]>
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/vmpressure.h> | |
23 | #include <linux/vmstat.h> | |
24 | #include <linux/file.h> | |
25 | #include <linux/writeback.h> | |
26 | #include <linux/blkdev.h> | |
27 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
28 | buffer_heads_over_limit */ | |
29 | #include <linux/mm_inline.h> | |
30 | #include <linux/backing-dev.h> | |
31 | #include <linux/rmap.h> | |
32 | #include <linux/topology.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/cpuset.h> | |
35 | #include <linux/compaction.h> | |
36 | #include <linux/notifier.h> | |
37 | #include <linux/rwsem.h> | |
38 | #include <linux/delay.h> | |
39 | #include <linux/kthread.h> | |
40 | #include <linux/freezer.h> | |
41 | #include <linux/memcontrol.h> | |
42 | #include <linux/delayacct.h> | |
43 | #include <linux/sysctl.h> | |
44 | #include <linux/oom.h> | |
45 | #include <linux/prefetch.h> | |
46 | ||
47 | #include <asm/tlbflush.h> | |
48 | #include <asm/div64.h> | |
49 | ||
50 | #include <linux/swapops.h> | |
51 | ||
52 | #include "internal.h" | |
53 | ||
54 | #define CREATE_TRACE_POINTS | |
55 | #include <trace/events/vmscan.h> | |
56 | ||
57 | struct scan_control { | |
58 | /* Incremented by the number of inactive pages that were scanned */ | |
59 | unsigned long nr_scanned; | |
60 | ||
61 | /* Number of pages freed so far during a call to shrink_zones() */ | |
62 | unsigned long nr_reclaimed; | |
63 | ||
64 | /* How many pages shrink_list() should reclaim */ | |
65 | unsigned long nr_to_reclaim; | |
66 | ||
67 | unsigned long hibernation_mode; | |
68 | ||
69 | /* This context's GFP mask */ | |
70 | gfp_t gfp_mask; | |
71 | ||
72 | int may_writepage; | |
73 | ||
74 | /* Can mapped pages be reclaimed? */ | |
75 | int may_unmap; | |
76 | ||
77 | /* Can pages be swapped as part of reclaim? */ | |
78 | int may_swap; | |
79 | ||
80 | int order; | |
81 | ||
82 | /* Scan (total_size >> priority) pages at once */ | |
83 | int priority; | |
84 | ||
85 | /* | |
86 | * The memory cgroup that hit its limit and as a result is the | |
87 | * primary target of this reclaim invocation. | |
88 | */ | |
89 | struct mem_cgroup *target_mem_cgroup; | |
90 | ||
91 | /* | |
92 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
93 | * are scanned. | |
94 | */ | |
95 | nodemask_t *nodemask; | |
96 | }; | |
97 | ||
98 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) | |
99 | ||
100 | #ifdef ARCH_HAS_PREFETCH | |
101 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
102 | do { \ | |
103 | if ((_page)->lru.prev != _base) { \ | |
104 | struct page *prev; \ | |
105 | \ | |
106 | prev = lru_to_page(&(_page->lru)); \ | |
107 | prefetch(&prev->_field); \ | |
108 | } \ | |
109 | } while (0) | |
110 | #else | |
111 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
112 | #endif | |
113 | ||
114 | #ifdef ARCH_HAS_PREFETCHW | |
115 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
116 | do { \ | |
117 | if ((_page)->lru.prev != _base) { \ | |
118 | struct page *prev; \ | |
119 | \ | |
120 | prev = lru_to_page(&(_page->lru)); \ | |
121 | prefetchw(&prev->_field); \ | |
122 | } \ | |
123 | } while (0) | |
124 | #else | |
125 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
126 | #endif | |
127 | ||
128 | /* | |
129 | * From 0 .. 100. Higher means more swappy. | |
130 | */ | |
131 | int vm_swappiness = 60; | |
132 | unsigned long vm_total_pages; /* The total number of pages which the VM controls */ | |
133 | ||
134 | static LIST_HEAD(shrinker_list); | |
135 | static DECLARE_RWSEM(shrinker_rwsem); | |
136 | ||
137 | #ifdef CONFIG_MEMCG | |
138 | static bool global_reclaim(struct scan_control *sc) | |
139 | { | |
140 | return !sc->target_mem_cgroup; | |
141 | } | |
142 | #else | |
143 | static bool global_reclaim(struct scan_control *sc) | |
144 | { | |
145 | return true; | |
146 | } | |
147 | #endif | |
148 | ||
149 | static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) | |
150 | { | |
151 | if (!mem_cgroup_disabled()) | |
152 | return mem_cgroup_get_lru_size(lruvec, lru); | |
153 | ||
154 | return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); | |
155 | } | |
156 | ||
157 | /* | |
158 | * Add a shrinker callback to be called from the vm | |
159 | */ | |
160 | void register_shrinker(struct shrinker *shrinker) | |
161 | { | |
162 | atomic_long_set(&shrinker->nr_in_batch, 0); | |
163 | down_write(&shrinker_rwsem); | |
164 | list_add_tail(&shrinker->list, &shrinker_list); | |
165 | up_write(&shrinker_rwsem); | |
166 | } | |
167 | EXPORT_SYMBOL(register_shrinker); | |
168 | ||
169 | /* | |
170 | * Remove one | |
171 | */ | |
172 | void unregister_shrinker(struct shrinker *shrinker) | |
173 | { | |
174 | down_write(&shrinker_rwsem); | |
175 | list_del(&shrinker->list); | |
176 | up_write(&shrinker_rwsem); | |
177 | } | |
178 | EXPORT_SYMBOL(unregister_shrinker); | |
179 | ||
180 | static inline int do_shrinker_shrink(struct shrinker *shrinker, | |
181 | struct shrink_control *sc, | |
182 | unsigned long nr_to_scan) | |
183 | { | |
184 | sc->nr_to_scan = nr_to_scan; | |
185 | return (*shrinker->shrink)(shrinker, sc); | |
186 | } | |
187 | ||
188 | #define SHRINK_BATCH 128 | |
189 | /* | |
190 | * Call the shrink functions to age shrinkable caches | |
191 | * | |
192 | * Here we assume it costs one seek to replace a lru page and that it also | |
193 | * takes a seek to recreate a cache object. With this in mind we age equal | |
194 | * percentages of the lru and ageable caches. This should balance the seeks | |
195 | * generated by these structures. | |
196 | * | |
197 | * If the vm encountered mapped pages on the LRU it increase the pressure on | |
198 | * slab to avoid swapping. | |
199 | * | |
200 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
201 | * | |
202 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
203 | * are eligible for the caller's allocation attempt. It is used for balancing | |
204 | * slab reclaim versus page reclaim. | |
205 | * | |
206 | * Returns the number of slab objects which we shrunk. | |
207 | */ | |
208 | unsigned long shrink_slab(struct shrink_control *shrink, | |
209 | unsigned long nr_pages_scanned, | |
210 | unsigned long lru_pages) | |
211 | { | |
212 | struct shrinker *shrinker; | |
213 | unsigned long ret = 0; | |
214 | ||
215 | if (nr_pages_scanned == 0) | |
216 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
217 | ||
218 | if (!down_read_trylock(&shrinker_rwsem)) { | |
219 | /* Assume we'll be able to shrink next time */ | |
220 | ret = 1; | |
221 | goto out; | |
222 | } | |
223 | ||
224 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
225 | unsigned long long delta; | |
226 | long total_scan; | |
227 | long max_pass; | |
228 | int shrink_ret = 0; | |
229 | long nr; | |
230 | long new_nr; | |
231 | long batch_size = shrinker->batch ? shrinker->batch | |
232 | : SHRINK_BATCH; | |
233 | ||
234 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); | |
235 | if (max_pass <= 0) | |
236 | continue; | |
237 | ||
238 | /* | |
239 | * copy the current shrinker scan count into a local variable | |
240 | * and zero it so that other concurrent shrinker invocations | |
241 | * don't also do this scanning work. | |
242 | */ | |
243 | nr = atomic_long_xchg(&shrinker->nr_in_batch, 0); | |
244 | ||
245 | total_scan = nr; | |
246 | delta = (4 * nr_pages_scanned) / shrinker->seeks; | |
247 | delta *= max_pass; | |
248 | do_div(delta, lru_pages + 1); | |
249 | total_scan += delta; | |
250 | if (total_scan < 0) { | |
251 | printk(KERN_ERR "shrink_slab: %pF negative objects to " | |
252 | "delete nr=%ld\n", | |
253 | shrinker->shrink, total_scan); | |
254 | total_scan = max_pass; | |
255 | } | |
256 | ||
257 | /* | |
258 | * We need to avoid excessive windup on filesystem shrinkers | |
259 | * due to large numbers of GFP_NOFS allocations causing the | |
260 | * shrinkers to return -1 all the time. This results in a large | |
261 | * nr being built up so when a shrink that can do some work | |
262 | * comes along it empties the entire cache due to nr >>> | |
263 | * max_pass. This is bad for sustaining a working set in | |
264 | * memory. | |
265 | * | |
266 | * Hence only allow the shrinker to scan the entire cache when | |
267 | * a large delta change is calculated directly. | |
268 | */ | |
269 | if (delta < max_pass / 4) | |
270 | total_scan = min(total_scan, max_pass / 2); | |
271 | ||
272 | /* | |
273 | * Avoid risking looping forever due to too large nr value: | |
274 | * never try to free more than twice the estimate number of | |
275 | * freeable entries. | |
276 | */ | |
277 | if (total_scan > max_pass * 2) | |
278 | total_scan = max_pass * 2; | |
279 | ||
280 | trace_mm_shrink_slab_start(shrinker, shrink, nr, | |
281 | nr_pages_scanned, lru_pages, | |
282 | max_pass, delta, total_scan); | |
283 | ||
284 | while (total_scan >= batch_size) { | |
285 | int nr_before; | |
286 | ||
287 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); | |
288 | shrink_ret = do_shrinker_shrink(shrinker, shrink, | |
289 | batch_size); | |
290 | if (shrink_ret == -1) | |
291 | break; | |
292 | if (shrink_ret < nr_before) | |
293 | ret += nr_before - shrink_ret; | |
294 | count_vm_events(SLABS_SCANNED, batch_size); | |
295 | total_scan -= batch_size; | |
296 | ||
297 | cond_resched(); | |
298 | } | |
299 | ||
300 | /* | |
301 | * move the unused scan count back into the shrinker in a | |
302 | * manner that handles concurrent updates. If we exhausted the | |
303 | * scan, there is no need to do an update. | |
304 | */ | |
305 | if (total_scan > 0) | |
306 | new_nr = atomic_long_add_return(total_scan, | |
307 | &shrinker->nr_in_batch); | |
308 | else | |
309 | new_nr = atomic_long_read(&shrinker->nr_in_batch); | |
310 | ||
311 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); | |
312 | } | |
313 | up_read(&shrinker_rwsem); | |
314 | out: | |
315 | cond_resched(); | |
316 | return ret; | |
317 | } | |
318 | ||
319 | static inline int is_page_cache_freeable(struct page *page) | |
320 | { | |
321 | /* | |
322 | * A freeable page cache page is referenced only by the caller | |
323 | * that isolated the page, the page cache radix tree and | |
324 | * optional buffer heads at page->private. | |
325 | */ | |
326 | return page_count(page) - page_has_private(page) == 2; | |
327 | } | |
328 | ||
329 | static int may_write_to_queue(struct backing_dev_info *bdi, | |
330 | struct scan_control *sc) | |
331 | { | |
332 | if (current->flags & PF_SWAPWRITE) | |
333 | return 1; | |
334 | if (!bdi_write_congested(bdi)) | |
335 | return 1; | |
336 | if (bdi == current->backing_dev_info) | |
337 | return 1; | |
338 | return 0; | |
339 | } | |
340 | ||
341 | /* | |
342 | * We detected a synchronous write error writing a page out. Probably | |
343 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
344 | * fsync(), msync() or close(). | |
345 | * | |
346 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
347 | * prevents it from being freed up. But we have a ref on the page and once | |
348 | * that page is locked, the mapping is pinned. | |
349 | * | |
350 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
351 | * __GFP_FS. | |
352 | */ | |
353 | static void handle_write_error(struct address_space *mapping, | |
354 | struct page *page, int error) | |
355 | { | |
356 | lock_page(page); | |
357 | if (page_mapping(page) == mapping) | |
358 | mapping_set_error(mapping, error); | |
359 | unlock_page(page); | |
360 | } | |
361 | ||
362 | /* possible outcome of pageout() */ | |
363 | typedef enum { | |
364 | /* failed to write page out, page is locked */ | |
365 | PAGE_KEEP, | |
366 | /* move page to the active list, page is locked */ | |
367 | PAGE_ACTIVATE, | |
368 | /* page has been sent to the disk successfully, page is unlocked */ | |
369 | PAGE_SUCCESS, | |
370 | /* page is clean and locked */ | |
371 | PAGE_CLEAN, | |
372 | } pageout_t; | |
373 | ||
374 | /* | |
375 | * pageout is called by shrink_page_list() for each dirty page. | |
376 | * Calls ->writepage(). | |
377 | */ | |
378 | static pageout_t pageout(struct page *page, struct address_space *mapping, | |
379 | struct scan_control *sc) | |
380 | { | |
381 | /* | |
382 | * If the page is dirty, only perform writeback if that write | |
383 | * will be non-blocking. To prevent this allocation from being | |
384 | * stalled by pagecache activity. But note that there may be | |
385 | * stalls if we need to run get_block(). We could test | |
386 | * PagePrivate for that. | |
387 | * | |
388 | * If this process is currently in __generic_file_aio_write() against | |
389 | * this page's queue, we can perform writeback even if that | |
390 | * will block. | |
391 | * | |
392 | * If the page is swapcache, write it back even if that would | |
393 | * block, for some throttling. This happens by accident, because | |
394 | * swap_backing_dev_info is bust: it doesn't reflect the | |
395 | * congestion state of the swapdevs. Easy to fix, if needed. | |
396 | */ | |
397 | if (!is_page_cache_freeable(page)) | |
398 | return PAGE_KEEP; | |
399 | if (!mapping) { | |
400 | /* | |
401 | * Some data journaling orphaned pages can have | |
402 | * page->mapping == NULL while being dirty with clean buffers. | |
403 | */ | |
404 | if (page_has_private(page)) { | |
405 | if (try_to_free_buffers(page)) { | |
406 | ClearPageDirty(page); | |
407 | printk("%s: orphaned page\n", __func__); | |
408 | return PAGE_CLEAN; | |
409 | } | |
410 | } | |
411 | return PAGE_KEEP; | |
412 | } | |
413 | if (mapping->a_ops->writepage == NULL) | |
414 | return PAGE_ACTIVATE; | |
415 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) | |
416 | return PAGE_KEEP; | |
417 | ||
418 | if (clear_page_dirty_for_io(page)) { | |
419 | int res; | |
420 | struct writeback_control wbc = { | |
421 | .sync_mode = WB_SYNC_NONE, | |
422 | .nr_to_write = SWAP_CLUSTER_MAX, | |
423 | .range_start = 0, | |
424 | .range_end = LLONG_MAX, | |
425 | .for_reclaim = 1, | |
426 | }; | |
427 | ||
428 | SetPageReclaim(page); | |
429 | res = mapping->a_ops->writepage(page, &wbc); | |
430 | if (res < 0) | |
431 | handle_write_error(mapping, page, res); | |
432 | if (res == AOP_WRITEPAGE_ACTIVATE) { | |
433 | ClearPageReclaim(page); | |
434 | return PAGE_ACTIVATE; | |
435 | } | |
436 | ||
437 | if (!PageWriteback(page)) { | |
438 | /* synchronous write or broken a_ops? */ | |
439 | ClearPageReclaim(page); | |
440 | } | |
441 | trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); | |
442 | inc_zone_page_state(page, NR_VMSCAN_WRITE); | |
443 | return PAGE_SUCCESS; | |
444 | } | |
445 | ||
446 | return PAGE_CLEAN; | |
447 | } | |
448 | ||
449 | /* | |
450 | * Same as remove_mapping, but if the page is removed from the mapping, it | |
451 | * gets returned with a refcount of 0. | |
452 | */ | |
453 | static int __remove_mapping(struct address_space *mapping, struct page *page) | |
454 | { | |
455 | BUG_ON(!PageLocked(page)); | |
456 | BUG_ON(mapping != page_mapping(page)); | |
457 | ||
458 | spin_lock_irq(&mapping->tree_lock); | |
459 | /* | |
460 | * The non racy check for a busy page. | |
461 | * | |
462 | * Must be careful with the order of the tests. When someone has | |
463 | * a ref to the page, it may be possible that they dirty it then | |
464 | * drop the reference. So if PageDirty is tested before page_count | |
465 | * here, then the following race may occur: | |
466 | * | |
467 | * get_user_pages(&page); | |
468 | * [user mapping goes away] | |
469 | * write_to(page); | |
470 | * !PageDirty(page) [good] | |
471 | * SetPageDirty(page); | |
472 | * put_page(page); | |
473 | * !page_count(page) [good, discard it] | |
474 | * | |
475 | * [oops, our write_to data is lost] | |
476 | * | |
477 | * Reversing the order of the tests ensures such a situation cannot | |
478 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
479 | * load is not satisfied before that of page->_count. | |
480 | * | |
481 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
482 | * and thus under tree_lock, then this ordering is not required. | |
483 | */ | |
484 | if (!page_freeze_refs(page, 2)) | |
485 | goto cannot_free; | |
486 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ | |
487 | if (unlikely(PageDirty(page))) { | |
488 | page_unfreeze_refs(page, 2); | |
489 | goto cannot_free; | |
490 | } | |
491 | ||
492 | if (PageSwapCache(page)) { | |
493 | swp_entry_t swap = { .val = page_private(page) }; | |
494 | __delete_from_swap_cache(page); | |
495 | spin_unlock_irq(&mapping->tree_lock); | |
496 | swapcache_free(swap, page); | |
497 | } else { | |
498 | void (*freepage)(struct page *); | |
499 | ||
500 | freepage = mapping->a_ops->freepage; | |
501 | ||
502 | __delete_from_page_cache(page); | |
503 | spin_unlock_irq(&mapping->tree_lock); | |
504 | mem_cgroup_uncharge_cache_page(page); | |
