zone *zone) | |
1737 | { | |
1738 | unsigned long active, inactive; | |
1739 | ||
1740 | active = zone_page_state(zone, NR_ACTIVE_FILE); | |
1741 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1742 | ||
1743 | return (active > inactive); | |
1744 | } | |
1745 | ||
1746 | /** | |
1747 | * inactive_file_is_low - check if file pages need to be deactivated | |
1748 | * @zone: zone to check | |
1749 | * @sc: scan control of this context | |
1750 | * | |
1751 | * When the system is doing streaming IO, memory pressure here | |
1752 | * ensures that active file pages get deactivated, until more | |
1753 | * than half of the file pages are on the inactive list. | |
1754 | * | |
1755 | * Once we get to that situation, protect the system's working | |
1756 | * set from being evicted by disabling active file page aging. | |
1757 | * | |
1758 | * This uses a different ratio than the anonymous pages, because | |
1759 | * the page cache uses a use-once replacement algorithm. | |
1760 | */ | |
1761 | static int inactive_file_is_low(struct zone *zone, struct scan_control *sc) | |
1762 | { | |
1763 | int low; | |
1764 | ||
1765 | if (scanning_global_lru(sc)) | |
1766 | low = inactive_file_is_low_global(zone); | |
1767 | else | |
1768 | low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup); | |
1769 | return low; | |
1770 | } | |
1771 | ||
1772 | static int inactive_list_is_low(struct zone *zone, struct scan_control *sc, | |
1773 | int file) | |
1774 | { | |
1775 | if (file) | |
1776 | return inactive_file_is_low(zone, sc); | |
1777 | else | |
1778 | return inactive_anon_is_low(zone, sc); | |
1779 | } | |
1780 | ||
1781 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, | |
1782 | struct zone *zone, struct scan_control *sc, int priority) | |
1783 | { | |
1784 | int file = is_file_lru(lru); | |
1785 | ||
1786 | if (is_active_lru(lru)) { | |
1787 | if (inactive_list_is_low(zone, sc, file)) | |
1788 | shrink_active_list(nr_to_scan, zone, sc, priority, file); | |
1789 | return 0; | |
1790 | } | |
1791 | ||
1792 | return shrink_inactive_list(nr_to_scan, zone, sc, priority, file); | |
1793 | } | |
1794 | ||
1795 | static int vmscan_swappiness(struct scan_control *sc) | |
1796 | { | |
1797 | if (scanning_global_lru(sc)) | |
1798 | return vm_swappiness; | |
1799 | return mem_cgroup_swappiness(sc->mem_cgroup); | |
1800 | } | |
1801 | ||
1802 | /* | |
1803 | * Determine how aggressively the anon and file LRU lists should be | |
1804 | * scanned. The relative value of each set of LRU lists is determined | |
1805 | * by looking at the fraction of the pages scanned we did rotate back | |
1806 | * onto the active list instead of evict. | |
1807 | * | |
1808 | * nr[0] = anon pages to scan; nr[1] = file pages to scan | |
1809 | */ | |
1810 | static void get_scan_count(struct zone *zone, struct scan_control *sc, | |
1811 | unsigned long *nr, int priority) | |
1812 | { | |
1813 | unsigned long anon, file, free; | |
1814 | unsigned long anon_prio, file_prio; | |
1815 | unsigned long ap, fp; | |
1816 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); | |
1817 | u64 fraction[2], denominator; | |
1818 | enum lru_list l; | |
1819 | int noswap = 0; | |
1820 | bool force_scan = false; | |
1821 | ||
1822 | /* | |
1823 | * If the zone or memcg is small, nr[l] can be 0. This | |
1824 | * results in no scanning on this priority and a potential | |
1825 | * priority drop. Global direct reclaim can go to the next | |
1826 | * zone and tends to have no problems. Global kswapd is for | |
1827 | * zone balancing and it needs to scan a minimum amount. When | |
1828 | * reclaiming for a memcg, a priority drop can cause high | |
1829 | * latencies, so it's better to scan a minimum amount there as | |
1830 | * well. | |
1831 | */ | |
1832 | if (scanning_global_lru(sc) && current_is_kswapd()) | |
1833 | force_scan = true; | |
1834 | if (!scanning_global_lru(sc)) | |
1835 | force_scan = true; | |
1836 | ||
1837 | /* If we have no swap space, do not bother scanning anon pages. */ | |
1838 | if (!sc->may_swap || (nr_swap_pages <= 0)) { | |
1839 | noswap = 1; | |
1840 | fraction[0] = 0; | |
1841 | fraction[1] = 1; | |
1842 | denominator = 1; | |
1843 | goto out; | |
1844 | } | |
1845 | ||
1846 | anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) + | |
1847 | zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON); | |
1848 | file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) + | |
1849 | zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE); | |
1850 | ||
1851 | if (scanning_global_lru(sc)) { | |
1852 | free = zone_page_state(zone, NR_FREE_PAGES); | |
1853 | /* If we have very few page cache pages, | |
1854 | force-scan anon pages. */ | |
1855 | if (unlikely(file + free <= high_wmark_pages(zone))) { | |
1856 | fraction[0] = 1; | |
1857 | fraction[1] = 0; | |
1858 | denominator = 1; | |
1859 | goto out; | |
1860 | } | |
1861 | } | |
1862 | ||
1863 | /* | |
1864 | * With swappiness at 100, anonymous and file have the same priority. | |
1865 | * This scanning priority is essentially the inverse of IO cost. | |
1866 | */ | |
1867 | anon_prio = vmscan_swappiness(sc); | |
1868 | file_prio = 200 - vmscan_swappiness(sc); | |
1869 | ||
1870 | /* | |
1871 | * OK, so we have swap space and a fair amount of page cache | |
1872 | * pages. We use the recently rotated / recently scanned | |
1873 | * ratios to determine how valuable each cache is. | |
1874 | * | |
1875 | * Because workloads change over time (and to avoid overflow) | |
1876 | * we keep these statistics as a floating average, which ends | |
1877 | * up weighing recent references more than old ones. | |
1878 | * | |
1879 | * anon in [0], file in [1] | |
1880 | */ | |
1881 | spin_lock_irq(&zone->lru_lock); | |
1882 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { | |
1883 | reclaim_stat->recent_scanned[0] /= 2; | |
1884 | reclaim_stat->recent_rotated[0] /= 2; | |
1885 | } | |
1886 | ||
1887 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { | |
1888 | reclaim_stat->recent_scanned[1] /= 2; | |
1889 | reclaim_stat->recent_rotated[1] /= 2; | |
1890 | } | |
1891 | ||
1892 | /* | |
1893 | * The amount of pressure on anon vs file pages is inversely | |
1894 | * proportional to the fraction of recently scanned pages on | |
1895 | * each list that were recently referenced and in active use. | |
1896 | */ | |
1897 | ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1); | |
1898 | ap /= reclaim_stat->recent_rotated[0] + 1; | |
1899 | ||
1900 | fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1); | |
1901 | fp /= reclaim_stat->recent_rotated[1] + 1; | |
1902 | spin_unlock_irq(&zone->lru_lock); | |
1903 | ||
1904 | fraction[0] = ap; | |
1905 | fraction[1] = fp; | |
1906 | denominator = ap + fp + 1; | |
1907 | out: | |
1908 | for_each_evictable_lru(l) { | |
1909 | int file = is_file_lru(l); | |
1910 | unsigned long scan; | |
1911 | ||
1912 | scan = zone_nr_lru_pages(zone, sc, l); | |
1913 | if (priority || noswap) { | |
1914 | scan >>= priority; | |
1915 | if (!scan && force_scan) | |
1916 | scan = SWAP_CLUSTER_MAX; | |
1917 | scan = div64_u64(scan * fraction[file], denominator); | |
1918 | } | |
1919 | nr[l] = scan; | |
1920 | } | |
1921 | } | |
1922 | ||
1923 | /* | |
1924 | * Reclaim/compaction depends on a number of pages being freed. To avoid | |
1925 | * disruption to the system, a small number of order-0 pages continue to be | |
1926 | * rotated and reclaimed in the normal fashion. However, by the time we get | |
1927 | * back to the allocator and call try_to_compact_zone(), we ensure that | |
1928 | * there are enough free pages for it to be likely successful | |
1929 | */ | |
1930 | static inline bool should_continue_reclaim(struct zone *zone, | |
1931 | unsigned long nr_reclaimed, | |
1932 | unsigned long nr_scanned, | |
1933 | struct scan_control *sc) | |
1934 | { | |
1935 | unsigned long pages_for_compaction; | |
1936 | unsigned long inactive_lru_pages; | |
1937 | ||
1938 | /* If not in reclaim/compaction mode, stop */ | |
1939 | if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION)) | |
1940 | return false; | |
1941 | ||
1942 | /* Consider stopping depending on scan and reclaim activity */ | |
1943 | if (sc->gfp_mask & __GFP_REPEAT) { | |
