Commit | Line | Data |
---|---|---|
1da177e4 LT |
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/slab.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> | |
e129b5c2 | 22 | #include <linux/vmstat.h> |
1da177e4 LT |
23 | #include <linux/file.h> |
24 | #include <linux/writeback.h> | |
25 | #include <linux/blkdev.h> | |
26 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
27 | buffer_heads_over_limit */ | |
28 | #include <linux/mm_inline.h> | |
29 | #include <linux/pagevec.h> | |
30 | #include <linux/backing-dev.h> | |
31 | #include <linux/rmap.h> | |
32 | #include <linux/topology.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/cpuset.h> | |
35 | #include <linux/notifier.h> | |
36 | #include <linux/rwsem.h> | |
248a0301 | 37 | #include <linux/delay.h> |
3218ae14 | 38 | #include <linux/kthread.h> |
7dfb7103 | 39 | #include <linux/freezer.h> |
66e1707b | 40 | #include <linux/memcontrol.h> |
873b4771 | 41 | #include <linux/delayacct.h> |
af936a16 | 42 | #include <linux/sysctl.h> |
1da177e4 LT |
43 | |
44 | #include <asm/tlbflush.h> | |
45 | #include <asm/div64.h> | |
46 | ||
47 | #include <linux/swapops.h> | |
48 | ||
0f8053a5 NP |
49 | #include "internal.h" |
50 | ||
1da177e4 | 51 | struct scan_control { |
1da177e4 LT |
52 | /* Incremented by the number of inactive pages that were scanned */ |
53 | unsigned long nr_scanned; | |
54 | ||
a79311c1 RR |
55 | /* Number of pages freed so far during a call to shrink_zones() */ |
56 | unsigned long nr_reclaimed; | |
57 | ||
1da177e4 | 58 | /* This context's GFP mask */ |
6daa0e28 | 59 | gfp_t gfp_mask; |
1da177e4 LT |
60 | |
61 | int may_writepage; | |
62 | ||
a6dc60f8 JW |
63 | /* Can mapped pages be reclaimed? */ |
64 | int may_unmap; | |
f1fd1067 | 65 | |
2e2e4259 KM |
66 | /* Can pages be swapped as part of reclaim? */ |
67 | int may_swap; | |
68 | ||
1da177e4 LT |
69 | /* This context's SWAP_CLUSTER_MAX. If freeing memory for |
70 | * suspend, we effectively ignore SWAP_CLUSTER_MAX. | |
71 | * In this context, it doesn't matter that we scan the | |
72 | * whole list at once. */ | |
73 | int swap_cluster_max; | |
d6277db4 RW |
74 | |
75 | int swappiness; | |
408d8544 NP |
76 | |
77 | int all_unreclaimable; | |
5ad333eb AW |
78 | |
79 | int order; | |
66e1707b BS |
80 | |
81 | /* Which cgroup do we reclaim from */ | |
82 | struct mem_cgroup *mem_cgroup; | |
83 | ||
327c0e96 KH |
84 | /* |
85 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
86 | * are scanned. | |
87 | */ | |
88 | nodemask_t *nodemask; | |
89 | ||
66e1707b BS |
90 | /* Pluggable isolate pages callback */ |
91 | unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst, | |
92 | unsigned long *scanned, int order, int mode, | |
93 | struct zone *z, struct mem_cgroup *mem_cont, | |
4f98a2fe | 94 | int active, int file); |
1da177e4 LT |
95 | }; |
96 | ||
1da177e4 LT |
97 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
98 | ||
99 | #ifdef ARCH_HAS_PREFETCH | |
100 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
101 | do { \ | |
102 | if ((_page)->lru.prev != _base) { \ | |
103 | struct page *prev; \ | |
104 | \ | |
105 | prev = lru_to_page(&(_page->lru)); \ | |
106 | prefetch(&prev->_field); \ | |
107 | } \ | |
108 | } while (0) | |
109 | #else | |
110 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
111 | #endif | |
112 | ||
113 | #ifdef ARCH_HAS_PREFETCHW | |
114 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
115 | do { \ | |
116 | if ((_page)->lru.prev != _base) { \ | |
117 | struct page *prev; \ | |
118 | \ | |
119 | prev = lru_to_page(&(_page->lru)); \ | |
120 | prefetchw(&prev->_field); \ | |
121 | } \ | |
122 | } while (0) | |
123 | #else | |
124 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
125 | #endif | |
126 | ||
127 | /* | |
128 | * From 0 .. 100. Higher means more swappy. | |
129 | */ | |
130 | int vm_swappiness = 60; | |
bd1e22b8 | 131 | long vm_total_pages; /* The total number of pages which the VM controls */ |
1da177e4 LT |
132 | |
133 | static LIST_HEAD(shrinker_list); | |
134 | static DECLARE_RWSEM(shrinker_rwsem); | |
135 | ||
00f0b825 | 136 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
e72e2bd6 | 137 | #define scanning_global_lru(sc) (!(sc)->mem_cgroup) |
91a45470 | 138 | #else |
e72e2bd6 | 139 | #define scanning_global_lru(sc) (1) |
91a45470 KH |
140 | #endif |
141 | ||
6e901571 KM |
142 | static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone, |
143 | struct scan_control *sc) | |
144 | { | |
e72e2bd6 | 145 | if (!scanning_global_lru(sc)) |
3e2f41f1 KM |
146 | return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone); |
147 | ||
6e901571 KM |
148 | return &zone->reclaim_stat; |
149 | } | |
150 | ||
0b217676 VL |
151 | static unsigned long zone_nr_lru_pages(struct zone *zone, |
152 | struct scan_control *sc, enum lru_list lru) | |
c9f299d9 | 153 | { |
e72e2bd6 | 154 | if (!scanning_global_lru(sc)) |
a3d8e054 KM |
155 | return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru); |
156 | ||
c9f299d9 KM |
157 | return zone_page_state(zone, NR_LRU_BASE + lru); |
158 | } | |
159 | ||
160 | ||
1da177e4 LT |
161 | /* |
162 | * Add a shrinker callback to be called from the vm | |
163 | */ | |
8e1f936b | 164 | void register_shrinker(struct shrinker *shrinker) |