505 | ||
506 | if (freepage != NULL) | |
507 | freepage(page); | |
508 | } | |
509 | ||
510 | return 1; | |
511 | ||
512 | cannot_free: | |
513 | spin_unlock_irq(&mapping->tree_lock); | |
514 | return 0; | |
515 | } | |
516 | ||
517 | /* | |
518 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
519 | * someone else has a ref on the page, abort and return 0. If it was | |
520 | * successfully detached, return 1. Assumes the caller has a single ref on | |
521 | * this page. | |
522 | */ | |
523 | int remove_mapping(struct address_space *mapping, struct page *page) | |
524 | { | |
525 | if (__remove_mapping(mapping, page)) { | |
526 | /* | |
527 | * Unfreezing the refcount with 1 rather than 2 effectively | |
528 | * drops the pagecache ref for us without requiring another | |
529 | * atomic operation. | |
530 | */ | |
531 | page_unfreeze_refs(page, 1); | |
532 | return 1; | |
533 | } | |
534 | return 0; | |
535 | } | |
536 | ||
537 | /** | |
538 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
539 | * @page: page to be put back to appropriate lru list | |
540 | * | |
541 | * Add previously isolated @page to appropriate LRU list. | |
542 | * Page may still be unevictable for other reasons. | |
543 | * | |
544 | * lru_lock must not be held, interrupts must be enabled. | |
545 | */ | |
546 | void putback_lru_page(struct page *page) | |
547 | { | |
548 | int lru; | |
549 | int active = !!TestClearPageActive(page); | |
550 | int was_unevictable = PageUnevictable(page); | |
551 | ||
552 | VM_BUG_ON(PageLRU(page)); | |
553 | ||
554 | redo: | |
555 | ClearPageUnevictable(page); | |
556 | ||
557 | if (page_evictable(page)) { | |
558 | /* | |
559 | * For evictable pages, we can use the cache. | |
560 | * In event of a race, worst case is we end up with an | |
561 | * unevictable page on [in]active list. | |
562 | * We know how to handle that. | |
563 | */ | |
564 | lru = active + page_lru_base_type(page); | |
565 | lru_cache_add_lru(page, lru); | |
566 | } else { | |
567 | /* | |
568 | * Put unevictable pages directly on zone's unevictable | |
569 | * list. | |
570 | */ | |
571 | lru = LRU_UNEVICTABLE; | |
572 | add_page_to_unevictable_list(page); | |
573 | /* | |
574 | * When racing with an mlock or AS_UNEVICTABLE clearing | |
575 | * (page is unlocked) make sure that if the other thread | |
576 | * does not observe our setting of PG_lru and fails | |
577 | * isolation/check_move_unevictable_pages, | |
578 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move | |
579 | * the page back to the evictable list. | |
580 | * | |
581 | * The other side is TestClearPageMlocked() or shmem_lock(). | |
582 | */ | |
583 | smp_mb(); | |
584 | } | |
585 | ||
586 | /* | |
587 | * page's status can change while we move it among lru. If an evictable | |
588 | * page is on unevictable list, it never be freed. To avoid that, | |
589 | * check after we added it to the list, again. | |
590 | */ | |
591 | if (lru == LRU_UNEVICTABLE && page_evictable(page)) { | |
592 | if (!isolate_lru_page(page)) { | |
593 | put_page(page); | |
594 | goto redo; | |
595 | } | |
596 | /* This means someone else dropped this page from LRU | |
597 | * So, it will be freed or putback to LRU again. There is | |
598 | * nothing to do here. | |
599 | */ | |
600 | } | |
601 | ||
602 | if (was_unevictable && lru != LRU_UNEVICTABLE) | |
603 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
604 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) | |
605 | count_vm_event(UNEVICTABLE_PGCULLED); | |
606 | ||
607 | put_page(page); /* drop ref from isolate */ | |
608 | } | |
609 | ||
610 | enum page_references { | |
611 | PAGEREF_RECLAIM, | |
612 | PAGEREF_RECLAIM_CLEAN, | |
613 | PAGEREF_KEEP, | |
614 | PAGEREF_ACTIVATE, | |
615 | }; | |
616 | ||
617 | static enum page_references page_check_references(struct page *page, | |
618 | struct scan_control *sc) | |
619 | { | |
620 | int referenced_ptes, referenced_page; | |
621 | unsigned long vm_flags; | |
622 | ||
623 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, | |
624 | &vm_flags); | |
625 | referenced_page = TestClearPageReferenced(page); | |
626 | ||
627 | /* | |
628 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
629 | * move the page to the unevictable list. | |
630 | */ | |
631 | if (vm_flags & VM_LOCKED) | |
632 | return PAGEREF_RECLAIM; | |
633 | ||
634 | if (referenced_ptes) { | |
635 | if (PageSwapBacked(page)) | |
636 | return PAGEREF_ACTIVATE; | |
637 | /* | |
638 | * All mapped pages start out with page table | |
639 | * references from the instantiating fault, so we need | |
640 | * to look twice if a mapped file page is used more | |
641 | * than once. | |
642 | * | |
643 | * Mark it and spare it for another trip around the | |
644 | * inactive list. Another page table reference will | |
645 | * lead to its activation. | |
646 | * | |
647 | * Note: the mark is set for activated pages as well | |
648 | * so that recently deactivated but used pages are | |
649 | * quickly recovered. | |
650 | */ | |
651 | SetPageReferenced(page); | |
652 | ||
653 | if (referenced_page || referenced_ptes > 1) | |
654 | return PAGEREF_ACTIVATE; | |
655 | ||
656 | /* | |
657 | * Activate file-backed executable pages after first usage. | |
658 | */ | |
659 | if (vm_flags & VM_EXEC) | |
660 | return PAGEREF_ACTIVATE; | |
661 | ||
662 | return PAGEREF_KEEP; | |
663 | } | |
664 | ||
665 | /* Reclaim if clean, defer dirty pages to writeback */ | |
666 | if (referenced_page && !PageSwapBacked(page)) | |
667 | return PAGEREF_RECLAIM_CLEAN; | |
668 | ||
669 | return PAGEREF_RECLAIM; | |
670 | } | |
671 | ||
672 | /* | |
673 | * shrink_page_list() returns the number of reclaimed pages | |
674 | */ | |
675 | static unsigned long shrink_page_list(struct list_head *page_list, | |
676 | struct zone *zone, | |
677 | struct scan_control *sc, | |
678 | enum ttu_flags ttu_flags, | |
679 | unsigned long *ret_nr_unqueued_dirty, | |
680 | unsigned long *ret_nr_writeback, | |
681 | bool force_reclaim) | |
682 | { | |
683 | LIST_HEAD(ret_pages); | |
684 | LIST_HEAD(free_pages); | |
685 | int pgactivate = 0; | |
686 | unsigned long nr_unqueued_dirty = 0; | |
687 | unsigned long nr_dirty = 0; | |
688 | unsigned long nr_congested = 0; | |
689 | unsigned long nr_reclaimed = 0; | |
690 | unsigned long nr_writeback = 0; | |
691 | ||
692 | cond_resched(); | |
693 | ||
694 | mem_cgroup_uncharge_start(); | |
695 | while (!list_empty(page_list)) { | |
696 | struct address_space *mapping; | |
697 | struct page *page; | |
698 | int may_enter_fs; | |
699 | enum page_references references = PAGEREF_RECLAIM_CLEAN; | |
700 | ||
701 | cond_resched(); | |
702 | ||
703 | page = lru_to_page(page_list); | |
704 | list_del(&page->lru); | |
705 | ||
706 | if (!trylock_page(page)) | |
707 | goto keep; | |
708 | ||
709 | VM_BUG_ON(PageActive(page)); | |
710 | VM_BUG_ON(page_zone(page) != zone); | |
711 | ||
712 | sc->nr_scanned++; | |
713 | ||
714 | if (unlikely(!page_evictable(page))) | |
715 | goto cull_mlocked; | |
716 | ||
717 | if (!sc->may_unmap && page_mapped(page)) | |
718 | goto keep_locked; | |
719 | ||
720 | /* Double the slab pressure for mapped and swapcache pages */ | |
721 | if (page_mapped(page) || PageSwapCache(page)) | |
722 | sc->nr_scanned++; | |
723 | ||
724 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || | |
725 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
726 | ||
727 | /* | |
728 | * If a page at the tail of the LRU is under writeback, there | |
729 | * are three cases to consider. | |
730 | * | |
731 | * 1) If reclaim is encountering an excessive number of pages | |
732 | * under writeback and this page is both under writeback and | |
733 | * PageReclaim then it indicates that pages are being queued | |
734 | * for IO but are being recycled through the LRU before the | |
735 | * IO can complete. Waiting on the page itself risks an | |
736 | * indefinite stall if it is impossible to writeback the | |
737 | * page due to IO error or disconnected storage so instead | |
738 | * block for HZ/10 or until some IO completes then clear the | |
739 | * ZONE_WRITEBACK flag to recheck if the condition exists. | |
740 | * | |
741 | * 2) Global reclaim encounters a page, memcg encounters a | |
742 | * page that is not marked for immediate reclaim or | |
743 | * the caller does not have __GFP_IO. In this case mark | |
744 | * the page for immediate reclaim and continue scanning. | |
745 | * | |
746 | * __GFP_IO is checked because a loop driver thread might | |
747 | * enter reclaim, and deadlock if it waits on a page for | |
748 | * which it is needed to do the write (loop masks off | |
749 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
750 | * would probably show more reasons. | |
751 | * | |
752 | * Don't require __GFP_FS, since we're not going into the | |
753 | * FS, just waiting on its writeback completion. Worryingly, | |
754 | * ext4 gfs2 and xfs allocate pages with | |
755 | * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing | |
756 | * may_enter_fs here is liable to OOM on them. | |
757 | * | |
758 | * 3) memcg encounters a page that is not already marked | |
759 | * PageReclaim. memcg does not have any dirty pages | |
760 | * throttling so we could easily OOM just because too many | |
761 | * pages are in writeback and there is nothing else to | |
762 | * reclaim. Wait for the writeback to complete. | |
763 | */ | |
764 | if (PageWriteback(page)) { | |
765 | /* Case 1 above */ | |
766 | if (current_is_kswapd() && | |
767 | PageReclaim(page) && | |
768 | zone_is_reclaim_writeback(zone)) { | |
769 | unlock_page(page); | |
770 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
771 | zone_clear_flag(zone, ZONE_WRITEBACK); | |
772 | goto keep; | |
773 | ||
774 | /* Case 2 above */ | |
775 | } else if (global_reclaim(sc) || | |
776 | !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) { | |
777 | /* | |
778 | * This is slightly racy - end_page_writeback() | |
779 | * might have just cleared PageReclaim, then | |
780 | * setting PageReclaim here end up interpreted | |
781 | * as PageReadahead - but that does not matter | |
782 | * enough to care. What we do want is for this | |
783 | * page to have PageReclaim set next time memcg | |
784 | * reclaim reaches the tests above, so it will | |
785 | * then wait_on_page_writeback() to avoid OOM; | |
786 | * and it's also appropriate in global reclaim. | |
787 | */ | |
788 | SetPageReclaim(page); | |
789 | nr_writeback++; | |
790 | ||
791 | goto keep_locked; | |
792 | ||
793 | /* Case 3 above */ | |
794 | } else { | |
795 | wait_on_page_writeback(page); | |
796 | } | |
797 | } | |
798 | ||
799 | if (!force_reclaim) | |
800 | references = page_check_references(page, sc); | |
801 | ||
802 | switch (references) { | |
803 | case PAGEREF_ACTIVATE: | |
804 | goto activate_locked; | |
805 | case PAGEREF_KEEP: | |
806 | goto keep_locked; | |
807 | case PAGEREF_RECLAIM: | |
808 | case PAGEREF_RECLAIM_CLEAN: | |
809 | ; /* try to reclaim the page below */ | |
810 | } | |
811 | ||
812 | /* | |
813 | * Anonymous process memory has backing store? | |
814 | * Try to allocate it some swap space here. | |
815 | */ | |
816 | if (PageAnon(page) && !PageSwapCache(page)) { | |
817 | if (!(sc->gfp_mask & __GFP_IO)) | |
818 | goto keep_locked; | |
819 | if (!add_to_swap(page, page_list)) | |
820 | goto activate_locked; | |
821 | may_enter_fs = 1; | |
822 | } | |
823 | ||
824 | mapping = page_mapping(page); | |
825 | ||
826 | /* | |
827 | * The page is mapped into the page tables of one or more | |
828 | * processes. Try to unmap it here. | |
829 | */ | |
830 | if (page_mapped(page) && mapping) { | |
831 | switch (try_to_unmap(page, ttu_flags)) { | |
832 | case SWAP_FAIL: | |
833 | goto activate_locked; | |
834 | case SWAP_AGAIN: | |
835 | goto keep_locked; | |
836 | case SWAP_MLOCK: | |
837 | goto cull_mlocked; | |
838 | case SWAP_SUCCESS: | |
839 | ; /* try to free the page below */ | |
840 | } | |
841 | } | |
842 | ||
843 | if (PageDirty(page)) { | |
844 | nr_dirty++; | |
845 | ||
846 | if (!PageWriteback(page)) | |
847 | nr_unqueued_dirty++; | |
848 | ||
849 | /* | |
850 | * Only kswapd can writeback filesystem pages to | |
851 | * avoid risk of stack overflow but only writeback | |
852 | * if many dirty pages have been encountered. | |
853 | */ | |
854 | if (page_is_file_cache(page) && | |
855 | (!current_is_kswapd() || | |
856 | !zone_is_reclaim_dirty(zone))) { | |
857 | /* | |
858 | * Immediately reclaim when written back. | |
859 | * Similar in principal to deactivate_page() | |
860 | * except we already have the page isolated | |
861 | * and know it's dirty | |
862 | */ | |
863 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
864 | SetPageReclaim(page); | |
865 | ||
866 | goto keep_locked; | |
867 | } | |
868 | ||
869 | if (references == PAGEREF_RECLAIM_CLEAN) | |
870 | goto keep_locked; | |
871 | if (!may_enter_fs) | |
872 | goto keep_locked; | |
873 | if (!sc->may_writepage) | |
874 | goto keep_locked; | |
875 | ||
876 | /* Page is dirty, try to write it out here */ | |
877 | switch (pageout(page, mapping, sc)) { | |
878 | case PAGE_KEEP: | |
879 | nr_congested++; | |
880 | goto keep_locked; | |
881 | case PAGE_ACTIVATE: | |
882 | goto activate_locked; | |
883 | case PAGE_SUCCESS: | |
884 | if (PageWriteback(page)) | |
885 | goto keep; | |
886 | if (PageDirty(page)) | |
887 | goto keep; | |
888 | ||
889 | /* | |
890 | * A synchronous write - probably a ramdisk. Go | |
891 | * ahead and try to reclaim the page. | |
892 | */ | |
893 | if (!trylock_page(page)) | |
894 | goto keep; | |
895 | if (PageDirty(page) || PageWriteback(page)) | |
896 | goto keep_locked; | |
897 | mapping = page_mapping(page); | |
898 | case PAGE_CLEAN: | |
899 | ; /* try to free the page below */ | |
900 | } | |
901 | } | |
902 | ||
903 | /* | |
904 | * If the page has buffers, try to free the buffer mappings | |
905 | * associated with this page. If we succeed we try to free | |
906 | * the page as well. | |
907 | * | |
908 | * We do this even if the page is PageDirty(). | |
909 | * try_to_release_page() does not perform I/O, but it is | |
910 | * possible for a page to have PageDirty set, but it is actually | |
911 | * clean (all its buffers are clean). This happens if the | |
912 | * buffers were written out directly, with submit_bh(). ext3 | |
913 | * will do this, as well as the blockdev mapping. | |
914 | * try_to_release_page() will discover that cleanness and will | |
915 | * drop the buffers and mark the page clean - it can be freed. | |
916 | * | |
917 | * Rarely, pages can have buffers and no ->mapping. These are | |
918 | * the pages which were not successfully invalidated in | |
919 | * truncate_complete_page(). We try to drop those buffers here | |
920 | * and if that worked, and the page is no longer mapped into | |
921 | * process address space (page_count == 1) it can be freed. | |
922 | * Otherwise, leave the page on the LRU so it is swappable. | |
923 | */ | |
924 | if (page_has_private(page)) { | |
925 | if (!try_to_release_page(page, sc->gfp_mask)) | |
926 | goto activate_locked; | |
927 | if (!mapping && page_count(page) == 1) { | |
928 | unlock_page(page); | |
929 | if (put_page_testzero(page)) | |
930 | goto free_it; | |
931 | else { | |
932 | /* | |
933 | * rare race with speculative reference. | |
934 | * the speculative reference will free | |
935 | * this page shortly, so we may | |
936 | * increment nr_reclaimed here (and | |
937 | * leave it off the LRU). | |
938 | */ | |
939 | nr_reclaimed++; | |
940 | continue; | |
941 | } | |
942 | } | |
943 | } | |
944 | ||
945 | if (!mapping || !__remove_mapping(mapping, page)) | |
946 | goto keep_locked; | |
947 | ||
948 | /* | |
949 | * At this point, we have no other references and there is | |
950 | * no way to pick any more up (removed from LRU, removed | |
951 | * from pagecache). Can use non-atomic bitops now (and | |
952 | * we obviously don't have to worry about waking up a process | |
953 | * waiting on the page lock, because there are no references. | |
954 | */ | |
955 | __clear_page_locked(page); | |
956 | free_it: | |
957 | nr_reclaimed++; | |
958 | ||
959 | /* | |
960 | * Is there need to periodically free_page_list? It would | |
961 | * appear not as the counts should be low | |
962 | */ | |
963 | list_add(&page->lru, &free_pages); | |
964 | continue; | |
965 | ||
966 | cull_mlocked: | |
967 | if (PageSwapCache(page)) | |
968 | try_to_free_swap(page); | |
969 | unlock_page(page); | |
970 | putback_lru_page(page); | |
971 | continue; | |
972 | ||
973 | activate_locked: | |
974 | /* Not a candidate for swapping, so reclaim swap space. */ | |
975 | if (PageSwapCache(page) && vm_swap_full()) | |
976 | try_to_free_swap(page); | |