1944 | /* | |
1945 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
1946 | * full LRU list has been scanned and we are still failing | |
1947 | * to reclaim pages. This full LRU scan is potentially | |
1948 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
1949 | */ | |
1950 | if (!nr_reclaimed && !nr_scanned) | |
1951 | return false; | |
1952 | } else { | |
1953 | /* | |
1954 | * For non-__GFP_REPEAT allocations which can presumably | |
1955 | * fail without consequence, stop if we failed to reclaim | |
1956 | * any pages from the last SWAP_CLUSTER_MAX number of | |
1957 | * pages that were scanned. This will return to the | |
1958 | * caller faster at the risk reclaim/compaction and | |
1959 | * the resulting allocation attempt fails | |
1960 | */ | |
1961 | if (!nr_reclaimed) | |
1962 | return false; | |
1963 | } | |
1964 | ||
1965 | /* | |
1966 | * If we have not reclaimed enough pages for compaction and the | |
1967 | * inactive lists are large enough, continue reclaiming | |
1968 | */ | |
1969 | pages_for_compaction = (2UL << sc->order); | |
1970 | inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) + | |
1971 | zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE); | |
1972 | if (sc->nr_reclaimed < pages_for_compaction && | |
1973 | inactive_lru_pages > pages_for_compaction) | |
1974 | return true; | |
1975 | ||
1976 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
1977 | switch (compaction_suitable(zone, sc->order)) { | |
1978 | case COMPACT_PARTIAL: | |
1979 | case COMPACT_CONTINUE: | |
1980 | return false; | |
1981 | default: | |
1982 | return true; | |
1983 | } | |
1984 | } | |
1985 | ||
1986 | /* | |
1987 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1988 | */ | |
1989 | static void shrink_zone(int priority, struct zone *zone, | |
1990 | struct scan_control *sc) | |
1991 | { | |
1992 | unsigned long nr[NR_LRU_LISTS]; | |
1993 | unsigned long nr_to_scan; | |
1994 | enum lru_list l; | |
1995 | unsigned long nr_reclaimed, nr_scanned; | |
1996 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
1997 | struct blk_plug plug; | |
1998 | ||
1999 | restart: | |
2000 | nr_reclaimed = 0; | |
2001 | nr_scanned = sc->nr_scanned; | |
2002 | get_scan_count(zone, sc, nr, priority); | |
2003 | ||
2004 | blk_start_plug(&plug); | |
2005 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2006 | nr[LRU_INACTIVE_FILE]) { | |
2007 | for_each_evictable_lru(l) { | |
2008 | if (nr[l]) { | |
2009 | nr_to_scan = min_t(unsigned long, | |
2010 | nr[l], SWAP_CLUSTER_MAX); | |
2011 | nr[l] -= nr_to_scan; | |
2012 | ||
2013 | nr_reclaimed += shrink_list(l, nr_to_scan, | |
2014 | zone, sc, priority); | |
2015 | } | |
2016 | } | |
2017 | /* | |
2018 | * On large memory systems, scan >> priority can become | |
2019 | * really large. This is fine for the starting priority; | |
2020 | * we want to put equal scanning pressure on each zone. | |
2021 | * However, if the VM has a harder time of freeing pages, | |
2022 | * with multiple processes reclaiming pages, the total | |
2023 | * freeing target can get unreasonably large. | |
2024 | */ | |
2025 | if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY) | |
2026 | break; | |
2027 | } | |
2028 | blk_finish_plug(&plug); | |
2029 | sc->nr_reclaimed += nr_reclaimed; | |
2030 | ||
2031 | /* | |
2032 | * Even if we did not try to evict anon pages at all, we want to | |
2033 | * rebalance the anon lru active/inactive ratio. | |
2034 | */ | |
2035 | if (inactive_anon_is_low(zone, sc)) | |
2036 | shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0); | |
2037 | ||
2038 | /* reclaim/compaction might need reclaim to continue */ | |
2039 | if (should_continue_reclaim(zone, nr_reclaimed, | |
2040 | sc->nr_scanned - nr_scanned, sc)) | |
2041 | goto restart; | |
2042 | ||
2043 | throttle_vm_writeout(sc->gfp_mask); | |
2044 | } | |
2045 | ||
2046 | /* | |
2047 | * This is the direct reclaim path, for page-allocating processes. We only | |
2048 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2049 | * request. | |
2050 | * | |
2051 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). | |
2052 | * Because: | |
2053 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order | |
2054 | * allocation or | |
2055 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones | |
2056 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2057 | * zone defense algorithm. | |
2058 | * | |
2059 | * If a zone is deemed to be full of pinned pages then just give it a light | |
2060 | * scan then give up on it. | |
2061 | */ | |
2062 | static void shrink_zones(int priority, struct zonelist *zonelist, | |
2063 | struct scan_control *sc) | |
2064 | { | |
2065 | struct zoneref *z; | |
2066 | struct zone *zone; | |
2067 | unsigned long nr_soft_reclaimed; | |
2068 | unsigned long nr_soft_scanned; | |
2069 | ||
2070 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2071 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2072 | if (!populated_zone(zone)) | |
2073 | continue; | |
2074 | /* | |
2075 | * Take care memory controller reclaiming has small influence | |
2076 | * to global LRU. | |
2077 | */ | |
2078 | if (scanning_global_lru(sc)) { | |
2079 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2080 | continue; | |
2081 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) | |
2082 | continue; /* Let kswapd poll it */ | |
2083 | /* | |
2084 | * This steals pages from memory cgroups over softlimit | |
2085 | * and returns the number of reclaimed pages and | |
2086 | * scanned pages. This works for global memory pressure | |
2087 | * and balancing, not for a memcg's limit. | |
2088 | */ | |
2089 | nr_soft_scanned = 0; | |
2090 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2091 | sc->order, sc->gfp_mask, | |
2092 | &nr_soft_scanned); | |
2093 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2094 | sc->nr_scanned += nr_soft_scanned; | |
2095 | /* need some check for avoid more shrink_zone() */ | |
2096 | } | |
2097 | ||
2098 | shrink_zone(priority, zone, sc); | |
2099 | } | |
2100 | } | |
2101 | ||
2102 | static bool zone_reclaimable(struct zone *zone) | |
2103 | { | |
2104 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
2105 | } | |
2106 | ||
2107 | /* All zones in zonelist are unreclaimable? */ | |
2108 | static bool all_unreclaimable(struct zonelist *zonelist, | |
2109 | struct scan_control *sc) | |
2110 | { | |
2111 | struct zoneref *z; | |
2112 | struct zone *zone; | |
2113 | ||
2114 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2115 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2116 | if (!populated_zone(zone)) | |
2117 | continue; | |
2118 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2119 | continue; | |
2120 | if (!zone->all_unreclaimable) | |
2121 | return false; | |
2122 | } | |
2123 | ||
2124 | return true; | |
2125 | } | |
2126 | ||
2127 | /* | |
2128 | * This is the main entry point to direct page reclaim. | |
2129 | * | |
2130 | * If a full scan of the inactive list fails to free enough memory then we | |
2131 | * are "out of memory" and something needs to be killed. | |
2132 | * | |
2133 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2134 | * high - the zone may be full of dirty or under-writeback pages, which this | |
2135 | * caller can't do much about. We kick the writeback threads and take explicit | |
2136 | * naps in the hope that some of these pages can be written. But if the | |
2137 | * allocating task holds filesystem locks which prevent writeout this might not | |
2138 | * work, and the allocation attempt will fail. | |
2139 | * | |
2140 | * returns: 0, if no pages reclaimed | |
2141 | * else, the number of pages reclaimed | |
2142 | */ | |
2143 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, | |
2144 | struct scan_control *sc, | |
2145 | struct shrink_control *shrink) | |
2146 | { | |
2147 | int priority; | |
2148 | unsigned long total_scanned = 0; | |
2149 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
2150 | struct zoneref *z; | |
2151 | struct zone *zone; | |
2152 | unsigned long writeback_threshold; | |
2153 | ||
2154 | get_mems_allowed(); | |
2155 | delayacct_freepages_start(); | |
2156 | ||
2157 | if (scanning_global_lru(sc)) | |