1da177e4 | 165 | { |
8e1f936b RR |
166 | shrinker->nr = 0; |
167 | down_write(&shrinker_rwsem); | |
168 | list_add_tail(&shrinker->list, &shrinker_list); | |
169 | up_write(&shrinker_rwsem); | |
1da177e4 | 170 | } |
8e1f936b | 171 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
172 | |
173 | /* | |
174 | * Remove one | |
175 | */ | |
8e1f936b | 176 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
177 | { |
178 | down_write(&shrinker_rwsem); | |
179 | list_del(&shrinker->list); | |
180 | up_write(&shrinker_rwsem); | |
1da177e4 | 181 | } |
8e1f936b | 182 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
183 | |
184 | #define SHRINK_BATCH 128 | |
185 | /* | |
186 | * Call the shrink functions to age shrinkable caches | |
187 | * | |
188 | * Here we assume it costs one seek to replace a lru page and that it also | |
189 | * takes a seek to recreate a cache object. With this in mind we age equal | |
190 | * percentages of the lru and ageable caches. This should balance the seeks | |
191 | * generated by these structures. | |
192 | * | |
183ff22b | 193 | * If the vm encountered mapped pages on the LRU it increase the pressure on |
1da177e4 LT |
194 | * slab to avoid swapping. |
195 | * | |
196 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
197 | * | |
198 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
199 | * are eligible for the caller's allocation attempt. It is used for balancing | |
200 | * slab reclaim versus page reclaim. | |
b15e0905 | 201 | * |
202 | * Returns the number of slab objects which we shrunk. | |
1da177e4 | 203 | */ |
69e05944 AM |
204 | unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask, |
205 | unsigned long lru_pages) | |
1da177e4 LT |
206 | { |
207 | struct shrinker *shrinker; | |
69e05944 | 208 | unsigned long ret = 0; |
1da177e4 LT |
209 | |
210 | if (scanned == 0) | |
211 | scanned = SWAP_CLUSTER_MAX; | |
212 | ||
213 | if (!down_read_trylock(&shrinker_rwsem)) | |
b15e0905 | 214 | return 1; /* Assume we'll be able to shrink next time */ |
1da177e4 LT |
215 | |
216 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
217 | unsigned long long delta; | |
218 | unsigned long total_scan; | |
8e1f936b | 219 | unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask); |
1da177e4 LT |
220 | |
221 | delta = (4 * scanned) / shrinker->seeks; | |
ea164d73 | 222 | delta *= max_pass; |
1da177e4 LT |
223 | do_div(delta, lru_pages + 1); |
224 | shrinker->nr += delta; | |
ea164d73 | 225 | if (shrinker->nr < 0) { |
88c3bd70 DR |
226 | printk(KERN_ERR "shrink_slab: %pF negative objects to " |
227 | "delete nr=%ld\n", | |
228 | shrinker->shrink, shrinker->nr); | |
ea164d73 AA |
229 | shrinker->nr = max_pass; |
230 | } | |
231 | ||
232 | /* | |
233 | * Avoid risking looping forever due to too large nr value: | |
234 | * never try to free more than twice the estimate number of | |
235 | * freeable entries. | |
236 | */ | |
237 | if (shrinker->nr > max_pass * 2) | |
238 | shrinker->nr = max_pass * 2; | |
1da177e4 LT |
239 | |
240 | total_scan = shrinker->nr; | |
241 | shrinker->nr = 0; | |
242 | ||
243 | while (total_scan >= SHRINK_BATCH) { | |
244 | long this_scan = SHRINK_BATCH; | |
245 | int shrink_ret; | |
b15e0905 | 246 | int nr_before; |
1da177e4 | 247 | |
8e1f936b RR |
248 | nr_before = (*shrinker->shrink)(0, gfp_mask); |
249 | shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask); | |
1da177e4 LT |
250 | if (shrink_ret == -1) |
251 | break; | |
b15e0905 | 252 | if (shrink_ret < nr_before) |
253 | ret += nr_before - shrink_ret; | |
f8891e5e | 254 | count_vm_events(SLABS_SCANNED, this_scan); |
1da177e4 LT |
255 | total_scan -= this_scan; |
256 | ||
257 | cond_resched(); | |
258 | } | |
259 | ||
260 | shrinker->nr += total_scan; | |
261 | } | |
262 | up_read(&shrinker_rwsem); | |
b15e0905 | 263 | return ret; |
1da177e4 LT |
264 | } |
265 | ||
266 | /* Called without lock on whether page is mapped, so answer is unstable */ | |
267 | static inline int page_mapping_inuse(struct page *page) | |
268 | { | |
269 | struct address_space *mapping; | |
270 | ||
271 | /* Page is in somebody's page tables. */ | |
272 | if (page_mapped(page)) | |
273 | return 1; | |
274 | ||
275 | /* Be more reluctant to reclaim swapcache than pagecache */ | |
276 | if (PageSwapCache(page)) | |
277 | return 1; | |
278 | ||
279 | mapping = page_mapping(page); | |
280 | if (!mapping) | |
281 | return 0; | |
282 | ||
283 | /* File is mmap'd by somebody? */ | |
284 | return mapping_mapped(mapping); | |
285 | } | |
286 | ||
287 | static inline int is_page_cache_freeable(struct page *page) | |
288 | { | |
ceddc3a5 JW |
289 | /* |
290 | * A freeable page cache page is referenced only by the caller | |
291 | * that isolated the page, the page cache radix tree and | |
292 | * optional buffer heads at page->private. | |
293 | */ | |
edcf4748 | 294 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
295 | } |
296 | ||
297 | static int may_write_to_queue(struct backing_dev_info *bdi) | |
298 | { | |
930d9152 | 299 | if (current->flags & PF_SWAPWRITE) |
1da177e4 LT |
300 | return 1; |
301 | if (!bdi_write_congested(bdi)) | |
302 | return 1; | |
303 | if (bdi == current->backing_dev_info) | |
304 | return 1; | |
305 | return 0; | |
306 | } | |
307 | ||
308 | /* | |
