]>
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/backing-dev.h> | |
30 | #include <linux/rmap.h> | |
31 | #include <linux/topology.h> | |
32 | #include <linux/cpu.h> | |
33 | #include <linux/cpuset.h> | |
34 | #include <linux/compaction.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 | #include <linux/oom.h> | |
44 | #include <linux/prefetch.h> | |
45 | ||
46 | #include <asm/tlbflush.h> | |
47 | #include <asm/div64.h> | |
48 | ||
49 | #include <linux/swapops.h> | |
50 | ||
51 | #include "internal.h" | |
52 | ||
53 | #define CREATE_TRACE_POINTS | |
54 | #include <trace/events/vmscan.h> | |
55 | ||
56 | struct scan_control { | |
57 | /* Incremented by the number of inactive pages that were scanned */ | |
58 | unsigned long nr_scanned; | |
59 | ||
60 | /* Number of pages freed so far during a call to shrink_zones() */ | |
61 | unsigned long nr_reclaimed; | |
62 | ||
63 | /* How many pages shrink_list() should reclaim */ | |
64 | unsigned long nr_to_reclaim; | |
65 | ||
66 | unsigned long hibernation_mode; | |
67 | ||
68 | /* This context's GFP mask */ | |
69 | gfp_t gfp_mask; | |
70 | ||
71 | int may_writepage; | |
72 | ||
73 | /* Can mapped pages be reclaimed? */ | |
74 | int may_unmap; | |
75 | ||
76 | /* Can pages be swapped as part of reclaim? */ | |
77 | int may_swap; | |
78 | ||
79 | int order; | |
80 | ||
81 | /* Scan (total_size >> priority) pages at once */ | |
82 | int priority; | |
83 | ||
84 | /* | |
85 | * The memory cgroup that hit its limit and as a result is the | |
86 | * primary target of this reclaim invocation. | |
87 | */ | |
88 | struct mem_cgroup *target_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_MEMCG | |
137 | static bool global_reclaim(struct scan_control *sc) | |
138 | { | |
139 | return !sc->target_mem_cgroup; | |
140 | } | |
141 | #else | |
142 | static bool global_reclaim(struct scan_control *sc) | |
143 | { | |
144 | return true; | |
145 | } | |
146 | #endif | |
147 | ||
148 | static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) | |
149 | { | |
150 | if (!mem_cgroup_disabled()) | |
151 | return mem_cgroup_get_lru_size(lruvec, lru); | |
152 | ||
153 | return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); | |
154 | } | |
155 | ||
156 | /* | |
157 | * Add a shrinker callback to be called from the vm | |
158 | */ | |
159 | void register_shrinker(struct shrinker *shrinker) | |
160 | { | |
161 | atomic_long_set(&shrinker->nr_in_batch, 0); | |
162 | down_write(&shrinker_rwsem); | |
163 | list_add_tail(&shrinker->list, &shrinker_list); | |
164 | up_write(&shrinker_rwsem); | |
165 | } | |
166 | EXPORT_SYMBOL(register_shrinker); | |
167 | ||
168 | /* | |
169 | * Remove one | |
170 | */ | |
171 | void unregister_shrinker(struct shrinker *shrinker) | |
172 | { | |
173 | down_write(&shrinker_rwsem); | |
174 | list_del(&shrinker->list); | |
175 | up_write(&shrinker_rwsem); | |
176 | } | |
177 | EXPORT_SYMBOL(unregister_shrinker); | |
178 | ||
179 | static inline int do_shrinker_shrink(struct shrinker *shrinker, | |
180 | struct shrink_control *sc, | |
181 | unsigned long nr_to_scan) | |
182 | { | |
183 | sc->nr_to_scan = nr_to_scan; | |
184 | return (*shrinker->shrink)(shrinker, sc); | |
185 | } | |
186 | ||
187 | #define SHRINK_BATCH 128 | |
188 | /* | |
189 | * Call the shrink functions to age shrinkable caches | |
190 | * | |
191 | * Here we assume it costs one seek to replace a lru page and that it also | |
192 | * takes a seek to recreate a cache object. With this in mind we age equal | |
193 | * percentages of the lru and ageable caches. This should balance the seeks | |
194 | * generated by these structures. | |
195 | * | |
196 | * If the vm encountered mapped pages on the LRU it increase the pressure on | |
197 | * slab to avoid swapping. | |
198 | * | |
199 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
200 | * | |
201 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
202 | * are eligible for the caller's allocation attempt. It is used for balancing | |
203 | * slab reclaim versus page reclaim. | |
204 | * | |
205 | * Returns the number of slab objects which we shrunk. | |
206 | */ | |
207 | unsigned long shrink_slab(struct shrink_control *shrink, | |
208 | unsigned long nr_pages_scanned, | |
209 | unsigned long lru_pages) | |
210 | { | |
211 | struct shrinker *shrinker; | |
212 | unsigned long ret = 0; | |
213 | ||
214 | if (nr_pages_scanned == 0) | |
215 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
216 | ||
217 | if (!down_read_trylock(&shrinker_rwsem)) { | |
218 | /* Assume we'll be able to shrink next time */ | |
219 | ret = 1; | |
220 | goto out; | |
221 | } | |
222 | ||
223 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
224 | unsigned long long delta; | |
225 | long total_scan; | |
226 | long max_pass; | |
227 | int shrink_ret = 0; | |
228 | long nr; | |
229 | long new_nr; | |
230 | long batch_size = shrinker->batch ? shrinker->batch | |
231 | : SHRINK_BATCH; | |
232 | ||
233 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); | |
234 | if (max_pass <= 0) | |
235 | continue; | |
236 | ||
237 | /* | |
238 | * copy the current shrinker scan count into a local variable | |
239 | * and zero it so that other concurrent shrinker invocations | |
240 | * don't also do this scanning work. | |
241 | */ | |
242 | nr = atomic_long_xchg(&shrinker->nr_in_batch, 0); | |
243 | ||
244 | total_scan = nr; | |
245 | delta = (4 * nr_pages_scanned) / shrinker->seeks; | |
246 | delta *= max_pass; | |
247 | do_div(delta, lru_pages + 1); | |
248 | total_scan += delta; | |
249 | if (total_scan < 0) { | |
250 | printk(KERN_ERR "shrink_slab: %pF negative objects to " | |
251 | "delete nr=%ld\n", | |
252 | shrinker->shrink, total_scan); | |
253 | total_scan = max_pass; | |
254 | } | |
255 | ||
256 | /* | |
257 | * We need to avoid excessive windup on filesystem shrinkers | |
258 | * due to large numbers of GFP_NOFS allocations causing the | |
259 | * shrinkers to return -1 all the time. This results in a large | |
260 | * nr being built up so when a shrink that can do some work | |
261 | * comes along it empties the entire cache due to nr >>> | |
262 | * max_pass. This is bad for sustaining a working set in | |
263 | * memory. | |
264 | * | |
265 | * Hence only allow the shrinker to scan the entire cache when | |
266 | * a large delta change is calculated directly. | |
267 | */ | |
268 | if (delta < max_pass / 4) | |
269 | total_scan = min(total_scan, max_pass / 2); | |
270 | ||
271 | /* | |
272 | * Avoid risking looping forever due to too large nr value: | |
273 | * never try to free more than twice the estimate number of | |
274 | * freeable entries. | |
275 | */ | |
276 | if (total_scan > max_pass * 2) | |
277 | total_scan = max_pass * 2; | |
278 | ||
279 | trace_mm_shrink_slab_start(shrinker, shrink, nr, | |
280 | nr_pages_scanned, lru_pages, | |
281 | max_pass, delta, total_scan); | |
282 | ||
283 | while (total_scan >= batch_size) { | |
284 | int nr_before; | |
285 | ||
286 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); | |
287 | shrink_ret = do_shrinker_shrink(shrinker, shrink, | |
288 | batch_size); | |
289 | if (shrink_ret == -1) | |
290 | break; | |
291 | if (shrink_ret < nr_before) | |
292 | ret += nr_before - shrink_ret; | |
293 | count_vm_events(SLABS_SCANNED, batch_size); | |
294 | total_scan -= batch_size; | |
295 | ||
296 | cond_resched(); | |
297 | } | |
298 | ||
299 | /* | |
300 | * move the unused scan count back into the shrinker in a | |
301 | * manner that handles concurrent updates. If we exhausted the | |
302 | * scan, there is no need to do an update. | |
303 | */ | |
304 | if (total_scan > 0) | |
305 | new_nr = atomic_long_add_return(total_scan, | |
306 | &shrinker->nr_in_batch); | |
307 | else | |
308 | new_nr = atomic_long_read(&shrinker->nr_in_batch); | |
309 | ||
310 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); | |
311 | } | |
312 | up_read(&shrinker_rwsem); | |
313 | out: | |
314 | cond_resched(); | |
315 | return ret; | |
316 | } | |
317 | ||
318 | static inline int is_page_cache_freeable(struct page *page) | |
319 | { | |
320 | /* | |
321 | * A freeable page cache page is referenced only by the caller | |
322 | * that isolated the page, the page cache radix tree and | |
323 | * optional buffer heads at page->private. | |
324 | */ | |
325 | return page_count(page) - page_has_private(page) == 2; | |
326 | } | |
327 | ||
328 | static int may_write_to_queue(struct backing_dev_info *bdi, | |
329 | struct scan_control *sc) | |
330 | { | |
331 | if (current->flags & PF_SWAPWRITE) | |
332 | return 1; | |
333 | if (!bdi_write_congested(bdi)) | |
334 | return 1; | |
335 | if (bdi == current->backing_dev_info) | |
336 | return 1; | |
337 | return 0; | |
338 | } | |
339 | ||
340 | /* | |
341 | * We detected a synchronous write error writing a page out. Probably | |
342 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
343 | * fsync(), msync() or close(). | |
344 | * | |
345 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
346 | * prevents it from being freed up. But we have a ref on the page and once | |
347 | * that page is locked, the mapping is pinned. | |
348 | * | |
349 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
350 | * __GFP_FS. | |
351 | */ | |
352 | static void handle_write_error(struct address_space *mapping, | |
353 | struct page *page, int error) | |
354 | { | |
355 | lock_page(page); | |
356 | if (page_mapping(page) == mapping) | |
357 | mapping_set_error(mapping, error); | |
358 | unlock_page(page); | |
359 | } | |
360 | ||
361 | /* possible outcome of pageout() */ | |
362 | typedef enum { | |
363 | /* failed to write page out, page is locked */ | |
364 | PAGE_KEEP, | |
365 | /* move page to the active list, page is locked */ | |
366 | PAGE_ACTIVATE, | |
367 | /* page has been sent to the disk successfully, page is unlocked */ | |
368 | PAGE_SUCCESS, | |
369 | /* page is clean and locked */ | |
370 | PAGE_CLEAN, | |
371 | } pageout_t; | |
372 | ||
373 | /* | |
374 | * pageout is called by shrink_page_list() for each dirty page. | |
375 | * Calls ->writepage(). | |
376 | */ | |
377 | static pageout_t pageout(struct page *page, struct address_space *mapping, | |
378 | struct scan_control *sc) | |
379 | { | |
380 | /* | |
381 | * If the page is dirty, only perform writeback if that write | |
382 | * will be non-blocking. To prevent this allocation from being | |
383 | * stalled by pagecache activity. But note that there may be | |
384 | * stalls if we need to run get_block(). We could test | |
385 | * PagePrivate for that. | |
386 | * | |
387 | * If this process is currently in __generic_file_aio_write() against | |
388 | * this page's queue, we can perform writeback even if that | |
389 | * will block. | |
390 | * | |
391 | * If the page is swapcache, write it back even if that would | |
392 | * block, for some throttling. This happens by accident, because | |
393 | * swap_backing_dev_info is bust: it doesn't reflect the | |
394 | * congestion state of the swapdevs. Easy to fix, if needed. | |
395 | */ | |
396 | if (!is_page_cache_freeable(page)) | |
397 | return PAGE_KEEP; | |
398 | if (!mapping) { | |
399 | /* | |
400 | * Some data journaling orphaned pages can have | |
401 | * page->mapping == NULL while being dirty with clean buffers. | |
402 | */ | |
403 | if (page_has_private(page)) { | |
404 | if (try_to_free_buffers(page)) { | |
405 | ClearPageDirty(page); | |
406 | printk("%s: orphaned page\n", __func__); | |
407 | return PAGE_CLEAN; | |
408 | } | |
409 | } | |
410 | return PAGE_KEEP; | |
411 | } | |
412 | if (mapping->a_ops->writepage == NULL) | |
413 | return PAGE_ACTIVATE; | |
414 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) | |
415 | return PAGE_KEEP; | |
416 | ||
417 | if (clear_page_dirty_for_io(page)) { | |
418 | int res; | |
419 | struct writeback_control wbc = { | |
420 | .sync_mode = WB_SYNC_NONE, | |
421 | .nr_to_write = SWAP_CLUSTER_MAX, | |
422 | .range_start = 0, | |
423 | .range_end = LLONG_MAX, | |
424 | .for_reclaim = 1, | |
425 | }; | |
426 | ||
427 | SetPageReclaim(page); | |
428 | res = mapping->a_ops->writepage(page, &wbc); | |
429 | if (res < 0) | |
430 | handle_write_error(mapping, page, res); | |
431 | if (res == AOP_WRITEPAGE_ACTIVATE) { | |
432 | ClearPageReclaim(page); | |
433 | return PAGE_ACTIVATE; | |
434 | } | |
435 | ||
436 | if (!PageWriteback(page)) { | |
437 | /* synchronous write or broken a_ops? */ | |
438 | ClearPageReclaim(page); | |
439 | } | |
440 | trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); | |
441 | inc_zone_page_state(page, NR_VMSCAN_WRITE); | |
442 | return PAGE_SUCCESS; | |
443 | } | |
444 | ||
445 | return PAGE_CLEAN; | |
446 | } | |
447 | ||
448 | /* | |
449 | * Same as remove_mapping, but if the page is removed from the mapping, it | |
450 | * gets returned with a refcount of 0. | |
451 | */ | |
452 | static int __remove_mapping(struct address_space *mapping, struct page *page) | |
453 | { | |
454 | BUG_ON(!PageLocked(page)); | |
455 | BUG_ON(mapping != page_mapping(page)); | |
456 | ||
457 | spin_lock_irq(&mapping->tree_lock); | |
458 | /* | |
459 | * The non racy check for a busy page. | |
460 | * | |
461 | * Must be careful with the order of the tests. When someone has | |
462 | * a ref to the page, it may be possible that they dirty it then | |
463 | * drop the reference. So if PageDirty is tested before page_count | |
464 | * here, then the following race may occur: | |
465 | * | |
466 | * get_user_pages(&page); | |
467 | * [user mapping goes away] | |
468 | * write_to(page); | |
469 | * !PageDirty(page) [good] | |
470 | * SetPageDirty(page); | |
471 | * put_page(page); | |
472 | * !page_count(page) [good, discard it] | |
473 | * | |
474 | * [oops, our write_to data is lost] | |
475 | * | |
476 | * Reversing the order of the tests ensures such a situation cannot | |
477 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
478 | * load is not satisfied before that of page->_count. | |
479 | * | |
480 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
481 | * and thus under tree_lock, then this ordering is not required. | |
482 | */ | |
483 | if (!page_freeze_refs(page, 2)) | |
484 | goto cannot_free; | |
485 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ | |
486 | if (unlikely(PageDirty(page))) { | |
487 | page_unfreeze_refs(page, 2); | |
488 | goto cannot_free; | |
489 | } | |
490 | ||
491 | if (PageSwapCache(page)) { | |
492 | swp_entry_t swap = { .val = page_private(page) }; | |
493 | __delete_from_swap_cache(page); | |
494 | spin_unlock_irq(&mapping->tree_lock); | |
495 | swapcache_free(swap, page); | |
496 | } else { | |
497 | void (*freepage)(struct page *); | |
498 | ||
499 | freepage = mapping->a_ops->freepage; | |
500 | ||
501 | __delete_from_page_cache(page); | |
502 | spin_unlock_irq(&mapping->tree_lock); | |
503 | mem_cgroup_uncharge_cache_page(page); | |
504 | ||
505 | if (freepage != NULL) | |
506 | freepage(page); | |
507 | } | |
508 | ||
509 | return 1; | |
510 | ||
511 | cannot_free: | |
512 | spin_unlock_irq(&mapping->tree_lock); | |
513 | return 0; | |
514 | } | |
515 | ||
516 | /* | |
517 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
518 | * someone else has a ref on the page, abort and return 0. If it was | |
519 | * successfully detached, return 1. Assumes the caller has a single ref on | |
520 | * this page. | |
521 | */ | |
522 | int remove_mapping(struct address_space *mapping, struct page *page) | |
523 | { | |
524 | if (__remove_mapping(mapping, page)) { | |
525 | /* | |
526 | * Unfreezing the refcount with 1 rather than 2 effectively | |
527 | * drops the pagecache ref for us without requiring another | |
528 | * atomic operation. | |
529 | */ | |
530 | page_unfreeze_refs(page, 1); | |
531 | return 1; | |
532 | } | |
533 | return 0; | |
534 | } | |
535 | ||
536 | /** | |
537 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
538 | * @page: page to be put back to appropriate lru list | |