2158 | count_vm_event(ALLOCSTALL); | |
2159 | ||
2160 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { | |
2161 | sc->nr_scanned = 0; | |
2162 | if (!priority) | |
2163 | disable_swap_token(sc->mem_cgroup); | |
2164 | shrink_zones(priority, zonelist, sc); | |
2165 | /* | |
2166 | * Don't shrink slabs when reclaiming memory from | |
2167 | * over limit cgroups | |
2168 | */ | |
2169 | if (scanning_global_lru(sc)) { | |
2170 | unsigned long lru_pages = 0; | |
2171 | for_each_zone_zonelist(zone, z, zonelist, | |
2172 | gfp_zone(sc->gfp_mask)) { | |
2173 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2174 | continue; | |
2175 | ||
2176 | lru_pages += zone_reclaimable_pages(zone); | |
2177 | } | |
2178 | ||
2179 | shrink_slab(shrink, sc->nr_scanned, lru_pages); | |
2180 | if (reclaim_state) { | |
2181 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2182 | reclaim_state->reclaimed_slab = 0; | |
2183 | } | |
2184 | } | |
2185 | total_scanned += sc->nr_scanned; | |
2186 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) | |
2187 | goto out; | |
2188 | ||
2189 | /* | |
2190 | * Try to write back as many pages as we just scanned. This | |
2191 | * tends to cause slow streaming writers to write data to the | |
2192 | * disk smoothly, at the dirtying rate, which is nice. But | |
2193 | * that's undesirable in laptop mode, where we *want* lumpy | |
2194 | * writeout. So in laptop mode, write out the whole world. | |
2195 | */ | |
2196 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; | |
2197 | if (total_scanned > writeback_threshold) { | |
2198 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned); | |
2199 | sc->may_writepage = 1; | |
2200 | } | |
2201 | ||
2202 | /* Take a nap, wait for some writeback to complete */ | |
2203 | if (!sc->hibernation_mode && sc->nr_scanned && | |
2204 | priority < DEF_PRIORITY - 2) { | |
2205 | struct zone *preferred_zone; | |
2206 | ||
2207 | first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), | |
2208 | &cpuset_current_mems_allowed, | |
2209 | &preferred_zone); | |
2210 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); | |
2211 | } | |
2212 | } | |
2213 | ||
2214 | out: | |
2215 | delayacct_freepages_end(); | |
2216 | put_mems_allowed(); | |
2217 | ||
2218 | if (sc->nr_reclaimed) | |
2219 | return sc->nr_reclaimed; | |
2220 | ||
2221 | /* | |
2222 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2223 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2224 | * check. | |
2225 | */ | |
2226 | if (oom_killer_disabled) | |
2227 | return 0; | |
2228 | ||
2229 | /* top priority shrink_zones still had more to do? don't OOM, then */ | |
2230 | if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc)) | |
2231 | return 1; | |
2232 | ||
2233 | return 0; | |
2234 | } | |
2235 | ||
2236 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, | |
2237 | gfp_t gfp_mask, nodemask_t *nodemask) | |
2238 | { | |
2239 | unsigned long nr_reclaimed; | |
2240 | struct scan_control sc = { | |
2241 | .gfp_mask = gfp_mask, | |
2242 | .may_writepage = !laptop_mode, | |
2243 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2244 | .may_unmap = 1, | |
2245 | .may_swap = 1, | |
2246 | .order = order, | |
2247 | .mem_cgroup = NULL, | |
2248 | .nodemask = nodemask, | |
2249 | }; | |
2250 | struct shrink_control shrink = { | |
2251 | .gfp_mask = sc.gfp_mask, | |
2252 | }; | |
2253 | ||
2254 | trace_mm_vmscan_direct_reclaim_begin(order, | |
2255 | sc.may_writepage, | |
2256 | gfp_mask); | |
2257 | ||
2258 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); | |
2259 | ||
2260 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2261 | ||
2262 | return nr_reclaimed; | |
2263 | } | |
2264 | ||
2265 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR | |
2266 | ||
2267 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem, | |
2268 | gfp_t gfp_mask, bool noswap, | |
2269 | struct zone *zone, | |
2270 | unsigned long *nr_scanned) | |
2271 | { | |
2272 | struct scan_control sc = { | |
2273 | .nr_scanned = 0, | |
2274 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2275 | .may_writepage = !laptop_mode, | |
2276 | .may_unmap = 1, | |
2277 | .may_swap = !noswap, | |
2278 | .order = 0, | |
2279 | .mem_cgroup = mem, | |
2280 | }; | |
2281 | ||
2282 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
2283 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
2284 | ||
2285 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(0, | |
2286 | sc.may_writepage, | |
2287 | sc.gfp_mask); | |
2288 | ||
2289 | /* | |
2290 | * NOTE: Although we can get the priority field, using it | |
2291 | * here is not a good idea, since it limits the pages we can scan. | |
2292 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2293 | * will pick up pages from other mem cgroup's as well. We hack | |
2294 | * the priority and make it zero. | |
2295 | */ | |
2296 | shrink_zone(0, zone, &sc); | |
2297 | ||
2298 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2299 | ||
2300 | *nr_scanned = sc.nr_scanned; | |
2301 | return sc.nr_reclaimed; | |
2302 | } | |
2303 | ||
2304 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont, | |
2305 | gfp_t gfp_mask, | |
2306 | bool noswap) | |
2307 | { | |
2308 | struct zonelist *zonelist; | |
2309 | unsigned long nr_reclaimed; | |
2310 | int nid; | |
2311 | struct scan_control sc = { | |
2312 | .may_writepage = !laptop_mode, | |
2313 | .may_unmap = 1, | |
2314 | .may_swap = !noswap, | |
2315 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2316 | .order = 0, | |
2317 | .mem_cgroup = mem_cont, | |
2318 | .nodemask = NULL, /* we don't care the placement */ | |
2319 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
2320 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2321 | }; | |
2322 | struct shrink_control shrink = { | |
2323 | .gfp_mask = sc.gfp_mask, | |
2324 | }; | |
2325 | ||
2326 | /* | |
2327 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2328 | * take care of from where we get pages. So the node where we start the | |
2329 | * scan does not need to be the current node. | |
2330 | */ | |
2331 | nid = mem_cgroup_select_victim_node(mem_cont); | |
2332 | ||
2333 | zonelist = NODE_DATA(nid)->node_zonelists; | |
2334 | ||
2335 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2336 | sc.may_writepage, | |
2337 | sc.gfp_mask); | |
2338 | ||
2339 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); | |
2340 | ||
2341 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2342 | ||
2343 | return nr_reclaimed; | |
2344 | } | |
2345 | #endif | |
2346 | ||
2347 | /* | |
2348 | * pgdat_balanced is used when checking if a node is balanced for high-order | |
2349 | * allocations. Only zones that meet watermarks and are in a zone allowed | |
2350 | * by the callers classzone_idx are added to balanced_pages. The total of | |
2351 | * balanced pages must be at least 25% of the zones allowed by classzone_idx | |
2352 | * for the node to be considered balanced. Forcing all zones to be balanced | |
2353 | * for high orders can cause excessive reclaim when there are imbalanced zones. | |
2354 | * The choice of 25% is due to | |
2355 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2356 | * reasonable sized machine | |
2357 | * o On all other machines, the top zone must be at least a reasonable | |
2358 | * percentage of the middle zones. For example, on 32-bit x86, highmem | |
2359 | * would need to be at least 256M for it to be balance a whole node. | |
2360 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2361 | * to balance a node on its own. These seemed like reasonable ratios. | |
2362 | */ | |
2363 | static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages, | |
2364 | int classzone_idx) | |
2365 | { | |
2366 | unsigned long present_pages = 0; | |
2367 | int i; | |
2368 | ||
2369 | for (i = 0; i <= classzone_idx; i++) | |
2370 | present_pages += pgdat->node_zones[i].present_pages; | |
2371 | ||
2372 | /* A special case here: if zone has no page, we think it's balanced */ | |
2373 | return balanced_pages >= (present_pages >> 2); | |
2374 | } | |
2375 | ||
2376 | /* is kswapd sleeping prematurely? */ | |
2377 | static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining, | |
2378 | int classzone_idx) | |