309 | * We detected a synchronous write error writing a page out. Probably | |
310 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
311 | * fsync(), msync() or close(). | |
312 | * | |
313 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
314 | * prevents it from being freed up. But we have a ref on the page and once | |
315 | * that page is locked, the mapping is pinned. | |
316 | * | |
317 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
318 | * __GFP_FS. | |
319 | */ | |
320 | static void handle_write_error(struct address_space *mapping, | |
321 | struct page *page, int error) | |
322 | { | |
323 | lock_page(page); | |
3e9f45bd GC |
324 | if (page_mapping(page) == mapping) |
325 | mapping_set_error(mapping, error); | |
1da177e4 LT |
326 | unlock_page(page); |
327 | } | |
328 | ||
c661b078 AW |
329 | /* Request for sync pageout. */ |
330 | enum pageout_io { | |
331 | PAGEOUT_IO_ASYNC, | |
332 | PAGEOUT_IO_SYNC, | |
333 | }; | |
334 | ||
04e62a29 CL |
335 | /* possible outcome of pageout() */ |
336 | typedef enum { | |
337 | /* failed to write page out, page is locked */ | |
338 | PAGE_KEEP, | |
339 | /* move page to the active list, page is locked */ | |
340 | PAGE_ACTIVATE, | |
341 | /* page has been sent to the disk successfully, page is unlocked */ | |
342 | PAGE_SUCCESS, | |
343 | /* page is clean and locked */ | |
344 | PAGE_CLEAN, | |
345 | } pageout_t; | |
346 | ||
1da177e4 | 347 | /* |
1742f19f AM |
348 | * pageout is called by shrink_page_list() for each dirty page. |
349 | * Calls ->writepage(). | |
1da177e4 | 350 | */ |
c661b078 AW |
351 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
352 | enum pageout_io sync_writeback) | |
1da177e4 LT |
353 | { |
354 | /* | |
355 | * If the page is dirty, only perform writeback if that write | |
356 | * will be non-blocking. To prevent this allocation from being | |
357 | * stalled by pagecache activity. But note that there may be | |
358 | * stalls if we need to run get_block(). We could test | |
359 | * PagePrivate for that. | |
360 | * | |
361 | * If this process is currently in generic_file_write() against | |
362 | * this page's queue, we can perform writeback even if that | |
363 | * will block. | |
364 | * | |
365 | * If the page is swapcache, write it back even if that would | |
366 | * block, for some throttling. This happens by accident, because | |
367 | * swap_backing_dev_info is bust: it doesn't reflect the | |
368 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
369 | */ |
370 | if (!is_page_cache_freeable(page)) | |
371 | return PAGE_KEEP; | |
372 | if (!mapping) { | |
373 | /* | |
374 | * Some data journaling orphaned pages can have | |
375 | * page->mapping == NULL while being dirty with clean buffers. | |
376 | */ | |
266cf658 | 377 | if (page_has_private(page)) { |
1da177e4 LT |
378 | if (try_to_free_buffers(page)) { |
379 | ClearPageDirty(page); | |
d40cee24 | 380 | printk("%s: orphaned page\n", __func__); |
1da177e4 LT |
381 | return PAGE_CLEAN; |
382 | } | |
383 | } | |
384 | return PAGE_KEEP; | |
385 | } | |
386 | if (mapping->a_ops->writepage == NULL) | |
387 | return PAGE_ACTIVATE; | |
388 | if (!may_write_to_queue(mapping->backing_dev_info)) | |
389 | return PAGE_KEEP; | |
390 | ||
391 | if (clear_page_dirty_for_io(page)) { | |
392 | int res; | |
393 | struct writeback_control wbc = { | |
394 | .sync_mode = WB_SYNC_NONE, | |
395 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
396 | .range_start = 0, |
397 | .range_end = LLONG_MAX, | |
1da177e4 LT |
398 | .nonblocking = 1, |
399 | .for_reclaim = 1, | |
400 | }; | |
401 | ||
402 | SetPageReclaim(page); | |
403 | res = mapping->a_ops->writepage(page, &wbc); | |
404 | if (res < 0) | |
405 | handle_write_error(mapping, page, res); | |
994fc28c | 406 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
407 | ClearPageReclaim(page); |
408 | return PAGE_ACTIVATE; | |
409 | } | |
c661b078 AW |
410 | |
411 | /* | |
412 | * Wait on writeback if requested to. This happens when | |
413 | * direct reclaiming a large contiguous area and the | |
414 | * first attempt to free a range of pages fails. | |
415 | */ | |
416 | if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC) | |
417 | wait_on_page_writeback(page); | |
418 | ||
1da177e4 LT |
419 | if (!PageWriteback(page)) { |
420 | /* synchronous write or broken a_ops? */ | |
421 | ClearPageReclaim(page); | |
422 | } | |
e129b5c2 | 423 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
424 | return PAGE_SUCCESS; |
425 | } | |
426 | ||
427 | return PAGE_CLEAN; | |
428 | } | |
429 | ||
a649fd92 | 430 | /* |
e286781d NP |
431 | * Same as remove_mapping, but if the page is removed from the mapping, it |
432 | * gets returned with a refcount of 0. | |
a649fd92 | 433 | */ |
e286781d | 434 | static int __remove_mapping(struct address_space *mapping, struct page *page) |
49d2e9cc | 435 | { |
28e4d965 NP |
436 | BUG_ON(!PageLocked(page)); |
437 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 438 | |
19fd6231 | 439 | spin_lock_irq(&mapping->tree_lock); |
49d2e9cc | 440 | /* |
0fd0e6b0 NP |
441 | * The non racy check for a busy page. |
442 | * | |
443 | * Must be careful with the order of the tests. When someone has | |
444 | * a ref to the page, it may be possible that they dirty it then | |