977 | VM_BUG_ON(PageActive(page)); | |
978 | SetPageActive(page); | |
979 | pgactivate++; | |
980 | keep_locked: | |
981 | unlock_page(page); | |
982 | keep: | |
983 | list_add(&page->lru, &ret_pages); | |
984 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); | |
985 | } | |
986 | ||
987 | /* | |
988 | * Tag a zone as congested if all the dirty pages encountered were | |
989 | * backed by a congested BDI. In this case, reclaimers should just | |
990 | * back off and wait for congestion to clear because further reclaim | |
991 | * will encounter the same problem | |
992 | */ | |
993 | if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc)) | |
994 | zone_set_flag(zone, ZONE_CONGESTED); | |
995 | ||
996 | free_hot_cold_page_list(&free_pages, 1); | |
997 | ||
998 | list_splice(&ret_pages, page_list); | |
999 | count_vm_events(PGACTIVATE, pgactivate); | |
1000 | mem_cgroup_uncharge_end(); | |
1001 | *ret_nr_unqueued_dirty += nr_unqueued_dirty; | |
1002 | *ret_nr_writeback += nr_writeback; | |
1003 | return nr_reclaimed; | |
1004 | } | |
1005 | ||
1006 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, | |
1007 | struct list_head *page_list) | |
1008 | { | |
1009 | struct scan_control sc = { | |
1010 | .gfp_mask = GFP_KERNEL, | |
1011 | .priority = DEF_PRIORITY, | |
1012 | .may_unmap = 1, | |
1013 | }; | |
1014 | unsigned long ret, dummy1, dummy2; | |
1015 | struct page *page, *next; | |
1016 | LIST_HEAD(clean_pages); | |
1017 | ||
1018 | list_for_each_entry_safe(page, next, page_list, lru) { | |
1019 | if (page_is_file_cache(page) && !PageDirty(page)) { | |
1020 | ClearPageActive(page); | |
1021 | list_move(&page->lru, &clean_pages); | |
1022 | } | |
1023 | } | |
1024 | ||
1025 | ret = shrink_page_list(&clean_pages, zone, &sc, | |
1026 | TTU_UNMAP|TTU_IGNORE_ACCESS, | |
1027 | &dummy1, &dummy2, true); | |
1028 | list_splice(&clean_pages, page_list); | |
1029 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); | |
1030 | return ret; | |
1031 | } | |
1032 | ||
1033 | /* | |
1034 | * Attempt to remove the specified page from its LRU. Only take this page | |
1035 | * if it is of the appropriate PageActive status. Pages which are being | |
1036 | * freed elsewhere are also ignored. | |
1037 | * | |
1038 | * page: page to consider | |
1039 | * mode: one of the LRU isolation modes defined above | |
1040 | * | |
1041 | * returns 0 on success, -ve errno on failure. | |
1042 | */ | |
1043 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) | |
1044 | { | |
1045 | int ret = -EINVAL; | |
1046 | ||
1047 | /* Only take pages on the LRU. */ | |
1048 | if (!PageLRU(page)) | |
1049 | return ret; | |
1050 | ||
1051 | /* Compaction should not handle unevictable pages but CMA can do so */ | |
1052 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
1053 | return ret; | |
1054 | ||
1055 | ret = -EBUSY; | |
1056 | ||
1057 | /* | |
1058 | * To minimise LRU disruption, the caller can indicate that it only | |
1059 | * wants to isolate pages it will be able to operate on without | |
1060 | * blocking - clean pages for the most part. | |
1061 | * | |
1062 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1063 | * is used by reclaim when it is cannot write to backing storage | |
1064 | * | |
1065 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1066 | * that it is possible to migrate without blocking | |
1067 | */ | |
1068 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1069 | /* All the caller can do on PageWriteback is block */ | |
1070 | if (PageWriteback(page)) | |
1071 | return ret; | |
1072 | ||
1073 | if (PageDirty(page)) { | |
1074 | struct address_space *mapping; | |
1075 | ||
1076 | /* ISOLATE_CLEAN means only clean pages */ | |
1077 | if (mode & ISOLATE_CLEAN) | |
1078 | return ret; | |
1079 | ||
1080 | /* | |
1081 | * Only pages without mappings or that have a | |
1082 | * ->migratepage callback are possible to migrate | |
1083 | * without blocking | |
1084 | */ | |
1085 | mapping = page_mapping(page); | |
1086 | if (mapping && !mapping->a_ops->migratepage) | |
1087 | return ret; | |
1088 | } | |
1089 | } | |
1090 | ||
1091 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) | |
1092 | return ret; | |
1093 | ||
1094 | if (likely(get_page_unless_zero(page))) { | |
1095 | /* | |
1096 | * Be careful not to clear PageLRU until after we're | |
1097 | * sure the page is not being freed elsewhere -- the | |
1098 | * page release code relies on it. | |
1099 | */ | |
1100 | ClearPageLRU(page); | |
1101 | ret = 0; | |
1102 | } | |
1103 | ||
1104 | return ret; | |
1105 | } | |
1106 | ||
1107 | /* | |
1108 | * zone->lru_lock is heavily contended. Some of the functions that | |
1109 | * shrink the lists perform better by taking out a batch of pages | |
1110 | * and working on them outside the LRU lock. | |
1111 | * | |
1112 | * For pagecache intensive workloads, this function is the hottest | |
1113 | * spot in the kernel (apart from copy_*_user functions). | |
1114 | * | |
1115 | * Appropriate locks must be held before calling this function. | |
1116 | * | |
1117 | * @nr_to_scan: The number of pages to look through on the list. | |
1118 | * @lruvec: The LRU vector to pull pages from. | |
1119 | * @dst: The temp list to put pages on to. | |
1120 | * @nr_scanned: The number of pages that were scanned. | |
1121 | * @sc: The scan_control struct for this reclaim session | |
1122 | * @mode: One of the LRU isolation modes | |
1123 | * @lru: LRU list id for isolating | |
1124 | * | |
1125 | * returns how many pages were moved onto *@dst. | |
1126 | */ | |
1127 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, | |
1128 | struct lruvec *lruvec, struct list_head *dst, | |
1129 | unsigned long *nr_scanned, struct scan_control *sc, | |
1130 | isolate_mode_t mode, enum lru_list lru) | |
1131 | { | |
1132 | struct list_head *src = &lruvec->lists[lru]; | |
1133 | unsigned long nr_taken = 0; | |
1134 | unsigned long scan; | |
1135 | ||
1136 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { | |
1137 | struct page *page; | |
1138 | int nr_pages; | |
1139 | ||
1140 | page = lru_to_page(src); | |
1141 | prefetchw_prev_lru_page(page, src, flags); | |
1142 | ||
1143 | VM_BUG_ON(!PageLRU(page)); | |
1144 | ||
1145 | switch (__isolate_lru_page(page, mode)) { | |
1146 | case 0: | |
1147 | nr_pages = hpage_nr_pages(page); | |
1148 | mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); | |
1149 | list_move(&page->lru, dst); | |
1150 | nr_taken += nr_pages; | |
1151 | break; | |
1152 | ||
1153 | case -EBUSY: | |
1154 | /* else it is being freed elsewhere */ | |
1155 | list_move(&page->lru, src); | |
1156 | continue; | |
1157 | ||
1158 | default: | |
1159 | BUG(); | |
1160 | } | |
1161 | } | |
1162 | ||
1163 | *nr_scanned = scan; | |
1164 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, | |
1165 | nr_taken, mode, is_file_lru(lru)); | |
1166 | return nr_taken; | |
1167 | } | |
1168 | ||
1169 | /** | |
1170 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1171 | * @page: page to isolate from its LRU list | |
1172 | * | |
1173 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1174 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1175 | * | |
1176 | * Returns 0 if the page was removed from an LRU list. | |
1177 | * Returns -EBUSY if the page was not on an LRU list. | |
1178 | * | |
1179 | * The returned page will have PageLRU() cleared. If it was found on | |
1180 | * the active list, it will have PageActive set. If it was found on | |
1181 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1182 | * may need to be cleared by the caller before letting the page go. | |
1183 | * | |
1184 | * The vmstat statistic corresponding to the list on which the page was | |
1185 | * found will be decremented. | |
1186 | * | |
1187 | * Restrictions: | |
1188 | * (1) Must be called with an elevated refcount on the page. This is a | |
1189 | * fundamentnal difference from isolate_lru_pages (which is called | |
1190 | * without a stable reference). | |
1191 | * (2) the lru_lock must not be held. | |
1192 | * (3) interrupts must be enabled. | |
1193 | */ | |
1194 | int isolate_lru_page(struct page *page) | |
1195 | { | |
1196 | int ret = -EBUSY; | |
1197 | ||
1198 | VM_BUG_ON(!page_count(page)); | |
1199 | ||
1200 | if (PageLRU(page)) { | |
1201 | struct zone *zone = page_zone(page); | |
1202 | struct lruvec *lruvec; | |
1203 | ||
1204 | spin_lock_irq(&zone->lru_lock); | |
1205 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1206 | if (PageLRU(page)) { | |
1207 | int lru = page_lru(page); | |
1208 | get_page(page); | |
1209 | ClearPageLRU(page); | |
1210 | del_page_from_lru_list(page, lruvec, lru); | |
1211 | ret = 0; | |
1212 | } | |
1213 | spin_unlock_irq(&zone->lru_lock); | |
1214 | } | |
1215 | return ret; | |
1216 | } | |
1217 | ||
1218 | /* | |
1219 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and | |
1220 | * then get resheduled. When there are massive number of tasks doing page | |
1221 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1222 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1223 | * unnecessary swapping, thrashing and OOM. | |
1224 | */ | |
1225 | static int too_many_isolated(struct zone *zone, int file, | |
1226 | struct scan_control *sc) | |
1227 | { | |
1228 | unsigned long inactive, isolated; | |
1229 | ||
1230 | if (current_is_kswapd()) | |
1231 | return 0; | |
1232 | ||
1233 | if (!global_reclaim(sc)) | |
1234 | return 0; | |
1235 | ||
1236 | if (file) { | |
1237 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1238 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1239 | } else { | |
1240 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1241 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1242 | } | |
1243 | ||
1244 | /* | |
1245 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1246 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1247 | * deadlock. | |
1248 | */ | |
1249 | if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS) | |
1250 | inactive >>= 3; | |
1251 | ||
1252 | return isolated > inactive; | |
1253 | } | |
1254 | ||
1255 | static noinline_for_stack void | |
1256 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) | |
1257 | { | |
1258 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1259 | struct zone *zone = lruvec_zone(lruvec); | |
1260 | LIST_HEAD(pages_to_free); | |
1261 | ||
1262 | /* | |
1263 | * Put back any unfreeable pages. | |
1264 | */ | |
1265 | while (!list_empty(page_list)) { | |
1266 | struct page *page = lru_to_page(page_list); | |
1267 | int lru; | |
1268 | ||
1269 | VM_BUG_ON(PageLRU(page)); | |
1270 | list_del(&page->lru); | |
1271 | if (unlikely(!page_evictable(page))) { | |
1272 | spin_unlock_irq(&zone->lru_lock); | |
1273 | putback_lru_page(page); | |
1274 | spin_lock_irq(&zone->lru_lock); | |
1275 | continue; | |
1276 | } | |
1277 | ||
1278 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1279 | ||
1280 | SetPageLRU(page); | |
1281 | lru = page_lru(page); | |
1282 | add_page_to_lru_list(page, lruvec, lru); | |
1283 | ||
1284 | if (is_active_lru(lru)) { | |
1285 | int file = is_file_lru(lru); | |
1286 | int numpages = hpage_nr_pages(page); | |
1287 | reclaim_stat->recent_rotated[file] += numpages; | |
1288 | } | |
1289 | if (put_page_testzero(page)) { | |
1290 | __ClearPageLRU(page); | |
1291 | __ClearPageActive(page); | |
1292 | del_page_from_lru_list(page, lruvec, lru); | |
1293 | ||
1294 | if (unlikely(PageCompound(page))) { | |
1295 | spin_unlock_irq(&zone->lru_lock); | |
1296 | (*get_compound_page_dtor(page))(page); | |
1297 | spin_lock_irq(&zone->lru_lock); | |
1298 | } else | |
1299 | list_add(&page->lru, &pages_to_free); | |
1300 | } | |
1301 | } | |
1302 | ||
1303 | /* | |
1304 | * To save our caller's stack, now use input list for pages to free. | |
1305 | */ | |
1306 | list_splice(&pages_to_free, page_list); | |
1307 | } | |
1308 | ||
1309 | /* | |
1310 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number | |
1311 | * of reclaimed pages | |
1312 | */ | |
1313 | static noinline_for_stack unsigned long | |
1314 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, | |
1315 | struct scan_control *sc, enum lru_list lru) | |
1316 | { | |
1317 | LIST_HEAD(page_list); | |
1318 | unsigned long nr_scanned; | |
1319 | unsigned long nr_reclaimed = 0; | |
1320 | unsigned long nr_taken; | |
1321 | unsigned long nr_dirty = 0; | |
1322 | unsigned long nr_writeback = 0; | |
1323 | isolate_mode_t isolate_mode = 0; | |
1324 | int file = is_file_lru(lru); | |
1325 | struct zone *zone = lruvec_zone(lruvec); | |
1326 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1327 | ||
1328 | while (unlikely(too_many_isolated(zone, file, sc))) { | |
1329 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
1330 | ||
1331 | /* We are about to die and free our memory. Return now. */ | |
1332 | if (fatal_signal_pending(current)) | |
1333 | return SWAP_CLUSTER_MAX; | |
1334 | } | |
1335 | ||
1336 | lru_add_drain(); | |
1337 | ||
1338 | if (!sc->may_unmap) | |
1339 | isolate_mode |= ISOLATE_UNMAPPED; | |
1340 | if (!sc->may_writepage) | |
1341 | isolate_mode |= ISOLATE_CLEAN; | |
1342 | ||
1343 | spin_lock_irq(&zone->lru_lock); | |
1344 | ||
1345 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, | |
1346 | &nr_scanned, sc, isolate_mode, lru); | |
1347 | ||
1348 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1349 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1350 | ||
1351 | if (global_reclaim(sc)) { | |
1352 | zone->pages_scanned += nr_scanned; | |
1353 | if (current_is_kswapd()) | |
1354 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); | |
1355 | else | |
1356 | __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); | |
1357 | } | |
1358 | spin_unlock_irq(&zone->lru_lock); | |
1359 | ||
1360 | if (nr_taken == 0) | |
1361 | return 0; | |
1362 | ||
1363 | nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, | |
1364 | &nr_dirty, &nr_writeback, false); | |
1365 | ||
1366 | spin_lock_irq(&zone->lru_lock); | |
1367 | ||
1368 | reclaim_stat->recent_scanned[file] += nr_taken; | |
1369 | ||
1370 | if (global_reclaim(sc)) { | |
1371 | if (current_is_kswapd()) | |
1372 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1373 | nr_reclaimed); | |
1374 | else | |
1375 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1376 | nr_reclaimed); | |
1377 | } | |
1378 | ||
1379 | putback_inactive_pages(lruvec, &page_list); | |
1380 | ||
1381 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); | |
1382 | ||
1383 | spin_unlock_irq(&zone->lru_lock); | |
1384 | ||
1385 | free_hot_cold_page_list(&page_list, 1); | |
1386 | ||
1387 | /* | |
1388 | * If reclaim is isolating dirty pages under writeback, it implies | |
1389 | * that the long-lived page allocation rate is exceeding the page | |
1390 | * laundering rate. Either the global limits are not being effective | |
1391 | * at throttling processes due to the page distribution throughout | |
1392 | * zones or there is heavy usage of a slow backing device. The | |
1393 | * only option is to throttle from reclaim context which is not ideal | |
1394 | * as there is no guarantee the dirtying process is throttled in the | |
1395 | * same way balance_dirty_pages() manages. | |
1396 | * | |
1397 | * This scales the number of dirty pages that must be under writeback | |
1398 | * before throttling depending on priority. It is a simple backoff | |
1399 | * function that has the most effect in the range DEF_PRIORITY to | |
1400 | * DEF_PRIORITY-2 which is the priority reclaim is considered to be | |
1401 | * in trouble and reclaim is considered to be in trouble. | |
1402 | * | |
1403 | * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle | |
1404 | * DEF_PRIORITY-1 50% must be PageWriteback | |
1405 | * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble | |
1406 | * ... | |
1407 | * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any | |
1408 | * isolated page is PageWriteback | |
1409 | */ | |
1410 | if (nr_writeback && nr_writeback >= | |
1411 | (nr_taken >> (DEF_PRIORITY - sc->priority))) { | |
1412 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); | |
1413 | zone_set_flag(zone, ZONE_WRITEBACK); | |
1414 | } | |
1415 | ||
1416 | /* | |
1417 | * Similarly, if many dirty pages are encountered that are not | |
1418 | * currently being written then flag that kswapd should start | |
1419 | * writing back pages. | |
1420 | */ | |
1421 | if (global_reclaim(sc) && nr_dirty && | |
1422 | nr_dirty >= (nr_taken >> (DEF_PRIORITY - sc->priority))) | |
1423 | zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY); | |
1424 | ||
1425 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, | |
1426 | zone_idx(zone), | |
1427 | nr_scanned, nr_reclaimed, | |
1428 | sc->priority, | |
1429 | trace_shrink_flags(file)); | |