2379 | { | |
2380 | int i; | |
2381 | unsigned long balanced = 0; | |
2382 | bool all_zones_ok = true; | |
2383 | ||
2384 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ | |
2385 | if (remaining) | |
2386 | return true; | |
2387 | ||
2388 | /* Check the watermark levels */ | |
2389 | for (i = 0; i <= classzone_idx; i++) { | |
2390 | struct zone *zone = pgdat->node_zones + i; | |
2391 | ||
2392 | if (!populated_zone(zone)) | |
2393 | continue; | |
2394 | ||
2395 | /* | |
2396 | * balance_pgdat() skips over all_unreclaimable after | |
2397 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2398 | * they must be considered balanced here as well if kswapd | |
2399 | * is to sleep | |
2400 | */ | |
2401 | if (zone->all_unreclaimable) { | |
2402 | balanced += zone->present_pages; | |
2403 | continue; | |
2404 | } | |
2405 | ||
2406 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone), | |
2407 | i, 0)) | |
2408 | all_zones_ok = false; | |
2409 | else | |
2410 | balanced += zone->present_pages; | |
2411 | } | |
2412 | ||
2413 | /* | |
2414 | * For high-order requests, the balanced zones must contain at least | |
2415 | * 25% of the nodes pages for kswapd to sleep. For order-0, all zones | |
2416 | * must be balanced | |
2417 | */ | |
2418 | if (order) | |
2419 | return !pgdat_balanced(pgdat, balanced, classzone_idx); | |
2420 | else | |
2421 | return !all_zones_ok; | |
2422 | } | |
2423 | ||
2424 | /* | |
2425 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
2426 | * they are all at high_wmark_pages(zone). | |
2427 | * | |
2428 | * Returns the final order kswapd was reclaiming at | |
2429 | * | |
2430 | * There is special handling here for zones which are full of pinned pages. | |
2431 | * This can happen if the pages are all mlocked, or if they are all used by | |
2432 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
2433 | * What we do is to detect the case where all pages in the zone have been | |
2434 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
2435 | * dead and from now on, only perform a short scan. Basically we're polling | |
2436 | * the zone for when the problem goes away. | |
2437 | * | |
2438 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
2439 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is | |
2440 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
2441 | * lower zones regardless of the number of free pages in the lower zones. This | |
2442 | * interoperates with the page allocator fallback scheme to ensure that aging | |
2443 | * of pages is balanced across the zones. | |
2444 | */ | |
2445 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, | |
2446 | int *classzone_idx) | |
2447 | { | |
2448 | int all_zones_ok; | |
2449 | unsigned long balanced; | |
2450 | int priority; | |
2451 | int i; | |
2452 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ | |
2453 | unsigned long total_scanned; | |
2454 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
2455 | unsigned long nr_soft_reclaimed; | |
2456 | unsigned long nr_soft_scanned; | |
2457 | struct scan_control sc = { | |
2458 | .gfp_mask = GFP_KERNEL, | |
2459 | .may_unmap = 1, | |
2460 | .may_swap = 1, | |
2461 | /* | |
2462 | * kswapd doesn't want to be bailed out while reclaim. because | |
2463 | * we want to put equal scanning pressure on each zone. | |
2464 | */ | |
2465 | .nr_to_reclaim = ULONG_MAX, | |
2466 | .order = order, | |
2467 | .mem_cgroup = NULL, | |
2468 | }; | |
2469 | struct shrink_control shrink = { | |
2470 | .gfp_mask = sc.gfp_mask, | |
2471 | }; | |
2472 | loop_again: | |
2473 | total_scanned = 0; | |
2474 | sc.nr_reclaimed = 0; | |
2475 | sc.may_writepage = !laptop_mode; | |
2476 | count_vm_event(PAGEOUTRUN); | |
2477 | ||
2478 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { | |
2479 | unsigned long lru_pages = 0; | |
2480 | int has_under_min_watermark_zone = 0; | |
2481 | ||
2482 | /* The swap token gets in the way of swapout... */ | |
2483 | if (!priority) | |
2484 | disable_swap_token(NULL); | |
2485 | ||
2486 | all_zones_ok = 1; | |
2487 | balanced = 0; | |
2488 | ||
2489 | /* | |
2490 | * Scan in the highmem->dma direction for the highest | |
2491 | * zone which needs scanning | |
2492 | */ | |
2493 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
2494 | struct zone *zone = pgdat->node_zones + i; | |
2495 | ||
2496 | if (!populated_zone(zone)) | |
2497 | continue; | |
2498 | ||
2499 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) | |
2500 | continue; | |
2501 | ||
2502 | /* | |
2503 | * Do some background aging of the anon list, to give | |
2504 | * pages a chance to be referenced before reclaiming. | |
2505 | */ | |
2506 | if (inactive_anon_is_low(zone, &sc)) | |
2507 | shrink_active_list(SWAP_CLUSTER_MAX, zone, | |
2508 | &sc, priority, 0); | |
2509 | ||
2510 | if (!zone_watermark_ok_safe(zone, order, | |
2511 | high_wmark_pages(zone), 0, 0)) { | |
2512 | end_zone = i; | |
2513 | break; | |
2514 | } else { | |
2515 | /* If balanced, clear the congested flag */ | |
2516 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2517 | } | |
2518 | } | |
2519 | if (i < 0) | |
2520 | goto out; | |
2521 | ||
2522 | for (i = 0; i <= end_zone; i++) { | |
2523 | struct zone *zone = pgdat->node_zones + i; | |
2524 | ||
2525 | lru_pages += zone_reclaimable_pages(zone); | |
2526 | } | |
2527 | ||
2528 | /* | |
2529 | * Now scan the zone in the dma->highmem direction, stopping | |
2530 | * at the last zone which needs scanning. | |
2531 | * | |
2532 | * We do this because the page allocator works in the opposite | |
2533 | * direction. This prevents the page allocator from allocating | |
2534 | * pages behind kswapd's direction of progress, which would | |
2535 | * cause too much scanning of the lower zones. | |
2536 | */ | |
2537 | for (i = 0; i <= end_zone; i++) { | |
2538 | struct zone *zone = pgdat->node_zones + i; | |
2539 | int nr_slab; | |
2540 | unsigned long balance_gap; | |
2541 | ||
2542 | if (!populated_zone(zone)) | |
2543 | continue; | |
2544 | ||
2545 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) | |
2546 | continue; | |
2547 | ||
2548 | sc.nr_scanned = 0; | |
2549 | ||
2550 | nr_soft_scanned = 0; | |
2551 | /* | |
2552 | * Call soft limit reclaim before calling shrink_zone. | |
2553 | */ | |
2554 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2555 | order, sc.gfp_mask, | |
2556 | &nr_soft_scanned); | |
2557 | sc.nr_reclaimed += nr_soft_reclaimed; | |
2558 | total_scanned += nr_soft_scanned; | |
2559 | ||
2560 | /* | |
2561 | * We put equal pressure on every zone, unless | |
2562 | * one zone has way too many pages free | |
2563 | * already. The "too many pages" is defined | |
2564 | * as the high wmark plus a "gap" where the | |
2565 | * gap is either the low watermark or 1% | |
2566 | * of the zone, whichever is smaller. | |
2567 | */ | |
2568 | balance_gap = min(low_wmark_pages(zone), | |
2569 | (zone->present_pages + | |
2570 | KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2571 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
2572 | if (!zone_watermark_ok_safe(zone, order, | |
2573 | high_wmark_pages(zone) + balance_gap, | |
2574 | end_zone, 0)) { | |
2575 | shrink_zone(priority, zone, &sc); | |
2576 | ||
2577 | reclaim_state->reclaimed_slab = 0; | |
2578 | nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages); | |
2579 | sc.nr_reclaimed += reclaim_state->reclaimed_slab; | |
2580 | total_scanned += sc.nr_scanned; | |
2581 | ||
2582 | if (nr_slab == 0 && !zone_reclaimable(zone)) | |
2583 | zone->all_unreclaimable = 1; | |
2584 | } | |
2585 | ||
2586 | /* | |
2587 | * If we've done a decent amount of scanning and | |
2588 | * the reclaim ratio is low, start doing writepage | |
2589 | * even in laptop mode | |
2590 | */ | |
2591 | if (total_scanned > SWAP_CLUSTER_MAX * 2 && | |
2592 | total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) | |
2593 | sc.may_writepage = 1; | |
2594 | ||
2595 | if (zone->all_unreclaimable) { | |
2596 | if (end_zone && end_zone == i) | |
2597 | end_zone--; | |
2598 | continue; | |
2599 | } | |
2600 | ||
2601 | if (!zone_watermark_ok_safe(zone, order, | |
2602 | high_wmark_pages(zone), end_zone, 0)) { | |
2603 | all_zones_ok = 0; | |
2604 | /* | |