445 | * drop the reference. So if PageDirty is tested before page_count | |
446 | * here, then the following race may occur: | |
447 | * | |
448 | * get_user_pages(&page); | |
449 | * [user mapping goes away] | |
450 | * write_to(page); | |
451 | * !PageDirty(page) [good] | |
452 | * SetPageDirty(page); | |
453 | * put_page(page); | |
454 | * !page_count(page) [good, discard it] | |
455 | * | |
456 | * [oops, our write_to data is lost] | |
457 | * | |
458 | * Reversing the order of the tests ensures such a situation cannot | |
459 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
460 | * load is not satisfied before that of page->_count. | |
461 | * | |
462 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
463 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 464 | */ |
e286781d | 465 | if (!page_freeze_refs(page, 2)) |
49d2e9cc | 466 | goto cannot_free; |
e286781d NP |
467 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
468 | if (unlikely(PageDirty(page))) { | |
469 | page_unfreeze_refs(page, 2); | |
49d2e9cc | 470 | goto cannot_free; |
e286781d | 471 | } |
49d2e9cc CL |
472 | |
473 | if (PageSwapCache(page)) { | |
474 | swp_entry_t swap = { .val = page_private(page) }; | |
475 | __delete_from_swap_cache(page); | |
19fd6231 | 476 | spin_unlock_irq(&mapping->tree_lock); |
cb4b86ba | 477 | swapcache_free(swap, page); |
e286781d NP |
478 | } else { |
479 | __remove_from_page_cache(page); | |
19fd6231 | 480 | spin_unlock_irq(&mapping->tree_lock); |
e767e056 | 481 | mem_cgroup_uncharge_cache_page(page); |
49d2e9cc CL |
482 | } |
483 | ||
49d2e9cc CL |
484 | return 1; |
485 | ||
486 | cannot_free: | |
19fd6231 | 487 | spin_unlock_irq(&mapping->tree_lock); |
49d2e9cc CL |
488 | return 0; |
489 | } | |
490 | ||
e286781d NP |
491 | /* |
492 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
493 | * someone else has a ref on the page, abort and return 0. If it was | |
494 | * successfully detached, return 1. Assumes the caller has a single ref on | |
495 | * this page. | |
496 | */ | |
497 | int remove_mapping(struct address_space *mapping, struct page *page) | |
498 | { | |
499 | if (__remove_mapping(mapping, page)) { | |
500 | /* | |
501 | * Unfreezing the refcount with 1 rather than 2 effectively | |
502 | * drops the pagecache ref for us without requiring another | |
503 | * atomic operation. | |
504 | */ | |
505 | page_unfreeze_refs(page, 1); | |
506 | return 1; | |
507 | } | |
508 | return 0; | |
509 | } | |
510 | ||
894bc310 LS |
511 | /** |
512 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
513 | * @page: page to be put back to appropriate lru list | |
514 | * | |
515 | * Add previously isolated @page to appropriate LRU list. | |
516 | * Page may still be unevictable for other reasons. | |
517 | * | |
518 | * lru_lock must not be held, interrupts must be enabled. | |
519 | */ | |
894bc310 LS |
520 | void putback_lru_page(struct page *page) |
521 | { | |
522 | int lru; | |
523 | int active = !!TestClearPageActive(page); | |
bbfd28ee | 524 | int was_unevictable = PageUnevictable(page); |
894bc310 LS |
525 | |
526 | VM_BUG_ON(PageLRU(page)); | |
527 | ||
528 | redo: | |
529 | ClearPageUnevictable(page); | |
530 | ||
531 | if (page_evictable(page, NULL)) { | |
532 | /* | |
533 | * For evictable pages, we can use the cache. | |
534 | * In event of a race, worst case is we end up with an | |
535 | * unevictable page on [in]active list. | |
536 | * We know how to handle that. | |
537 | */ | |
401a8e1c | 538 | lru = active + page_lru_base_type(page); |
894bc310 LS |
539 | lru_cache_add_lru(page, lru); |
540 | } else { | |
541 | /* | |
542 | * Put unevictable pages directly on zone's unevictable | |
543 | * list. | |
544 | */ | |
545 | lru = LRU_UNEVICTABLE; | |
546 | add_page_to_unevictable_list(page); | |
547 | } | |
894bc310 LS |
548 | |
549 | /* | |
550 | * page's status can change while we move it among lru. If an evictable | |
551 | * page is on unevictable list, it never be freed. To avoid that, | |
552 | * check after we added it to the list, again. | |
553 | */ | |
554 | if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { | |
555 | if (!isolate_lru_page(page)) { | |
556 | put_page(page); | |
557 | goto redo; | |
558 | } | |
559 | /* This means someone else dropped this page from LRU | |
560 | * So, it will be freed or putback to LRU again. There is | |
561 | * nothing to do here. | |
562 | */ | |
563 | } | |
564 | ||
bbfd28ee LS |
565 | if (was_unevictable && lru != LRU_UNEVICTABLE) |
566 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
567 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) | |
568 | count_vm_event(UNEVICTABLE_PGCULLED); | |
569 | ||
894bc310 LS |
570 | put_page(page); /* drop ref from isolate */ |
571 | } | |
572 | ||
1da177e4 | 573 | /* |
1742f19f | 574 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 575 | */ |
1742f19f | 576 | static unsigned long shrink_page_list(struct list_head *page_list, |
c661b078 AW |
577 | struct scan_control *sc, |
578 | enum pageout_io sync_writeback) | |
1da177e4 LT |
579 | { |
580 | LIST_HEAD(ret_pages); | |
581 | struct pagevec freed_pvec; | |
582 | int pgactivate = 0; | |
05ff5137 | 583 | unsigned long nr_reclaimed = 0; |