1430 | return nr_reclaimed; | |
1431 | } | |
1432 | ||
1433 | /* | |
1434 | * This moves pages from the active list to the inactive list. | |
1435 | * | |
1436 | * We move them the other way if the page is referenced by one or more | |
1437 | * processes, from rmap. | |
1438 | * | |
1439 | * If the pages are mostly unmapped, the processing is fast and it is | |
1440 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1441 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1442 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1443 | * this, so instead we remove the pages from the LRU while processing them. | |
1444 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1445 | * nobody will play with that bit on a non-LRU page. | |
1446 | * | |
1447 | * The downside is that we have to touch page->_count against each page. | |
1448 | * But we had to alter page->flags anyway. | |
1449 | */ | |
1450 | ||
1451 | static void move_active_pages_to_lru(struct lruvec *lruvec, | |
1452 | struct list_head *list, | |
1453 | struct list_head *pages_to_free, | |
1454 | enum lru_list lru) | |
1455 | { | |
1456 | struct zone *zone = lruvec_zone(lruvec); | |
1457 | unsigned long pgmoved = 0; | |
1458 | struct page *page; | |
1459 | int nr_pages; | |
1460 | ||
1461 | while (!list_empty(list)) { | |
1462 | page = lru_to_page(list); | |
1463 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1464 | ||
1465 | VM_BUG_ON(PageLRU(page)); | |
1466 | SetPageLRU(page); | |
1467 | ||
1468 | nr_pages = hpage_nr_pages(page); | |
1469 | mem_cgroup_update_lru_size(lruvec, lru, nr_pages); | |
1470 | list_move(&page->lru, &lruvec->lists[lru]); | |
1471 | pgmoved += nr_pages; | |
1472 | ||
1473 | if (put_page_testzero(page)) { | |
1474 | __ClearPageLRU(page); | |
1475 | __ClearPageActive(page); | |
1476 | del_page_from_lru_list(page, lruvec, lru); | |
1477 | ||
1478 | if (unlikely(PageCompound(page))) { | |
1479 | spin_unlock_irq(&zone->lru_lock); | |
1480 | (*get_compound_page_dtor(page))(page); | |
1481 | spin_lock_irq(&zone->lru_lock); | |
1482 | } else | |
1483 | list_add(&page->lru, pages_to_free); | |
1484 | } | |
1485 | } | |
1486 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1487 | if (!is_active_lru(lru)) | |
1488 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1489 | } | |
1490 | ||
1491 | static void shrink_active_list(unsigned long nr_to_scan, | |
1492 | struct lruvec *lruvec, | |
1493 | struct scan_control *sc, | |
1494 | enum lru_list lru) | |
1495 | { | |
1496 | unsigned long nr_taken; | |
1497 | unsigned long nr_scanned; | |
1498 | unsigned long vm_flags; | |
1499 | LIST_HEAD(l_hold); /* The pages which were snipped off */ | |
1500 | LIST_HEAD(l_active); | |
1501 | LIST_HEAD(l_inactive); | |
1502 | struct page *page; | |
1503 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1504 | unsigned long nr_rotated = 0; | |
1505 | isolate_mode_t isolate_mode = 0; | |
1506 | int file = is_file_lru(lru); | |
1507 | struct zone *zone = lruvec_zone(lruvec); | |
1508 | ||
1509 | lru_add_drain(); | |
1510 | ||
1511 | if (!sc->may_unmap) | |
1512 | isolate_mode |= ISOLATE_UNMAPPED; | |
1513 | if (!sc->may_writepage) | |
1514 | isolate_mode |= ISOLATE_CLEAN; | |
1515 | ||
1516 | spin_lock_irq(&zone->lru_lock); | |
1517 | ||
1518 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, | |
1519 | &nr_scanned, sc, isolate_mode, lru); | |
1520 | if (global_reclaim(sc)) | |
1521 | zone->pages_scanned += nr_scanned; | |
1522 | ||
1523 | reclaim_stat->recent_scanned[file] += nr_taken; | |
1524 | ||
1525 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); | |
1526 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1527 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1528 | spin_unlock_irq(&zone->lru_lock); | |
1529 | ||
1530 | while (!list_empty(&l_hold)) { | |
1531 | cond_resched(); | |
1532 | page = lru_to_page(&l_hold); | |
1533 | list_del(&page->lru); | |
1534 | ||
1535 | if (unlikely(!page_evictable(page))) { | |
1536 | putback_lru_page(page); | |
1537 | continue; | |
1538 | } | |
1539 | ||
1540 | if (unlikely(buffer_heads_over_limit)) { | |
1541 | if (page_has_private(page) && trylock_page(page)) { | |
1542 | if (page_has_private(page)) | |
1543 | try_to_release_page(page, 0); | |
1544 | unlock_page(page); | |
1545 | } | |
1546 | } | |
1547 | ||
1548 | if (page_referenced(page, 0, sc->target_mem_cgroup, | |
1549 | &vm_flags)) { | |
1550 | nr_rotated += hpage_nr_pages(page); | |
1551 | /* | |
1552 | * Identify referenced, file-backed active pages and | |
1553 | * give them one more trip around the active list. So | |
1554 | * that executable code get better chances to stay in | |
1555 | * memory under moderate memory pressure. Anon pages | |
1556 | * are not likely to be evicted by use-once streaming | |
1557 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1558 | * so we ignore them here. | |
1559 | */ | |
1560 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { | |
1561 | list_add(&page->lru, &l_active); | |
1562 | continue; | |
1563 | } | |
1564 | } | |
1565 | ||
1566 | ClearPageActive(page); /* we are de-activating */ | |
1567 | list_add(&page->lru, &l_inactive); | |
1568 | } | |
1569 | ||
1570 | /* | |
1571 | * Move pages back to the lru list. | |
1572 | */ | |
1573 | spin_lock_irq(&zone->lru_lock); | |
1574 | /* | |
1575 | * Count referenced pages from currently used mappings as rotated, | |
1576 | * even though only some of them are actually re-activated. This | |
1577 | * helps balance scan pressure between file and anonymous pages in | |
1578 | * get_scan_ratio. | |
1579 | */ | |
1580 | reclaim_stat->recent_rotated[file] += nr_rotated; | |
1581 | ||
1582 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); | |
1583 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
1584 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); | |
1585 | spin_unlock_irq(&zone->lru_lock); | |
1586 | ||
1587 | free_hot_cold_page_list(&l_hold, 1); | |
1588 | } | |
1589 | ||
1590 | #ifdef CONFIG_SWAP | |
1591 | static int inactive_anon_is_low_global(struct zone *zone) | |
1592 | { | |
1593 | unsigned long active, inactive; | |
1594 | ||
1595 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1596 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1597 | ||
1598 | if (inactive * zone->inactive_ratio < active) | |
1599 | return 1; | |
1600 | ||
1601 | return 0; | |
1602 | } | |
1603 | ||
1604 | /** | |
1605 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
1606 | * @lruvec: LRU vector to check | |
1607 | * | |
1608 | * Returns true if the zone does not have enough inactive anon pages, | |
1609 | * meaning some active anon pages need to be deactivated. | |
1610 | */ | |
1611 | static int inactive_anon_is_low(struct lruvec *lruvec) | |
1612 | { | |
1613 | /* | |
1614 | * If we don't have swap space, anonymous page deactivation | |
1615 | * is pointless. | |
1616 | */ | |
1617 | if (!total_swap_pages) | |
1618 | return 0; | |
1619 | ||
1620 | if (!mem_cgroup_disabled()) | |
1621 | return mem_cgroup_inactive_anon_is_low(lruvec); | |
1622 | ||
1623 | return inactive_anon_is_low_global(lruvec_zone(lruvec)); | |
1624 | } | |
1625 | #else | |
1626 | static inline int inactive_anon_is_low(struct lruvec *lruvec) | |
1627 | { | |
1628 | return 0; | |
1629 | } | |
1630 | #endif | |
1631 | ||
1632 | /** | |
1633 | * inactive_file_is_low - check if file pages need to be deactivated | |
1634 | * @lruvec: LRU vector to check | |
1635 | * | |
1636 | * When the system is doing streaming IO, memory pressure here | |
1637 | * ensures that active file pages get deactivated, until more | |
1638 | * than half of the file pages are on the inactive list. | |
1639 | * | |
1640 | * Once we get to that situation, protect the system's working | |
1641 | * set from being evicted by disabling active file page aging. | |
1642 | * | |
1643 | * This uses a different ratio than the anonymous pages, because | |
1644 | * the page cache uses a use-once replacement algorithm. | |
1645 | */ | |
1646 | static int inactive_file_is_low(struct lruvec *lruvec) | |
1647 | { | |
1648 | unsigned long inactive; | |
1649 | unsigned long active; | |
1650 | ||
1651 | inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
1652 | active = get_lru_size(lruvec, LRU_ACTIVE_FILE); | |
1653 | ||
1654 | return active > inactive; | |
1655 | } | |
1656 | ||
1657 | static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) | |
1658 | { | |
1659 | if (is_file_lru(lru)) | |
1660 | return inactive_file_is_low(lruvec); | |
1661 | else | |
1662 | return inactive_anon_is_low(lruvec); | |
1663 | } | |
1664 | ||
1665 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, | |
1666 | struct lruvec *lruvec, struct scan_control *sc) | |
1667 | { | |
1668 | if (is_active_lru(lru)) { | |
1669 | if (inactive_list_is_low(lruvec, lru)) | |
1670 | shrink_active_list(nr_to_scan, lruvec, sc, lru); | |
1671 | return 0; | |
1672 | } | |
1673 | ||
1674 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); | |
1675 | } | |
1676 | ||
1677 | static int vmscan_swappiness(struct scan_control *sc) | |
1678 | { | |
1679 | if (global_reclaim(sc)) | |
1680 | return vm_swappiness; | |
1681 | return mem_cgroup_swappiness(sc->target_mem_cgroup); | |
1682 | } | |
1683 | ||
1684 | enum scan_balance { | |
1685 | SCAN_EQUAL, | |
1686 | SCAN_FRACT, | |
1687 | SCAN_ANON, | |
1688 | SCAN_FILE, | |
1689 | }; | |
1690 | ||
1691 | /* | |
1692 | * Determine how aggressively the anon and file LRU lists should be | |
1693 | * scanned. The relative value of each set of LRU lists is determined | |
1694 | * by looking at the fraction of the pages scanned we did rotate back | |
1695 | * onto the active list instead of evict. | |
1696 | * | |
1697 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan | |
1698 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
1699 | */ | |
1700 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, | |
1701 | unsigned long *nr) | |
1702 | { | |
1703 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
1704 | u64 fraction[2]; | |
1705 | u64 denominator = 0; /* gcc */ | |
1706 | struct zone *zone = lruvec_zone(lruvec); | |
1707 | unsigned long anon_prio, file_prio; | |
1708 | enum scan_balance scan_balance; | |
1709 | unsigned long anon, file, free; | |
1710 | bool force_scan = false; | |
1711 | unsigned long ap, fp; | |
1712 | enum lru_list lru; | |
1713 | ||
1714 | /* | |
1715 | * If the zone or memcg is small, nr[l] can be 0. This | |
1716 | * results in no scanning on this priority and a potential | |
1717 | * priority drop. Global direct reclaim can go to the next | |
1718 | * zone and tends to have no problems. Global kswapd is for | |
1719 | * zone balancing and it needs to scan a minimum amount. When | |
1720 | * reclaiming for a memcg, a priority drop can cause high | |
1721 | * latencies, so it's better to scan a minimum amount there as | |
1722 | * well. | |
1723 | */ | |
1724 | if (current_is_kswapd() && zone->all_unreclaimable) | |
1725 | force_scan = true; | |
1726 | if (!global_reclaim(sc)) | |
1727 | force_scan = true; | |
1728 | ||
1729 | /* If we have no swap space, do not bother scanning anon pages. */ | |
1730 | if (!sc->may_swap || (get_nr_swap_pages() <= 0)) { | |
1731 | scan_balance = SCAN_FILE; | |
1732 | goto out; | |
1733 | } | |
1734 | ||
1735 | /* | |
1736 | * Global reclaim will swap to prevent OOM even with no | |
1737 | * swappiness, but memcg users want to use this knob to | |
1738 | * disable swapping for individual groups completely when | |
1739 | * using the memory controller's swap limit feature would be | |
1740 | * too expensive. | |
1741 | */ | |
1742 | if (!global_reclaim(sc) && !vmscan_swappiness(sc)) { | |
1743 | scan_balance = SCAN_FILE; | |
1744 | goto out; | |
1745 | } | |
1746 | ||
1747 | /* | |
1748 | * Do not apply any pressure balancing cleverness when the | |
1749 | * system is close to OOM, scan both anon and file equally | |
1750 | * (unless the swappiness setting disagrees with swapping). | |
1751 | */ | |
1752 | if (!sc->priority && vmscan_swappiness(sc)) { | |
1753 | scan_balance = SCAN_EQUAL; | |
1754 | goto out; | |
1755 | } | |
1756 | ||
1757 | anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + | |
1758 | get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
1759 | file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
1760 | get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
1761 | ||
1762 | /* | |
1763 | * If it's foreseeable that reclaiming the file cache won't be | |
1764 | * enough to get the zone back into a desirable shape, we have | |
1765 | * to swap. Better start now and leave the - probably heavily | |
1766 | * thrashing - remaining file pages alone. | |
1767 | */ | |
1768 | if (global_reclaim(sc)) { | |
1769 | free = zone_page_state(zone, NR_FREE_PAGES); | |
1770 | if (unlikely(file + free <= high_wmark_pages(zone))) { | |
1771 | scan_balance = SCAN_ANON; | |
1772 | goto out; | |
1773 | } | |
1774 | } | |
1775 | ||
1776 | /* | |
1777 | * There is enough inactive page cache, do not reclaim | |
1778 | * anything from the anonymous working set right now. | |
1779 | */ | |
1780 | if (!inactive_file_is_low(lruvec)) { | |
1781 | scan_balance = SCAN_FILE; | |
1782 | goto out; | |
1783 | } | |
1784 | ||
1785 | scan_balance = SCAN_FRACT; | |
1786 | ||
1787 | /* | |
1788 | * With swappiness at 100, anonymous and file have the same priority. | |
1789 | * This scanning priority is essentially the inverse of IO cost. | |
1790 | */ | |
1791 | anon_prio = vmscan_swappiness(sc); | |
1792 | file_prio = 200 - anon_prio; | |
1793 | ||
1794 | /* | |
1795 | * OK, so we have swap space and a fair amount of page cache | |
1796 | * pages. We use the recently rotated / recently scanned | |
1797 | * ratios to determine how valuable each cache is. | |
1798 | * | |
1799 | * Because workloads change over time (and to avoid overflow) | |
1800 | * we keep these statistics as a floating average, which ends | |
1801 | * up weighing recent references more than old ones. | |
1802 | * | |
1803 | * anon in [0], file in [1] | |
1804 | */ | |
1805 | spin_lock_irq(&zone->lru_lock); | |
1806 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { | |
1807 | reclaim_stat->recent_scanned[0] /= 2; | |
1808 | reclaim_stat->recent_rotated[0] /= 2; | |
1809 | } | |
1810 | ||
1811 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { | |
1812 | reclaim_stat->recent_scanned[1] /= 2; | |
1813 | reclaim_stat->recent_rotated[1] /= 2; | |
1814 | } | |
1815 | ||
1816 | /* | |
1817 | * The amount of pressure on anon vs file pages is inversely | |
1818 | * proportional to the fraction of recently scanned pages on | |
1819 | * each list that were recently referenced and in active use. | |
1820 | */ | |
1821 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); | |
1822 | ap /= reclaim_stat->recent_rotated[0] + 1; | |
1823 | ||
1824 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); | |
1825 | fp /= reclaim_stat->recent_rotated[1] + 1; | |
1826 | spin_unlock_irq(&zone->lru_lock); | |
1827 | ||
1828 | fraction[0] = ap; | |
1829 | fraction[1] = fp; | |
1830 | denominator = ap + fp + 1; | |
1831 | out: | |
1832 | for_each_evictable_lru(lru) { | |
1833 | int file = is_file_lru(lru); | |
1834 | unsigned long size; | |
1835 | unsigned long scan; | |
1836 | ||
1837 | size = get_lru_size(lruvec, lru); | |
1838 | scan = size >> sc->priority; | |
1839 | ||
1840 | if (!scan && force_scan) | |
1841 | scan = min(size, SWAP_CLUSTER_MAX); | |
1842 | ||
1843 | switch (scan_balance) { | |
1844 | case SCAN_EQUAL: | |
1845 | /* Scan lists relative to size */ | |
1846 | break; | |
1847 | case SCAN_FRACT: | |
1848 | /* | |
1849 | * Scan types proportional to swappiness and | |
1850 | * their relative recent reclaim efficiency. | |
1851 | */ | |
1852 | scan = div64_u64(scan * fraction[file], denominator); | |
1853 | break; | |
1854 | case SCAN_FILE: | |
1855 | case SCAN_ANON: | |
1856 | /* Scan one type exclusively */ | |
1857 | if ((scan_balance == SCAN_FILE) != file) | |
1858 | scan = 0; | |
1859 | break; | |
1860 | default: | |
1861 | /* Look ma, no brain */ | |
1862 | BUG(); | |
1863 | } | |
1864 | nr[lru] = scan; | |
1865 | } | |
1866 | } | |
1867 | ||
1868 | /* | |
1869 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1870 | */ | |
1871 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) | |
1872 | { | |
1873 | unsigned long nr[NR_LRU_LISTS]; | |
1874 | unsigned long targets[NR_LRU_LISTS]; | |
1875 | unsigned long nr_to_scan; | |
1876 | enum lru_list lru; | |
1877 | unsigned long nr_reclaimed = 0; | |
1878 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
1879 | struct blk_plug plug; | |
1880 | bool scan_adjusted = false; | |
1881 | ||
1882 | get_scan_count(lruvec, sc, nr); | |
1883 | ||