2605 | * We are still under min water mark. This | |
2606 | * means that we have a GFP_ATOMIC allocation | |
2607 | * failure risk. Hurry up! | |
2608 | */ | |
2609 | if (!zone_watermark_ok_safe(zone, order, | |
2610 | min_wmark_pages(zone), end_zone, 0)) | |
2611 | has_under_min_watermark_zone = 1; | |
2612 | } else { | |
2613 | /* | |
2614 | * If a zone reaches its high watermark, | |
2615 | * consider it to be no longer congested. It's | |
2616 | * possible there are dirty pages backed by | |
2617 | * congested BDIs but as pressure is relieved, | |
2618 | * spectulatively avoid congestion waits | |
2619 | */ | |
2620 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2621 | if (i <= *classzone_idx) | |
2622 | balanced += zone->present_pages; | |
2623 | } | |
2624 | ||
2625 | } | |
2626 | if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx))) | |
2627 | break; /* kswapd: all done */ | |
2628 | /* | |
2629 | * OK, kswapd is getting into trouble. Take a nap, then take | |
2630 | * another pass across the zones. | |
2631 | */ | |
2632 | if (total_scanned && (priority < DEF_PRIORITY - 2)) { | |
2633 | if (has_under_min_watermark_zone) | |
2634 | count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT); | |
2635 | else | |
2636 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2637 | } | |
2638 | ||
2639 | /* | |
2640 | * We do this so kswapd doesn't build up large priorities for | |
2641 | * example when it is freeing in parallel with allocators. It | |
2642 | * matches the direct reclaim path behaviour in terms of impact | |
2643 | * on zone->*_priority. | |
2644 | */ | |
2645 | if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) | |
2646 | break; | |
2647 | } | |
2648 | out: | |
2649 | ||
2650 | /* | |
2651 | * order-0: All zones must meet high watermark for a balanced node | |
2652 | * high-order: Balanced zones must make up at least 25% of the node | |
2653 | * for the node to be balanced | |
2654 | */ | |
2655 | if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) { | |
2656 | cond_resched(); | |
2657 | ||
2658 | try_to_freeze(); | |
2659 | ||
2660 | /* | |
2661 | * Fragmentation may mean that the system cannot be | |
2662 | * rebalanced for high-order allocations in all zones. | |
2663 | * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, | |
2664 | * it means the zones have been fully scanned and are still | |
2665 | * not balanced. For high-order allocations, there is | |
2666 | * little point trying all over again as kswapd may | |
2667 | * infinite loop. | |
2668 | * | |
2669 | * Instead, recheck all watermarks at order-0 as they | |
2670 | * are the most important. If watermarks are ok, kswapd will go | |
2671 | * back to sleep. High-order users can still perform direct | |
2672 | * reclaim if they wish. | |
2673 | */ | |
2674 | if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) | |
2675 | order = sc.order = 0; | |
2676 | ||
2677 | goto loop_again; | |
2678 | } | |
2679 | ||
2680 | /* | |
2681 | * If kswapd was reclaiming at a higher order, it has the option of | |
2682 | * sleeping without all zones being balanced. Before it does, it must | |
2683 | * ensure that the watermarks for order-0 on *all* zones are met and | |
2684 | * that the congestion flags are cleared. The congestion flag must | |
2685 | * be cleared as kswapd is the only mechanism that clears the flag | |
2686 | * and it is potentially going to sleep here. | |
2687 | */ | |
2688 | if (order) { | |
2689 | for (i = 0; i <= end_zone; i++) { | |
2690 | struct zone *zone = pgdat->node_zones + i; | |
2691 | ||
2692 | if (!populated_zone(zone)) | |
2693 | continue; | |
2694 | ||
2695 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) | |
2696 | continue; | |
2697 | ||
2698 | /* Confirm the zone is balanced for order-0 */ | |
2699 | if (!zone_watermark_ok(zone, 0, | |
2700 | high_wmark_pages(zone), 0, 0)) { | |
2701 | order = sc.order = 0; | |
2702 | goto loop_again; | |
2703 | } | |
2704 | ||
2705 | /* If balanced, clear the congested flag */ | |
2706 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2707 | } | |
2708 | } | |
2709 | ||
2710 | /* | |
2711 | * Return the order we were reclaiming at so sleeping_prematurely() | |
2712 | * makes a decision on the order we were last reclaiming at. However, | |
2713 | * if another caller entered the allocator slow path while kswapd | |
2714 | * was awake, order will remain at the higher level | |
2715 | */ | |
2716 | *classzone_idx = end_zone; | |
2717 | return order; | |
2718 | } | |
2719 | ||
2720 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) | |
2721 | { | |
2722 | long remaining = 0; | |
2723 | DEFINE_WAIT(wait); | |
2724 | ||
2725 | if (freezing(current) || kthread_should_stop()) | |
2726 | return; | |
2727 | ||
2728 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2729 | ||
2730 | /* Try to sleep for a short interval */ | |
2731 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { | |
2732 | remaining = schedule_timeout(HZ/10); | |
2733 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2734 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2735 | } | |
2736 | ||
2737 | /* | |
2738 | * After a short sleep, check if it was a premature sleep. If not, then | |
2739 | * go fully to sleep until explicitly woken up. | |
2740 | */ | |
2741 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { | |
2742 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); | |
2743 | ||
2744 | /* | |
2745 | * vmstat counters are not perfectly accurate and the estimated | |
2746 | * value for counters such as NR_FREE_PAGES can deviate from the | |
2747 | * true value by nr_online_cpus * threshold. To avoid the zone | |
2748 | * watermarks being breached while under pressure, we reduce the | |
2749 | * per-cpu vmstat threshold while kswapd is awake and restore | |
2750 | * them before going back to sleep. | |
2751 | */ | |
2752 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
2753 | schedule(); | |
2754 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); | |
2755 | } else { | |
2756 | if (remaining) | |
2757 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
2758 | else | |
2759 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
2760 | } | |
2761 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2762 | } | |
2763 | ||
2764 | /* | |
2765 | * The background pageout daemon, started as a kernel thread | |
2766 | * from the init process. | |
2767 | * | |
2768 | * This basically trickles out pages so that we have _some_ | |
2769 | * free memory available even if there is no other activity | |
2770 | * that frees anything up. This is needed for things like routing | |
2771 | * etc, where we otherwise might have all activity going on in | |
2772 | * asynchronous contexts that cannot page things out. | |
2773 | * | |
2774 | * If there are applications that are active memory-allocators | |
2775 | * (most normal use), this basically shouldn't matter. | |
2776 | */ | |
2777 | static int kswapd(void *p) | |
2778 | { | |
2779 | unsigned long order, new_order; | |
2780 | int classzone_idx, new_classzone_idx; | |
2781 | pg_data_t *pgdat = (pg_data_t*)p; | |
2782 | struct task_struct *tsk = current; | |
2783 | ||
2784 | struct reclaim_state reclaim_state = { | |
2785 | .reclaimed_slab = 0, | |
2786 | }; | |
2787 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2788 | ||
2789 | lockdep_set_current_reclaim_state(GFP_KERNEL); | |
2790 | ||
2791 | if (!cpumask_empty(cpumask)) | |
2792 | set_cpus_allowed_ptr(tsk, cpumask); | |
2793 | current->reclaim_state = &reclaim_state; | |
2794 | ||
2795 | /* | |
2796 | * Tell the memory management that we're a "memory allocator", | |
2797 | * and that if we need more memory we should get access to it | |
2798 | * regardless (see "__alloc_pages()"). "kswapd" should | |
2799 | * never get caught in the normal page freeing logic. | |
2800 | * | |
2801 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
2802 | * you need a small amount of memory in order to be able to | |
2803 | * page out something else, and this flag essentially protects | |
2804 | * us from recursively trying to free more memory as we're | |
2805 | * trying to free the first piece of memory in the first place). | |
2806 | */ | |