539 | * | |
540 | * Add previously isolated @page to appropriate LRU list. | |
541 | * Page may still be unevictable for other reasons. | |
542 | * | |
543 | * lru_lock must not be held, interrupts must be enabled. | |
544 | */ | |
545 | void putback_lru_page(struct page *page) | |
546 | { | |
547 | int lru; | |
548 | int active = !!TestClearPageActive(page); | |
549 | int was_unevictable = PageUnevictable(page); | |
550 | ||
551 | VM_BUG_ON(PageLRU(page)); | |
552 | ||
553 | redo: | |
554 | ClearPageUnevictable(page); | |
555 | ||
556 | if (page_evictable(page, NULL)) { | |
557 | /* | |
558 | * For evictable pages, we can use the cache. | |
559 | * In event of a race, worst case is we end up with an | |
560 | * unevictable page on [in]active list. | |
561 | * We know how to handle that. | |
562 | */ | |
563 | lru = active + page_lru_base_type(page); | |
564 | lru_cache_add_lru(page, lru); | |
565 | } else { | |
566 | /* | |
567 | * Put unevictable pages directly on zone's unevictable | |
568 | * list. | |
569 | */ | |
570 | lru = LRU_UNEVICTABLE; | |
571 | add_page_to_unevictable_list(page); | |
572 | /* | |
573 | * When racing with an mlock or AS_UNEVICTABLE clearing | |
574 | * (page is unlocked) make sure that if the other thread | |
575 | * does not observe our setting of PG_lru and fails | |
576 | * isolation/check_move_unevictable_pages, | |
577 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move | |
578 | * the page back to the evictable list. | |
579 | * | |
580 | * The other side is TestClearPageMlocked() or shmem_lock(). | |
581 | */ | |
582 | smp_mb(); | |
583 | } | |
584 | ||
585 | /* | |
586 | * page's status can change while we move it among lru. If an evictable | |
587 | * page is on unevictable list, it never be freed. To avoid that, | |
588 | * check after we added it to the list, again. | |
589 | */ | |
590 | if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { | |
591 | if (!isolate_lru_page(page)) { | |
592 | put_page(page); | |
593 | goto redo; | |
594 | } | |
595 | /* This means someone else dropped this page from LRU | |
596 | * So, it will be freed or putback to LRU again. There is | |
597 | * nothing to do here. | |
598 | */ | |
599 | } | |
600 | ||
601 | if (was_unevictable && lru != LRU_UNEVICTABLE) | |
602 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
603 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) | |
604 | count_vm_event(UNEVICTABLE_PGCULLED); | |
605 | ||
606 | put_page(page); /* drop ref from isolate */ | |
607 | } | |
608 | ||
609 | enum page_references { | |
610 | PAGEREF_RECLAIM, | |
611 | PAGEREF_RECLAIM_CLEAN, | |
612 | PAGEREF_KEEP, | |
613 | PAGEREF_ACTIVATE, | |
614 | }; | |
615 | ||
616 | static enum page_references page_check_references(struct page *page, | |
617 | struct scan_control *sc) | |
618 | { | |
619 | int referenced_ptes, referenced_page; | |
620 | unsigned long vm_flags; | |
621 | ||
622 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, | |
623 | &vm_flags); | |
624 | referenced_page = TestClearPageReferenced(page); | |
625 | ||
626 | /* | |
627 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
628 | * move the page to the unevictable list. | |
629 | */ | |
630 | if (vm_flags & VM_LOCKED) | |
631 | return PAGEREF_RECLAIM; | |
632 | ||
633 | if (referenced_ptes) { | |
634 | if (PageSwapBacked(page)) | |
635 | return PAGEREF_ACTIVATE; | |
636 | /* | |
637 | * All mapped pages start out with page table | |
638 | * references from the instantiating fault, so we need | |
639 | * to look twice if a mapped file page is used more | |
640 | * than once. | |
641 | * | |
642 | * Mark it and spare it for another trip around the | |
643 | * inactive list. Another page table reference will | |
644 | * lead to its activation. | |
645 | * | |
646 | * Note: the mark is set for activated pages as well | |
647 | * so that recently deactivated but used pages are | |
648 | * quickly recovered. | |
649 | */ | |
650 | SetPageReferenced(page); | |
651 | ||
652 | if (referenced_page || referenced_ptes > 1) | |
653 | return PAGEREF_ACTIVATE; | |
654 | ||
655 | /* | |
656 | * Activate file-backed executable pages after first usage. | |
657 | */ | |
658 | if (vm_flags & VM_EXEC) | |
659 | return PAGEREF_ACTIVATE; | |
660 | ||
661 | return PAGEREF_KEEP; | |
662 | } | |
663 | ||
664 | /* Reclaim if clean, defer dirty pages to writeback */ | |
665 | if (referenced_page && !PageSwapBacked(page)) | |
666 | return PAGEREF_RECLAIM_CLEAN; | |
667 | ||
668 | return PAGEREF_RECLAIM; | |
669 | } | |
670 | ||
671 | /* | |
672 | * shrink_page_list() returns the number of reclaimed pages | |
673 | */ | |
674 | static unsigned long shrink_page_list(struct list_head *page_list, | |
675 | struct zone *zone, | |
676 | struct scan_control *sc, | |
677 | enum ttu_flags ttu_flags, | |
678 | unsigned long *ret_nr_dirty, | |
679 | unsigned long *ret_nr_writeback, | |
680 | bool force_reclaim) | |
681 | { | |
682 | LIST_HEAD(ret_pages); | |
683 | LIST_HEAD(free_pages); | |
684 | int pgactivate = 0; | |
685 | unsigned long nr_dirty = 0; | |
686 | unsigned long nr_congested = 0; | |
687 | unsigned long nr_reclaimed = 0; | |
688 | unsigned long nr_writeback = 0; | |
689 | ||
690 | cond_resched(); | |
691 | ||
692 | mem_cgroup_uncharge_start(); | |
693 | while (!list_empty(page_list)) { | |
694 | struct address_space *mapping; | |
695 | struct page *page; | |
696 | int may_enter_fs; | |
697 | enum page_references references = PAGEREF_RECLAIM_CLEAN; | |
698 | ||
699 | cond_resched(); | |
700 | ||
701 | page = lru_to_page(page_list); | |
702 | list_del(&page->lru); | |
703 | ||
704 | if (!trylock_page(page)) | |
705 | goto keep; | |
706 | ||
707 | VM_BUG_ON(PageActive(page)); | |
708 | VM_BUG_ON(page_zone(page) != zone); | |
709 | ||
710 | sc->nr_scanned++; | |
711 | ||
712 | if (unlikely(!page_evictable(page, NULL))) | |
713 | goto cull_mlocked; | |
714 | ||
715 | if (!sc->may_unmap && page_mapped(page)) | |
716 | goto keep_locked; | |
717 | ||
718 | /* Double the slab pressure for mapped and swapcache pages */ | |
719 | if (page_mapped(page) || PageSwapCache(page)) | |
720 | sc->nr_scanned++; | |
721 | ||
722 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || | |
723 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
724 | ||
725 | if (PageWriteback(page)) { | |
726 | /* | |
727 | * memcg doesn't have any dirty pages throttling so we | |
728 | * could easily OOM just because too many pages are in | |
729 | * writeback and there is nothing else to reclaim. | |
730 | * | |
731 | * Check __GFP_IO, certainly because a loop driver | |
732 | * thread might enter reclaim, and deadlock if it waits | |
733 | * on a page for which it is needed to do the write | |
734 | * (loop masks off __GFP_IO|__GFP_FS for this reason); | |
735 | * but more thought would probably show more reasons. | |
736 | * | |
737 | * Don't require __GFP_FS, since we're not going into | |
738 | * the FS, just waiting on its writeback completion. | |
739 | * Worryingly, ext4 gfs2 and xfs allocate pages with | |
740 | * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so | |
741 | * testing may_enter_fs here is liable to OOM on them. | |
742 | */ | |
743 | if (global_reclaim(sc) || | |
744 | !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) { | |
745 | /* | |
746 | * This is slightly racy - end_page_writeback() | |
747 | * might have just cleared PageReclaim, then | |
748 | * setting PageReclaim here end up interpreted | |
749 | * as PageReadahead - but that does not matter | |
750 | * enough to care. What we do want is for this | |
751 | * page to have PageReclaim set next time memcg | |
752 | * reclaim reaches the tests above, so it will | |
753 | * then wait_on_page_writeback() to avoid OOM; | |
754 | * and it's also appropriate in global reclaim. | |
755 | */ | |
756 | SetPageReclaim(page); | |
757 | nr_writeback++; | |
758 | goto keep_locked; | |
759 | } | |
760 | wait_on_page_writeback(page); | |
761 | } | |
762 | ||
763 | if (!force_reclaim) | |
764 | references = page_check_references(page, sc); | |
765 | ||
766 | switch (references) { | |
767 | case PAGEREF_ACTIVATE: | |
768 | goto activate_locked; | |
769 | case PAGEREF_KEEP: | |
770 | goto keep_locked; | |
771 | case PAGEREF_RECLAIM: | |
772 | case PAGEREF_RECLAIM_CLEAN: | |
773 | ; /* try to reclaim the page below */ | |
774 | } | |
775 | ||
776 | /* | |
777 | * Anonymous process memory has backing store? | |
778 | * Try to allocate it some swap space here. | |
779 | */ | |
780 | if (PageAnon(page) && !PageSwapCache(page)) { | |
781 | if (!(sc->gfp_mask & __GFP_IO)) | |
782 | goto keep_locked; | |
783 | if (!add_to_swap(page)) | |
784 | goto activate_locked; | |
785 | may_enter_fs = 1; | |
786 | } | |
787 | ||
788 | mapping = page_mapping(page); | |
789 | ||
790 | /* | |
791 | * The page is mapped into the page tables of one or more | |
792 | * processes. Try to unmap it here. | |
793 | */ | |
794 | if (page_mapped(page) && mapping) { | |
795 | switch (try_to_unmap(page, ttu_flags)) { | |
796 | case SWAP_FAIL: | |
797 | goto activate_locked; | |
798 | case SWAP_AGAIN: | |
799 | goto keep_locked; | |
800 | case SWAP_MLOCK: | |
801 | goto cull_mlocked; | |
802 | case SWAP_SUCCESS: | |
803 | ; /* try to free the page below */ | |
804 | } | |
805 | } | |
806 | ||
807 | if (PageDirty(page)) { | |
808 | nr_dirty++; | |
809 | ||
810 | /* | |
811 | * Only kswapd can writeback filesystem pages to | |
812 | * avoid risk of stack overflow but do not writeback | |
813 | * unless under significant pressure. | |
814 | */ | |
815 | if (page_is_file_cache(page) && | |
816 | (!current_is_kswapd() || | |
817 | sc->priority >= DEF_PRIORITY - 2)) { | |
818 | /* | |
819 | * Immediately reclaim when written back. | |
820 | * Similar in principal to deactivate_page() | |
821 | * except we already have the page isolated | |
822 | * and know it's dirty | |
823 | */ | |
824 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
825 | SetPageReclaim(page); | |
826 | ||
827 | goto keep_locked; | |
828 | } | |
829 | ||
830 | if (references == PAGEREF_RECLAIM_CLEAN) | |
831 | goto keep_locked; | |
832 | if (!may_enter_fs) | |
833 | goto keep_locked; | |
834 | if (!sc->may_writepage) | |
835 | goto keep_locked; | |
836 | ||
837 | /* Page is dirty, try to write it out here */ | |
838 | switch (pageout(page, mapping, sc)) { | |
839 | case PAGE_KEEP: | |
840 | nr_congested++; | |
841 | goto keep_locked; | |
842 | case PAGE_ACTIVATE: | |
843 | goto activate_locked; | |
844 | case PAGE_SUCCESS: | |
845 | if (PageWriteback(page)) | |
846 | goto keep; | |
847 | if (PageDirty(page)) | |
848 | goto keep; | |
849 | ||
850 | /* | |
851 | * A synchronous write - probably a ramdisk. Go | |
852 | * ahead and try to reclaim the page. | |
853 | */ | |
854 | if (!trylock_page(page)) | |
855 | goto keep; | |
856 | if (PageDirty(page) || PageWriteback(page)) | |
857 | goto keep_locked; | |
858 | mapping = page_mapping(page); | |
859 | case PAGE_CLEAN: | |
860 | ; /* try to free the page below */ | |
861 | } | |
862 | } | |
863 | ||
864 | /* | |
865 | * If the page has buffers, try to free the buffer mappings | |
866 | * associated with this page. If we succeed we try to free | |
867 | * the page as well. | |
868 | * | |
869 | * We do this even if the page is PageDirty(). | |
870 | * try_to_release_page() does not perform I/O, but it is | |
871 | * possible for a page to have PageDirty set, but it is actually | |
872 | * clean (all its buffers are clean). This happens if the | |
873 | * buffers were written out directly, with submit_bh(). ext3 | |
874 | * will do this, as well as the blockdev mapping. | |
875 | * try_to_release_page() will discover that cleanness and will | |
876 | * drop the buffers and mark the page clean - it can be freed. | |
877 | * | |
878 | * Rarely, pages can have buffers and no ->mapping. These are | |
879 | * the pages which were not successfully invalidated in | |
880 | * truncate_complete_page(). We try to drop those buffers here | |
881 | * and if that worked, and the page is no longer mapped into | |
882 | * process address space (page_count == 1) it can be freed. | |
883 | * Otherwise, leave the page on the LRU so it is swappable. | |
884 | */ | |
885 | if (page_has_private(page)) { | |
886 | if (!try_to_release_page(page, sc->gfp_mask)) | |
887 | goto activate_locked; | |
888 | if (!mapping && page_count(page) == 1) { | |
889 | unlock_page(page); | |
890 | if (put_page_testzero(page)) | |
891 | goto free_it; | |
892 | else { | |
893 | /* | |
894 | * rare race with speculative reference. | |
895 | * the speculative reference will free | |
896 | * this page shortly, so we may | |
897 | * increment nr_reclaimed here (and | |
898 | * leave it off the LRU). | |
899 | */ | |
900 | nr_reclaimed++; | |
901 | continue; | |
902 | } | |
903 | } | |
904 | } | |
905 | ||
906 | if (!mapping || !__remove_mapping(mapping, page)) | |
907 | goto keep_locked; | |
908 | ||
909 | /* | |
910 | * At this point, we have no other references and there is | |
911 | * no way to pick any more up (removed from LRU, removed | |
912 | * from pagecache). Can use non-atomic bitops now (and | |
913 | * we obviously don't have to worry about waking up a process | |
914 | * waiting on the page lock, because there are no references. | |
915 | */ | |
916 | __clear_page_locked(page); | |
917 | free_it: | |
918 | nr_reclaimed++; | |
919 | ||
920 | /* | |
921 | * Is there need to periodically free_page_list? It would | |
922 | * appear not as the counts should be low | |
923 | */ | |
924 | list_add(&page->lru, &free_pages); | |
925 | continue; | |
926 | ||
927 | cull_mlocked: | |
928 | if (PageSwapCache(page)) | |
929 | try_to_free_swap(page); | |
930 | unlock_page(page); | |
931 | putback_lru_page(page); | |
932 | continue; | |
933 | ||
934 | activate_locked: | |
935 | /* Not a candidate for swapping, so reclaim swap space. */ | |
936 | if (PageSwapCache(page) && vm_swap_full()) | |
937 | try_to_free_swap(page); | |
938 | VM_BUG_ON(PageActive(page)); | |
939 | SetPageActive(page); | |
940 | pgactivate++; | |
941 | keep_locked: | |
942 | unlock_page(page); | |
943 | keep: | |
944 | list_add(&page->lru, &ret_pages); | |
945 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); | |
946 | } | |
947 | ||
948 | /* | |
949 | * Tag a zone as congested if all the dirty pages encountered were | |
950 | * backed by a congested BDI. In this case, reclaimers should just | |
951 | * back off and wait for congestion to clear because further reclaim | |
952 | * will encounter the same problem | |
953 | */ | |
954 | if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc)) | |
955 | zone_set_flag(zone, ZONE_CONGESTED); | |
956 | ||
957 | free_hot_cold_page_list(&free_pages, 1); | |
958 | ||
959 | list_splice(&ret_pages, page_list); | |
960 | count_vm_events(PGACTIVATE, pgactivate); | |
961 | mem_cgroup_uncharge_end(); | |
962 | *ret_nr_dirty += nr_dirty; | |
963 | *ret_nr_writeback += nr_writeback; | |
964 | return nr_reclaimed; | |
965 | } | |
966 | ||
967 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, | |
968 | struct list_head *page_list) | |
969 | { | |
970 | struct scan_control sc = { | |
971 | .gfp_mask = GFP_KERNEL, | |
972 | .priority = DEF_PRIORITY, | |
973 | .may_unmap = 1, | |
974 | }; | |
975 | unsigned long ret, dummy1, dummy2; | |
976 | struct page *page, *next; | |
977 | LIST_HEAD(clean_pages); | |
978 | ||
979 | list_for_each_entry_safe(page, next, page_list, lru) { | |
980 | if (page_is_file_cache(page) && !PageDirty(page)) { | |
981 | ClearPageActive(page); | |
982 | list_move(&page->lru, &clean_pages); | |
983 | } | |
984 | } | |
985 | ||
986 | ret = shrink_page_list(&clean_pages, zone, &sc, | |
987 | TTU_UNMAP|TTU_IGNORE_ACCESS, | |
988 | &dummy1, &dummy2, true); | |
989 | list_splice(&clean_pages, page_list); | |
990 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); | |
991 | return ret; | |
992 | } | |
993 | ||
994 | /* | |