6fe6b7e3 | 584 | unsigned long vm_flags; |
1da177e4 LT |
585 | |
586 | cond_resched(); | |
587 | ||
588 | pagevec_init(&freed_pvec, 1); | |
589 | while (!list_empty(page_list)) { | |
590 | struct address_space *mapping; | |
591 | struct page *page; | |
592 | int may_enter_fs; | |
593 | int referenced; | |
594 | ||
595 | cond_resched(); | |
596 | ||
597 | page = lru_to_page(page_list); | |
598 | list_del(&page->lru); | |
599 | ||
529ae9aa | 600 | if (!trylock_page(page)) |
1da177e4 LT |
601 | goto keep; |
602 | ||
725d704e | 603 | VM_BUG_ON(PageActive(page)); |
1da177e4 LT |
604 | |
605 | sc->nr_scanned++; | |
80e43426 | 606 | |
b291f000 NP |
607 | if (unlikely(!page_evictable(page, NULL))) |
608 | goto cull_mlocked; | |
894bc310 | 609 | |
a6dc60f8 | 610 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
611 | goto keep_locked; |
612 | ||
1da177e4 LT |
613 | /* Double the slab pressure for mapped and swapcache pages */ |
614 | if (page_mapped(page) || PageSwapCache(page)) | |
615 | sc->nr_scanned++; | |
616 | ||
c661b078 AW |
617 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
618 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
619 | ||
620 | if (PageWriteback(page)) { | |
621 | /* | |
622 | * Synchronous reclaim is performed in two passes, | |
623 | * first an asynchronous pass over the list to | |
624 | * start parallel writeback, and a second synchronous | |
625 | * pass to wait for the IO to complete. Wait here | |
626 | * for any page for which writeback has already | |
627 | * started. | |
628 | */ | |
629 | if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs) | |
630 | wait_on_page_writeback(page); | |
4dd4b920 | 631 | else |
c661b078 AW |
632 | goto keep_locked; |
633 | } | |
1da177e4 | 634 | |
6fe6b7e3 WF |
635 | referenced = page_referenced(page, 1, |
636 | sc->mem_cgroup, &vm_flags); | |
03ef83af MK |
637 | /* |
638 | * In active use or really unfreeable? Activate it. | |
639 | * If page which have PG_mlocked lost isoltation race, | |
640 | * try_to_unmap moves it to unevictable list | |
641 | */ | |
5ad333eb | 642 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && |
03ef83af MK |
643 | referenced && page_mapping_inuse(page) |
644 | && !(vm_flags & VM_LOCKED)) | |
1da177e4 LT |
645 | goto activate_locked; |
646 | ||
1da177e4 LT |
647 | /* |
648 | * Anonymous process memory has backing store? | |
649 | * Try to allocate it some swap space here. | |
650 | */ | |
b291f000 | 651 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
652 | if (!(sc->gfp_mask & __GFP_IO)) |
653 | goto keep_locked; | |
ac47b003 | 654 | if (!add_to_swap(page)) |
1da177e4 | 655 | goto activate_locked; |
63eb6b93 | 656 | may_enter_fs = 1; |
b291f000 | 657 | } |
1da177e4 LT |
658 | |
659 | mapping = page_mapping(page); | |
1da177e4 LT |
660 | |
661 | /* | |
662 | * The page is mapped into the page tables of one or more | |
663 | * processes. Try to unmap it here. | |
664 | */ | |
665 | if (page_mapped(page) && mapping) { | |
a48d07af | 666 | switch (try_to_unmap(page, 0)) { |
1da177e4 LT |
667 | case SWAP_FAIL: |
668 | goto activate_locked; | |
669 | case SWAP_AGAIN: | |
670 | goto keep_locked; | |
b291f000 NP |
671 | case SWAP_MLOCK: |
672 | goto cull_mlocked; | |
1da177e4 LT |
673 | case SWAP_SUCCESS: |
674 | ; /* try to free the page below */ | |
675 | } | |
676 | } | |
677 | ||
678 | if (PageDirty(page)) { | |
5ad333eb | 679 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced) |
1da177e4 | 680 | goto keep_locked; |
4dd4b920 | 681 | if (!may_enter_fs) |
1da177e4 | 682 | goto keep_locked; |
52a8363e | 683 | if (!sc->may_writepage) |
1da177e4 LT |
684 | goto keep_locked; |
685 | ||
686 | /* Page is dirty, try to write it out here */ | |
c661b078 | 687 | switch (pageout(page, mapping, sync_writeback)) { |
1da177e4 LT |
688 | case PAGE_KEEP: |
689 | goto keep_locked; | |
690 | case PAGE_ACTIVATE: | |
691 | goto activate_locked; | |
692 | case PAGE_SUCCESS: | |
4dd4b920 | 693 | if (PageWriteback(page) || PageDirty(page)) |
1da177e4 LT |
694 | goto keep; |
695 | /* | |
696 | * A synchronous write - probably a ramdisk. Go | |
697 | * ahead and try to reclaim the page. | |
698 | */ | |
529ae9aa | 699 | if (!trylock_page(page)) |
1da177e4 LT |
700 | goto keep; |
701 | if (PageDirty(page) || PageWriteback(page)) | |
702 | goto keep_locked; | |
703 | mapping = page_mapping(page); | |
704 | case PAGE_CLEAN: | |
705 | ; /* try to free the page below */ | |
706 | } | |
707 | } | |
708 | ||
709 | /* | |
710 | * If the page has buffers, try to free the buffer mappings | |
711 | * associated with this page. If we succeed we try to free | |
712 | * the page as well. | |
713 | * | |
714 | * We do this even if the page is PageDirty(). | |
715 | * try_to_release_page() does not perform I/O, but it is | |
716 | * possible for a page to have PageDirty set, but it is actually | |
717 | * clean (all its buffers are clean). This happens if the | |
718 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 719 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
720 | * try_to_release_page() will discover that cleanness and will |
721 | * drop the buffers and mark the page clean - it can be freed. | |
722 | * | |
723 | * Rarely, pages can have buffers and no ->mapping. These are | |
