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