995 | * Attempt to remove the specified page from its LRU. Only take this page | |
996 | * if it is of the appropriate PageActive status. Pages which are being | |
997 | * freed elsewhere are also ignored. | |
998 | * | |
999 | * page: page to consider | |
1000 | * mode: one of the LRU isolation modes defined above | |
1001 | * | |
1002 | * returns 0 on success, -ve errno on failure. | |
1003 | */ | |
1004 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) | |
1005 | { | |
1006 | int ret = -EINVAL; | |
1007 | ||
1008 | /* Only take pages on the LRU. */ | |
1009 | if (!PageLRU(page)) | |
1010 | return ret; | |
1011 | ||
1012 | /* Do not give back unevictable pages for compaction */ | |
1013 | if (PageUnevictable(page)) | |
1014 | return ret; | |
1015 | ||
1016 | ret = -EBUSY; | |
1017 | ||
1018 | /* | |
1019 | * To minimise LRU disruption, the caller can indicate that it only | |
1020 | * wants to isolate pages it will be able to operate on without | |
1021 | * blocking - clean pages for the most part. | |
1022 | * | |
1023 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1024 | * is used by reclaim when it is cannot write to backing storage | |
1025 | * | |
1026 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1027 | * that it is possible to migrate without blocking | |
1028 | */ | |
1029 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1030 | /* All the caller can do on PageWriteback is block */ | |
1031 | if (PageWriteback(page)) | |
1032 | return ret; | |
1033 | ||
1034 | if (PageDirty(page)) { | |
1035 | struct address_space *mapping; | |
1036 | ||
1037 | /* ISOLATE_CLEAN means only clean pages */ | |
1038 | if (mode & ISOLATE_CLEAN) | |
1039 | return ret; | |
1040 | ||
1041 | /* | |
1042 | * Only pages without mappings or that have a | |
1043 | * ->migratepage callback are possible to migrate | |
1044 | * without blocking | |
1045 | */ | |
1046 | mapping = page_mapping(page); | |
1047 | if (mapping && !mapping->a_ops->migratepage) | |
1048 | return ret; | |
1049 | } | |
1050 | } | |
1051 | ||
1052 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) | |
1053 | return ret; | |
1054 | ||
1055 | if (likely(get_page_unless_zero(page))) { | |
1056 | /* | |
1057 | * Be careful not to clear PageLRU until after we're | |
1058 | * sure the page is not being freed elsewhere -- the | |
1059 | * page release code relies on it. | |
1060 | */ | |
1061 | ClearPageLRU(page); | |
1062 | ret = 0; | |
1063 | } | |
1064 | ||
1065 | return ret; | |
1066 | } | |
1067 | ||
1068 | /* | |
1069 | * zone->lru_lock is heavily contended. Some of the functions that | |
1070 | * shrink the lists perform better by taking out a batch of pages | |
1071 | * and working on them outside the LRU lock. | |
1072 | * | |
1073 | * For pagecache intensive workloads, this function is the hottest | |
1074 | * spot in the kernel (apart from copy_*_user functions). | |
1075 | * | |
1076 | * Appropriate locks must be held before calling this function. | |
1077 | * | |
1078 | * @nr_to_scan: The number of pages to look through on the list. | |
1079 | * @lruvec: The LRU vector to pull pages from. | |
1080 | * @dst: The temp list to put pages on to. | |
1081 | * @nr_scanned: The number of pages that were scanned. | |
1082 | * @sc: The scan_control struct for this reclaim session | |
1083 | * @mode: One of the LRU isolation modes | |
1084 | * @lru: LRU list id for isolating | |
1085 | * | |
1086 | * returns how many pages were moved onto *@dst. | |
1087 | */ | |
1088 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, | |
1089 | struct lruvec *lruvec, struct list_head *dst, | |
1090 | unsigned long *nr_scanned, struct scan_control *sc, | |
1091 | isolate_mode_t mode, enum lru_list lru) | |
1092 | { | |
1093 | struct list_head *src = &lruvec->lists[lru]; | |
1094 | unsigned long nr_taken = 0; | |
1095 | unsigned long scan; | |
1096 | ||
1097 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { | |
1098 | struct page *page; | |
1099 | int nr_pages; | |
1100 | ||
1101 | page = lru_to_page(src); | |
1102 | prefetchw_prev_lru_page(page, src, flags); | |
1103 | ||
1104 | VM_BUG_ON(!PageLRU(page)); | |
1105 | ||
1106 | switch (__isolate_lru_page(page, mode)) { | |
1107 | case 0: | |
1108 | nr_pages = hpage_nr_pages(page); | |
1109 | mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); | |
1110 | list_move(&page->lru, dst); | |
1111 | nr_taken += nr_pages; | |
1112 | break; | |
1113 | ||
1114 | case -EBUSY: | |
1115 | /* else it is being freed elsewhere */ | |
1116 | list_move(&page->lru, src); | |
1117 | continue; | |
1118 | ||
1119 | default: | |
1120 | BUG(); | |
1121 | } | |
1122 | } | |
1123 | ||
1124 | *nr_scanned = scan; | |
1125 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, | |
1126 | nr_taken, mode, is_file_lru(lru)); | |
1127 | return nr_taken; | |
1128 | } | |
1129 | ||
1130 | /** | |
1131 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1132 | * @page: page to isolate from its LRU list | |
1133 | * | |
1134 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1135 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1136 | * | |
1137 | * Returns 0 if the page was removed from an LRU list. | |
1138 | * Returns -EBUSY if the page was not on an LRU list. | |
1139 | * | |
1140 | * The returned page will have PageLRU() cleared. If it was found on | |
1141 | * the active list, it will have PageActive set. If it was found on | |
1142 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1143 | * may need to be cleared by the caller before letting the page go. | |
1144 | * | |
1145 | * The vmstat statistic corresponding to the list on which the page was | |
1146 | * found will be decremented. | |
1147 | * | |
1148 | * Restrictions: | |
1149 | * (1) Must be called with an elevated refcount on the page. This is a | |
1150 | * fundamentnal difference from isolate_lru_pages (which is called | |
1151 | * without a stable reference). | |
1152 | * (2) the lru_lock must not be held. | |
1153 | * (3) interrupts must be enabled. | |
1154 | */ | |
1155 | int isolate_lru_page(struct page *page) | |
1156 | { | |
1157 | int ret = -EBUSY; | |
1158 | ||
1159 | VM_BUG_ON(!page_count(page)); | |
1160 | ||
1161 | if (PageLRU(page)) { | |
1162 | struct zone *zone = page_zone(page); | |
1163 | struct lruvec *lruvec; | |
1164 | ||
1165 | spin_lock_irq(&zone->lru_lock); | |
1166 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1167 | if (PageLRU(page)) { | |
1168 | int lru = page_lru(page); | |
1169 | get_page(page); | |
1170 | ClearPageLRU(page); | |
1171 | del_page_from_lru_list(page, lruvec, lru); | |
1172 | ret = 0; | |
1173 | } | |
1174 | spin_unlock_irq(&zone->lru_lock); | |
1175 | } | |
1176 | return ret; | |
1177 | } | |
1178 | ||
1179 | /* | |
1180 | * Are there way too many processes in the direct reclaim path already? | |
1181 | */ | |
1182 | static int too_many_isolated(struct zone *zone, int file, | |
1183 | struct scan_control *sc) | |
1184 | { | |
1185 | unsigned long inactive, isolated; | |
1186 | ||
1187 | if (current_is_kswapd()) | |
1188 | return 0; | |
1189 | ||
1190 | if (!global_reclaim(sc)) | |
1191 | return 0; | |
1192 | ||
1193 | if (file) { | |
1194 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1195 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1196 | } else { | |
1197 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1198 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1199 | } | |
1200 | ||
1201 | return isolated > inactive; | |
1202 | } | |
1203 | ||
1204 | static noinline_for_stack void | |
1205 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) | |
1206 | { | |
1207 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1208 | struct zone *zone = lruvec_zone(lruvec); | |
1209 | LIST_HEAD(pages_to_free); | |
1210 | ||
1211 | /* | |
1212 | * Put back any unfreeable pages. | |
1213 | */ | |
1214 | while (!list_empty(page_list)) { | |
1215 | struct page *page = lru_to_page(page_list); | |
1216 | int lru; | |
1217 | ||
1218 | VM_BUG_ON(PageLRU(page)); | |
1219 | list_del(&page->lru); | |
1220 | if (unlikely(!page_evictable(page, NULL))) { | |
1221 | spin_unlock_irq(&zone->lru_lock); | |
1222 | putback_lru_page(page); | |
1223 | spin_lock_irq(&zone->lru_lock); | |
1224 | continue; | |
1225 | } | |
1226 | ||
1227 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1228 | ||
1229 | SetPageLRU(page); | |
1230 | lru = page_lru(page); | |
1231 | add_page_to_lru_list(page, lruvec, lru); | |
1232 | ||
1233 | if (is_active_lru(lru)) { | |
1234 | int file = is_file_lru(lru); | |
1235 | int numpages = hpage_nr_pages(page); | |
1236 | reclaim_stat->recent_rotated[file] += numpages; | |
1237 | } | |
1238 | if (put_page_testzero(page)) { | |
1239 | __ClearPageLRU(page); | |
1240 | __ClearPageActive(page); | |
1241 | del_page_from_lru_list(page, lruvec, lru); | |
1242 | ||
1243 | if (unlikely(PageCompound(page))) { | |
1244 | spin_unlock_irq(&zone->lru_lock); | |
1245 | (*get_compound_page_dtor(page))(page); | |
1246 | spin_lock_irq(&zone->lru_lock); | |
1247 | } else | |
1248 | list_add(&page->lru, &pages_to_free); | |
1249 | } | |
1250 | } | |
1251 | ||
1252 | /* | |
1253 | * To save our caller's stack, now use input list for pages to free. | |
1254 | */ | |
1255 | list_splice(&pages_to_free, page_list); | |
1256 | } | |
1257 | ||
1258 | /* | |
1259 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number | |
1260 | * of reclaimed pages | |
1261 | */ | |
1262 | static noinline_for_stack unsigned long | |
1263 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, | |
1264 | struct scan_control *sc, enum lru_list lru) | |
1265 | { | |
1266 | LIST_HEAD(page_list); | |
1267 | unsigned long nr_scanned; | |
1268 | unsigned long nr_reclaimed = 0; | |
1269 | unsigned long nr_taken; | |
1270 | unsigned long nr_dirty = 0; | |
1271 | unsigned long nr_writeback = 0; | |
1272 | isolate_mode_t isolate_mode = 0; | |
1273 | int file = is_file_lru(lru); | |
1274 | struct zone *zone = lruvec_zone(lruvec); | |
1275 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1276 | ||
1277 | while (unlikely(too_many_isolated(zone, file, sc))) { | |
1278 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
1279 | ||
1280 | /* We are about to die and free our memory. Return now. */ | |
1281 | if (fatal_signal_pending(current)) | |
1282 | return SWAP_CLUSTER_MAX; | |
1283 | } | |
1284 | ||
1285 | lru_add_drain(); | |
1286 | ||
1287 | if (!sc->may_unmap) | |
1288 | isolate_mode |= ISOLATE_UNMAPPED; | |
1289 | if (!sc->may_writepage) | |
1290 | isolate_mode |= ISOLATE_CLEAN; | |
1291 | ||
1292 | spin_lock_irq(&zone->lru_lock); | |
1293 | ||
1294 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, | |
1295 | &nr_scanned, sc, isolate_mode, lru); | |
1296 | ||
1297 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1298 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1299 | ||
1300 | if (global_reclaim(sc)) { | |
1301 | zone->pages_scanned += nr_scanned; | |
1302 | if (current_is_kswapd()) | |
1303 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); | |
1304 | else | |
1305 | __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); | |
1306 | } | |
1307 | spin_unlock_irq(&zone->lru_lock); | |
1308 | ||
1309 | if (nr_taken == 0) | |
1310 | return 0; | |
1311 | ||
1312 | nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, | |
1313 | &nr_dirty, &nr_writeback, false); | |
1314 | ||
1315 | spin_lock_irq(&zone->lru_lock); | |
1316 | ||
1317 | reclaim_stat->recent_scanned[file] += nr_taken; | |
1318 | ||
1319 | if (global_reclaim(sc)) { | |
1320 | if (current_is_kswapd()) | |
1321 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1322 | nr_reclaimed); | |
1323 | else | |
1324 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1325 | nr_reclaimed); | |
1326 | } | |
1327 | ||
1328 | putback_inactive_pages(lruvec, &page_list); | |
1329 | ||
1330 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); | |
1331 | ||
1332 | spin_unlock_irq(&zone->lru_lock); | |
1333 | ||
1334 | free_hot_cold_page_list(&page_list, 1); | |
1335 | ||
1336 | /* | |
1337 | * If reclaim is isolating dirty pages under writeback, it implies | |
1338 | * that the long-lived page allocation rate is exceeding the page | |
1339 | * laundering rate. Either the global limits are not being effective | |
1340 | * at throttling processes due to the page distribution throughout | |
1341 | * zones or there is heavy usage of a slow backing device. The | |
1342 | * only option is to throttle from reclaim context which is not ideal | |
1343 | * as there is no guarantee the dirtying process is throttled in the | |
1344 | * same way balance_dirty_pages() manages. | |
1345 | * | |
1346 | * This scales the number of dirty pages that must be under writeback | |
1347 | * before throttling depending on priority. It is a simple backoff | |
1348 | * function that has the most effect in the range DEF_PRIORITY to | |
1349 | * DEF_PRIORITY-2 which is the priority reclaim is considered to be | |
1350 | * in trouble and reclaim is considered to be in trouble. | |
1351 | * | |
1352 | * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle | |
1353 | * DEF_PRIORITY-1 50% must be PageWriteback | |
1354 | * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble | |
1355 | * ... | |
1356 | * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any | |
1357 | * isolated page is PageWriteback | |
1358 | */ | |
1359 | if (nr_writeback && nr_writeback >= | |
1360 | (nr_taken >> (DEF_PRIORITY - sc->priority))) | |
1361 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); | |
1362 | ||
1363 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, | |
1364 | zone_idx(zone), | |
1365 | nr_scanned, nr_reclaimed, | |
1366 | sc->priority, | |
1367 | trace_shrink_flags(file)); | |
1368 | return nr_reclaimed; | |
1369 | } | |
1370 | ||
1371 | /* | |
1372 | * This moves pages from the active list to the inactive list. | |
1373 | * | |
1374 | * We move them the other way if the page is referenced by one or more | |
1375 | * processes, from rmap. | |
1376 | * | |
1377 | * If the pages are mostly unmapped, the processing is fast and it is | |
1378 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1379 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1380 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1381 | * this, so instead we remove the pages from the LRU while processing them. | |
1382 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1383 | * nobody will play with that bit on a non-LRU page. | |
1384 | * | |
1385 | * The downside is that we have to touch page->_count against each page. | |
1386 | * But we had to alter page->flags anyway. | |
1387 | */ | |
1388 | ||
1389 | static void move_active_pages_to_lru(struct lruvec *lruvec, | |
1390 | struct list_head *list, | |
1391 | struct list_head *pages_to_free, | |
1392 | enum lru_list lru) | |
1393 | { | |
1394 | struct zone *zone = lruvec_zone(lruvec); | |
1395 | unsigned long pgmoved = 0; | |
1396 | struct page *page; | |
1397 | int nr_pages; | |
1398 | ||
1399 | while (!list_empty(list)) { | |
1400 | page = lru_to_page(list); | |
1401 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1402 | ||
1403 | VM_BUG_ON(PageLRU(page)); | |
1404 | SetPageLRU(page); | |
1405 | ||
1406 | nr_pages = hpage_nr_pages(page); | |
1407 | mem_cgroup_update_lru_size(lruvec, lru, nr_pages); | |
1408 | list_move(&page->lru, &lruvec->lists[lru]); | |
1409 | pgmoved += nr_pages; | |
1410 | ||
1411 | if (put_page_testzero(page)) { | |
1412 | __ClearPageLRU(page); | |
1413 | __ClearPageActive(page); | |
1414 | del_page_from_lru_list(page, lruvec, lru); | |
1415 | ||
1416 | if (unlikely(PageCompound(page))) { | |
1417 | spin_unlock_irq(&zone->lru_lock); | |
1418 | (*get_compound_page_dtor(page))(page); | |
1419 | spin_lock_irq(&zone->lru_lock); | |
1420 | } else | |
1421 | list_add(&page->lru, pages_to_free); | |
1422 | } | |
1423 | } | |
1424 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1425 | if (!is_active_lru(lru)) | |