724 | * the pages which were not successfully invalidated in | |
725 | * truncate_complete_page(). We try to drop those buffers here | |
726 | * and if that worked, and the page is no longer mapped into | |
727 | * process address space (page_count == 1) it can be freed. | |
728 | * Otherwise, leave the page on the LRU so it is swappable. | |
729 | */ | |
266cf658 | 730 | if (page_has_private(page)) { |
1da177e4 LT |
731 | if (!try_to_release_page(page, sc->gfp_mask)) |
732 | goto activate_locked; | |
e286781d NP |
733 | if (!mapping && page_count(page) == 1) { |
734 | unlock_page(page); | |
735 | if (put_page_testzero(page)) | |
736 | goto free_it; | |
737 | else { | |
738 | /* | |
739 | * rare race with speculative reference. | |
740 | * the speculative reference will free | |
741 | * this page shortly, so we may | |
742 | * increment nr_reclaimed here (and | |
743 | * leave it off the LRU). | |
744 | */ | |
745 | nr_reclaimed++; | |
746 | continue; | |
747 | } | |
748 | } | |
1da177e4 LT |
749 | } |
750 | ||
e286781d | 751 | if (!mapping || !__remove_mapping(mapping, page)) |
49d2e9cc | 752 | goto keep_locked; |
1da177e4 | 753 | |
a978d6f5 NP |
754 | /* |
755 | * At this point, we have no other references and there is | |
756 | * no way to pick any more up (removed from LRU, removed | |
757 | * from pagecache). Can use non-atomic bitops now (and | |
758 | * we obviously don't have to worry about waking up a process | |
759 | * waiting on the page lock, because there are no references. | |
760 | */ | |
761 | __clear_page_locked(page); | |
e286781d | 762 | free_it: |
05ff5137 | 763 | nr_reclaimed++; |
e286781d NP |
764 | if (!pagevec_add(&freed_pvec, page)) { |
765 | __pagevec_free(&freed_pvec); | |
766 | pagevec_reinit(&freed_pvec); | |
767 | } | |
1da177e4 LT |
768 | continue; |
769 | ||
b291f000 | 770 | cull_mlocked: |
63d6c5ad HD |
771 | if (PageSwapCache(page)) |
772 | try_to_free_swap(page); | |
b291f000 NP |
773 | unlock_page(page); |
774 | putback_lru_page(page); | |
775 | continue; | |
776 | ||
1da177e4 | 777 | activate_locked: |
68a22394 RR |
778 | /* Not a candidate for swapping, so reclaim swap space. */ |
779 | if (PageSwapCache(page) && vm_swap_full()) | |
a2c43eed | 780 | try_to_free_swap(page); |
894bc310 | 781 | VM_BUG_ON(PageActive(page)); |
1da177e4 LT |
782 | SetPageActive(page); |
783 | pgactivate++; | |
784 | keep_locked: | |
785 | unlock_page(page); | |
786 | keep: | |
787 | list_add(&page->lru, &ret_pages); | |
b291f000 | 788 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); |
1da177e4 LT |
789 | } |
790 | list_splice(&ret_pages, page_list); | |
791 | if (pagevec_count(&freed_pvec)) | |
e286781d | 792 | __pagevec_free(&freed_pvec); |
f8891e5e | 793 | count_vm_events(PGACTIVATE, pgactivate); |
05ff5137 | 794 | return nr_reclaimed; |
1da177e4 LT |
795 | } |
796 | ||
5ad333eb AW |
797 | /* LRU Isolation modes. */ |
798 | #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */ | |
799 | #define ISOLATE_ACTIVE 1 /* Isolate active pages. */ | |
800 | #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */ | |
801 | ||
802 | /* | |
803 | * Attempt to remove the specified page from its LRU. Only take this page | |
804 | * if it is of the appropriate PageActive status. Pages which are being | |
805 | * freed elsewhere are also ignored. | |
806 | * | |
807 | * page: page to consider | |
808 | * mode: one of the LRU isolation modes defined above | |
809 | * | |
810 | * returns 0 on success, -ve errno on failure. | |
811 | */ | |
4f98a2fe | 812 | int __isolate_lru_page(struct page *page, int mode, int file) |
5ad333eb AW |
813 | { |
814 | int ret = -EINVAL; | |
815 | ||
816 | /* Only take pages on the LRU. */ | |
817 | if (!PageLRU(page)) | |
818 | return ret; | |
819 | ||
820 | /* | |
821 | * When checking the active state, we need to be sure we are | |
822 | * dealing with comparible boolean values. Take the logical not | |
823 | * of each. | |
824 | */ | |
825 | if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode)) | |
826 | return ret; | |
827 | ||
6c0b1351 | 828 | if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file) |
4f98a2fe RR |
829 | return ret; |
830 | ||
894bc310 LS |
831 | /* |
832 | * When this function is being called for lumpy reclaim, we | |
833 | * initially look into all LRU pages, active, inactive and | |
834 | * unevictable; only give shrink_page_list evictable pages. | |
835 | */ | |
836 | if (PageUnevictable(page)) | |
837 | return ret; | |
838 | ||
5ad333eb | 839 | ret = -EBUSY; |
08e552c6 | 840 | |
5ad333eb AW |
841 | if (likely(get_page_unless_zero(page))) { |
842 | /* | |
843 | * Be careful not to clear PageLRU until after we're | |
844 | * sure the page is not being freed elsewhere -- the | |
845 | * page release code relies on it. | |
846 | */ | |
847 | ClearPageLRU(page); | |
848 | ret = 0; | |
849 | } | |
850 | ||
851 | return ret; | |
852 | } | |
853 | ||
1da177e4 LT |
854 | /* |
855 | * zone->lru_lock is heavily contended. Some of the functions that | |
856 | * shrink the lists perform better by taking out a batch of pages | |
857 | * and working on them outside the LRU lock. | |
858 | * | |
859 | * For pagecache intensive workloads, this function is the hottest | |
860 | * spot in the kernel (apart from copy_*_user functions). | |