1884 | /* Record the original scan target for proportional adjustments later */ | |
1885 | memcpy(targets, nr, sizeof(nr)); | |
1886 | ||
1887 | blk_start_plug(&plug); | |
1888 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
1889 | nr[LRU_INACTIVE_FILE]) { | |
1890 | unsigned long nr_anon, nr_file, percentage; | |
1891 | unsigned long nr_scanned; | |
1892 | ||
1893 | for_each_evictable_lru(lru) { | |
1894 | if (nr[lru]) { | |
1895 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
1896 | nr[lru] -= nr_to_scan; | |
1897 | ||
1898 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
1899 | lruvec, sc); | |
1900 | } | |
1901 | } | |
1902 | ||
1903 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) | |
1904 | continue; | |
1905 | ||
1906 | /* | |
1907 | * For global direct reclaim, reclaim only the number of pages | |
1908 | * requested. Less care is taken to scan proportionally as it | |
1909 | * is more important to minimise direct reclaim stall latency | |
1910 | * than it is to properly age the LRU lists. | |
1911 | */ | |
1912 | if (global_reclaim(sc) && !current_is_kswapd()) | |
1913 | break; | |
1914 | ||
1915 | /* | |
1916 | * For kswapd and memcg, reclaim at least the number of pages | |
1917 | * requested. Ensure that the anon and file LRUs shrink | |
1918 | * proportionally what was requested by get_scan_count(). We | |
1919 | * stop reclaiming one LRU and reduce the amount scanning | |
1920 | * proportional to the original scan target. | |
1921 | */ | |
1922 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
1923 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
1924 | ||
1925 | if (nr_file > nr_anon) { | |
1926 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
1927 | targets[LRU_ACTIVE_ANON] + 1; | |
1928 | lru = LRU_BASE; | |
1929 | percentage = nr_anon * 100 / scan_target; | |
1930 | } else { | |
1931 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
1932 | targets[LRU_ACTIVE_FILE] + 1; | |
1933 | lru = LRU_FILE; | |
1934 | percentage = nr_file * 100 / scan_target; | |
1935 | } | |
1936 | ||
1937 | /* Stop scanning the smaller of the LRU */ | |
1938 | nr[lru] = 0; | |
1939 | nr[lru + LRU_ACTIVE] = 0; | |
1940 | ||
1941 | /* | |
1942 | * Recalculate the other LRU scan count based on its original | |
1943 | * scan target and the percentage scanning already complete | |
1944 | */ | |
1945 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
1946 | nr_scanned = targets[lru] - nr[lru]; | |
1947 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
1948 | nr[lru] -= min(nr[lru], nr_scanned); | |
1949 | ||
1950 | lru += LRU_ACTIVE; | |
1951 | nr_scanned = targets[lru] - nr[lru]; | |
1952 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
1953 | nr[lru] -= min(nr[lru], nr_scanned); | |
1954 | ||
1955 | scan_adjusted = true; | |
1956 | } | |
1957 | blk_finish_plug(&plug); | |
1958 | sc->nr_reclaimed += nr_reclaimed; | |
1959 | ||
1960 | /* | |
1961 | * Even if we did not try to evict anon pages at all, we want to | |
1962 | * rebalance the anon lru active/inactive ratio. | |
1963 | */ | |
1964 | if (inactive_anon_is_low(lruvec)) | |
1965 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
1966 | sc, LRU_ACTIVE_ANON); | |
1967 | ||
1968 | throttle_vm_writeout(sc->gfp_mask); | |
1969 | } | |
1970 | ||
1971 | /* Use reclaim/compaction for costly allocs or under memory pressure */ | |
1972 | static bool in_reclaim_compaction(struct scan_control *sc) | |
1973 | { | |
1974 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && | |
1975 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || | |
1976 | sc->priority < DEF_PRIORITY - 2)) | |
1977 | return true; | |
1978 | ||
1979 | return false; | |
1980 | } | |
1981 | ||
1982 | /* | |
1983 | * Reclaim/compaction is used for high-order allocation requests. It reclaims | |
1984 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
1985 | * true if more pages should be reclaimed such that when the page allocator | |
1986 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
1987 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
1988 | */ | |
1989 | static inline bool should_continue_reclaim(struct zone *zone, | |
1990 | unsigned long nr_reclaimed, | |
1991 | unsigned long nr_scanned, | |
1992 | struct scan_control *sc) | |
1993 | { | |
1994 | unsigned long pages_for_compaction; | |
1995 | unsigned long inactive_lru_pages; | |
1996 | ||
1997 | /* If not in reclaim/compaction mode, stop */ | |
1998 | if (!in_reclaim_compaction(sc)) | |
1999 | return false; | |
2000 | ||
2001 | /* Consider stopping depending on scan and reclaim activity */ | |
2002 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2003 | /* | |
2004 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2005 | * full LRU list has been scanned and we are still failing | |
2006 | * to reclaim pages. This full LRU scan is potentially | |
2007 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2008 | */ | |
2009 | if (!nr_reclaimed && !nr_scanned) | |
2010 | return false; | |
2011 | } else { | |
2012 | /* | |
2013 | * For non-__GFP_REPEAT allocations which can presumably | |
2014 | * fail without consequence, stop if we failed to reclaim | |
2015 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2016 | * pages that were scanned. This will return to the | |
2017 | * caller faster at the risk reclaim/compaction and | |
2018 | * the resulting allocation attempt fails | |
2019 | */ | |
2020 | if (!nr_reclaimed) | |
2021 | return false; | |
2022 | } | |
2023 | ||
2024 | /* | |
2025 | * If we have not reclaimed enough pages for compaction and the | |
2026 | * inactive lists are large enough, continue reclaiming | |
2027 | */ | |
2028 | pages_for_compaction = (2UL << sc->order); | |
2029 | inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); | |
2030 | if (get_nr_swap_pages() > 0) | |
2031 | inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); | |
2032 | if (sc->nr_reclaimed < pages_for_compaction && | |
2033 | inactive_lru_pages > pages_for_compaction) | |
2034 | return true; | |
2035 | ||
2036 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
2037 | switch (compaction_suitable(zone, sc->order)) { | |
2038 | case COMPACT_PARTIAL: | |
2039 | case COMPACT_CONTINUE: | |
2040 | return false; | |
2041 | default: | |
2042 | return true; | |
2043 | } | |
2044 | } | |
2045 | ||
2046 | static void shrink_zone(struct zone *zone, struct scan_control *sc) | |
2047 | { | |
2048 | unsigned long nr_reclaimed, nr_scanned; | |
2049 | ||
2050 | do { | |
2051 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2052 | struct mem_cgroup_reclaim_cookie reclaim = { | |
2053 | .zone = zone, | |
2054 | .priority = sc->priority, | |
2055 | }; | |
2056 | struct mem_cgroup *memcg; | |
2057 | ||
2058 | nr_reclaimed = sc->nr_reclaimed; | |
2059 | nr_scanned = sc->nr_scanned; | |
2060 | ||
2061 | memcg = mem_cgroup_iter(root, NULL, &reclaim); | |
2062 | do { | |
2063 | struct lruvec *lruvec; | |
2064 | ||
2065 | lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
2066 | ||
2067 | shrink_lruvec(lruvec, sc); | |
2068 | ||
2069 | /* | |
2070 | * Direct reclaim and kswapd have to scan all memory | |
2071 | * cgroups to fulfill the overall scan target for the | |
2072 | * zone. | |
2073 | * | |
2074 | * Limit reclaim, on the other hand, only cares about | |
2075 | * nr_to_reclaim pages to be reclaimed and it will | |
2076 | * retry with decreasing priority if one round over the | |
2077 | * whole hierarchy is not sufficient. | |
2078 | */ | |
2079 | if (!global_reclaim(sc) && | |
2080 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
2081 | mem_cgroup_iter_break(root, memcg); | |
2082 | break; | |
2083 | } | |
2084 | memcg = mem_cgroup_iter(root, memcg, &reclaim); | |
2085 | } while (memcg); | |
2086 | ||
2087 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, | |
2088 | sc->nr_scanned - nr_scanned, | |
2089 | sc->nr_reclaimed - nr_reclaimed); | |
2090 | ||
2091 | } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, | |
2092 | sc->nr_scanned - nr_scanned, sc)); | |
2093 | } | |
2094 | ||
2095 | /* Returns true if compaction should go ahead for a high-order request */ | |
2096 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
2097 | { | |
2098 | unsigned long balance_gap, watermark; | |
2099 | bool watermark_ok; | |
2100 | ||
2101 | /* Do not consider compaction for orders reclaim is meant to satisfy */ | |
2102 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
2103 | return false; | |
2104 | ||
2105 | /* | |
2106 | * Compaction takes time to run and there are potentially other | |
2107 | * callers using the pages just freed. Continue reclaiming until | |
2108 | * there is a buffer of free pages available to give compaction | |
2109 | * a reasonable chance of completing and allocating the page | |
2110 | */ | |
2111 | balance_gap = min(low_wmark_pages(zone), | |
2112 | (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2113 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
2114 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order); | |
2115 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); | |
2116 | ||
2117 | /* | |
2118 | * If compaction is deferred, reclaim up to a point where | |
2119 | * compaction will have a chance of success when re-enabled | |
2120 | */ | |
2121 | if (compaction_deferred(zone, sc->order)) | |
2122 | return watermark_ok; | |
2123 | ||
2124 | /* If compaction is not ready to start, keep reclaiming */ | |
2125 | if (!compaction_suitable(zone, sc->order)) | |
2126 | return false; | |
2127 | ||
2128 | return watermark_ok; | |
2129 | } | |
2130 | ||
2131 | /* | |
2132 | * This is the direct reclaim path, for page-allocating processes. We only | |
2133 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2134 | * request. | |
2135 | * | |
2136 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). | |
2137 | * Because: | |
2138 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order | |
2139 | * allocation or | |
2140 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones | |
2141 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2142 | * zone defense algorithm. | |
2143 | * | |
2144 | * If a zone is deemed to be full of pinned pages then just give it a light | |
2145 | * scan then give up on it. | |
2146 | * | |
2147 | * This function returns true if a zone is being reclaimed for a costly | |
2148 | * high-order allocation and compaction is ready to begin. This indicates to | |
2149 | * the caller that it should consider retrying the allocation instead of | |
2150 | * further reclaim. | |
2151 | */ | |
2152 | static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) | |
2153 | { | |
2154 | struct zoneref *z; | |
2155 | struct zone *zone; | |
2156 | unsigned long nr_soft_reclaimed; | |
2157 | unsigned long nr_soft_scanned; | |
2158 | bool aborted_reclaim = false; | |
2159 | ||
2160 | /* | |
2161 | * If the number of buffer_heads in the machine exceeds the maximum | |
2162 | * allowed level, force direct reclaim to scan the highmem zone as | |
2163 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2164 | */ | |
2165 | if (buffer_heads_over_limit) | |
2166 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2167 | ||
2168 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2169 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2170 | if (!populated_zone(zone)) | |
2171 | continue; | |
2172 | /* | |
2173 | * Take care memory controller reclaiming has small influence | |
2174 | * to global LRU. | |
2175 | */ | |
2176 | if (global_reclaim(sc)) { | |
2177 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2178 | continue; | |
2179 | if (zone->all_unreclaimable && | |
2180 | sc->priority != DEF_PRIORITY) | |
2181 | continue; /* Let kswapd poll it */ | |
2182 | if (IS_ENABLED(CONFIG_COMPACTION)) { | |
2183 | /* | |
2184 | * If we already have plenty of memory free for | |
2185 | * compaction in this zone, don't free any more. | |
2186 | * Even though compaction is invoked for any | |
2187 | * non-zero order, only frequent costly order | |
2188 | * reclamation is disruptive enough to become a | |
2189 | * noticeable problem, like transparent huge | |
2190 | * page allocations. | |
2191 | */ | |
2192 | if (compaction_ready(zone, sc)) { | |
2193 | aborted_reclaim = true; | |
2194 | continue; | |
2195 | } | |
2196 | } | |
2197 | /* | |
2198 | * This steals pages from memory cgroups over softlimit | |
2199 | * and returns the number of reclaimed pages and | |
2200 | * scanned pages. This works for global memory pressure | |
2201 | * and balancing, not for a memcg's limit. | |
2202 | */ | |
2203 | nr_soft_scanned = 0; | |
2204 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2205 | sc->order, sc->gfp_mask, | |
2206 | &nr_soft_scanned); | |
2207 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2208 | sc->nr_scanned += nr_soft_scanned; | |
2209 | /* need some check for avoid more shrink_zone() */ | |
2210 | } | |
2211 | ||
2212 | shrink_zone(zone, sc); | |
2213 | } | |
2214 | ||
2215 | return aborted_reclaim; | |
2216 | } | |
2217 | ||
2218 | static bool zone_reclaimable(struct zone *zone) | |
2219 | { | |
2220 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
2221 | } | |
2222 | ||
2223 | /* All zones in zonelist are unreclaimable? */ | |
2224 | static bool all_unreclaimable(struct zonelist *zonelist, | |
2225 | struct scan_control *sc) | |
2226 | { | |
2227 | struct zoneref *z; | |
2228 | struct zone *zone; | |
2229 | ||
2230 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2231 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2232 | if (!populated_zone(zone)) | |
2233 | continue; | |
2234 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2235 | continue; | |
2236 | if (!zone->all_unreclaimable) | |
2237 | return false; | |
2238 | } | |
2239 | ||
2240 | return true; | |
2241 | } | |
2242 | ||
2243 | /* | |
2244 | * This is the main entry point to direct page reclaim. | |
2245 | * | |
2246 | * If a full scan of the inactive list fails to free enough memory then we | |
2247 | * are "out of memory" and something needs to be killed. | |
2248 | * | |
2249 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2250 | * high - the zone may be full of dirty or under-writeback pages, which this | |
2251 | * caller can't do much about. We kick the writeback threads and take explicit | |
2252 | * naps in the hope that some of these pages can be written. But if the | |
2253 | * allocating task holds filesystem locks which prevent writeout this might not | |
2254 | * work, and the allocation attempt will fail. | |
2255 | * | |
2256 | * returns: 0, if no pages reclaimed | |
2257 | * else, the number of pages reclaimed | |
2258 | */ | |
2259 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, | |
2260 | struct scan_control *sc, | |
2261 | struct shrink_control *shrink) | |
2262 | { | |
2263 | unsigned long total_scanned = 0; | |
2264 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
2265 | struct zoneref *z; | |
2266 | struct zone *zone; | |
2267 | unsigned long writeback_threshold; | |
2268 | bool aborted_reclaim; | |
2269 | ||
2270 | delayacct_freepages_start(); | |
2271 | ||
2272 | if (global_reclaim(sc)) | |
2273 | count_vm_event(ALLOCSTALL); | |
2274 | ||
2275 | do { | |
2276 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, | |
2277 | sc->priority); | |
2278 | sc->nr_scanned = 0; | |
2279 | aborted_reclaim = shrink_zones(zonelist, sc); | |
2280 | ||
2281 | /* | |
2282 | * Don't shrink slabs when reclaiming memory from | |
2283 | * over limit cgroups | |
2284 | */ | |
2285 | if (global_reclaim(sc)) { | |
2286 | unsigned long lru_pages = 0; | |
2287 | for_each_zone_zonelist(zone, z, zonelist, | |
2288 | gfp_zone(sc->gfp_mask)) { | |
2289 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2290 | continue; | |
2291 | ||
2292 | lru_pages += zone_reclaimable_pages(zone); | |
2293 | } | |
2294 | ||
2295 | shrink_slab(shrink, sc->nr_scanned, lru_pages); | |
2296 | if (reclaim_state) { | |
2297 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2298 | reclaim_state->reclaimed_slab = 0; | |
2299 | } | |
2300 | } | |
2301 | total_scanned += sc->nr_scanned; | |
2302 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) | |
2303 | goto out; | |
2304 | ||
2305 | /* | |
2306 | * If we're getting trouble reclaiming, start doing | |
2307 | * writepage even in laptop mode. | |
2308 | */ | |
2309 | if (sc->priority < DEF_PRIORITY - 2) | |
2310 | sc->may_writepage = 1; | |
2311 | ||
2312 | /* | |
2313 | * Try to write back as many pages as we just scanned. This | |
2314 | * tends to cause slow streaming writers to write data to the | |
2315 | * disk smoothly, at the dirtying rate, which is nice. But | |
2316 | * that's undesirable in laptop mode, where we *want* lumpy | |