1426 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1427 | } | |
1428 | ||
1429 | static void shrink_active_list(unsigned long nr_to_scan, | |
1430 | struct lruvec *lruvec, | |
1431 | struct scan_control *sc, | |
1432 | enum lru_list lru) | |
1433 | { | |
1434 | unsigned long nr_taken; | |
1435 | unsigned long nr_scanned; | |
1436 | unsigned long vm_flags; | |
1437 | LIST_HEAD(l_hold); /* The pages which were snipped off */ | |
1438 | LIST_HEAD(l_active); | |
1439 | LIST_HEAD(l_inactive); | |
1440 | struct page *page; | |
1441 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1442 | unsigned long nr_rotated = 0; | |
1443 | isolate_mode_t isolate_mode = 0; | |
1444 | int file = is_file_lru(lru); | |
1445 | struct zone *zone = lruvec_zone(lruvec); | |
1446 | ||
1447 | lru_add_drain(); | |
1448 | ||
1449 | if (!sc->may_unmap) | |
1450 | isolate_mode |= ISOLATE_UNMAPPED; | |
1451 | if (!sc->may_writepage) | |
1452 | isolate_mode |= ISOLATE_CLEAN; | |
1453 | ||
1454 | spin_lock_irq(&zone->lru_lock); | |
1455 | ||
1456 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, | |
1457 | &nr_scanned, sc, isolate_mode, lru); | |
1458 | if (global_reclaim(sc)) | |
1459 | zone->pages_scanned += nr_scanned; | |
1460 | ||
1461 | reclaim_stat->recent_scanned[file] += nr_taken; | |
1462 | ||
1463 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); | |
1464 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1465 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1466 | spin_unlock_irq(&zone->lru_lock); | |
1467 | ||
1468 | while (!list_empty(&l_hold)) { | |
1469 | cond_resched(); | |
1470 | page = lru_to_page(&l_hold); | |
1471 | list_del(&page->lru); | |
1472 | ||
1473 | if (unlikely(!page_evictable(page, NULL))) { | |
1474 | putback_lru_page(page); | |
1475 | continue; | |
1476 | } | |
1477 | ||
1478 | if (unlikely(buffer_heads_over_limit)) { | |
1479 | if (page_has_private(page) && trylock_page(page)) { | |
1480 | if (page_has_private(page)) | |
1481 | try_to_release_page(page, 0); | |
1482 | unlock_page(page); | |
1483 | } | |
1484 | } | |
1485 | ||
1486 | if (page_referenced(page, 0, sc->target_mem_cgroup, | |
1487 | &vm_flags)) { | |
1488 | nr_rotated += hpage_nr_pages(page); | |
1489 | /* | |
1490 | * Identify referenced, file-backed active pages and | |
1491 | * give them one more trip around the active list. So | |
1492 | * that executable code get better chances to stay in | |
1493 | * memory under moderate memory pressure. Anon pages | |
1494 | * are not likely to be evicted by use-once streaming | |
1495 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1496 | * so we ignore them here. | |
1497 | */ | |
1498 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { | |
1499 | list_add(&page->lru, &l_active); | |
1500 | continue; | |
1501 | } | |
1502 | } | |
1503 | ||
1504 | ClearPageActive(page); /* we are de-activating */ | |
1505 | list_add(&page->lru, &l_inactive); | |
1506 | } | |
1507 | ||
1508 | /* | |
1509 | * Move pages back to the lru list. | |
1510 | */ | |
1511 | spin_lock_irq(&zone->lru_lock); | |
1512 | /* | |
1513 | * Count referenced pages from currently used mappings as rotated, | |
1514 | * even though only some of them are actually re-activated. This | |
1515 | * helps balance scan pressure between file and anonymous pages in | |
1516 | * get_scan_ratio. | |
1517 | */ | |
1518 | reclaim_stat->recent_rotated[file] += nr_rotated; | |
1519 | ||
1520 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); | |
1521 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
1522 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); | |
1523 | spin_unlock_irq(&zone->lru_lock); | |
1524 | ||
1525 | free_hot_cold_page_list(&l_hold, 1); | |
1526 | } | |
1527 | ||
1528 | #ifdef CONFIG_SWAP | |
1529 | static int inactive_anon_is_low_global(struct zone *zone) | |
1530 | { | |
1531 | unsigned long active, inactive; | |
1532 | ||
1533 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1534 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1535 | ||
1536 | if (inactive * zone->inactive_ratio < active) | |
1537 | return 1; | |
1538 | ||
1539 | return 0; | |
1540 | } | |
1541 | ||
1542 | /** | |
1543 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
1544 | * @lruvec: LRU vector to check | |
1545 | * | |
1546 | * Returns true if the zone does not have enough inactive anon pages, | |
1547 | * meaning some active anon pages need to be deactivated. | |
1548 | */ | |
1549 | static int inactive_anon_is_low(struct lruvec *lruvec) | |
1550 | { | |
1551 | /* | |
1552 | * If we don't have swap space, anonymous page deactivation | |
1553 | * is pointless. | |
1554 | */ | |
1555 | if (!total_swap_pages) | |
1556 | return 0; | |
1557 | ||
1558 | if (!mem_cgroup_disabled()) | |
1559 | return mem_cgroup_inactive_anon_is_low(lruvec); | |
1560 | ||
1561 | return inactive_anon_is_low_global(lruvec_zone(lruvec)); | |
1562 | } | |
1563 | #else | |
1564 | static inline int inactive_anon_is_low(struct lruvec *lruvec) | |
1565 | { | |
1566 | return 0; | |
1567 | } | |
1568 | #endif | |
1569 | ||
1570 | static int inactive_file_is_low_global(struct zone *zone) | |
1571 | { | |
1572 | unsigned long active, inactive; | |
1573 | ||
1574 | active = zone_page_state(zone, NR_ACTIVE_FILE); | |
1575 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1576 | ||
1577 | return (active > inactive); | |
1578 | } | |
1579 | ||
1580 | /** | |
1581 | * inactive_file_is_low - check if file pages need to be deactivated | |
1582 | * @lruvec: LRU vector to check | |
1583 | * | |
1584 | * When the system is doing streaming IO, memory pressure here | |
1585 | * ensures that active file pages get deactivated, until more | |
1586 | * than half of the file pages are on the inactive list. | |
1587 | * | |
1588 | * Once we get to that situation, protect the system's working | |
1589 | * set from being evicted by disabling active file page aging. | |
1590 | * | |
1591 | * This uses a different ratio than the anonymous pages, because | |
1592 | * the page cache uses a use-once replacement algorithm. | |
1593 | */ | |
1594 | static int inactive_file_is_low(struct lruvec *lruvec) | |
1595 | { | |
1596 | if (!mem_cgroup_disabled()) | |
1597 | return mem_cgroup_inactive_file_is_low(lruvec); | |
1598 | ||
1599 | return inactive_file_is_low_global(lruvec_zone(lruvec)); | |
1600 | } | |
1601 | ||
1602 | static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) | |
1603 | { | |
1604 | if (is_file_lru(lru)) | |
1605 | return inactive_file_is_low(lruvec); | |
1606 | else | |
1607 | return inactive_anon_is_low(lruvec); | |
1608 | } | |
1609 | ||
1610 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, | |
1611 | struct lruvec *lruvec, struct scan_control *sc) | |
1612 | { | |
1613 | if (is_active_lru(lru)) { | |
1614 | if (inactive_list_is_low(lruvec, lru)) | |
1615 | shrink_active_list(nr_to_scan, lruvec, sc, lru); | |
1616 | return 0; | |
1617 | } | |
1618 | ||
1619 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); | |
1620 | } | |
1621 | ||
1622 | static int vmscan_swappiness(struct scan_control *sc) | |
1623 | { | |
1624 | if (global_reclaim(sc)) | |
1625 | return vm_swappiness; | |
1626 | return mem_cgroup_swappiness(sc->target_mem_cgroup); | |
1627 | } | |
1628 | ||
1629 | /* | |
1630 | * Determine how aggressively the anon and file LRU lists should be | |
1631 | * scanned. The relative value of each set of LRU lists is determined | |
1632 | * by looking at the fraction of the pages scanned we did rotate back | |
1633 | * onto the active list instead of evict. | |
1634 | * | |
1635 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan | |
1636 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
1637 | */ | |
1638 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, | |
1639 | unsigned long *nr) | |
1640 | { | |
1641 | unsigned long anon, file, free; | |
1642 | unsigned long anon_prio, file_prio; | |
1643 | unsigned long ap, fp; | |
1644 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1645 | u64 fraction[2], denominator; | |
1646 | enum lru_list lru; | |
1647 | int noswap = 0; | |
1648 | bool force_scan = false; | |
1649 | struct zone *zone = lruvec_zone(lruvec); | |
1650 | ||
1651 | /* | |
1652 | * If the zone or memcg is small, nr[l] can be 0. This | |
1653 | * results in no scanning on this priority and a potential | |
1654 | * priority drop. Global direct reclaim can go to the next | |
1655 | * zone and tends to have no problems. Global kswapd is for | |
1656 | * zone balancing and it needs to scan a minimum amount. When | |
1657 | * reclaiming for a memcg, a priority drop can cause high | |
1658 | * latencies, so it's better to scan a minimum amount there as | |
1659 | * well. | |
1660 | */ | |
1661 | if (current_is_kswapd() && zone->all_unreclaimable) | |
1662 | force_scan = true; | |
1663 | if (!global_reclaim(sc)) | |
1664 | force_scan = true; | |
1665 | ||
1666 | /* If we have no swap space, do not bother scanning anon pages. */ | |
1667 | if (!sc->may_swap || (nr_swap_pages <= 0)) { | |
1668 | noswap = 1; | |
1669 | fraction[0] = 0; | |
1670 | fraction[1] = 1; | |
1671 | denominator = 1; | |
1672 | goto out; | |
1673 | } | |
1674 | ||
1675 | anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + | |
1676 | get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
1677 | file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
1678 | get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
1679 | ||
1680 | if (global_reclaim(sc)) { | |
1681 | free = zone_page_state(zone, NR_FREE_PAGES); | |
1682 | /* If we have very few page cache pages, | |
1683 | force-scan anon pages. */ | |
1684 | if (unlikely(file + free <= high_wmark_pages(zone))) { | |
1685 | fraction[0] = 1; | |
1686 | fraction[1] = 0; | |
1687 | denominator = 1; | |
1688 | goto out; | |
1689 | } | |
1690 | } | |
1691 | ||
1692 | /* | |
1693 | * With swappiness at 100, anonymous and file have the same priority. | |
1694 | * This scanning priority is essentially the inverse of IO cost. | |
1695 | */ | |
1696 | anon_prio = vmscan_swappiness(sc); | |
1697 | file_prio = 200 - anon_prio; | |
1698 | ||
1699 | /* | |
1700 | * OK, so we have swap space and a fair amount of page cache | |
1701 | * pages. We use the recently rotated / recently scanned | |
1702 | * ratios to determine how valuable each cache is. | |
1703 | * | |
1704 | * Because workloads change over time (and to avoid overflow) | |
1705 | * we keep these statistics as a floating average, which ends | |
1706 | * up weighing recent references more than old ones. | |
1707 | * | |
1708 | * anon in [0], file in [1] | |
1709 | */ | |
1710 | spin_lock_irq(&zone->lru_lock); | |
1711 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { | |
1712 | reclaim_stat->recent_scanned[0] /= 2; | |
1713 | reclaim_stat->recent_rotated[0] /= 2; | |
1714 | } | |
1715 | ||
1716 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { | |
1717 | reclaim_stat->recent_scanned[1] /= 2; | |
1718 | reclaim_stat->recent_rotated[1] /= 2; | |
1719 | } | |
1720 | ||
1721 | /* | |
1722 | * The amount of pressure on anon vs file pages is inversely | |
1723 | * proportional to the fraction of recently scanned pages on | |
1724 | * each list that were recently referenced and in active use. | |
1725 | */ | |
1726 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); | |
1727 | ap /= reclaim_stat->recent_rotated[0] + 1; | |
1728 | ||
1729 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); | |
1730 | fp /= reclaim_stat->recent_rotated[1] + 1; | |
1731 | spin_unlock_irq(&zone->lru_lock); | |
1732 | ||
1733 | fraction[0] = ap; | |
1734 | fraction[1] = fp; | |
1735 | denominator = ap + fp + 1; | |
1736 | out: | |
1737 | for_each_evictable_lru(lru) { | |
1738 | int file = is_file_lru(lru); | |
1739 | unsigned long scan; | |
1740 | ||
1741 | scan = get_lru_size(lruvec, lru); | |
1742 | if (sc->priority || noswap || !vmscan_swappiness(sc)) { | |
1743 | scan >>= sc->priority; | |
1744 | if (!scan && force_scan) | |
1745 | scan = SWAP_CLUSTER_MAX; | |
1746 | scan = div64_u64(scan * fraction[file], denominator); | |
1747 | } | |
1748 | nr[lru] = scan; | |
1749 | } | |
1750 | } | |
1751 | ||
1752 | /* Use reclaim/compaction for costly allocs or under memory pressure */ | |
1753 | static bool in_reclaim_compaction(struct scan_control *sc) | |
1754 | { | |
1755 | if (COMPACTION_BUILD && sc->order && | |
1756 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || | |
1757 | sc->priority < DEF_PRIORITY - 2)) | |
1758 | return true; | |
1759 | ||
1760 | return false; | |
1761 | } | |
1762 | ||
1763 | #ifdef CONFIG_COMPACTION | |
1764 | /* | |
1765 | * If compaction is deferred for sc->order then scale the number of pages | |
1766 | * reclaimed based on the number of consecutive allocation failures | |
1767 | */ | |
1768 | static unsigned long scale_for_compaction(unsigned long pages_for_compaction, | |
1769 | struct lruvec *lruvec, struct scan_control *sc) | |
1770 | { | |
1771 | struct zone *zone = lruvec_zone(lruvec); | |
1772 | ||
1773 | if (zone->compact_order_failed <= sc->order) | |
1774 | pages_for_compaction <<= zone->compact_defer_shift; | |
1775 | return pages_for_compaction; | |
1776 | } | |
1777 | #else | |
1778 | static unsigned long scale_for_compaction(unsigned long pages_for_compaction, | |
1779 | struct lruvec *lruvec, struct scan_control *sc) | |
1780 | { | |
1781 | return pages_for_compaction; | |
1782 | } | |
1783 | #endif | |
1784 | ||
1785 | /* | |
1786 | * Reclaim/compaction is used for high-order allocation requests. It reclaims | |
1787 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
1788 | * true if more pages should be reclaimed such that when the page allocator | |
1789 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
1790 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
1791 | */ | |
1792 | static inline bool should_continue_reclaim(struct lruvec *lruvec, | |
1793 | unsigned long nr_reclaimed, | |
1794 | unsigned long nr_scanned, | |
1795 | struct scan_control *sc) | |
1796 | { | |
1797 | unsigned long pages_for_compaction; | |
1798 | unsigned long inactive_lru_pages; | |
1799 | ||
1800 | /* If not in reclaim/compaction mode, stop */ | |
1801 | if (!in_reclaim_compaction(sc)) | |
1802 | return false; | |
1803 | ||
1804 | /* Consider stopping depending on scan and reclaim activity */ | |
1805 | if (sc->gfp_mask & __GFP_REPEAT) { | |
1806 | /* | |
1807 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
1808 | * full LRU list has been scanned and we are still failing | |
1809 | * to reclaim pages. This full LRU scan is potentially | |
1810 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
1811 | */ | |
1812 | if (!nr_reclaimed && !nr_scanned) | |
1813 | return false; | |
1814 | } else { | |
1815 | /* | |
1816 | * For non-__GFP_REPEAT allocations which can presumably | |
1817 | * fail without consequence, stop if we failed to reclaim | |
1818 | * any pages from the last SWAP_CLUSTER_MAX number of | |
1819 | * pages that were scanned. This will return to the | |
1820 | * caller faster at the risk reclaim/compaction and | |
1821 | * the resulting allocation attempt fails | |
1822 | */ | |
1823 | if (!nr_reclaimed) | |
1824 | return false; | |
1825 | } | |
1826 | ||
1827 | /* | |
1828 | * If we have not reclaimed enough pages for compaction and the | |
1829 | * inactive lists are large enough, continue reclaiming | |
1830 | */ | |
1831 | pages_for_compaction = (2UL << sc->order); | |
1832 | ||
1833 | pages_for_compaction = scale_for_compaction(pages_for_compaction, | |
1834 | lruvec, sc); | |
1835 | inactive_lru_pages = get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
1836 | if (nr_swap_pages > 0) | |
1837 | inactive_lru_pages += get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
1838 | if (sc->nr_reclaimed < pages_for_compaction && | |
1839 | inactive_lru_pages > pages_for_compaction) | |
1840 | return true; | |
1841 | ||