2807 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; | |
2808 | set_freezable(); | |
2809 | ||
2810 | order = new_order = 0; | |
2811 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; | |
2812 | for ( ; ; ) { | |
2813 | int ret; | |
2814 | ||
2815 | /* | |
2816 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
2817 | * new request of a similar or harder type will succeed soon | |
2818 | * so consider going to sleep on the basis we reclaimed at | |
2819 | */ | |
2820 | if (classzone_idx >= new_classzone_idx && order == new_order) { | |
2821 | new_order = pgdat->kswapd_max_order; | |
2822 | new_classzone_idx = pgdat->classzone_idx; | |
2823 | pgdat->kswapd_max_order = 0; | |
2824 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
2825 | } | |
2826 | ||
2827 | if (order < new_order || classzone_idx > new_classzone_idx) { | |
2828 | /* | |
2829 | * Don't sleep if someone wants a larger 'order' | |
2830 | * allocation or has tigher zone constraints | |
2831 | */ | |
2832 | order = new_order; | |
2833 | classzone_idx = new_classzone_idx; | |
2834 | } else { | |
2835 | kswapd_try_to_sleep(pgdat, order, classzone_idx); | |
2836 | order = pgdat->kswapd_max_order; | |
2837 | classzone_idx = pgdat->classzone_idx; | |
2838 | pgdat->kswapd_max_order = 0; | |
2839 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
2840 | } | |
2841 | ||
2842 | ret = try_to_freeze(); | |
2843 | if (kthread_should_stop()) | |
2844 | break; | |
2845 | ||
2846 | /* | |
2847 | * We can speed up thawing tasks if we don't call balance_pgdat | |
2848 | * after returning from the refrigerator | |
2849 | */ | |
2850 | if (!ret) { | |
2851 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
2852 | order = balance_pgdat(pgdat, order, &classzone_idx); | |
2853 | } | |
2854 | } | |
2855 | return 0; | |
2856 | } | |
2857 | ||
2858 | /* | |
2859 | * A zone is low on free memory, so wake its kswapd task to service it. | |
2860 | */ | |
2861 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) | |
2862 | { | |
2863 | pg_data_t *pgdat; | |
2864 | ||
2865 | if (!populated_zone(zone)) | |
2866 | return; | |
2867 | ||
2868 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2869 | return; | |
2870 | pgdat = zone->zone_pgdat; | |
2871 | if (pgdat->kswapd_max_order < order) { | |
2872 | pgdat->kswapd_max_order = order; | |
2873 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); | |
2874 | } | |
2875 | if (!waitqueue_active(&pgdat->kswapd_wait)) | |
2876 | return; | |
2877 | if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) | |
2878 | return; | |
2879 | ||
2880 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
2881 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2882 | } | |
2883 | ||
2884 | /* | |
2885 | * The reclaimable count would be mostly accurate. | |
2886 | * The less reclaimable pages may be | |
2887 | * - mlocked pages, which will be moved to unevictable list when encountered | |
2888 | * - mapped pages, which may require several travels to be reclaimed | |
2889 | * - dirty pages, which is not "instantly" reclaimable | |
2890 | */ | |
2891 | unsigned long global_reclaimable_pages(void) | |
2892 | { | |
2893 | int nr; | |
2894 | ||
2895 | nr = global_page_state(NR_ACTIVE_FILE) + | |
2896 | global_page_state(NR_INACTIVE_FILE); | |
2897 | ||
2898 | if (nr_swap_pages > 0) | |
2899 | nr += global_page_state(NR_ACTIVE_ANON) + | |
2900 | global_page_state(NR_INACTIVE_ANON); | |
2901 | ||
2902 | return nr; | |
2903 | } | |
2904 | ||
2905 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
2906 | { | |
2907 | int nr; | |
2908 | ||
2909 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
2910 | zone_page_state(zone, NR_INACTIVE_FILE); | |
2911 | ||
2912 | if (nr_swap_pages > 0) | |
2913 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
2914 | zone_page_state(zone, NR_INACTIVE_ANON); | |
2915 | ||
2916 | return nr; | |
2917 | } | |
2918 | ||
2919 | #ifdef CONFIG_HIBERNATION | |
2920 | /* | |
2921 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of | |
2922 | * freed pages. | |
2923 | * | |
2924 | * Rather than trying to age LRUs the aim is to preserve the overall | |
2925 | * LRU order by reclaiming preferentially | |
2926 | * inactive > active > active referenced > active mapped | |
2927 | */ | |
2928 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) | |
2929 | { | |
2930 | struct reclaim_state reclaim_state; | |
2931 | struct scan_control sc = { | |
2932 | .gfp_mask = GFP_HIGHUSER_MOVABLE, | |
2933 | .may_swap = 1, | |
2934 | .may_unmap = 1, | |
2935 | .may_writepage = 1, | |
2936 | .nr_to_reclaim = nr_to_reclaim, | |
2937 | .hibernation_mode = 1, | |
2938 | .order = 0, | |
2939 | }; | |
2940 | struct shrink_control shrink = { | |
2941 | .gfp_mask = sc.gfp_mask, | |
2942 | }; | |
2943 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
2944 | struct task_struct *p = current; | |
2945 | unsigned long nr_reclaimed; | |
2946 | ||
2947 | p->flags |= PF_MEMALLOC; | |
2948 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
2949 | reclaim_state.reclaimed_slab = 0; | |
2950 | p->reclaim_state = &reclaim_state; | |
2951 | ||
2952 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); | |
2953 | ||
2954 | p->reclaim_state = NULL; | |
2955 | lockdep_clear_current_reclaim_state(); | |
2956 | p->flags &= ~PF_MEMALLOC; | |
2957 | ||
2958 | return nr_reclaimed; | |
2959 | } | |
2960 | #endif /* CONFIG_HIBERNATION */ | |
2961 | ||
2962 | /* It's optimal to keep kswapds on the same CPUs as their memory, but | |
2963 | not required for correctness. So if the last cpu in a node goes | |
2964 | away, we get changed to run anywhere: as the first one comes back, | |
2965 | restore their cpu bindings. */ | |
2966 | static int __devinit cpu_callback(struct notifier_block *nfb, | |
2967 | unsigned long action, void *hcpu) | |
2968 | { | |
2969 | int nid; | |
2970 | ||
2971 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { | |
2972 | for_each_node_state(nid, N_HIGH_MEMORY) { | |
2973 | pg_data_t *pgdat = NODE_DATA(nid); | |
2974 | const struct cpumask *mask; | |
2975 | ||
2976 | mask = cpumask_of_node(pgdat->node_id); | |
2977 | ||
2978 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) | |
2979 | /* One of our CPUs online: restore mask */ | |
2980 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
2981 | } | |
2982 | } | |
2983 | return NOTIFY_OK; | |
2984 | } | |
2985 | ||
2986 | /* | |
2987 | * This kswapd start function will be called by init and node-hot-add. | |
2988 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
2989 | */ | |
2990 | int kswapd_run(int nid) | |
2991 | { | |
2992 | pg_data_t *pgdat = NODE_DATA(nid); | |
2993 | int ret = 0; | |
2994 | ||
2995 | if (pgdat->kswapd) | |
2996 | return 0; | |
2997 | ||
2998 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
2999 | if (IS_ERR(pgdat->kswapd)) { | |
3000 | /* failure at boot is fatal */ | |
3001 | BUG_ON(system_state == SYSTEM_BOOTING); | |
3002 | printk("Failed to start kswapd on node %d\n",nid); | |
3003 | ret = -1; | |
3004 | } | |
3005 | return ret; | |
3006 | } | |
3007 | ||
3008 | /* | |
3009 | * Called by memory hotplug when all memory in a node is offlined. | |
3010 | */ | |
3011 | void kswapd_stop(int nid) | |
3012 | { | |
3013 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3014 | ||
3015 | if (kswapd) | |
3016 | kthread_stop(kswapd); | |
3017 | } | |
3018 | ||
3019 | static int __init kswapd_init(void) | |
3020 | { | |
3021 | int nid; | |
3022 | ||
3023 | swap_setup(); | |
3024 | for_each_node_state(nid, N_HIGH_MEMORY) | |
3025 | kswapd_run(nid); | |
3026 | hotcpu_notifier(cpu_callback, 0); | |
3027 | return 0; | |
3028 | } | |
3029 | ||
3030 | module_init(kswapd_init) | |
3031 | ||
3032 | #ifdef CONFIG_NUMA | |
3033 | /* | |
3034 | * Zone reclaim mode | |
3035 | * | |
3036 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3037 | * the watermarks. | |
3038 | */ | |
3039 | int zone_reclaim_mode __read_mostly; | |
3040 | ||
3041 | #define RECLAIM_OFF 0 | |
3042 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ | |
3043 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ | |
3044 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3045 | ||
3046 | /* | |