2317 | * writeout. So in laptop mode, write out the whole world. | |
2318 | */ | |
2319 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; | |
2320 | if (total_scanned > writeback_threshold) { | |
2321 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, | |
2322 | WB_REASON_TRY_TO_FREE_PAGES); | |
2323 | sc->may_writepage = 1; | |
2324 | } | |
2325 | ||
2326 | /* Take a nap, wait for some writeback to complete */ | |
2327 | if (!sc->hibernation_mode && sc->nr_scanned && | |
2328 | sc->priority < DEF_PRIORITY - 2) { | |
2329 | struct zone *preferred_zone; | |
2330 | ||
2331 | first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), | |
2332 | &cpuset_current_mems_allowed, | |
2333 | &preferred_zone); | |
2334 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); | |
2335 | } | |
2336 | } while (--sc->priority >= 0); | |
2337 | ||
2338 | out: | |
2339 | delayacct_freepages_end(); | |
2340 | ||
2341 | if (sc->nr_reclaimed) | |
2342 | return sc->nr_reclaimed; | |
2343 | ||
2344 | /* | |
2345 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2346 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2347 | * check. | |
2348 | */ | |
2349 | if (oom_killer_disabled) | |
2350 | return 0; | |
2351 | ||
2352 | /* Aborted reclaim to try compaction? don't OOM, then */ | |
2353 | if (aborted_reclaim) | |
2354 | return 1; | |
2355 | ||
2356 | /* top priority shrink_zones still had more to do? don't OOM, then */ | |
2357 | if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc)) | |
2358 | return 1; | |
2359 | ||
2360 | return 0; | |
2361 | } | |
2362 | ||
2363 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) | |
2364 | { | |
2365 | struct zone *zone; | |
2366 | unsigned long pfmemalloc_reserve = 0; | |
2367 | unsigned long free_pages = 0; | |
2368 | int i; | |
2369 | bool wmark_ok; | |
2370 | ||
2371 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2372 | zone = &pgdat->node_zones[i]; | |
2373 | pfmemalloc_reserve += min_wmark_pages(zone); | |
2374 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2375 | } | |
2376 | ||
2377 | wmark_ok = free_pages > pfmemalloc_reserve / 2; | |
2378 | ||
2379 | /* kswapd must be awake if processes are being throttled */ | |
2380 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2381 | pgdat->classzone_idx = min(pgdat->classzone_idx, | |
2382 | (enum zone_type)ZONE_NORMAL); | |
2383 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2384 | } | |
2385 | ||
2386 | return wmark_ok; | |
2387 | } | |
2388 | ||
2389 | /* | |
2390 | * Throttle direct reclaimers if backing storage is backed by the network | |
2391 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2392 | * depleted. kswapd will continue to make progress and wake the processes | |
2393 | * when the low watermark is reached. | |
2394 | * | |
2395 | * Returns true if a fatal signal was delivered during throttling. If this | |
2396 | * happens, the page allocator should not consider triggering the OOM killer. | |
2397 | */ | |
2398 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, | |
2399 | nodemask_t *nodemask) | |
2400 | { | |
2401 | struct zone *zone; | |
2402 | int high_zoneidx = gfp_zone(gfp_mask); | |
2403 | pg_data_t *pgdat; | |
2404 | ||
2405 | /* | |
2406 | * Kernel threads should not be throttled as they may be indirectly | |
2407 | * responsible for cleaning pages necessary for reclaim to make forward | |
2408 | * progress. kjournald for example may enter direct reclaim while | |
2409 | * committing a transaction where throttling it could forcing other | |
2410 | * processes to block on log_wait_commit(). | |
2411 | */ | |
2412 | if (current->flags & PF_KTHREAD) | |
2413 | goto out; | |
2414 | ||
2415 | /* | |
2416 | * If a fatal signal is pending, this process should not throttle. | |
2417 | * It should return quickly so it can exit and free its memory | |
2418 | */ | |
2419 | if (fatal_signal_pending(current)) | |
2420 | goto out; | |
2421 | ||
2422 | /* Check if the pfmemalloc reserves are ok */ | |
2423 | first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone); | |
2424 | pgdat = zone->zone_pgdat; | |
2425 | if (pfmemalloc_watermark_ok(pgdat)) | |
2426 | goto out; | |
2427 | ||
2428 | /* Account for the throttling */ | |
2429 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2430 | ||
2431 | /* | |
2432 | * If the caller cannot enter the filesystem, it's possible that it | |
2433 | * is due to the caller holding an FS lock or performing a journal | |
2434 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2435 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2436 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2437 | * second before continuing. | |
2438 | */ | |
2439 | if (!(gfp_mask & __GFP_FS)) { | |
2440 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2441 | pfmemalloc_watermark_ok(pgdat), HZ); | |
2442 | ||
2443 | goto check_pending; | |
2444 | } | |
2445 | ||
2446 | /* Throttle until kswapd wakes the process */ | |
2447 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2448 | pfmemalloc_watermark_ok(pgdat)); | |
2449 | ||
2450 | check_pending: | |
2451 | if (fatal_signal_pending(current)) | |
2452 | return true; | |
2453 | ||
2454 | out: | |
2455 | return false; | |
2456 | } | |
2457 | ||
2458 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, | |
2459 | gfp_t gfp_mask, nodemask_t *nodemask) | |
2460 | { | |
2461 | unsigned long nr_reclaimed; | |
2462 | struct scan_control sc = { | |
2463 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), | |
2464 | .may_writepage = !laptop_mode, | |
2465 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2466 | .may_unmap = 1, | |
2467 | .may_swap = 1, | |
2468 | .order = order, | |
2469 | .priority = DEF_PRIORITY, | |
2470 | .target_mem_cgroup = NULL, | |
2471 | .nodemask = nodemask, | |
2472 | }; | |
2473 | struct shrink_control shrink = { | |
2474 | .gfp_mask = sc.gfp_mask, | |
2475 | }; | |
2476 | ||
2477 | /* | |
2478 | * Do not enter reclaim if fatal signal was delivered while throttled. | |
2479 | * 1 is returned so that the page allocator does not OOM kill at this | |
2480 | * point. | |
2481 | */ | |
2482 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) | |
2483 | return 1; | |
2484 | ||
2485 | trace_mm_vmscan_direct_reclaim_begin(order, | |
2486 | sc.may_writepage, | |
2487 | gfp_mask); | |
2488 | ||
2489 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); | |
2490 | ||
2491 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2492 | ||
2493 | return nr_reclaimed; | |
2494 | } | |
2495 | ||
2496 | #ifdef CONFIG_MEMCG | |
2497 | ||
2498 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, | |
2499 | gfp_t gfp_mask, bool noswap, | |
2500 | struct zone *zone, | |
2501 | unsigned long *nr_scanned) | |
2502 | { | |
2503 | struct scan_control sc = { | |
2504 | .nr_scanned = 0, | |
2505 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2506 | .may_writepage = !laptop_mode, | |
2507 | .may_unmap = 1, | |
2508 | .may_swap = !noswap, | |
2509 | .order = 0, | |
2510 | .priority = 0, | |
2511 | .target_mem_cgroup = memcg, | |
2512 | }; | |
2513 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
2514 | ||
2515 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
2516 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
2517 | ||
2518 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, | |
2519 | sc.may_writepage, | |
2520 | sc.gfp_mask); | |
2521 | ||
2522 | /* | |
2523 | * NOTE: Although we can get the priority field, using it | |
2524 | * here is not a good idea, since it limits the pages we can scan. | |
2525 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2526 | * will pick up pages from other mem cgroup's as well. We hack | |
2527 | * the priority and make it zero. | |
2528 | */ | |
2529 | shrink_lruvec(lruvec, &sc); | |
2530 | ||
2531 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2532 | ||
2533 | *nr_scanned = sc.nr_scanned; | |
2534 | return sc.nr_reclaimed; | |
2535 | } | |
2536 | ||
2537 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, | |
2538 | gfp_t gfp_mask, | |
2539 | bool noswap) | |
2540 | { | |
2541 | struct zonelist *zonelist; | |
2542 | unsigned long nr_reclaimed; | |
2543 | int nid; | |
2544 | struct scan_control sc = { | |
2545 | .may_writepage = !laptop_mode, | |
2546 | .may_unmap = 1, | |
2547 | .may_swap = !noswap, | |
2548 | .nr_to_reclaim = SWAP_CLUSTER_MAX, | |
2549 | .order = 0, | |
2550 | .priority = DEF_PRIORITY, | |
2551 | .target_mem_cgroup = memcg, | |
2552 | .nodemask = NULL, /* we don't care the placement */ | |
2553 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
2554 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2555 | }; | |
2556 | struct shrink_control shrink = { | |
2557 | .gfp_mask = sc.gfp_mask, | |
2558 | }; | |
2559 | ||
2560 | /* | |
2561 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2562 | * take care of from where we get pages. So the node where we start the | |
2563 | * scan does not need to be the current node. | |
2564 | */ | |
2565 | nid = mem_cgroup_select_victim_node(memcg); | |
2566 | ||
2567 | zonelist = NODE_DATA(nid)->node_zonelists; | |
2568 | ||
2569 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2570 | sc.may_writepage, | |
2571 | sc.gfp_mask); | |
2572 | ||
2573 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); | |
2574 | ||
2575 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2576 | ||
2577 | return nr_reclaimed; | |
2578 | } | |
2579 | #endif | |
2580 | ||
2581 | static void age_active_anon(struct zone *zone, struct scan_control *sc) | |
2582 | { | |
2583 | struct mem_cgroup *memcg; | |
2584 | ||
2585 | if (!total_swap_pages) | |
2586 | return; | |
2587 | ||
2588 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2589 | do { | |
2590 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
2591 | ||
2592 | if (inactive_anon_is_low(lruvec)) | |
2593 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
2594 | sc, LRU_ACTIVE_ANON); | |
2595 | ||
2596 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2597 | } while (memcg); | |
2598 | } | |
2599 | ||
2600 | static bool zone_balanced(struct zone *zone, int order, | |
2601 | unsigned long balance_gap, int classzone_idx) | |
2602 | { | |
2603 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + | |
2604 | balance_gap, classzone_idx, 0)) | |
2605 | return false; | |
2606 | ||
2607 | if (IS_ENABLED(CONFIG_COMPACTION) && order && | |
2608 | !compaction_suitable(zone, order)) | |
2609 | return false; | |
2610 | ||
2611 | return true; | |
2612 | } | |
2613 | ||
2614 | /* | |
2615 | * pgdat_balanced() is used when checking if a node is balanced. | |
2616 | * | |
2617 | * For order-0, all zones must be balanced! | |
2618 | * | |
2619 | * For high-order allocations only zones that meet watermarks and are in a | |
2620 | * zone allowed by the callers classzone_idx are added to balanced_pages. The | |
2621 | * total of balanced pages must be at least 25% of the zones allowed by | |
2622 | * classzone_idx for the node to be considered balanced. Forcing all zones to | |
2623 | * be balanced for high orders can cause excessive reclaim when there are | |
2624 | * imbalanced zones. | |
2625 | * The choice of 25% is due to | |
2626 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2627 | * reasonable sized machine | |
2628 | * o On all other machines, the top zone must be at least a reasonable | |
2629 | * percentage of the middle zones. For example, on 32-bit x86, highmem | |
2630 | * would need to be at least 256M for it to be balance a whole node. | |
2631 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2632 | * to balance a node on its own. These seemed like reasonable ratios. | |
2633 | */ | |
2634 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
2635 | { | |
2636 | unsigned long managed_pages = 0; | |
2637 | unsigned long balanced_pages = 0; | |
2638 | int i; | |
2639 | ||
2640 | /* Check the watermark levels */ | |
2641 | for (i = 0; i <= classzone_idx; i++) { | |
2642 | struct zone *zone = pgdat->node_zones + i; | |
2643 | ||
2644 | if (!populated_zone(zone)) | |
2645 | continue; | |
2646 | ||
2647 | managed_pages += zone->managed_pages; | |
2648 | ||
2649 | /* | |
2650 | * A special case here: | |
2651 | * | |
2652 | * balance_pgdat() skips over all_unreclaimable after | |
2653 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2654 | * they must be considered balanced here as well! | |
2655 | */ | |
2656 | if (zone->all_unreclaimable) { | |
2657 | balanced_pages += zone->managed_pages; | |
2658 | continue; | |
2659 | } | |
2660 | ||
2661 | if (zone_balanced(zone, order, 0, i)) | |
2662 | balanced_pages += zone->managed_pages; | |
2663 | else if (!order) | |
2664 | return false; | |
2665 | } | |
2666 | ||
2667 | if (order) | |
2668 | return balanced_pages >= (managed_pages >> 2); | |
2669 | else | |
2670 | return true; | |
2671 | } | |
2672 | ||
2673 | /* | |
2674 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
2675 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
2676 | * | |
2677 | * Returns true if kswapd is ready to sleep | |
2678 | */ | |
2679 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, | |
2680 | int classzone_idx) | |
2681 | { | |
2682 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ | |
2683 | if (remaining) | |
2684 | return false; | |
2685 | ||
2686 | /* | |
2687 | * There is a potential race between when kswapd checks its watermarks | |
2688 | * and a process gets throttled. There is also a potential race if | |
2689 | * processes get throttled, kswapd wakes, a large process exits therby | |
2690 | * balancing the zones that causes kswapd to miss a wakeup. If kswapd | |
2691 | * is going to sleep, no process should be sleeping on pfmemalloc_wait | |
2692 | * so wake them now if necessary. If necessary, processes will wake | |
2693 | * kswapd and get throttled again | |
2694 | */ | |
2695 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) { | |
2696 | wake_up(&pgdat->pfmemalloc_wait); | |
2697 | return false; | |
2698 | } | |
2699 | ||
2700 | return pgdat_balanced(pgdat, order, classzone_idx); | |
2701 | } | |
2702 | ||
2703 | /* | |
2704 | * kswapd shrinks the zone by the number of pages required to reach | |
2705 | * the high watermark. | |
2706 | * | |
2707 | * Returns true if kswapd scanned at least the requested number of pages to | |
2708 | * reclaim or if the lack of progress was due to pages under writeback. | |
2709 | * This is used to determine if the scanning priority needs to be raised. | |
2710 | */ | |
2711 | static bool kswapd_shrink_zone(struct zone *zone, | |
2712 | struct scan_control *sc, | |
2713 | unsigned long lru_pages, | |
2714 | unsigned long *nr_attempted) | |
2715 | { | |
2716 | unsigned long nr_slab; | |
2717 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
2718 | struct shrink_control shrink = { | |
2719 | .gfp_mask = sc->gfp_mask, | |
2720 | }; | |
2721 | ||
2722 | /* Reclaim above the high watermark. */ | |
2723 | sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone)); | |
2724 | shrink_zone(zone, sc); | |
2725 | ||
2726 | reclaim_state->reclaimed_slab = 0; | |
2727 | nr_slab = shrink_slab(&shrink, sc->nr_scanned, lru_pages); | |
2728 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2729 | ||
2730 | /* Account for the number of pages attempted to reclaim */ | |
2731 | *nr_attempted += sc->nr_to_reclaim; | |
2732 | ||
2733 | if (nr_slab == 0 && !zone_reclaimable(zone)) | |
2734 | zone->all_unreclaimable = 1; | |
2735 | ||
2736 | zone_clear_flag(zone, ZONE_WRITEBACK); | |
2737 | ||
2738 | return sc->nr_scanned >= sc->nr_to_reclaim; | |
2739 | } | |
2740 | ||
2741 | /* | |
2742 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
2743 | * they are all at high_wmark_pages(zone). | |
2744 | * | |
2745 | * Returns the final order kswapd was reclaiming at | |
2746 | * | |
2747 | * There is special handling here for zones which are full of pinned pages. | |
2748 | * This can happen if the pages are all mlocked, or if they are all used by | |
2749 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
2750 | * What we do is to detect the case where all pages in the zone have been | |
2751 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
2752 | * dead and from now on, only perform a short scan. Basically we're polling | |
2753 | * the zone for when the problem goes away. | |
2754 | * | |
2755 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
2756 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is | |
2757 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
2758 | * lower zones regardless of the number of free pages in the lower zones. This | |
2759 | * interoperates with the page allocator fallback scheme to ensure that aging | |
2760 | * of pages is balanced across the zones. | |
2761 | */ | |