1842 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
1843 | switch (compaction_suitable(lruvec_zone(lruvec), sc->order)) { | |
1844 | case COMPACT_PARTIAL: | |
1845 | case COMPACT_CONTINUE: | |
1846 | return false; | |
1847 | default: | |
1848 | return true; | |
1849 | } | |
1850 | } | |
1851 | ||
1852 | /* | |
1853 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1854 | */ | |
1855 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) | |
1856 | { | |
1857 | unsigned long nr[NR_LRU_LISTS]; | |
1858 | unsigned long nr_to_scan; | |
1859 | enum lru_list lru; | |
1860 | unsigned long nr_reclaimed, nr_scanned; | |
1861 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
1862 | struct blk_plug plug; | |
1863 | ||
1864 | restart: | |
1865 | nr_reclaimed = 0; | |
1866 | nr_scanned = sc->nr_scanned; | |
1867 | get_scan_count(lruvec, sc, nr); | |
1868 | ||
1869 | blk_start_plug(&plug); | |
1870 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
1871 | nr[LRU_INACTIVE_FILE]) { | |
1872 | for_each_evictable_lru(lru) { | |
1873 | if (nr[lru]) { | |
1874 | nr_to_scan = min_t(unsigned long, | |
1875 | nr[lru], SWAP_CLUSTER_MAX); | |
1876 | nr[lru] -= nr_to_scan; | |
1877 | ||
1878 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
1879 | lruvec, sc); | |
1880 | } | |
1881 | } | |
1882 | /* | |
1883 | * On large memory systems, scan >> priority can become | |
1884 | * really large. This is fine for the starting priority; | |
1885 | * we want to put equal scanning pressure on each zone. | |
1886 | * However, if the VM has a harder time of freeing pages, | |
1887 | * with multiple processes reclaiming pages, the total | |
1888 | * freeing target can get unreasonably large. | |
1889 | */ | |
1890 | if (nr_reclaimed >= nr_to_reclaim && | |
1891 | sc->priority < DEF_PRIORITY) | |
1892 | break; | |
1893 | } | |
1894 | blk_finish_plug(&plug); | |
1895 | sc->nr_reclaimed += nr_reclaimed; | |
1896 | ||
1897 | /* | |
1898 | * Even if we did not try to evict anon pages at all, we want to | |
1899 | * rebalance the anon lru active/inactive ratio. | |
1900 | */ | |
1901 | if (inactive_anon_is_low(lruvec)) | |
1902 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
1903 | sc, LRU_ACTIVE_ANON); | |
1904 | ||
1905 | /* reclaim/compaction might need reclaim to continue */ | |
1906 | if (should_continue_reclaim(lruvec, nr_reclaimed, | |
1907 | sc->nr_scanned - nr_scanned, sc)) | |
1908 | goto restart; | |
1909 | ||
1910 | throttle_vm_writeout(sc->gfp_mask); | |
1911 | } | |
1912 | ||
1913 | static void shrink_zone(struct zone *zone, struct scan_control *sc) | |
1914 | { | |
1915 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
1916 | struct mem_cgroup_reclaim_cookie reclaim = { | |
1917 | .zone = zone, | |
1918 | .priority = sc->priority, | |
1919 | }; | |
1920 | struct mem_cgroup *memcg; | |
1921 | ||
1922 | memcg = mem_cgroup_iter(root, NULL, &reclaim); | |
1923 | do { | |
1924 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
1925 | ||
1926 | shrink_lruvec(lruvec, sc); | |
1927 | ||
1928 | /* | |
1929 | * Limit reclaim has historically picked one memcg and | |
1930 | * scanned it with decreasing priority levels until | |
1931 | * nr_to_reclaim had been reclaimed. This priority | |
1932 | * cycle is thus over after a single memcg. | |
1933 | * | |
1934 | * Direct reclaim and kswapd, on the other hand, have | |
1935 | * to scan all memory cgroups to fulfill the overall | |
1936 | * scan target for the zone. | |
1937 | */ | |
1938 | if (!global_reclaim(sc)) { | |
1939 | mem_cgroup_iter_break(root, memcg); | |
1940 | break; | |
1941 | } | |
1942 | memcg = mem_cgroup_iter(root, memcg, &reclaim); | |
1943 | } while (memcg); | |
1944 | } | |
1945 | ||
1946 | /* Returns true if compaction should go ahead for a high-order request */ | |
1947 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
1948 | { | |
1949 | unsigned long balance_gap, watermark; | |
1950 | bool watermark_ok; | |
1951 | ||
1952 | /* Do not consider compaction for orders reclaim is meant to satisfy */ | |
1953 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
1954 | return false; | |
1955 | ||
1956 | /* | |
1957 | * Compaction takes time to run and there are potentially other | |
1958 | * callers using the pages just freed. Continue reclaiming until | |
1959 | * there is a buffer of free pages available to give compaction | |
1960 | * a reasonable chance of completing and allocating the page | |
1961 | */ | |
1962 | balance_gap = min(low_wmark_pages(zone), | |
1963 | (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
1964 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
1965 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order); | |
1966 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); | |
1967 | ||
1968 | /* | |
1969 | * If compaction is deferred, reclaim up to a point where | |
1970 | * compaction will have a chance of success when re-enabled | |
1971 | */ | |
1972 | if (compaction_deferred(zone, sc->order)) | |
1973 | return watermark_ok; | |
1974 | ||
1975 | /* If compaction is not ready to start, keep reclaiming */ | |
1976 | if (!compaction_suitable(zone, sc->order)) | |
1977 | return false; | |
1978 | ||
1979 | return watermark_ok; | |
1980 | } | |
1981 | ||
1982 | /* | |
1983 | * This is the direct reclaim path, for page-allocating processes. We only | |
1984 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
1985 | * request. | |
1986 | * | |
1987 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). | |
1988 | * Because: | |
1989 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order | |
1990 | * allocation or | |
1991 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones | |
1992 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
1993 | * zone defense algorithm. | |
1994 | * | |
1995 | * If a zone is deemed to be full of pinned pages then just give it a light | |
1996 | * scan then give up on it. | |
1997 | * | |
1998 | * This function returns true if a zone is being reclaimed for a costly | |
1999 | * high-order allocation and compaction is ready to begin. This indicates to | |
2000 | * the caller that it should consider retrying the allocation instead of | |
2001 | * further reclaim. | |
2002 | */ | |
2003 | static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) | |
2004 | { | |
2005 | struct zoneref *z; | |
2006 | struct zone *zone; | |
2007 | unsigned long nr_soft_reclaimed; | |
2008 | unsigned long nr_soft_scanned; | |
2009 | bool aborted_reclaim = false; | |
2010 | ||
2011 | /* | |
2012 | * If the number of buffer_heads in the machine exceeds the maximum | |
2013 | * allowed level, force direct reclaim to scan the highmem zone as | |
2014 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2015 | */ | |
2016 | if (buffer_heads_over_limit) | |
2017 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2018 | ||
2019 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2020 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2021 | if (!populated_zone(zone)) | |
2022 | continue; | |
2023 | /* | |
2024 | * Take care memory controller reclaiming has small influence | |
2025 | * to global LRU. | |
2026 | */ | |
2027 | if (global_reclaim(sc)) { | |
2028 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2029 | continue; | |
2030 | if (zone->all_unreclaimable && | |
2031 | sc->priority != DEF_PRIORITY) | |
2032 | continue; /* Let kswapd poll it */ | |
2033 | if (COMPACTION_BUILD) { | |
2034 | /* | |
2035 | * If we already have plenty of memory free for | |
2036 | * compaction in this zone, don't free any more. | |
2037 | * Even though compaction is invoked for any | |
2038 | * non-zero order, only frequent costly order | |
2039 | * reclamation is disruptive enough to become a | |
2040 | * noticeable problem, like transparent huge | |
2041 | * page allocations. | |
2042 | */ | |
2043 | if (compaction_ready(zone, sc)) { | |
2044 | aborted_reclaim = true; | |
2045 | continue; | |
2046 | } | |
2047 | } | |
2048 | /* | |
2049 | * This steals pages from memory cgroups over softlimit | |
2050 | * and returns the number of reclaimed pages and | |
2051 | * scanned pages. This works for global memory pressure | |
2052 | * and balancing, not for a memcg's limit. | |
2053 | */ | |
2054 | nr_soft_scanned = 0; | |
2055 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2056 | sc->order, sc->gfp_mask, | |
2057 | &nr_soft_scanned); | |
2058 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2059 | sc->nr_scanned += nr_soft_scanned; | |
2060 | /* need some check for avoid more shrink_zone() */ | |
2061 | } | |
2062 | ||
2063 | shrink_zone(zone, sc); | |
2064 | } | |
2065 | ||
2066 | return aborted_reclaim; | |
2067 | } | |
2068 | ||
2069 | static bool zone_reclaimable(struct zone *zone) | |
2070 | { | |
2071 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
2072 | } | |
2073 | ||
2074 | /* All zones in zonelist are unreclaimable? */ | |
2075 | static bool all_unreclaimable(struct zonelist *zonelist, | |
2076 | struct scan_control *sc) | |
2077 | { | |
2078 | struct zoneref *z; | |
2079 | struct zone *zone; | |
2080 | ||
2081 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2082 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2083 | if (!populated_zone(zone)) | |
2084 | continue; | |
2085 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2086 | continue; | |
2087 | if (!zone->all_unreclaimable) | |
2088 | return false; | |
2089 | } | |
2090 | ||
2091 | return true; | |
2092 | } | |
2093 | ||
2094 | /* | |
2095 | * This is the main entry point to direct page reclaim. | |
2096 | * | |
2097 | * If a full scan of the inactive list fails to free enough memory then we | |
2098 | * are "out of memory" and something needs to be killed. | |
2099 | * | |
2100 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2101 | * high - the zone may be full of dirty or under-writeback pages, which this | |
2102 | * caller can't do much about. We kick the writeback threads and take explicit | |
2103 | * naps in the hope that some of these pages can be written. But if the | |
2104 | * allocating task holds filesystem locks which prevent writeout this might not | |
2105 | * work, and the allocation attempt will fail. | |
2106 | * | |
2107 | * returns: 0, if no pages reclaimed | |
2108 | * else, the number of pages reclaimed | |
2109 | */ | |
2110 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, | |
2111 | struct scan_control *sc, | |
2112 | struct shrink_control *shrink) | |
2113 | { | |
2114 | unsigned long total_scanned = 0; | |
2115 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
2116 | struct zoneref *z; | |
2117 | struct zone *zone; | |
2118 | unsigned long writeback_threshold; | |
2119 | bool aborted_reclaim; | |
2120 | ||
2121 | delayacct_freepages_start(); | |
2122 | ||
2123 | if (global_reclaim(sc)) | |
2124 | count_vm_event(ALLOCSTALL); | |
2125 | ||
2126 | do { | |
2127 | sc->nr_scanned = 0; | |
2128 | aborted_reclaim = shrink_zones(zonelist, sc); | |
2129 | ||
2130 | /* | |
2131 | * Don't shrink slabs when reclaiming memory from | |
2132 | * over limit cgroups | |
2133 | */ | |
2134 | if (global_reclaim(sc)) { | |
2135 | unsigned long lru_pages = 0; | |
2136 | for_each_zone_zonelist(zone, z, zonelist, | |
2137 | gfp_zone(sc->gfp_mask)) { | |
2138 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2139 | continue; | |
2140 | ||
2141 | lru_pages += zone_reclaimable_pages(zone); | |
2142 | } | |
2143 | ||
2144 | shrink_slab(shrink, sc->nr_scanned, lru_pages); | |
2145 | if (reclaim_state) { | |
2146 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2147 | reclaim_state->reclaimed_slab = 0; | |
2148 | } | |
2149 | } | |
2150 | total_scanned += sc->nr_scanned; | |
2151 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) | |
2152 | goto out; | |
2153 | ||
2154 | /* | |
2155 | * Try to write back as many pages as we just scanned. This | |
2156 | * tends to cause slow streaming writers to write data to the | |
2157 | * disk smoothly, at the dirtying rate, which is nice. But | |
2158 | * that's undesirable in laptop mode, where we *want* lumpy | |
2159 | * writeout. So in laptop mode, write out the whole world. | |
2160 | */ | |
2161 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; | |
2162 | if (total_scanned > writeback_threshold) { | |
2163 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, | |
2164 | WB_REASON_TRY_TO_FREE_PAGES); | |
2165 | sc->may_writepage = 1; | |
2166 | } | |
2167 | ||
2168 | /* Take a nap, wait for some writeback to complete */ | |
2169 | if (!sc->hibernation_mode && sc->nr_scanned && | |
2170 | sc->priority < DEF_PRIORITY - 2) { | |
2171 | struct zone *preferred_zone; | |
2172 | ||
2173 | first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), | |
2174 | &cpuset_current_mems_allowed, | |
2175 | &preferred_zone); | |
2176 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); | |
2177 | } | |
2178 | } while (--sc->priority >= 0); | |
2179 | ||
2180 | out: | |
2181 | delayacct_freepages_end(); | |
2182 | ||
2183 | if (sc->nr_reclaimed) | |
2184 | return sc->nr_reclaimed; | |
2185 | ||
2186 | /* | |
2187 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2188 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2189 | * check. | |
2190 | */ | |
2191 | if (oom_killer_disabled) | |
2192 | return 0; | |
2193 | ||
2194 | /* Aborted reclaim to try compaction? don't OOM, then */ | |
2195 | if (aborted_reclaim) | |
2196 | return 1; | |
2197 | ||
2198 | /* top priority shrink_zones still had more to do? don't OOM, then */ | |
2199 | if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc)) | |
2200 | return 1; | |
2201 | ||
2202 | return 0; | |
2203 | } | |
2204 | ||
2205 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) | |
2206 | { | |
2207 | struct zone *zone; | |
2208 | unsigned long pfmemalloc_reserve = 0; | |
2209 | unsigned long free_pages = 0; | |
2210 | int i; | |
2211 | bool wmark_ok; | |
2212 | ||
2213 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2214 | zone = &pgdat->node_zones[i]; | |
2215 | pfmemalloc_reserve += min_wmark_pages(zone); | |
2216 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2217 | } | |
2218 | ||
2219 | wmark_ok = free_pages > pfmemalloc_reserve / 2; | |
2220 | ||
2221 | /* kswapd must be awake if processes are being throttled */ | |
2222 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2223 | pgdat->classzone_idx = min(pgdat->classzone_idx, | |
2224 | (enum zone_type)ZONE_NORMAL); | |
2225 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2226 | } | |
2227 | ||
2228 | return wmark_ok; | |
2229 | } | |
2230 | ||
2231 | /* | |
2232 | * Throttle direct reclaimers if backing storage is backed by the network | |
2233 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2234 | * depleted. kswapd will continue to make progress and wake the processes | |
2235 | * when the low watermark is reached | |
2236 | */ | |
2237 | static void throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, | |
2238 | nodemask_t *nodemask) | |
2239 | { | |
2240 | struct zone *zone; | |
2241 | int high_zoneidx = gfp_zone(gfp_mask); | |
2242 | pg_data_t *pgdat; | |
2243 | ||
2244 | /* | |
2245 | * Kernel threads should not be throttled as they may be indirectly | |
2246 | * responsible for cleaning pages necessary for reclaim to make forward | |
2247 | * progress. kjournald for example may enter direct reclaim while | |
2248 | * committing a transaction where throttling it could forcing other | |
2249 | * processes to block on log_wait_commit(). | |
2250 | */ | |
2251 | if (current->flags & PF_KTHREAD) | |
2252 | return; | |
2253 | ||
2254 | /* Check if the pfmemalloc reserves are ok */ | |
2255 | first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone); | |
2256 | pgdat = zone->zone_pgdat; | |
2257 | if (pfmemalloc_watermark_ok(pgdat)) | |
2258 | return; | |
2259 | ||
2260 | /* Account for the throttling */ | |
2261 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2262 | ||