3047 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3048 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3049 | * a zone. | |
3050 | */ | |
3051 | #define ZONE_RECLAIM_PRIORITY 4 | |
3052 | ||
3053 | /* | |
3054 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3055 | * occur. | |
3056 | */ | |
3057 | int sysctl_min_unmapped_ratio = 1; | |
3058 | ||
3059 | /* | |
3060 | * If the number of slab pages in a zone grows beyond this percentage then | |
3061 | * slab reclaim needs to occur. | |
3062 | */ | |
3063 | int sysctl_min_slab_ratio = 5; | |
3064 | ||
3065 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) | |
3066 | { | |
3067 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3068 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3069 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3070 | ||
3071 | /* | |
3072 | * It's possible for there to be more file mapped pages than | |
3073 | * accounted for by the pages on the file LRU lists because | |
3074 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3075 | */ | |
3076 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3077 | } | |
3078 | ||
3079 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3080 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3081 | { | |
3082 | long nr_pagecache_reclaimable; | |
3083 | long delta = 0; | |
3084 | ||
3085 | /* | |
3086 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3087 | * potentially reclaimable. Otherwise, we have to worry about | |
3088 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3089 | * a better estimate | |
3090 | */ | |
3091 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3092 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3093 | else | |
3094 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3095 | ||
3096 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3097 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3098 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3099 | ||
3100 | /* Watch for any possible underflows due to delta */ | |
3101 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3102 | delta = nr_pagecache_reclaimable; | |
3103 | ||
3104 | return nr_pagecache_reclaimable - delta; | |
3105 | } | |
3106 | ||
3107 | /* | |
3108 | * Try to free up some pages from this zone through reclaim. | |
3109 | */ | |
3110 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3111 | { | |
3112 | /* Minimum pages needed in order to stay on node */ | |
3113 | const unsigned long nr_pages = 1 << order; | |
3114 | struct task_struct *p = current; | |
3115 | struct reclaim_state reclaim_state; | |
3116 | int priority; | |
3117 | struct scan_control sc = { | |
3118 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
3119 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), | |
3120 | .may_swap = 1, | |
3121 | .nr_to_reclaim = max_t(unsigned long, nr_pages, | |
3122 | SWAP_CLUSTER_MAX), | |
3123 | .gfp_mask = gfp_mask, | |
3124 | .order = order, | |
3125 | }; | |
3126 | struct shrink_control shrink = { | |
3127 | .gfp_mask = sc.gfp_mask, | |
3128 | }; | |
3129 | unsigned long nr_slab_pages0, nr_slab_pages1; | |
3130 | ||
3131 | cond_resched(); | |
3132 | /* | |
3133 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3134 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3135 | * and RECLAIM_SWAP. | |
3136 | */ | |
3137 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
3138 | lockdep_set_current_reclaim_state(gfp_mask); | |
3139 | reclaim_state.reclaimed_slab = 0; | |
3140 | p->reclaim_state = &reclaim_state; | |
3141 | ||
3142 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { | |
3143 | /* | |
3144 | * Free memory by calling shrink zone with increasing | |
3145 | * priorities until we have enough memory freed. | |
3146 | */ | |
3147 | priority = ZONE_RECLAIM_PRIORITY; | |
3148 | do { | |
3149 | shrink_zone(priority, zone, &sc); | |
3150 | priority--; | |
3151 | } while (priority >= 0 && sc.nr_reclaimed < nr_pages); | |
3152 | } | |
3153 | ||
3154 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); | |
3155 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
3156 | /* | |
3157 | * shrink_slab() does not currently allow us to determine how | |
3158 | * many pages were freed in this zone. So we take the current | |
3159 | * number of slab pages and shake the slab until it is reduced | |
3160 | * by the same nr_pages that we used for reclaiming unmapped | |
3161 | * pages. | |
3162 | * | |
3163 | * Note that shrink_slab will free memory on all zones and may | |
3164 | * take a long time. | |
3165 | */ | |
3166 | for (;;) { | |
3167 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3168 | ||
3169 | /* No reclaimable slab or very low memory pressure */ | |
3170 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) | |
3171 | break; | |
3172 | ||
3173 | /* Freed enough memory */ | |
3174 | nr_slab_pages1 = zone_page_state(zone, | |
3175 | NR_SLAB_RECLAIMABLE); | |
3176 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3177 | break; | |
3178 | } | |
3179 | ||
3180 | /* | |
3181 | * Update nr_reclaimed by the number of slab pages we | |
3182 | * reclaimed from this zone. | |
3183 | */ | |
3184 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); | |
3185 | if (nr_slab_pages1 < nr_slab_pages0) | |
3186 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
3187 | } | |
3188 | ||
3189 | p->reclaim_state = NULL; | |
3190 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); | |
3191 | lockdep_clear_current_reclaim_state(); | |
3192 | return sc.nr_reclaimed >= nr_pages; | |
3193 | } | |
3194 | ||
3195 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3196 | { | |
3197 | int node_id; | |
3198 | int ret; | |
3199 | ||
3200 | /* | |
3201 | * Zone reclaim reclaims unmapped file backed pages and | |
3202 | * slab pages if we are over the defined limits. | |
3203 | * | |
3204 | * A small portion of unmapped file backed pages is needed for | |
3205 | * file I/O otherwise pages read by file I/O will be immediately | |
3206 | * thrown out if the zone is overallocated. So we do not reclaim | |
3207 | * if less than a specified percentage of the zone is used by | |
3208 | * unmapped file backed pages. | |
3209 | */ | |
3210 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && | |
3211 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
3212 | return ZONE_RECLAIM_FULL; | |
3213 | ||
3214 | if (zone->all_unreclaimable) | |
3215 | return ZONE_RECLAIM_FULL; | |
3216 | ||
3217 | /* | |
3218 | * Do not scan if the allocation should not be delayed. | |
3219 | */ | |
3220 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) | |
3221 | return ZONE_RECLAIM_NOSCAN; | |
3222 | ||
3223 | /* | |
3224 | * Only run zone reclaim on the local zone or on zones that do not | |
3225 | * have associated processors. This will favor the local processor | |
3226 | * over remote processors and spread off node memory allocations | |
3227 | * as wide as possible. | |
3228 | */ | |
3229 | node_id = zone_to_nid(zone); | |
3230 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) | |
3231 | return ZONE_RECLAIM_NOSCAN; | |
3232 | ||
3233 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
3234 | return ZONE_RECLAIM_NOSCAN; | |
3235 | ||
3236 | ret = __zone_reclaim(zone, gfp_mask, order); | |
3237 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3238 | ||
3239 | if (!ret) | |
3240 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3241 | ||
3242 | return ret; | |
3243 | } | |
3244 | #endif | |
3245 | ||
3246 | /* | |
3247 | * page_evictable - test whether a page is evictable | |
3248 | * @page: the page to test | |
3249 | * @vma: the VMA in which the page is or will be mapped, may be NULL | |
3250 | * | |
3251 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
3252 | * lists vs unevictable list. The vma argument is !NULL when called from the | |
3253 | * fault path to determine how to instantate a new page. | |
3254 | * | |
3255 | * Reasons page might not be evictable: | |
3256 | * (1) page's mapping marked unevictable | |
3257 | * (2) page is part of an mlocked VMA | |
3258 | * | |
3259 | */ | |
3260 | int page_evictable(struct page *page, struct vm_area_struct *vma) | |
3261 | { | |
3262 | ||
3263 | if (mapping_unevictable(page_mapping(page))) | |