2762 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, | |
2763 | int *classzone_idx) | |
2764 | { | |
2765 | int i; | |
2766 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ | |
2767 | unsigned long nr_soft_reclaimed; | |
2768 | unsigned long nr_soft_scanned; | |
2769 | struct scan_control sc = { | |
2770 | .gfp_mask = GFP_KERNEL, | |
2771 | .priority = DEF_PRIORITY, | |
2772 | .may_unmap = 1, | |
2773 | .may_swap = 1, | |
2774 | .may_writepage = !laptop_mode, | |
2775 | .order = order, | |
2776 | .target_mem_cgroup = NULL, | |
2777 | }; | |
2778 | count_vm_event(PAGEOUTRUN); | |
2779 | ||
2780 | do { | |
2781 | unsigned long lru_pages = 0; | |
2782 | unsigned long nr_attempted = 0; | |
2783 | bool raise_priority = true; | |
2784 | bool pgdat_needs_compaction = (order > 0); | |
2785 | ||
2786 | sc.nr_reclaimed = 0; | |
2787 | ||
2788 | /* | |
2789 | * Scan in the highmem->dma direction for the highest | |
2790 | * zone which needs scanning | |
2791 | */ | |
2792 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
2793 | struct zone *zone = pgdat->node_zones + i; | |
2794 | ||
2795 | if (!populated_zone(zone)) | |
2796 | continue; | |
2797 | ||
2798 | if (zone->all_unreclaimable && | |
2799 | sc.priority != DEF_PRIORITY) | |
2800 | continue; | |
2801 | ||
2802 | /* | |
2803 | * Do some background aging of the anon list, to give | |
2804 | * pages a chance to be referenced before reclaiming. | |
2805 | */ | |
2806 | age_active_anon(zone, &sc); | |
2807 | ||
2808 | /* | |
2809 | * If the number of buffer_heads in the machine | |
2810 | * exceeds the maximum allowed level and this node | |
2811 | * has a highmem zone, force kswapd to reclaim from | |
2812 | * it to relieve lowmem pressure. | |
2813 | */ | |
2814 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
2815 | end_zone = i; | |
2816 | break; | |
2817 | } | |
2818 | ||
2819 | if (!zone_balanced(zone, order, 0, 0)) { | |
2820 | end_zone = i; | |
2821 | break; | |
2822 | } else { | |
2823 | /* | |
2824 | * If balanced, clear the dirty and congested | |
2825 | * flags | |
2826 | */ | |
2827 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2828 | zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY); | |
2829 | } | |
2830 | } | |
2831 | ||
2832 | if (i < 0) | |
2833 | goto out; | |
2834 | ||
2835 | for (i = 0; i <= end_zone; i++) { | |
2836 | struct zone *zone = pgdat->node_zones + i; | |
2837 | ||
2838 | if (!populated_zone(zone)) | |
2839 | continue; | |
2840 | ||
2841 | lru_pages += zone_reclaimable_pages(zone); | |
2842 | ||
2843 | /* | |
2844 | * If any zone is currently balanced then kswapd will | |
2845 | * not call compaction as it is expected that the | |
2846 | * necessary pages are already available. | |
2847 | */ | |
2848 | if (pgdat_needs_compaction && | |
2849 | zone_watermark_ok(zone, order, | |
2850 | low_wmark_pages(zone), | |
2851 | *classzone_idx, 0)) | |
2852 | pgdat_needs_compaction = false; | |
2853 | } | |
2854 | ||
2855 | /* | |
2856 | * If we're getting trouble reclaiming, start doing writepage | |
2857 | * even in laptop mode. | |
2858 | */ | |
2859 | if (sc.priority < DEF_PRIORITY - 2) | |
2860 | sc.may_writepage = 1; | |
2861 | ||
2862 | /* | |
2863 | * Now scan the zone in the dma->highmem direction, stopping | |
2864 | * at the last zone which needs scanning. | |
2865 | * | |
2866 | * We do this because the page allocator works in the opposite | |
2867 | * direction. This prevents the page allocator from allocating | |
2868 | * pages behind kswapd's direction of progress, which would | |
2869 | * cause too much scanning of the lower zones. | |
2870 | */ | |
2871 | for (i = 0; i <= end_zone; i++) { | |
2872 | struct zone *zone = pgdat->node_zones + i; | |
2873 | int testorder; | |
2874 | unsigned long balance_gap; | |
2875 | ||
2876 | if (!populated_zone(zone)) | |
2877 | continue; | |
2878 | ||
2879 | if (zone->all_unreclaimable && | |
2880 | sc.priority != DEF_PRIORITY) | |
2881 | continue; | |
2882 | ||
2883 | sc.nr_scanned = 0; | |
2884 | ||
2885 | nr_soft_scanned = 0; | |
2886 | /* | |
2887 | * Call soft limit reclaim before calling shrink_zone. | |
2888 | */ | |
2889 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2890 | order, sc.gfp_mask, | |
2891 | &nr_soft_scanned); | |
2892 | sc.nr_reclaimed += nr_soft_reclaimed; | |
2893 | ||
2894 | /* | |
2895 | * We put equal pressure on every zone, unless | |
2896 | * one zone has way too many pages free | |
2897 | * already. The "too many pages" is defined | |
2898 | * as the high wmark plus a "gap" where the | |
2899 | * gap is either the low watermark or 1% | |
2900 | * of the zone, whichever is smaller. | |
2901 | */ | |
2902 | balance_gap = min(low_wmark_pages(zone), | |
2903 | (zone->managed_pages + | |
2904 | KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2905 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
2906 | /* | |
2907 | * Kswapd reclaims only single pages with compaction | |
2908 | * enabled. Trying too hard to reclaim until contiguous | |
2909 | * free pages have become available can hurt performance | |
2910 | * by evicting too much useful data from memory. | |
2911 | * Do not reclaim more than needed for compaction. | |
2912 | */ | |
2913 | testorder = order; | |
2914 | if (IS_ENABLED(CONFIG_COMPACTION) && order && | |
2915 | compaction_suitable(zone, order) != | |
2916 | COMPACT_SKIPPED) | |
2917 | testorder = 0; | |
2918 | ||
2919 | if ((buffer_heads_over_limit && is_highmem_idx(i)) || | |
2920 | !zone_balanced(zone, testorder, | |
2921 | balance_gap, end_zone)) { | |
2922 | /* | |
2923 | * There should be no need to raise the | |
2924 | * scanning priority if enough pages are | |
2925 | * already being scanned that high | |
2926 | * watermark would be met at 100% efficiency. | |
2927 | */ | |
2928 | if (kswapd_shrink_zone(zone, &sc, lru_pages, | |
2929 | &nr_attempted)) | |
2930 | raise_priority = false; | |
2931 | } | |
2932 | ||
2933 | if (zone->all_unreclaimable) { | |
2934 | if (end_zone && end_zone == i) | |
2935 | end_zone--; | |
2936 | continue; | |
2937 | } | |
2938 | ||
2939 | if (zone_balanced(zone, testorder, 0, end_zone)) | |
2940 | /* | |
2941 | * If a zone reaches its high watermark, | |
2942 | * consider it to be no longer congested. It's | |
2943 | * possible there are dirty pages backed by | |
2944 | * congested BDIs but as pressure is relieved, | |
2945 | * speculatively avoid congestion waits | |
2946 | * or writing pages from kswapd context. | |
2947 | */ | |
2948 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2949 | zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY); | |
2950 | } | |
2951 | ||
2952 | /* | |
2953 | * If the low watermark is met there is no need for processes | |
2954 | * to be throttled on pfmemalloc_wait as they should not be | |
2955 | * able to safely make forward progress. Wake them | |
2956 | */ | |
2957 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
2958 | pfmemalloc_watermark_ok(pgdat)) | |
2959 | wake_up(&pgdat->pfmemalloc_wait); | |
2960 | ||
2961 | /* | |
2962 | * Fragmentation may mean that the system cannot be rebalanced | |
2963 | * for high-order allocations in all zones. If twice the | |
2964 | * allocation size has been reclaimed and the zones are still | |
2965 | * not balanced then recheck the watermarks at order-0 to | |
2966 | * prevent kswapd reclaiming excessively. Assume that a | |
2967 | * process requested a high-order can direct reclaim/compact. | |
2968 | */ | |
2969 | if (order && sc.nr_reclaimed >= 2UL << order) | |
2970 | order = sc.order = 0; | |
2971 | ||
2972 | /* Check if kswapd should be suspending */ | |
2973 | if (try_to_freeze() || kthread_should_stop()) | |
2974 | break; | |
2975 | ||
2976 | /* | |
2977 | * Compact if necessary and kswapd is reclaiming at least the | |
2978 | * high watermark number of pages as requsted | |
2979 | */ | |
2980 | if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted) | |
2981 | compact_pgdat(pgdat, order); | |
2982 | ||
2983 | /* | |
2984 | * Raise priority if scanning rate is too low or there was no | |
2985 | * progress in reclaiming pages | |
2986 | */ | |
2987 | if (raise_priority || !sc.nr_reclaimed) | |
2988 | sc.priority--; | |
2989 | } while (sc.priority >= 1 && | |
2990 | !pgdat_balanced(pgdat, order, *classzone_idx)); | |
2991 | ||
2992 | out: | |
2993 | /* | |
2994 | * Return the order we were reclaiming at so prepare_kswapd_sleep() | |
2995 | * makes a decision on the order we were last reclaiming at. However, | |
2996 | * if another caller entered the allocator slow path while kswapd | |
2997 | * was awake, order will remain at the higher level | |
2998 | */ | |
2999 | *classzone_idx = end_zone; | |
3000 | return order; | |
3001 | } | |
3002 | ||
3003 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) | |
3004 | { | |
3005 | long remaining = 0; | |
3006 | DEFINE_WAIT(wait); | |
3007 | ||
3008 | if (freezing(current) || kthread_should_stop()) | |
3009 | return; | |
3010 | ||
3011 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3012 | ||
3013 | /* Try to sleep for a short interval */ | |
3014 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { | |
3015 | remaining = schedule_timeout(HZ/10); | |
3016 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3017 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3018 | } | |
3019 | ||
3020 | /* | |
3021 | * After a short sleep, check if it was a premature sleep. If not, then | |
3022 | * go fully to sleep until explicitly woken up. | |
3023 | */ | |
3024 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { | |
3025 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); | |
3026 | ||
3027 | /* | |
3028 | * vmstat counters are not perfectly accurate and the estimated | |
3029 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3030 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3031 | * watermarks being breached while under pressure, we reduce the | |
3032 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3033 | * them before going back to sleep. | |
3034 | */ | |
3035 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
3036 | ||
3037 | /* | |
3038 | * Compaction records what page blocks it recently failed to | |
3039 | * isolate pages from and skips them in the future scanning. | |
3040 | * When kswapd is going to sleep, it is reasonable to assume | |
3041 | * that pages and compaction may succeed so reset the cache. | |
3042 | */ | |
3043 | reset_isolation_suitable(pgdat); | |
3044 | ||
3045 | if (!kthread_should_stop()) | |
3046 | schedule(); | |
3047 | ||
3048 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); | |
3049 | } else { | |
3050 | if (remaining) | |
3051 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3052 | else | |
3053 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3054 | } | |
3055 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3056 | } | |
3057 | ||
3058 | /* | |
3059 | * The background pageout daemon, started as a kernel thread | |
3060 | * from the init process. | |
3061 | * | |
3062 | * This basically trickles out pages so that we have _some_ | |
3063 | * free memory available even if there is no other activity | |
3064 | * that frees anything up. This is needed for things like routing | |
3065 | * etc, where we otherwise might have all activity going on in | |
3066 | * asynchronous contexts that cannot page things out. | |
3067 | * | |
3068 | * If there are applications that are active memory-allocators | |
3069 | * (most normal use), this basically shouldn't matter. | |
3070 | */ | |
3071 | static int kswapd(void *p) | |
3072 | { | |
3073 | unsigned long order, new_order; | |
3074 | unsigned balanced_order; | |
3075 | int classzone_idx, new_classzone_idx; | |
3076 | int balanced_classzone_idx; | |
3077 | pg_data_t *pgdat = (pg_data_t*)p; | |
3078 | struct task_struct *tsk = current; | |
3079 | ||
3080 | struct reclaim_state reclaim_state = { | |
3081 | .reclaimed_slab = 0, | |
3082 | }; | |
3083 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); | |
3084 | ||
3085 | lockdep_set_current_reclaim_state(GFP_KERNEL); | |
3086 | ||
3087 | if (!cpumask_empty(cpumask)) | |
3088 | set_cpus_allowed_ptr(tsk, cpumask); | |
3089 | current->reclaim_state = &reclaim_state; | |
3090 | ||
3091 | /* | |
3092 | * Tell the memory management that we're a "memory allocator", | |
3093 | * and that if we need more memory we should get access to it | |
3094 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3095 | * never get caught in the normal page freeing logic. | |
3096 | * | |
3097 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3098 | * you need a small amount of memory in order to be able to | |
3099 | * page out something else, and this flag essentially protects | |
3100 | * us from recursively trying to free more memory as we're | |
3101 | * trying to free the first piece of memory in the first place). | |
3102 | */ | |
3103 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; | |
3104 | set_freezable(); | |
3105 | ||
3106 | order = new_order = 0; | |
3107 | balanced_order = 0; | |
3108 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; | |
3109 | balanced_classzone_idx = classzone_idx; | |
3110 | for ( ; ; ) { | |
3111 | bool ret; | |
3112 | ||
3113 | /* | |
3114 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
3115 | * new request of a similar or harder type will succeed soon | |
3116 | * so consider going to sleep on the basis we reclaimed at | |
3117 | */ | |
3118 | if (balanced_classzone_idx >= new_classzone_idx && | |
3119 | balanced_order == new_order) { | |
3120 | new_order = pgdat->kswapd_max_order; | |
3121 | new_classzone_idx = pgdat->classzone_idx; | |
3122 | pgdat->kswapd_max_order = 0; | |
3123 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3124 | } | |
3125 | ||
3126 | if (order < new_order || classzone_idx > new_classzone_idx) { | |
3127 | /* | |
3128 | * Don't sleep if someone wants a larger 'order' | |
3129 | * allocation or has tigher zone constraints | |
3130 | */ | |
3131 | order = new_order; | |
3132 | classzone_idx = new_classzone_idx; | |
3133 | } else { | |
3134 | kswapd_try_to_sleep(pgdat, balanced_order, | |
3135 | balanced_classzone_idx); | |
3136 | order = pgdat->kswapd_max_order; | |
3137 | classzone_idx = pgdat->classzone_idx; | |
3138 | new_order = order; | |
3139 | new_classzone_idx = classzone_idx; | |
3140 | pgdat->kswapd_max_order = 0; | |
3141 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3142 | } | |
3143 | ||
3144 | ret = try_to_freeze(); | |
3145 | if (kthread_should_stop()) | |
3146 | break; | |
3147 | ||
3148 | /* | |
3149 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3150 | * after returning from the refrigerator | |
3151 | */ | |
3152 | if (!ret) { | |
3153 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
3154 | balanced_classzone_idx = classzone_idx; | |
3155 | balanced_order = balance_pgdat(pgdat, order, | |
3156 | &balanced_classzone_idx); | |
3157 | } | |
3158 | } | |
3159 | ||
3160 | current->reclaim_state = NULL; | |
3161 | return 0; | |
3162 | } | |
3163 | ||
3164 | /* | |
3165 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3166 | */ | |
3167 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) | |
3168 | { | |
3169 | pg_data_t *pgdat; | |
3170 | ||
3171 | if (!populated_zone(zone)) | |
3172 | return; | |
3173 | ||
3174 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
3175 | return; | |
3176 | pgdat = zone->zone_pgdat; | |
3177 | if (pgdat->kswapd_max_order < order) { | |
3178 | pgdat->kswapd_max_order = order; | |
3179 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); | |
3180 | } | |
3181 | if (!waitqueue_active(&pgdat->kswapd_wait)) | |
3182 | return; | |
3183 | if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) | |
3184 | return; | |
3185 | ||
3186 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
3187 | wake_up_interruptible(&pgdat->kswapd_wait); | |
3188 | } | |
3189 | ||
3190 | /* | |
3191 | * The reclaimable count would be mostly accurate. | |
3192 | * The less reclaimable pages may be | |
3193 | * - mlocked pages, which will be moved to unevictable list when encountered | |
3194 | * - mapped pages, which may require several travels to be reclaimed | |
3195 | * - dirty pages, which is not "instantly" reclaimable | |
3196 | */ | |
3197 | unsigned long global_reclaimable_pages(void) | |
3198 | { | |
3199 | int nr; | |
3200 | ||
3201 | nr = global_page_state(NR_ACTIVE_FILE) + | |
3202 | global_page_state(NR_INACTIVE_FILE); | |
3203 | ||
3204 | if (get_nr_swap_pages() > 0) | |
3205 | nr += global_page_state(NR_ACTIVE_ANON) + | |
3206 | global_page_state(NR_INACTIVE_ANON); | |
3207 | ||
3208 | return nr; | |
3209 | } | |
3210 | ||
3211 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
3212 | { | |
3213 | int nr; | |
3214 | ||