2263 | /* | |
2264 | * If the caller cannot enter the filesystem, it's possible that it | |
2265 | * is due to the caller holding an FS lock or performing a journal | |
2266 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2267 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2268 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2269 | * second before continuing. | |
2270 | */ | |
2271 | if (!(gfp_mask & __GFP_FS)) { | |
2272 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2273 | pfmemalloc_watermark_ok(pgdat), HZ); | |
2274 | return; | |
2275 | } | |
2276 | ||
2277 | /* Throttle until kswapd wakes the process */ | |
2278 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2279 | pfmemalloc_watermark_ok(pgdat)); | |
2280 | } | |
2281 | ||
2282 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, | |
2283 | gfp_t gfp_mask, nodemask_t *nodemask) | |
2284 | { | |
2285 | unsigned long nr_reclaimed; | |
2286 | struct scan_control sc = { | |
2287 | .gfp_mask = gfp_mask, | |
2288 | .may_writepage = !laptop_mode, | |
2289 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2290 | .may_unmap = 1, | |
2291 | .may_swap = 1, | |
2292 | .order = order, | |
2293 | .priority = DEF_PRIORITY, | |
2294 | .target_mem_cgroup = NULL, | |
2295 | .nodemask = nodemask, | |
2296 | }; | |
2297 | struct shrink_control shrink = { | |
2298 | .gfp_mask = sc.gfp_mask, | |
2299 | }; | |
2300 | ||
2301 | throttle_direct_reclaim(gfp_mask, zonelist, nodemask); | |
2302 | ||
2303 | /* | |
2304 | * Do not enter reclaim if fatal signal is pending. 1 is returned so | |
2305 | * that the page allocator does not consider triggering OOM | |
2306 | */ | |
2307 | if (fatal_signal_pending(current)) | |
2308 | return 1; | |
2309 | ||
2310 | trace_mm_vmscan_direct_reclaim_begin(order, | |
2311 | sc.may_writepage, | |
2312 | gfp_mask); | |
2313 | ||
2314 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); | |
2315 | ||
2316 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2317 | ||
2318 | return nr_reclaimed; | |
2319 | } | |
2320 | ||
2321 | #ifdef CONFIG_MEMCG | |
2322 | ||
2323 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, | |
2324 | gfp_t gfp_mask, bool noswap, | |
2325 | struct zone *zone, | |
2326 | unsigned long *nr_scanned) | |
2327 | { | |
2328 | struct scan_control sc = { | |
2329 | .nr_scanned = 0, | |
2330 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2331 | .may_writepage = !laptop_mode, | |
2332 | .may_unmap = 1, | |
2333 | .may_swap = !noswap, | |
2334 | .order = 0, | |
2335 | .priority = 0, | |
2336 | .target_mem_cgroup = memcg, | |
2337 | }; | |
2338 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
2339 | ||
2340 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
2341 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
2342 | ||
2343 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, | |
2344 | sc.may_writepage, | |
2345 | sc.gfp_mask); | |
2346 | ||
2347 | /* | |
2348 | * NOTE: Although we can get the priority field, using it | |
2349 | * here is not a good idea, since it limits the pages we can scan. | |
2350 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2351 | * will pick up pages from other mem cgroup's as well. We hack | |
2352 | * the priority and make it zero. | |
2353 | */ | |
2354 | shrink_lruvec(lruvec, &sc); | |
2355 | ||
2356 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2357 | ||
2358 | *nr_scanned = sc.nr_scanned; | |
2359 | return sc.nr_reclaimed; | |
2360 | } | |
2361 | ||
2362 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, | |
2363 | gfp_t gfp_mask, | |
2364 | bool noswap) | |
2365 | { | |
2366 | struct zonelist *zonelist; | |
2367 | unsigned long nr_reclaimed; | |
2368 | int nid; | |
2369 | struct scan_control sc = { | |
2370 | .may_writepage = !laptop_mode, | |
2371 | .may_unmap = 1, | |
2372 | .may_swap = !noswap, | |
2373 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2374 | .order = 0, | |
2375 | .priority = DEF_PRIORITY, | |
2376 | .target_mem_cgroup = memcg, | |
2377 | .nodemask = NULL, /* we don't care the placement */ | |
2378 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
2379 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2380 | }; | |
2381 | struct shrink_control shrink = { | |
2382 | .gfp_mask = sc.gfp_mask, | |
2383 | }; | |
2384 | ||
2385 | /* | |
2386 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2387 | * take care of from where we get pages. So the node where we start the | |
2388 | * scan does not need to be the current node. | |
2389 | */ | |
2390 | nid = mem_cgroup_select_victim_node(memcg); | |
2391 | ||
2392 | zonelist = NODE_DATA(nid)->node_zonelists; | |
2393 | ||
2394 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2395 | sc.may_writepage, | |
2396 | sc.gfp_mask); | |
2397 | ||
2398 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); | |
2399 | ||
2400 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2401 | ||
2402 | return nr_reclaimed; | |
2403 | } | |
2404 | #endif | |
2405 | ||
2406 | static void age_active_anon(struct zone *zone, struct scan_control *sc) | |
2407 | { | |
2408 | struct mem_cgroup *memcg; | |
2409 | ||
2410 | if (!total_swap_pages) | |
2411 | return; | |
2412 | ||
2413 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2414 | do { | |
2415 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
2416 | ||
2417 | if (inactive_anon_is_low(lruvec)) | |
2418 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
2419 | sc, LRU_ACTIVE_ANON); | |
2420 | ||
2421 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2422 | } while (memcg); | |
2423 | } | |
2424 | ||
2425 | /* | |
2426 | * pgdat_balanced is used when checking if a node is balanced for high-order | |
2427 | * allocations. Only zones that meet watermarks and are in a zone allowed | |
2428 | * by the callers classzone_idx are added to balanced_pages. The total of | |
2429 | * balanced pages must be at least 25% of the zones allowed by classzone_idx | |
2430 | * for the node to be considered balanced. Forcing all zones to be balanced | |
2431 | * for high orders can cause excessive reclaim when there are imbalanced zones. | |
2432 | * The choice of 25% is due to | |
2433 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2434 | * reasonable sized machine | |
2435 | * o On all other machines, the top zone must be at least a reasonable | |
2436 | * percentage of the middle zones. For example, on 32-bit x86, highmem | |
2437 | * would need to be at least 256M for it to be balance a whole node. | |
2438 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2439 | * to balance a node on its own. These seemed like reasonable ratios. | |
2440 | */ | |
2441 | static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages, | |
2442 | int classzone_idx) | |
2443 | { | |
2444 | unsigned long present_pages = 0; | |
2445 | int i; | |
2446 | ||
2447 | for (i = 0; i <= classzone_idx; i++) | |
2448 | present_pages += pgdat->node_zones[i].present_pages; | |
2449 | ||
2450 | /* A special case here: if zone has no page, we think it's balanced */ | |
2451 | return balanced_pages >= (present_pages >> 2); | |
2452 | } | |
2453 | ||
2454 | /* | |
2455 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
2456 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
2457 | * | |
2458 | * Returns true if kswapd is ready to sleep | |
2459 | */ | |
2460 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, | |
2461 | int classzone_idx) | |
2462 | { | |
2463 | int i; | |
2464 | unsigned long balanced = 0; | |
2465 | bool all_zones_ok = true; | |
2466 | ||
2467 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ | |
2468 | if (remaining) | |
2469 | return false; | |
2470 | ||
2471 | /* | |
2472 | * There is a potential race between when kswapd checks its watermarks | |
2473 | * and a process gets throttled. There is also a potential race if | |
2474 | * processes get throttled, kswapd wakes, a large process exits therby | |
2475 | * balancing the zones that causes kswapd to miss a wakeup. If kswapd | |
2476 | * is going to sleep, no process should be sleeping on pfmemalloc_wait | |
2477 | * so wake them now if necessary. If necessary, processes will wake | |
2478 | * kswapd and get throttled again | |
2479 | */ | |
2480 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) { | |
2481 | wake_up(&pgdat->pfmemalloc_wait); | |
2482 | return false; | |
2483 | } | |
2484 | ||
2485 | /* Check the watermark levels */ | |
2486 | for (i = 0; i <= classzone_idx; i++) { | |
2487 | struct zone *zone = pgdat->node_zones + i; | |
2488 | ||
2489 | if (!populated_zone(zone)) | |
2490 | continue; | |
2491 | ||
2492 | /* | |
2493 | * balance_pgdat() skips over all_unreclaimable after | |
2494 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2495 | * they must be considered balanced here as well if kswapd | |
2496 | * is to sleep | |
2497 | */ | |
2498 | if (zone->all_unreclaimable) { | |
2499 | balanced += zone->present_pages; | |
2500 | continue; | |
2501 | } | |
2502 | ||
2503 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone), | |
2504 | i, 0)) | |
2505 | all_zones_ok = false; | |
2506 | else | |
2507 | balanced += zone->present_pages; | |
2508 | } | |
2509 | ||
2510 | /* | |
2511 | * For high-order requests, the balanced zones must contain at least | |
2512 | * 25% of the nodes pages for kswapd to sleep. For order-0, all zones | |
2513 | * must be balanced | |
2514 | */ | |
2515 | if (order) | |
2516 | return pgdat_balanced(pgdat, balanced, classzone_idx); | |
2517 | else | |
2518 | return all_zones_ok; | |
2519 | } | |
2520 | ||
2521 | /* | |
2522 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
2523 | * they are all at high_wmark_pages(zone). | |
2524 | * | |
2525 | * Returns the final order kswapd was reclaiming at | |
2526 | * | |
2527 | * There is special handling here for zones which are full of pinned pages. | |
2528 | * This can happen if the pages are all mlocked, or if they are all used by | |
2529 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
2530 | * What we do is to detect the case where all pages in the zone have been | |
2531 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
2532 | * dead and from now on, only perform a short scan. Basically we're polling | |
2533 | * the zone for when the problem goes away. | |
2534 | * | |
2535 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
2536 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is | |
2537 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
2538 | * lower zones regardless of the number of free pages in the lower zones. This | |
2539 | * interoperates with the page allocator fallback scheme to ensure that aging | |
2540 | * of pages is balanced across the zones. | |
2541 | */ | |
2542 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, | |
2543 | int *classzone_idx) | |
2544 | { | |
2545 | int all_zones_ok; | |
2546 | unsigned long balanced; | |
2547 | int i; | |
2548 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ | |
2549 | unsigned long total_scanned; | |
2550 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
2551 | unsigned long nr_soft_reclaimed; | |
2552 | unsigned long nr_soft_scanned; | |
2553 | struct scan_control sc = { | |
2554 | .gfp_mask = GFP_KERNEL, | |
2555 | .may_unmap = 1, | |
2556 | .may_swap = 1, | |
2557 | /* | |
2558 | * kswapd doesn't want to be bailed out while reclaim. because | |
2559 | * we want to put equal scanning pressure on each zone. | |
2560 | */ | |
2561 | .nr_to_reclaim = ULONG_MAX, | |
2562 | .order = order, | |
2563 | .target_mem_cgroup = NULL, | |
2564 | }; | |
2565 | struct shrink_control shrink = { | |
2566 | .gfp_mask = sc.gfp_mask, | |
2567 | }; | |
2568 | loop_again: | |
2569 | total_scanned = 0; | |
2570 | sc.priority = DEF_PRIORITY; | |
2571 | sc.nr_reclaimed = 0; | |
2572 | sc.may_writepage = !laptop_mode; | |
2573 | count_vm_event(PAGEOUTRUN); | |
2574 | ||
2575 | do { | |
2576 | unsigned long lru_pages = 0; | |
2577 | int has_under_min_watermark_zone = 0; | |
2578 | ||
2579 | all_zones_ok = 1; | |
2580 | balanced = 0; | |
2581 | ||
2582 | /* | |
2583 | * Scan in the highmem->dma direction for the highest | |
2584 | * zone which needs scanning | |
2585 | */ | |
2586 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
2587 | struct zone *zone = pgdat->node_zones + i; | |
2588 | ||
2589 | if (!populated_zone(zone)) | |
2590 | continue; | |
2591 | ||
2592 | if (zone->all_unreclaimable && | |
2593 | sc.priority != DEF_PRIORITY) | |
2594 | continue; | |
2595 | ||
2596 | /* | |
2597 | * Do some background aging of the anon list, to give | |
2598 | * pages a chance to be referenced before reclaiming. | |
2599 | */ | |
2600 | age_active_anon(zone, &sc); | |
2601 | ||
2602 | /* | |
2603 | * If the number of buffer_heads in the machine | |
2604 | * exceeds the maximum allowed level and this node | |
2605 | * has a highmem zone, force kswapd to reclaim from | |
2606 | * it to relieve lowmem pressure. | |
2607 | */ | |
2608 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
2609 | end_zone = i; | |
2610 | break; | |
2611 | } | |
2612 | ||
2613 | if (!zone_watermark_ok_safe(zone, order, | |
2614 | high_wmark_pages(zone), 0, 0)) { | |
2615 | end_zone = i; | |
2616 | break; | |
2617 | } else { | |
2618 | /* If balanced, clear the congested flag */ | |
2619 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2620 | } | |
2621 | } | |
2622 | if (i < 0) | |
2623 | goto out; | |
2624 | ||
2625 | for (i = 0; i <= end_zone; i++) { | |
2626 | struct zone *zone = pgdat->node_zones + i; | |
2627 | ||
2628 | lru_pages += zone_reclaimable_pages(zone); | |
2629 | } | |
2630 | ||
2631 | /* | |
2632 | * Now scan the zone in the dma->highmem direction, stopping | |
2633 | * at the last zone which needs scanning. | |
2634 | * | |
2635 | * We do this because the page allocator works in the opposite | |
2636 | * direction. This prevents the page allocator from allocating | |
2637 | * pages behind kswapd's direction of progress, which would | |
2638 | * cause too much scanning of the lower zones. | |
2639 | */ | |
2640 | for (i = 0; i <= end_zone; i++) { | |
2641 | struct zone *zone = pgdat->node_zones + i; | |
2642 | int nr_slab, testorder; | |
2643 | unsigned long balance_gap; | |
2644 | ||
2645 | if (!populated_zone(zone)) | |
2646 | continue; | |
2647 | ||
2648 | if (zone->all_unreclaimable && | |
2649 | sc.priority != DEF_PRIORITY) | |
2650 | continue; | |
2651 | ||
2652 | sc.nr_scanned = 0; | |
2653 | ||
2654 | nr_soft_scanned = 0; | |
2655 | /* | |
2656 | * Call soft limit reclaim before calling shrink_zone. | |
2657 | */ | |
2658 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2659 | order, sc.gfp_mask, | |
2660 | &nr_soft_scanned); | |
2661 | sc.nr_reclaimed += nr_soft_reclaimed; | |
2662 | total_scanned += nr_soft_scanned; | |
2663 | ||
2664 | /* | |
2665 | * We put equal pressure on every zone, unless | |
2666 | * one zone has way too many pages free | |
2667 | * already. The "too many pages" is defined | |
2668 | * as the high wmark plus a "gap" where the | |
2669 | * gap is either the low watermark or 1% | |
2670 | * of the zone, whichever is smaller. | |
2671 | */ | |
2672 | balance_gap = min(low_wmark_pages(zone), | |
2673 | (zone->present_pages + | |
2674 | KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2675 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
2676 | /* | |
2677 | * Kswapd reclaims only single pages with compaction | |
2678 | * enabled. Trying too hard to reclaim until contiguous | |
2679 | * free pages have become available can hurt performance | |
2680 | * by evicting too much useful data from memory. | |
2681 | * Do not reclaim more than needed for compaction. | |
2682 | */ | |
2683 | testorder = order; | |
2684 | if (COMPACTION_BUILD && order && | |
2685 | compaction_suitable(zone, order) != | |
2686 | COMPACT_SKIPPED) | |
2687 | testorder = 0; | |
2688 | ||
2689 | if ((buffer_heads_over_limit && is_highmem_idx(i)) || | |