861 | * | |
862 | * Appropriate locks must be held before calling this function. | |
863 | * | |
864 | * @nr_to_scan: The number of pages to look through on the list. | |
865 | * @src: The LRU list to pull pages off. | |
866 | * @dst: The temp list to put pages on to. | |
867 | * @scanned: The number of pages that were scanned. | |
5ad333eb AW |
868 | * @order: The caller's attempted allocation order |
869 | * @mode: One of the LRU isolation modes | |
4f98a2fe | 870 | * @file: True [1] if isolating file [!anon] pages |
1da177e4 LT |
871 | * |
872 | * returns how many pages were moved onto *@dst. | |
873 | */ | |
69e05944 AM |
874 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
875 | struct list_head *src, struct list_head *dst, | |
4f98a2fe | 876 | unsigned long *scanned, int order, int mode, int file) |
1da177e4 | 877 | { |
69e05944 | 878 | unsigned long nr_taken = 0; |
c9b02d97 | 879 | unsigned long scan; |
1da177e4 | 880 | |
c9b02d97 | 881 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
5ad333eb AW |
882 | struct page *page; |
883 | unsigned long pfn; | |
884 | unsigned long end_pfn; | |
885 | unsigned long page_pfn; | |
886 | int zone_id; | |
887 | ||
1da177e4 LT |
888 | page = lru_to_page(src); |
889 | prefetchw_prev_lru_page(page, src, flags); | |
890 | ||
725d704e | 891 | VM_BUG_ON(!PageLRU(page)); |
8d438f96 | 892 | |
4f98a2fe | 893 | switch (__isolate_lru_page(page, mode, file)) { |
5ad333eb AW |
894 | case 0: |
895 | list_move(&page->lru, dst); | |
2ffebca6 | 896 | mem_cgroup_del_lru(page); |
7c8ee9a8 | 897 | nr_taken++; |
5ad333eb AW |
898 | break; |
899 | ||
900 | case -EBUSY: | |
901 | /* else it is being freed elsewhere */ | |
902 | list_move(&page->lru, src); | |
2ffebca6 | 903 | mem_cgroup_rotate_lru_list(page, page_lru(page)); |
5ad333eb | 904 | continue; |
46453a6e | 905 | |
5ad333eb AW |
906 | default: |
907 | BUG(); | |
908 | } | |
909 | ||
910 | if (!order) | |
911 | continue; | |
912 | ||
913 | /* | |
914 | * Attempt to take all pages in the order aligned region | |
915 | * surrounding the tag page. Only take those pages of | |
916 | * the same active state as that tag page. We may safely | |
917 | * round the target page pfn down to the requested order | |
918 | * as the mem_map is guarenteed valid out to MAX_ORDER, | |
919 | * where that page is in a different zone we will detect | |
920 | * it from its zone id and abort this block scan. | |
921 | */ | |
922 | zone_id = page_zone_id(page); | |
923 | page_pfn = page_to_pfn(page); | |
924 | pfn = page_pfn & ~((1 << order) - 1); | |
925 | end_pfn = pfn + (1 << order); | |
926 | for (; pfn < end_pfn; pfn++) { | |
927 | struct page *cursor_page; | |
928 | ||
929 | /* The target page is in the block, ignore it. */ | |
930 | if (unlikely(pfn == page_pfn)) | |
931 | continue; | |
932 | ||
933 | /* Avoid holes within the zone. */ | |
934 | if (unlikely(!pfn_valid_within(pfn))) | |
935 | break; | |
936 | ||
937 | cursor_page = pfn_to_page(pfn); | |
4f98a2fe | 938 | |
5ad333eb AW |
939 | /* Check that we have not crossed a zone boundary. */ |
940 | if (unlikely(page_zone_id(cursor_page) != zone_id)) | |
941 | continue; | |
de2e7567 MK |
942 | |
943 | /* | |
944 | * If we don't have enough swap space, reclaiming of | |
945 | * anon page which don't already have a swap slot is | |
946 | * pointless. | |
947 | */ | |
948 | if (nr_swap_pages <= 0 && PageAnon(cursor_page) && | |
949 | !PageSwapCache(cursor_page)) | |
950 | continue; | |
951 | ||
ee993b13 | 952 | if (__isolate_lru_page(cursor_page, mode, file) == 0) { |
5ad333eb | 953 | list_move(&cursor_page->lru, dst); |
cb4cbcf6 | 954 | mem_cgroup_del_lru(cursor_page); |
5ad333eb AW |
955 | nr_taken++; |
956 | scan++; | |
5ad333eb AW |
957 | } |
958 | } | |
1da177e4 LT |
959 | } |
960 | ||
961 | *scanned = scan; | |
962 | return nr_taken; | |
963 | } | |
964 | ||
66e1707b BS |
965 | static unsigned long isolate_pages_global(unsigned long nr, |
966 | struct list_head *dst, | |
967 | unsigned long *scanned, int order, | |
968 | int mode, struct zone *z, | |
969 | struct mem_cgroup *mem_cont, | |
4f98a2fe | 970 | int active, int file) |
66e1707b | 971 | { |
4f98a2fe | 972 | int lru = LRU_BASE; |
66e1707b | 973 | if (active) |
4f98a2fe RR |
974 | lru += LRU_ACTIVE; |
975 | if (file) | |
976 | lru += LRU_FILE; | |
977 | return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order, | |
b7c46d15 | 978 | mode, file); |
66e1707b BS |
979 | } |
980 | ||
5ad333eb AW |
981 | /* |
982 | * clear_active_flags() is a helper for shrink_active_list(), clearing | |
983 | * any active bits from the pages in the list. | |
984 | */ | |
4f98a2fe RR |
985 | static unsigned long clear_active_flags(struct list_head *page_list, |
986 | unsigned int *count) | |
5ad333eb AW |
987 | { |
988 | int nr_active = 0; | |
4f98a2fe | 989 | int lru; |
5ad333eb AW |
990 | struct page *page; |
991 | ||
4f98a2fe | 992 | list_for_each_entry(page, page_list, lru) { |
401a8e1c | 993 | lru = page_lru_base_type(page); |
5ad333eb | 994 | if (PageActive(page)) { |
4f98a2fe | 995 | lru += LRU_ACTIVE; |
5ad333eb AW |
996 | ClearPageActive(page); |
997 | nr_active++; | |
998 | } | |
4f98a2fe RR |
999 | count[lru]++; |
1000 | } | |
5ad333eb AW |
1001 | |