3264 | return 0; | |
3265 | ||
3266 | if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page))) | |
3267 | return 0; | |
3268 | ||
3269 | return 1; | |
3270 | } | |
3271 | ||
3272 | /** | |
3273 | * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list | |
3274 | * @page: page to check evictability and move to appropriate lru list | |
3275 | * @zone: zone page is in | |
3276 | * | |
3277 | * Checks a page for evictability and moves the page to the appropriate | |
3278 | * zone lru list. | |
3279 | * | |
3280 | * Restrictions: zone->lru_lock must be held, page must be on LRU and must | |
3281 | * have PageUnevictable set. | |
3282 | */ | |
3283 | static void check_move_unevictable_page(struct page *page, struct zone *zone) | |
3284 | { | |
3285 | VM_BUG_ON(PageActive(page)); | |
3286 | ||
3287 | retry: | |
3288 | ClearPageUnevictable(page); | |
3289 | if (page_evictable(page, NULL)) { | |
3290 | enum lru_list l = page_lru_base_type(page); | |
3291 | ||
3292 | __dec_zone_state(zone, NR_UNEVICTABLE); | |
3293 | list_move(&page->lru, &zone->lru[l].list); | |
3294 | mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l); | |
3295 | __inc_zone_state(zone, NR_INACTIVE_ANON + l); | |
3296 | __count_vm_event(UNEVICTABLE_PGRESCUED); | |
3297 | } else { | |
3298 | /* | |
3299 | * rotate unevictable list | |
3300 | */ | |
3301 | SetPageUnevictable(page); | |
3302 | list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list); | |
3303 | mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE); | |
3304 | if (page_evictable(page, NULL)) | |
3305 | goto retry; | |
3306 | } | |
3307 | } | |
3308 | ||
3309 | /** | |
3310 | * scan_mapping_unevictable_pages - scan an address space for evictable pages | |
3311 | * @mapping: struct address_space to scan for evictable pages | |
3312 | * | |
3313 | * Scan all pages in mapping. Check unevictable pages for | |
3314 | * evictability and move them to the appropriate zone lru list. | |
3315 | */ | |
3316 | void scan_mapping_unevictable_pages(struct address_space *mapping) | |
3317 | { | |
3318 | pgoff_t next = 0; | |
3319 | pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >> | |
3320 | PAGE_CACHE_SHIFT; | |
3321 | struct zone *zone; | |
3322 | struct pagevec pvec; | |
3323 | ||
3324 | if (mapping->nrpages == 0) | |
3325 | return; | |
3326 | ||
3327 | pagevec_init(&pvec, 0); | |
3328 | while (next < end && | |
3329 | pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { | |
3330 | int i; | |
3331 | int pg_scanned = 0; | |
3332 | ||
3333 | zone = NULL; | |
3334 | ||
3335 | for (i = 0; i < pagevec_count(&pvec); i++) { | |
3336 | struct page *page = pvec.pages[i]; | |
3337 | pgoff_t page_index = page->index; | |
3338 | struct zone *pagezone = page_zone(page); | |
3339 | ||
3340 | pg_scanned++; | |
3341 | if (page_index > next) | |
3342 | next = page_index; | |
3343 | next++; | |
3344 | ||
3345 | if (pagezone != zone) { | |
3346 | if (zone) | |
3347 | spin_unlock_irq(&zone->lru_lock); | |
3348 | zone = pagezone; | |
3349 | spin_lock_irq(&zone->lru_lock); | |
3350 | } | |
3351 | ||
3352 | if (PageLRU(page) && PageUnevictable(page)) | |
3353 | check_move_unevictable_page(page, zone); | |
3354 | } | |
3355 | if (zone) | |
3356 | spin_unlock_irq(&zone->lru_lock); | |
3357 | pagevec_release(&pvec); | |
3358 | ||
3359 | count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned); | |
3360 | } | |
3361 | ||
3362 | } | |
3363 | ||
3364 | /** | |
3365 | * scan_zone_unevictable_pages - check unevictable list for evictable pages | |
3366 | * @zone - zone of which to scan the unevictable list | |
3367 | * | |
3368 | * Scan @zone's unevictable LRU lists to check for pages that have become | |
3369 | * evictable. Move those that have to @zone's inactive list where they | |
3370 | * become candidates for reclaim, unless shrink_inactive_zone() decides | |
3371 | * to reactivate them. Pages that are still unevictable are rotated | |
3372 | * back onto @zone's unevictable list. | |
3373 | */ | |
3374 | #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */ | |
3375 | static void scan_zone_unevictable_pages(struct zone *zone) | |
3376 | { | |
3377 | struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list; | |
3378 | unsigned long scan; | |
3379 | unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE); | |
3380 | ||
3381 | while (nr_to_scan > 0) { | |
3382 | unsigned long batch_size = min(nr_to_scan, | |
3383 | SCAN_UNEVICTABLE_BATCH_SIZE); | |
3384 | ||
3385 | spin_lock_irq(&zone->lru_lock); | |
3386 | for (scan = 0; scan < batch_size; scan++) { | |
3387 | struct page *page = lru_to_page(l_unevictable); | |
3388 | ||
3389 | if (!trylock_page(page)) | |
3390 | continue; | |
3391 | ||
3392 | prefetchw_prev_lru_page(page, l_unevictable, flags); | |
3393 | ||
3394 | if (likely(PageLRU(page) && PageUnevictable(page))) | |
3395 | check_move_unevictable_page(page, zone); | |
3396 | ||
3397 | unlock_page(page); | |
3398 | } | |
3399 | spin_unlock_irq(&zone->lru_lock); | |
3400 | ||
3401 | nr_to_scan -= batch_size; | |
3402 | } | |
3403 | } | |
3404 | ||
3405 | ||
3406 | /** | |
3407 | * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages | |
3408 | * | |
3409 | * A really big hammer: scan all zones' unevictable LRU lists to check for | |
3410 | * pages that have become evictable. Move those back to the zones' | |
3411 | * inactive list where they become candidates for reclaim. | |
3412 | * This occurs when, e.g., we have unswappable pages on the unevictable lists, | |
3413 | * and we add swap to the system. As such, it runs in the context of a task | |
3414 | * that has possibly/probably made some previously unevictable pages | |
3415 | * evictable. | |
3416 | */ | |
3417 | static void scan_all_zones_unevictable_pages(void) | |
3418 | { | |
3419 | struct zone *zone; | |
3420 | ||
3421 | for_each_zone(zone) { | |
3422 | scan_zone_unevictable_pages(zone); | |
3423 | } | |
3424 | } | |
3425 | ||
3426 | /* | |
3427 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3428 | * all nodes' unevictable lists for evictable pages | |
3429 | */ | |
3430 | unsigned long scan_unevictable_pages; | |
3431 | ||
3432 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
3433 | void __user *buffer, | |
3434 | size_t *length, loff_t *ppos) | |
3435 | { | |
3436 | proc_doulongvec_minmax(table, write, buffer, length, ppos); | |
3437 | ||
3438 | if (write && *(unsigned long *)table->data) | |
3439 | scan_all_zones_unevictable_pages(); | |
3440 | ||
3441 | scan_unevictable_pages = 0; | |
3442 | return 0; | |
3443 | } | |
3444 | ||
3445 | #ifdef CONFIG_NUMA | |
3446 | /* | |
3447 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3448 | * a specified node's per zone unevictable lists for evictable pages. | |
3449 | */ | |
3450 | ||
3451 | static ssize_t read_scan_unevictable_node(struct sys_device *dev, | |
3452 | struct sysdev_attribute *attr, | |
3453 | char *buf) | |
3454 | { | |
3455 | return sprintf(buf, "0\n"); /* always zero; should fit... */ | |
3456 | } | |
3457 | ||
3458 | static ssize_t write_scan_unevictable_node(struct sys_device *dev, | |
3459 | struct sysdev_attribute *attr, | |
3460 | const char *buf, size_t count) | |
3461 | { | |
3462 | struct zone *node_zones = NODE_DATA(dev->id)->node_zones; | |
3463 | struct zone *zone; | |
3464 | unsigned long res; | |
3465 | unsigned long req = strict_strtoul(buf, 10, &res); | |
3466 | ||
3467 | if (!req) | |
3468 | return 1; /* zero is no-op */ | |
3469 | ||
3470 | for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { | |
3471 | if (!populated_zone(zone)) | |
3472 | continue; | |
3473 | scan_zone_unevictable_pages(zone); | |
3474 | } | |
3475 | return 1; | |
3476 | } | |
3477 | ||
3478 | ||
3479 | static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, | |
3480 | read_scan_unevictable_node, | |
3481 | write_scan_unevictable_node); | |
3482 | ||
3483 | int scan_unevictable_register_node(struct node *node) | |
3484 | { | |
3485 | return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages); | |
3486 | } | |
3487 | ||
3488 | void scan_unevictable_unregister_node(struct node *node) | |
3489 | { | |
3490 | sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages); | |
3491 | } | |
3492 | #endif |