2690 | !zone_watermark_ok_safe(zone, testorder, | |
2691 | high_wmark_pages(zone) + balance_gap, | |
2692 | end_zone, 0)) { | |
2693 | shrink_zone(zone, &sc); | |
2694 | ||
2695 | reclaim_state->reclaimed_slab = 0; | |
2696 | nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages); | |
2697 | sc.nr_reclaimed += reclaim_state->reclaimed_slab; | |
2698 | total_scanned += sc.nr_scanned; | |
2699 | ||
2700 | if (nr_slab == 0 && !zone_reclaimable(zone)) | |
2701 | zone->all_unreclaimable = 1; | |
2702 | } | |
2703 | ||
2704 | /* | |
2705 | * If we've done a decent amount of scanning and | |
2706 | * the reclaim ratio is low, start doing writepage | |
2707 | * even in laptop mode | |
2708 | */ | |
2709 | if (total_scanned > SWAP_CLUSTER_MAX * 2 && | |
2710 | total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) | |
2711 | sc.may_writepage = 1; | |
2712 | ||
2713 | if (zone->all_unreclaimable) { | |
2714 | if (end_zone && end_zone == i) | |
2715 | end_zone--; | |
2716 | continue; | |
2717 | } | |
2718 | ||
2719 | if (!zone_watermark_ok_safe(zone, testorder, | |
2720 | high_wmark_pages(zone), end_zone, 0)) { | |
2721 | all_zones_ok = 0; | |
2722 | /* | |
2723 | * We are still under min water mark. This | |
2724 | * means that we have a GFP_ATOMIC allocation | |
2725 | * failure risk. Hurry up! | |
2726 | */ | |
2727 | if (!zone_watermark_ok_safe(zone, order, | |
2728 | min_wmark_pages(zone), end_zone, 0)) | |
2729 | has_under_min_watermark_zone = 1; | |
2730 | } else { | |
2731 | /* | |
2732 | * If a zone reaches its high watermark, | |
2733 | * consider it to be no longer congested. It's | |
2734 | * possible there are dirty pages backed by | |
2735 | * congested BDIs but as pressure is relieved, | |
2736 | * speculatively avoid congestion waits | |
2737 | */ | |
2738 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2739 | if (i <= *classzone_idx) | |
2740 | balanced += zone->present_pages; | |
2741 | } | |
2742 | ||
2743 | } | |
2744 | ||
2745 | /* | |
2746 | * If the low watermark is met there is no need for processes | |
2747 | * to be throttled on pfmemalloc_wait as they should not be | |
2748 | * able to safely make forward progress. Wake them | |
2749 | */ | |
2750 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
2751 | pfmemalloc_watermark_ok(pgdat)) | |
2752 | wake_up(&pgdat->pfmemalloc_wait); | |
2753 | ||
2754 | if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx))) | |
2755 | break; /* kswapd: all done */ | |
2756 | /* | |
2757 | * OK, kswapd is getting into trouble. Take a nap, then take | |
2758 | * another pass across the zones. | |
2759 | */ | |
2760 | if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) { | |
2761 | if (has_under_min_watermark_zone) | |
2762 | count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT); | |
2763 | else | |
2764 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2765 | } | |
2766 | ||
2767 | /* | |
2768 | * We do this so kswapd doesn't build up large priorities for | |
2769 | * example when it is freeing in parallel with allocators. It | |
2770 | * matches the direct reclaim path behaviour in terms of impact | |
2771 | * on zone->*_priority. | |
2772 | */ | |
2773 | if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) | |
2774 | break; | |
2775 | } while (--sc.priority >= 0); | |
2776 | out: | |
2777 | ||
2778 | /* | |
2779 | * order-0: All zones must meet high watermark for a balanced node | |
2780 | * high-order: Balanced zones must make up at least 25% of the node | |
2781 | * for the node to be balanced | |
2782 | */ | |
2783 | if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) { | |
2784 | cond_resched(); | |
2785 | ||
2786 | try_to_freeze(); | |
2787 | ||
2788 | /* | |
2789 | * Fragmentation may mean that the system cannot be | |
2790 | * rebalanced for high-order allocations in all zones. | |
2791 | * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, | |
2792 | * it means the zones have been fully scanned and are still | |
2793 | * not balanced. For high-order allocations, there is | |
2794 | * little point trying all over again as kswapd may | |
2795 | * infinite loop. | |
2796 | * | |
2797 | * Instead, recheck all watermarks at order-0 as they | |
2798 | * are the most important. If watermarks are ok, kswapd will go | |
2799 | * back to sleep. High-order users can still perform direct | |
2800 | * reclaim if they wish. | |
2801 | */ | |
2802 | if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) | |
2803 | order = sc.order = 0; | |
2804 | ||
2805 | goto loop_again; | |
2806 | } | |
2807 | ||
2808 | /* | |
2809 | * If kswapd was reclaiming at a higher order, it has the option of | |
2810 | * sleeping without all zones being balanced. Before it does, it must | |
2811 | * ensure that the watermarks for order-0 on *all* zones are met and | |
2812 | * that the congestion flags are cleared. The congestion flag must | |
2813 | * be cleared as kswapd is the only mechanism that clears the flag | |
2814 | * and it is potentially going to sleep here. | |
2815 | */ | |
2816 | if (order) { | |
2817 | int zones_need_compaction = 1; | |
2818 | ||
2819 | for (i = 0; i <= end_zone; i++) { | |
2820 | struct zone *zone = pgdat->node_zones + i; | |
2821 | ||
2822 | if (!populated_zone(zone)) | |
2823 | continue; | |
2824 | ||
2825 | if (zone->all_unreclaimable && | |
2826 | sc.priority != DEF_PRIORITY) | |
2827 | continue; | |
2828 | ||
2829 | /* Would compaction fail due to lack of free memory? */ | |
2830 | if (COMPACTION_BUILD && | |
2831 | compaction_suitable(zone, order) == COMPACT_SKIPPED) | |
2832 | goto loop_again; | |
2833 | ||
2834 | /* Confirm the zone is balanced for order-0 */ | |
2835 | if (!zone_watermark_ok(zone, 0, | |
2836 | high_wmark_pages(zone), 0, 0)) { | |
2837 | order = sc.order = 0; | |
2838 | goto loop_again; | |
2839 | } | |
2840 | ||
2841 | /* Check if the memory needs to be defragmented. */ | |
2842 | if (zone_watermark_ok(zone, order, | |
2843 | low_wmark_pages(zone), *classzone_idx, 0)) | |
2844 | zones_need_compaction = 0; | |
2845 | ||
2846 | /* If balanced, clear the congested flag */ | |
2847 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2848 | } | |
2849 | ||
2850 | if (zones_need_compaction) | |
2851 | compact_pgdat(pgdat, order); | |
2852 | } | |
2853 | ||
2854 | /* | |
2855 | * Return the order we were reclaiming at so prepare_kswapd_sleep() | |
2856 | * makes a decision on the order we were last reclaiming at. However, | |
2857 | * if another caller entered the allocator slow path while kswapd | |
2858 | * was awake, order will remain at the higher level | |
2859 | */ | |
2860 | *classzone_idx = end_zone; | |
2861 | return order; | |
2862 | } | |
2863 | ||
2864 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) | |
2865 | { | |
2866 | long remaining = 0; | |
2867 | DEFINE_WAIT(wait); | |
2868 | ||
2869 | if (freezing(current) || kthread_should_stop()) | |
2870 | return; | |
2871 | ||
2872 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2873 | ||
2874 | /* Try to sleep for a short interval */ | |
2875 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { | |
2876 | remaining = schedule_timeout(HZ/10); | |
2877 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2878 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2879 | } | |
2880 | ||
2881 | /* | |
2882 | * After a short sleep, check if it was a premature sleep. If not, then | |
2883 | * go fully to sleep until explicitly woken up. | |
2884 | */ | |
2885 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { | |
2886 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); | |
2887 | ||
2888 | /* | |
2889 | * vmstat counters are not perfectly accurate and the estimated | |
2890 | * value for counters such as NR_FREE_PAGES can deviate from the | |
2891 | * true value by nr_online_cpus * threshold. To avoid the zone | |
2892 | * watermarks being breached while under pressure, we reduce the | |
2893 | * per-cpu vmstat threshold while kswapd is awake and restore | |
2894 | * them before going back to sleep. | |
2895 | */ | |
2896 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
2897 | ||
2898 | /* | |
2899 | * Compaction records what page blocks it recently failed to | |
2900 | * isolate pages from and skips them in the future scanning. | |
2901 | * When kswapd is going to sleep, it is reasonable to assume | |
2902 | * that pages and compaction may succeed so reset the cache. | |
2903 | */ | |
2904 | reset_isolation_suitable(pgdat); | |
2905 | ||
2906 | if (!kthread_should_stop()) | |
2907 | schedule(); | |
2908 | ||
2909 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); | |
2910 | } else { | |
2911 | if (remaining) | |
2912 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
2913 | else | |
2914 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
2915 | } | |
2916 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2917 | } | |
2918 | ||
2919 | /* | |
2920 | * The background pageout daemon, started as a kernel thread | |
2921 | * from the init process. | |
2922 | * | |
2923 | * This basically trickles out pages so that we have _some_ | |
2924 | * free memory available even if there is no other activity | |
2925 | * that frees anything up. This is needed for things like routing | |
2926 | * etc, where we otherwise might have all activity going on in | |
2927 | * asynchronous contexts that cannot page things out. | |
2928 | * | |
2929 | * If there are applications that are active memory-allocators | |
2930 | * (most normal use), this basically shouldn't matter. | |
2931 | */ | |
2932 | static int kswapd(void *p) | |
2933 | { | |
2934 | unsigned long order, new_order; | |
2935 | unsigned balanced_order; | |
2936 | int classzone_idx, new_classzone_idx; | |
2937 | int balanced_classzone_idx; | |
2938 | pg_data_t *pgdat = (pg_data_t*)p; | |
2939 | struct task_struct *tsk = current; | |
2940 | ||
2941 | struct reclaim_state reclaim_state = { | |
2942 | .reclaimed_slab = 0, | |
2943 | }; | |
2944 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
2945 | ||
2946 | lockdep_set_current_reclaim_state(GFP_KERNEL); | |
2947 | ||
2948 | if (!cpumask_empty(cpumask)) | |
2949 | set_cpus_allowed_ptr(tsk, cpumask); | |
2950 | current->reclaim_state = &reclaim_state; | |
2951 | ||
2952 | /* | |
2953 | * Tell the memory management that we're a "memory allocator", | |
2954 | * and that if we need more memory we should get access to it | |
2955 | * regardless (see "__alloc_pages()"). "kswapd" should | |
2956 | * never get caught in the normal page freeing logic. | |
2957 | * | |
2958 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
2959 | * you need a small amount of memory in order to be able to | |
2960 | * page out something else, and this flag essentially protects | |
2961 | * us from recursively trying to free more memory as we're | |
2962 | * trying to free the first piece of memory in the first place). | |
2963 | */ | |
2964 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; | |
2965 | set_freezable(); | |
2966 | ||
2967 | order = new_order = 0; | |
2968 | balanced_order = 0; | |
2969 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; | |
2970 | balanced_classzone_idx = classzone_idx; | |
2971 | for ( ; ; ) { | |
2972 | int ret; | |
2973 | ||
2974 | /* | |
2975 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
2976 | * new request of a similar or harder type will succeed soon | |
2977 | * so consider going to sleep on the basis we reclaimed at | |
2978 | */ | |
2979 | if (balanced_classzone_idx >= new_classzone_idx && | |
2980 | balanced_order == new_order) { | |
2981 | new_order = pgdat->kswapd_max_order; | |
2982 | new_classzone_idx = pgdat->classzone_idx; | |
2983 | pgdat->kswapd_max_order = 0; | |
2984 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
2985 | } | |
2986 | ||
2987 | if (order < new_order || classzone_idx > new_classzone_idx) { | |
2988 | /* | |
2989 | * Don't sleep if someone wants a larger 'order' | |
2990 | * allocation or has tigher zone constraints | |
2991 | */ | |
2992 | order = new_order; | |
2993 | classzone_idx = new_classzone_idx; | |
2994 | } else { | |
2995 | kswapd_try_to_sleep(pgdat, balanced_order, | |
2996 | balanced_classzone_idx); | |
2997 | order = pgdat->kswapd_max_order; | |
2998 | classzone_idx = pgdat->classzone_idx; | |
2999 | new_order = order; | |
3000 | new_classzone_idx = classzone_idx; | |
3001 | pgdat->kswapd_max_order = 0; | |
3002 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3003 | } | |
3004 | ||
3005 | ret = try_to_freeze(); | |
3006 | if (kthread_should_stop()) | |
3007 | break; | |
3008 | ||
3009 | /* | |
3010 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3011 | * after returning from the refrigerator | |
3012 | */ | |
3013 | if (!ret) { | |
3014 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
3015 | balanced_classzone_idx = classzone_idx; | |
3016 | balanced_order = balance_pgdat(pgdat, order, | |
3017 | &balanced_classzone_idx); | |
3018 | } | |
3019 | } | |
3020 | return 0; | |
3021 | } | |
3022 | ||
3023 | /* | |
3024 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3025 | */ | |
3026 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) | |
3027 | { | |
3028 | pg_data_t *pgdat; | |
3029 | ||
3030 | if (!populated_zone(zone)) | |
3031 | return; | |
3032 | ||
3033 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
3034 | return; | |
3035 | pgdat = zone->zone_pgdat; | |
3036 | if (pgdat->kswapd_max_order < order) { | |
3037 | pgdat->kswapd_max_order = order; | |
3038 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); | |
3039 | } | |
3040 | if (!waitqueue_active(&pgdat->kswapd_wait)) | |
3041 | return; | |
3042 | if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) | |
3043 | return; | |
3044 | ||
3045 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
3046 | wake_up_interruptible(&pgdat->kswapd_wait); | |
3047 | } | |
3048 | ||
3049 | /* | |
3050 | * The reclaimable count would be mostly accurate. | |
3051 | * The less reclaimable pages may be | |
3052 | * - mlocked pages, which will be moved to unevictable list when encountered | |
3053 | * - mapped pages, which may require several travels to be reclaimed | |
3054 | * - dirty pages, which is not "instantly" reclaimable | |
3055 | */ | |
3056 | unsigned long global_reclaimable_pages(void) | |
3057 | { | |
3058 | int nr; | |
3059 | ||
3060 | nr = global_page_state(NR_ACTIVE_FILE) + | |
3061 | global_page_state(NR_INACTIVE_FILE); | |
3062 | ||
3063 | if (nr_swap_pages > 0) | |
3064 | nr += global_page_state(NR_ACTIVE_ANON) + | |
3065 | global_page_state(NR_INACTIVE_ANON); | |
3066 | ||
3067 | return nr; | |
3068 | } | |
3069 | ||
3070 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
3071 | { | |
3072 | int nr; | |
3073 | ||
3074 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
3075 | zone_page_state(zone, NR_INACTIVE_FILE); | |
3076 | ||
3077 | if (nr_swap_pages > 0) | |
3078 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
3079 | zone_page_state(zone, NR_INACTIVE_ANON); | |
3080 | ||
3081 | return nr; | |
3082 | } | |
3083 | ||
3084 | #ifdef CONFIG_HIBERNATION | |
3085 | /* | |
3086 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of | |
3087 | * freed pages. | |
3088 | * | |
3089 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3090 | * LRU order by reclaiming preferentially | |
3091 | * inactive > active > active referenced > active mapped | |
3092 | */ | |
3093 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) | |
3094 | { | |
3095 | struct reclaim_state reclaim_state; | |
3096 | struct scan_control sc = { | |
3097 | .gfp_mask = GFP_HIGHUSER_MOVABLE, | |
3098 | .may_swap = 1, | |
3099 | .may_unmap = 1, | |
3100 | .may_writepage = 1, | |
3101 | .nr_to_reclaim = nr_to_reclaim, | |
3102 | .hibernation_mode = 1, | |
3103 | .order = 0, | |
3104 | .priority = DEF_PRIORITY, | |
3105 | }; | |
3106 | struct shrink_control shrink = { | |
3107 | .gfp_mask = sc.gfp_mask, | |
3108 | }; | |