3215 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
3216 | zone_page_state(zone, NR_INACTIVE_FILE); | |
3217 | ||
3218 | if (get_nr_swap_pages() > 0) | |
3219 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
3220 | zone_page_state(zone, NR_INACTIVE_ANON); | |
3221 | ||
3222 | return nr; | |
3223 | } | |
3224 | ||
3225 | #ifdef CONFIG_HIBERNATION | |
3226 | /* | |
3227 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of | |
3228 | * freed pages. | |
3229 | * | |
3230 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3231 | * LRU order by reclaiming preferentially | |
3232 | * inactive > active > active referenced > active mapped | |
3233 | */ | |
3234 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) | |
3235 | { | |
3236 | struct reclaim_state reclaim_state; | |
3237 | struct scan_control sc = { | |
3238 | .gfp_mask = GFP_HIGHUSER_MOVABLE, | |
3239 | .may_swap = 1, | |
3240 | .may_unmap = 1, | |
3241 | .may_writepage = 1, | |
3242 | .nr_to_reclaim = nr_to_reclaim, | |
3243 | .hibernation_mode = 1, | |
3244 | .order = 0, | |
3245 | .priority = DEF_PRIORITY, | |
3246 | }; | |
3247 | struct shrink_control shrink = { | |
3248 | .gfp_mask = sc.gfp_mask, | |
3249 | }; | |
3250 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
3251 | struct task_struct *p = current; | |
3252 | unsigned long nr_reclaimed; | |
3253 | ||
3254 | p->flags |= PF_MEMALLOC; | |
3255 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3256 | reclaim_state.reclaimed_slab = 0; | |
3257 | p->reclaim_state = &reclaim_state; | |
3258 | ||
3259 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); | |
3260 | ||
3261 | p->reclaim_state = NULL; | |
3262 | lockdep_clear_current_reclaim_state(); | |
3263 | p->flags &= ~PF_MEMALLOC; | |
3264 | ||
3265 | return nr_reclaimed; | |
3266 | } | |
3267 | #endif /* CONFIG_HIBERNATION */ | |
3268 | ||
3269 | /* It's optimal to keep kswapds on the same CPUs as their memory, but | |
3270 | not required for correctness. So if the last cpu in a node goes | |
3271 | away, we get changed to run anywhere: as the first one comes back, | |
3272 | restore their cpu bindings. */ | |
3273 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, | |
3274 | void *hcpu) | |
3275 | { | |
3276 | int nid; | |
3277 | ||
3278 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { | |
3279 | for_each_node_state(nid, N_MEMORY) { | |
3280 | pg_data_t *pgdat = NODE_DATA(nid); | |
3281 | const struct cpumask *mask; | |
3282 | ||
3283 | mask = cpumask_of_node(pgdat->node_id); | |
3284 | ||
3285 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) | |
3286 | /* One of our CPUs online: restore mask */ | |
3287 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
3288 | } | |
3289 | } | |
3290 | return NOTIFY_OK; | |
3291 | } | |
3292 | ||
3293 | /* | |
3294 | * This kswapd start function will be called by init and node-hot-add. | |
3295 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3296 | */ | |
3297 | int kswapd_run(int nid) | |
3298 | { | |
3299 | pg_data_t *pgdat = NODE_DATA(nid); | |
3300 | int ret = 0; | |
3301 | ||
3302 | if (pgdat->kswapd) | |
3303 | return 0; | |
3304 | ||
3305 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3306 | if (IS_ERR(pgdat->kswapd)) { | |
3307 | /* failure at boot is fatal */ | |
3308 | BUG_ON(system_state == SYSTEM_BOOTING); | |
3309 | pr_err("Failed to start kswapd on node %d\n", nid); | |
3310 | ret = PTR_ERR(pgdat->kswapd); | |
3311 | pgdat->kswapd = NULL; | |
3312 | } | |
3313 | return ret; | |
3314 | } | |
3315 | ||
3316 | /* | |
3317 | * Called by memory hotplug when all memory in a node is offlined. Caller must | |
3318 | * hold lock_memory_hotplug(). | |
3319 | */ | |
3320 | void kswapd_stop(int nid) | |
3321 | { | |
3322 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3323 | ||
3324 | if (kswapd) { | |
3325 | kthread_stop(kswapd); | |
3326 | NODE_DATA(nid)->kswapd = NULL; | |
3327 | } | |
3328 | } | |
3329 | ||
3330 | static int __init kswapd_init(void) | |
3331 | { | |
3332 | int nid; | |
3333 | ||
3334 | swap_setup(); | |
3335 | for_each_node_state(nid, N_MEMORY) | |
3336 | kswapd_run(nid); | |
3337 | hotcpu_notifier(cpu_callback, 0); | |
3338 | return 0; | |
3339 | } | |
3340 | ||
3341 | module_init(kswapd_init) | |
3342 | ||
3343 | #ifdef CONFIG_NUMA | |
3344 | /* | |
3345 | * Zone reclaim mode | |
3346 | * | |
3347 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3348 | * the watermarks. | |
3349 | */ | |
3350 | int zone_reclaim_mode __read_mostly; | |
3351 | ||
3352 | #define RECLAIM_OFF 0 | |
3353 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ | |
3354 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ | |
3355 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3356 | ||
3357 | /* | |
3358 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3359 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3360 | * a zone. | |
3361 | */ | |
3362 | #define ZONE_RECLAIM_PRIORITY 4 | |
3363 | ||
3364 | /* | |
3365 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3366 | * occur. | |
3367 | */ | |
3368 | int sysctl_min_unmapped_ratio = 1; | |
3369 | ||
3370 | /* | |
3371 | * If the number of slab pages in a zone grows beyond this percentage then | |
3372 | * slab reclaim needs to occur. | |
3373 | */ | |
3374 | int sysctl_min_slab_ratio = 5; | |
3375 | ||
3376 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) | |
3377 | { | |
3378 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3379 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3380 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3381 | ||
3382 | /* | |
3383 | * It's possible for there to be more file mapped pages than | |
3384 | * accounted for by the pages on the file LRU lists because | |
3385 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3386 | */ | |
3387 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3388 | } | |
3389 | ||
3390 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3391 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3392 | { | |
3393 | long nr_pagecache_reclaimable; | |
3394 | long delta = 0; | |
3395 | ||
3396 | /* | |
3397 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3398 | * potentially reclaimable. Otherwise, we have to worry about | |
3399 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3400 | * a better estimate | |
3401 | */ | |
3402 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3403 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3404 | else | |
3405 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3406 | ||
3407 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3408 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3409 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3410 | ||
3411 | /* Watch for any possible underflows due to delta */ | |
3412 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3413 | delta = nr_pagecache_reclaimable; | |
3414 | ||
3415 | return nr_pagecache_reclaimable - delta; | |
3416 | } | |
3417 | ||
3418 | /* | |
3419 | * Try to free up some pages from this zone through reclaim. | |
3420 | */ | |
3421 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3422 | { | |
3423 | /* Minimum pages needed in order to stay on node */ | |
3424 | const unsigned long nr_pages = 1 << order; | |
3425 | struct task_struct *p = current; | |
3426 | struct reclaim_state reclaim_state; | |
3427 | struct scan_control sc = { | |
3428 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
3429 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), | |
3430 | .may_swap = 1, | |
3431 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), | |
3432 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), | |
3433 | .order = order, | |
3434 | .priority = ZONE_RECLAIM_PRIORITY, | |
3435 | }; | |
3436 | struct shrink_control shrink = { | |
3437 | .gfp_mask = sc.gfp_mask, | |
3438 | }; | |
3439 | unsigned long nr_slab_pages0, nr_slab_pages1; | |
3440 | ||
3441 | cond_resched(); | |
3442 | /* | |
3443 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3444 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3445 | * and RECLAIM_SWAP. | |
3446 | */ | |
3447 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
3448 | lockdep_set_current_reclaim_state(gfp_mask); | |
3449 | reclaim_state.reclaimed_slab = 0; | |
3450 | p->reclaim_state = &reclaim_state; | |
3451 | ||
3452 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { | |
3453 | /* | |
3454 | * Free memory by calling shrink zone with increasing | |
3455 | * priorities until we have enough memory freed. | |
3456 | */ | |
3457 | do { | |
3458 | shrink_zone(zone, &sc); | |
3459 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); | |
3460 | } | |
3461 | ||
3462 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); | |
3463 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
3464 | /* | |
3465 | * shrink_slab() does not currently allow us to determine how | |
3466 | * many pages were freed in this zone. So we take the current | |
3467 | * number of slab pages and shake the slab until it is reduced | |
3468 | * by the same nr_pages that we used for reclaiming unmapped | |
3469 | * pages. | |
3470 | * | |
3471 | * Note that shrink_slab will free memory on all zones and may | |
3472 | * take a long time. | |
3473 | */ | |
3474 | for (;;) { | |
3475 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3476 | ||
3477 | /* No reclaimable slab or very low memory pressure */ | |
3478 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) | |
3479 | break; | |
3480 | ||
3481 | /* Freed enough memory */ | |
3482 | nr_slab_pages1 = zone_page_state(zone, | |
3483 | NR_SLAB_RECLAIMABLE); | |
3484 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3485 | break; | |
3486 | } | |
3487 | ||
3488 | /* | |
3489 | * Update nr_reclaimed by the number of slab pages we | |
3490 | * reclaimed from this zone. | |
3491 | */ | |
3492 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); | |
3493 | if (nr_slab_pages1 < nr_slab_pages0) | |
3494 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
3495 | } | |
3496 | ||
3497 | p->reclaim_state = NULL; | |
3498 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); | |
3499 | lockdep_clear_current_reclaim_state(); | |
3500 | return sc.nr_reclaimed >= nr_pages; | |
3501 | } | |
3502 | ||
3503 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3504 | { | |
3505 | int node_id; | |
3506 | int ret; | |
3507 | ||
3508 | /* | |
3509 | * Zone reclaim reclaims unmapped file backed pages and | |
3510 | * slab pages if we are over the defined limits. | |
3511 | * | |
3512 | * A small portion of unmapped file backed pages is needed for | |
3513 | * file I/O otherwise pages read by file I/O will be immediately | |
3514 | * thrown out if the zone is overallocated. So we do not reclaim | |
3515 | * if less than a specified percentage of the zone is used by | |
3516 | * unmapped file backed pages. | |
3517 | */ | |
3518 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && | |
3519 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
3520 | return ZONE_RECLAIM_FULL; | |
3521 | ||
3522 | if (zone->all_unreclaimable) | |
3523 | return ZONE_RECLAIM_FULL; | |
3524 | ||
3525 | /* | |
3526 | * Do not scan if the allocation should not be delayed. | |
3527 | */ | |
3528 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) | |
3529 | return ZONE_RECLAIM_NOSCAN; | |
3530 | ||
3531 | /* | |
3532 | * Only run zone reclaim on the local zone or on zones that do not | |
3533 | * have associated processors. This will favor the local processor | |
3534 | * over remote processors and spread off node memory allocations | |
3535 | * as wide as possible. | |
3536 | */ | |
3537 | node_id = zone_to_nid(zone); | |
3538 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) | |
3539 | return ZONE_RECLAIM_NOSCAN; | |
3540 | ||
3541 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
3542 | return ZONE_RECLAIM_NOSCAN; | |
3543 | ||
3544 | ret = __zone_reclaim(zone, gfp_mask, order); | |
3545 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3546 | ||
3547 | if (!ret) | |
3548 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3549 | ||
3550 | return ret; | |
3551 | } | |
3552 | #endif | |
3553 | ||
3554 | /* | |
3555 | * page_evictable - test whether a page is evictable | |
3556 | * @page: the page to test | |
3557 | * | |
3558 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
3559 | * lists vs unevictable list. | |
3560 | * | |
3561 | * Reasons page might not be evictable: | |
3562 | * (1) page's mapping marked unevictable | |
3563 | * (2) page is part of an mlocked VMA | |
3564 | * | |
3565 | */ | |
3566 | int page_evictable(struct page *page) | |
3567 | { | |
3568 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); | |
3569 | } | |
3570 | ||
3571 | #ifdef CONFIG_SHMEM | |
3572 | /** | |
3573 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list | |
3574 | * @pages: array of pages to check | |
3575 | * @nr_pages: number of pages to check | |
3576 | * | |
3577 | * Checks pages for evictability and moves them to the appropriate lru list. | |
3578 | * | |
3579 | * This function is only used for SysV IPC SHM_UNLOCK. | |
3580 | */ | |
3581 | void check_move_unevictable_pages(struct page **pages, int nr_pages) | |
3582 | { | |
3583 | struct lruvec *lruvec; | |
3584 | struct zone *zone = NULL; | |
3585 | int pgscanned = 0; | |
3586 | int pgrescued = 0; | |
3587 | int i; | |
3588 | ||
3589 | for (i = 0; i < nr_pages; i++) { | |
3590 | struct page *page = pages[i]; | |
3591 | struct zone *pagezone; | |
3592 | ||
3593 | pgscanned++; | |
3594 | pagezone = page_zone(page); | |
3595 | if (pagezone != zone) { | |
3596 | if (zone) | |
3597 | spin_unlock_irq(&zone->lru_lock); | |
3598 | zone = pagezone; | |
3599 | spin_lock_irq(&zone->lru_lock); | |
3600 | } | |
3601 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
3602 | ||
3603 | if (!PageLRU(page) || !PageUnevictable(page)) | |
3604 | continue; | |
3605 | ||
3606 | if (page_evictable(page)) { | |
3607 | enum lru_list lru = page_lru_base_type(page); | |
3608 | ||
3609 | VM_BUG_ON(PageActive(page)); | |
3610 | ClearPageUnevictable(page); | |
3611 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); | |
3612 | add_page_to_lru_list(page, lruvec, lru); | |
3613 | pgrescued++; | |
3614 | } | |
3615 | } | |
3616 | ||
3617 | if (zone) { | |
3618 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3619 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3620 | spin_unlock_irq(&zone->lru_lock); | |
3621 | } | |
3622 | } | |
3623 | #endif /* CONFIG_SHMEM */ | |
3624 | ||
3625 | static void warn_scan_unevictable_pages(void) | |
3626 | { | |
3627 | printk_once(KERN_WARNING | |
3628 | "%s: The scan_unevictable_pages sysctl/node-interface has been " | |
3629 | "disabled for lack of a legitimate use case. If you have " | |
3630 | "one, please send an email to linux-mm@kvack.org.\n", | |
3631 | current->comm); | |
3632 | } | |
3633 | ||
3634 | /* | |
3635 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3636 | * all nodes' unevictable lists for evictable pages | |
3637 | */ | |
3638 | unsigned long scan_unevictable_pages; | |
3639 | ||
3640 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
3641 | void __user *buffer, | |
3642 | size_t *length, loff_t *ppos) | |
3643 | { | |
3644 | warn_scan_unevictable_pages(); | |
3645 | proc_doulongvec_minmax(table, write, buffer, length, ppos); | |
3646 | scan_unevictable_pages = 0; | |
3647 | return 0; | |
3648 | } | |
3649 | ||
3650 | #ifdef CONFIG_NUMA | |
3651 | /* | |
3652 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3653 | * a specified node's per zone unevictable lists for evictable pages. | |
3654 | */ | |
3655 | ||
3656 | static ssize_t read_scan_unevictable_node(struct device *dev, | |
3657 | struct device_attribute *attr, | |
3658 | char *buf) | |
3659 | { | |
3660 | warn_scan_unevictable_pages(); | |
3661 | return sprintf(buf, "0\n"); /* always zero; should fit... */ | |
3662 | } | |
3663 | ||
3664 | static ssize_t write_scan_unevictable_node(struct device *dev, | |
3665 | struct device_attribute *attr, | |
3666 | const char *buf, size_t count) | |
3667 | { | |
3668 | warn_scan_unevictable_pages(); | |
3669 | return 1; | |
3670 | } | |
3671 | ||
3672 | ||
3673 | static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, | |
3674 | read_scan_unevictable_node, | |
3675 | write_scan_unevictable_node); | |
3676 | ||
3677 | int scan_unevictable_register_node(struct node *node) | |
3678 | { | |
3679 | return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages); | |
3680 | } | |
3681 | ||
3682 | void scan_unevictable_unregister_node(struct node *node) | |
3683 | { | |
3684 | device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages); | |
3685 | } | |
3686 | #endif |