3109 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
3110 | struct task_struct *p = current; | |
3111 | unsigned long nr_reclaimed; | |
3112 | ||
3113 | p->flags |= PF_MEMALLOC; | |
3114 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3115 | reclaim_state.reclaimed_slab = 0; | |
3116 | p->reclaim_state = &reclaim_state; | |
3117 | ||
3118 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); | |
3119 | ||
3120 | p->reclaim_state = NULL; | |
3121 | lockdep_clear_current_reclaim_state(); | |
3122 | p->flags &= ~PF_MEMALLOC; | |
3123 | ||
3124 | return nr_reclaimed; | |
3125 | } | |
3126 | #endif /* CONFIG_HIBERNATION */ | |
3127 | ||
3128 | /* It's optimal to keep kswapds on the same CPUs as their memory, but | |
3129 | not required for correctness. So if the last cpu in a node goes | |
3130 | away, we get changed to run anywhere: as the first one comes back, | |
3131 | restore their cpu bindings. */ | |
3132 | static int __devinit cpu_callback(struct notifier_block *nfb, | |
3133 | unsigned long action, void *hcpu) | |
3134 | { | |
3135 | int nid; | |
3136 | ||
3137 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { | |
3138 | for_each_node_state(nid, N_HIGH_MEMORY) { | |
3139 | pg_data_t *pgdat = NODE_DATA(nid); | |
3140 | const struct cpumask *mask; | |
3141 | ||
3142 | mask = cpumask_of_node(pgdat->node_id); | |
3143 | ||
3144 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) | |
3145 | /* One of our CPUs online: restore mask */ | |
3146 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
3147 | } | |
3148 | } | |
3149 | return NOTIFY_OK; | |
3150 | } | |
3151 | ||
3152 | /* | |
3153 | * This kswapd start function will be called by init and node-hot-add. | |
3154 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3155 | */ | |
3156 | int kswapd_run(int nid) | |
3157 | { | |
3158 | pg_data_t *pgdat = NODE_DATA(nid); | |
3159 | int ret = 0; | |
3160 | ||
3161 | if (pgdat->kswapd) | |
3162 | return 0; | |
3163 | ||
3164 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3165 | if (IS_ERR(pgdat->kswapd)) { | |
3166 | /* failure at boot is fatal */ | |
3167 | BUG_ON(system_state == SYSTEM_BOOTING); | |
3168 | pgdat->kswapd = NULL; | |
3169 | pr_err("Failed to start kswapd on node %d\n", nid); | |
3170 | ret = PTR_ERR(pgdat->kswapd); | |
3171 | } | |
3172 | return ret; | |
3173 | } | |
3174 | ||
3175 | /* | |
3176 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
3177 | * hold lock_memory_hotplug(). | |
3178 | */ | |
3179 | void kswapd_stop(int nid) | |
3180 | { | |
3181 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3182 | ||
3183 | if (kswapd) { | |
3184 | kthread_stop(kswapd); | |
3185 | NODE_DATA(nid)->kswapd = NULL; | |
3186 | } | |
3187 | } | |
3188 | ||
3189 | static int __init kswapd_init(void) | |
3190 | { | |
3191 | int nid; | |
3192 | ||
3193 | swap_setup(); | |
3194 | for_each_node_state(nid, N_HIGH_MEMORY) | |
3195 | kswapd_run(nid); | |
3196 | hotcpu_notifier(cpu_callback, 0); | |
3197 | return 0; | |
3198 | } | |
3199 | ||
3200 | module_init(kswapd_init) | |
3201 | ||
3202 | #ifdef CONFIG_NUMA | |
3203 | /* | |
3204 | * Zone reclaim mode | |
3205 | * | |
3206 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3207 | * the watermarks. | |
3208 | */ | |
3209 | int zone_reclaim_mode __read_mostly; | |
3210 | ||
3211 | #define RECLAIM_OFF 0 | |
3212 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ | |
3213 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ | |
3214 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3215 | ||
3216 | /* | |
3217 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3218 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3219 | * a zone. | |
3220 | */ | |
3221 | #define ZONE_RECLAIM_PRIORITY 4 | |
3222 | ||
3223 | /* | |
3224 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3225 | * occur. | |
3226 | */ | |
3227 | int sysctl_min_unmapped_ratio = 1; | |
3228 | ||
3229 | /* | |
3230 | * If the number of slab pages in a zone grows beyond this percentage then | |
3231 | * slab reclaim needs to occur. | |
3232 | */ | |
3233 | int sysctl_min_slab_ratio = 5; | |
3234 | ||
3235 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) | |
3236 | { | |
3237 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3238 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3239 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3240 | ||
3241 | /* | |
3242 | * It's possible for there to be more file mapped pages than | |
3243 | * accounted for by the pages on the file LRU lists because | |
3244 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3245 | */ | |
3246 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3247 | } | |
3248 | ||
3249 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3250 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3251 | { | |
3252 | long nr_pagecache_reclaimable; | |
3253 | long delta = 0; | |
3254 | ||
3255 | /* | |
3256 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3257 | * potentially reclaimable. Otherwise, we have to worry about | |
3258 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3259 | * a better estimate | |
3260 | */ | |
3261 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3262 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3263 | else | |
3264 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3265 | ||
3266 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3267 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3268 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3269 | ||
3270 | /* Watch for any possible underflows due to delta */ | |
3271 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3272 | delta = nr_pagecache_reclaimable; | |
3273 | ||
3274 | return nr_pagecache_reclaimable - delta; | |
3275 | } | |
3276 | ||
3277 | /* | |
3278 | * Try to free up some pages from this zone through reclaim. | |
3279 | */ | |
3280 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3281 | { | |
3282 | /* Minimum pages needed in order to stay on node */ | |
3283 | const unsigned long nr_pages = 1 << order; | |
3284 | struct task_struct *p = current; | |
3285 | struct reclaim_state reclaim_state; | |
3286 | struct scan_control sc = { | |
3287 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
3288 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), | |
3289 | .may_swap = 1, | |
3290 | .nr_to_reclaim = max_t(unsigned long, nr_pages, | |
3291 | SWAP_CLUSTER_MAX), | |
3292 | .gfp_mask = gfp_mask, | |
3293 | .order = order, | |
3294 | .priority = ZONE_RECLAIM_PRIORITY, | |
3295 | }; | |
3296 | struct shrink_control shrink = { | |
3297 | .gfp_mask = sc.gfp_mask, | |
3298 | }; | |
3299 | unsigned long nr_slab_pages0, nr_slab_pages1; | |
3300 | ||
3301 | cond_resched(); | |
3302 | /* | |
3303 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3304 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3305 | * and RECLAIM_SWAP. | |
3306 | */ | |
3307 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
3308 | lockdep_set_current_reclaim_state(gfp_mask); | |
3309 | reclaim_state.reclaimed_slab = 0; | |
3310 | p->reclaim_state = &reclaim_state; | |
3311 | ||
3312 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { | |
3313 | /* | |
3314 | * Free memory by calling shrink zone with increasing | |
3315 | * priorities until we have enough memory freed. | |
3316 | */ | |
3317 | do { | |
3318 | shrink_zone(zone, &sc); | |
3319 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); | |
3320 | } | |
3321 | ||
3322 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); | |
3323 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
3324 | /* | |
3325 | * shrink_slab() does not currently allow us to determine how | |
3326 | * many pages were freed in this zone. So we take the current | |
3327 | * number of slab pages and shake the slab until it is reduced | |
3328 | * by the same nr_pages that we used for reclaiming unmapped | |
3329 | * pages. | |
3330 | * | |
3331 | * Note that shrink_slab will free memory on all zones and may | |
3332 | * take a long time. | |
3333 | */ | |
3334 | for (;;) { | |
3335 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3336 | ||
3337 | /* No reclaimable slab or very low memory pressure */ | |
3338 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) | |
3339 | break; | |
3340 | ||
3341 | /* Freed enough memory */ | |
3342 | nr_slab_pages1 = zone_page_state(zone, | |
3343 | NR_SLAB_RECLAIMABLE); | |
3344 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3345 | break; | |
3346 | } | |
3347 | ||
3348 | /* | |
3349 | * Update nr_reclaimed by the number of slab pages we | |
3350 | * reclaimed from this zone. | |
3351 | */ | |
3352 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); | |
3353 | if (nr_slab_pages1 < nr_slab_pages0) | |
3354 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
3355 | } | |
3356 | ||
3357 | p->reclaim_state = NULL; | |
3358 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); | |
3359 | lockdep_clear_current_reclaim_state(); | |
3360 | return sc.nr_reclaimed >= nr_pages; | |
3361 | } | |
3362 | ||
3363 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3364 | { | |
3365 | int node_id; | |
3366 | int ret; | |
3367 | ||
3368 | /* | |
3369 | * Zone reclaim reclaims unmapped file backed pages and | |
3370 | * slab pages if we are over the defined limits. | |
3371 | * | |
3372 | * A small portion of unmapped file backed pages is needed for | |
3373 | * file I/O otherwise pages read by file I/O will be immediately | |
3374 | * thrown out if the zone is overallocated. So we do not reclaim | |
3375 | * if less than a specified percentage of the zone is used by | |
3376 | * unmapped file backed pages. | |
3377 | */ | |
3378 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && | |
3379 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
3380 | return ZONE_RECLAIM_FULL; | |
3381 | ||
3382 | if (zone->all_unreclaimable) | |
3383 | return ZONE_RECLAIM_FULL; | |
3384 | ||
3385 | /* | |
3386 | * Do not scan if the allocation should not be delayed. | |
3387 | */ | |
3388 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) | |
3389 | return ZONE_RECLAIM_NOSCAN; | |
3390 | ||
3391 | /* | |
3392 | * Only run zone reclaim on the local zone or on zones that do not | |
3393 | * have associated processors. This will favor the local processor | |
3394 | * over remote processors and spread off node memory allocations | |
3395 | * as wide as possible. | |
3396 | */ | |
3397 | node_id = zone_to_nid(zone); | |
3398 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) | |
3399 | return ZONE_RECLAIM_NOSCAN; | |
3400 | ||
3401 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
3402 | return ZONE_RECLAIM_NOSCAN; | |
3403 | ||
3404 | ret = __zone_reclaim(zone, gfp_mask, order); | |
3405 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3406 | ||
3407 | if (!ret) | |
3408 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3409 | ||
3410 | return ret; | |
3411 | } | |
3412 | #endif | |
3413 | ||
3414 | /* | |
3415 | * page_evictable - test whether a page is evictable | |
3416 | * @page: the page to test | |
3417 | * @vma: the VMA in which the page is or will be mapped, may be NULL | |
3418 | * | |
3419 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
3420 | * lists vs unevictable list. The vma argument is !NULL when called from the | |
3421 | * fault path to determine how to instantate a new page. | |
3422 | * | |
3423 | * Reasons page might not be evictable: | |
3424 | * (1) page's mapping marked unevictable | |
3425 | * (2) page is part of an mlocked VMA | |
3426 | * | |
3427 | */ | |
3428 | int page_evictable(struct page *page, struct vm_area_struct *vma) | |
3429 | { | |
3430 | ||
3431 | if (mapping_unevictable(page_mapping(page))) | |
3432 | return 0; | |
3433 | ||
3434 | if (PageMlocked(page) || (vma && mlocked_vma_newpage(vma, page))) | |
3435 | return 0; | |
3436 | ||
3437 | return 1; | |
3438 | } | |
3439 | ||
3440 | #ifdef CONFIG_SHMEM | |
3441 | /** | |
3442 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list | |
3443 | * @pages: array of pages to check | |
3444 | * @nr_pages: number of pages to check | |
3445 | * | |
3446 | * Checks pages for evictability and moves them to the appropriate lru list. | |
3447 | * | |
3448 | * This function is only used for SysV IPC SHM_UNLOCK. | |
3449 | */ | |
3450 | void check_move_unevictable_pages(struct page **pages, int nr_pages) | |
3451 | { | |
3452 | struct lruvec *lruvec; | |
3453 | struct zone *zone = NULL; | |
3454 | int pgscanned = 0; | |
3455 | int pgrescued = 0; | |
3456 | int i; | |
3457 | ||
3458 | for (i = 0; i < nr_pages; i++) { | |
3459 | struct page *page = pages[i]; | |
3460 | struct zone *pagezone; | |
3461 | ||
3462 | pgscanned++; | |
3463 | pagezone = page_zone(page); | |
3464 | if (pagezone != zone) { | |
3465 | if (zone) | |
3466 | spin_unlock_irq(&zone->lru_lock); | |
3467 | zone = pagezone; | |
3468 | spin_lock_irq(&zone->lru_lock); | |
3469 | } | |
3470 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
3471 | ||
3472 | if (!PageLRU(page) || !PageUnevictable(page)) | |
3473 | continue; | |
3474 | ||
3475 | if (page_evictable(page, NULL)) { | |
3476 | enum lru_list lru = page_lru_base_type(page); | |
3477 | ||
3478 | VM_BUG_ON(PageActive(page)); | |
3479 | ClearPageUnevictable(page); | |
3480 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); | |
3481 | add_page_to_lru_list(page, lruvec, lru); | |
3482 | pgrescued++; | |
3483 | } | |
3484 | } | |
3485 | ||
3486 | if (zone) { | |
3487 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3488 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3489 | spin_unlock_irq(&zone->lru_lock); | |
3490 | } | |
3491 | } | |
3492 | #endif /* CONFIG_SHMEM */ | |
3493 | ||
3494 | static void warn_scan_unevictable_pages(void) | |
3495 | { | |
3496 | printk_once(KERN_WARNING | |
3497 | "%s: The scan_unevictable_pages sysctl/node-interface has been " | |
3498 | "disabled for lack of a legitimate use case. If you have " | |
3499 | "one, please send an email to linux-mm@kvack.org.\n", | |
3500 | current->comm); | |
3501 | } | |
3502 | ||
3503 | /* | |
3504 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3505 | * all nodes' unevictable lists for evictable pages | |
3506 | */ | |
3507 | unsigned long scan_unevictable_pages; | |
3508 | ||
3509 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
3510 | void __user *buffer, | |
3511 | size_t *length, loff_t *ppos) | |
3512 | { | |
3513 | warn_scan_unevictable_pages(); | |
3514 | proc_doulongvec_minmax(table, write, buffer, length, ppos); | |
3515 | scan_unevictable_pages = 0; | |
3516 | return 0; | |
3517 | } | |
3518 | ||
3519 | #ifdef CONFIG_NUMA | |
3520 | /* | |
3521 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3522 | * a specified node's per zone unevictable lists for evictable pages. | |
3523 | */ | |
3524 | ||
3525 | static ssize_t read_scan_unevictable_node(struct device *dev, | |
3526 | struct device_attribute *attr, | |
3527 | char *buf) | |
3528 | { | |
3529 | warn_scan_unevictable_pages(); | |
3530 | return sprintf(buf, "0\n"); /* always zero; should fit... */ | |
3531 | } | |
3532 | ||
3533 | static ssize_t write_scan_unevictable_node(struct device *dev, | |
3534 | struct device_attribute *attr, | |
3535 | const char *buf, size_t count) | |
3536 | { | |
3537 | warn_scan_unevictable_pages(); | |
3538 | return 1; | |
3539 | } | |
3540 | ||
3541 | ||
3542 | static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, | |
3543 | read_scan_unevictable_node, | |
3544 | write_scan_unevictable_node); | |
3545 | ||
3546 | int scan_unevictable_register_node(struct node *node) | |
3547 | { | |
3548 | return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages); | |
3549 | } | |
3550 | ||
3551 | void scan_unevictable_unregister_node(struct node *node) | |
3552 | { | |
3553 | device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages); | |
3554 | } | |
3555 | #endif |