1002 | return nr_active; | |
1003 | } | |
1004 | ||
62695a84 NP |
1005 | /** |
1006 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1007 | * @page: page to isolate from its LRU list | |
1008 | * | |
1009 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1010 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1011 | * | |
1012 | * Returns 0 if the page was removed from an LRU list. | |
1013 | * Returns -EBUSY if the page was not on an LRU list. | |
1014 | * | |
1015 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1016 | * the active list, it will have PageActive set. If it was found on |
1017 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1018 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1019 | * |
1020 | * The vmstat statistic corresponding to the list on which the page was | |
1021 | * found will be decremented. | |
1022 | * | |
1023 | * Restrictions: | |
1024 | * (1) Must be called with an elevated refcount on the page. This is a | |
1025 | * fundamentnal difference from isolate_lru_pages (which is called | |
1026 | * without a stable reference). | |
1027 | * (2) the lru_lock must not be held. | |
1028 | * (3) interrupts must be enabled. | |
1029 | */ | |
1030 | int isolate_lru_page(struct page *page) | |
1031 | { | |
1032 | int ret = -EBUSY; | |
1033 | ||
1034 | if (PageLRU(page)) { | |
1035 | struct zone *zone = page_zone(page); | |
1036 | ||
1037 | spin_lock_irq(&zone->lru_lock); | |
1038 | if (PageLRU(page) && get_page_unless_zero(page)) { | |
894bc310 | 1039 | int lru = page_lru(page); |
62695a84 NP |
1040 | ret = 0; |
1041 | ClearPageLRU(page); | |
4f98a2fe | 1042 | |
4f98a2fe | 1043 | del_page_from_lru_list(zone, page, lru); |
62695a84 NP |
1044 | } |
1045 | spin_unlock_irq(&zone->lru_lock); | |
1046 | } | |
1047 | return ret; | |
1048 | } | |
1049 | ||
35cd7815 RR |
1050 | /* |
1051 | * Are there way too many processes in the direct reclaim path already? | |
1052 | */ | |
1053 | static int too_many_isolated(struct zone *zone, int file, | |
1054 | struct scan_control *sc) | |
1055 | { | |
1056 | unsigned long inactive, isolated; | |
1057 | ||
1058 | if (current_is_kswapd()) | |
1059 | return 0; | |
1060 | ||
1061 | if (!scanning_global_lru(sc)) | |
1062 | return 0; | |
1063 | ||
1064 | if (file) { | |
1065 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1066 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1067 | } else { | |
1068 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1069 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1070 | } | |
1071 | ||
1072 | return isolated > inactive; | |
1073 | } | |
1074 | ||
1da177e4 | 1075 | /* |
1742f19f AM |
1076 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1077 | * of reclaimed pages | |
1da177e4 | 1078 | */ |
1742f19f | 1079 | static unsigned long shrink_inactive_list(unsigned long max_scan, |
33c120ed RR |
1080 | struct zone *zone, struct scan_control *sc, |
1081 | int priority, int file) | |
1da177e4 LT |
1082 | { |
1083 | LIST_HEAD(page_list); | |
1084 | struct pagevec pvec; | |
69e05944 | 1085 | unsigned long nr_scanned = 0; |
05ff5137 | 1086 | unsigned long nr_reclaimed = 0; |
6e901571 | 1087 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); |
78dc583d KM |
1088 | int lumpy_reclaim = 0; |
1089 | ||
35cd7815 RR |
1090 | while (unlikely(too_many_isolated(zone, file, sc))) { |
1091 | congestion_wait(WRITE, HZ/10); | |
1092 | ||
1093 | /* We are about to die and free our memory. Return now. */ | |
1094 | if (fatal_signal_pending(current)) | |
1095 | return SWAP_CLUSTER_MAX; | |
1096 | } | |
1097 | ||
78dc583d KM |
1098 | /* |
1099 | * If we need a large contiguous chunk of memory, or have | |
1100 | * trouble getting a small set of contiguous pages, we | |
1101 | * will reclaim both active and inactive pages. | |
1102 | * | |
1103 | * We use the same threshold as pageout congestion_wait below. | |
1104 | */ | |
1105 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) | |
1106 | lumpy_reclaim = 1; | |
1107 | else if (sc->order && priority < DEF_PRIORITY - 2) | |
1108 | lumpy_reclaim = 1; | |
1da177e4 LT |
1109 | |
1110 | pagevec_init(&pvec, 1); | |
1111 | ||
1112 | lru_add_drain(); | |
1113 | spin_lock_irq(&zone->lru_lock); | |
69e05944 | 1114 | do { |
1da177e4 | 1115 | struct page *page; |
69e05944 AM |
1116 | unsigned long nr_taken; |
1117 | unsigned long nr_scan; | |
1118 | unsigned long nr_freed; | |
5ad333eb | 1119 | unsigned long nr_active; |
4f98a2fe | 1120 | unsigned int count[NR_LRU_LISTS] = { 0, }; |
78dc583d | 1121 | int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE; |
a731286d KM |
1122 | unsigned long nr_anon; |
1123 | unsigned long nr_file; | |
1da177e4 | 1124 | |
66e1707b | 1125 | nr_taken = sc->isolate_pages(sc->swap_cluster_max, |
4f98a2fe RR |
1126 | &page_list, &nr_scan, sc->order, mode, |
1127 | zone, sc->mem_cgroup, 0, file); | |
b35ea17b KM |
1128 | |
1129 | if (scanning_global_lru(sc)) { | |
1130 | zone->pages_scanned += nr_scan; | |
1131 | if (current_is_kswapd()) | |
1132 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, | |
1133 | nr_scan); | |
1134 | else | |
1135 | __count_zone_vm_events(PGSCAN_DIRECT, zone, | |
1136 | nr_scan); | |
1137 | } | |
1138 | ||
1139 | if (nr_taken == 0) | |
1140 | goto done; | |
1141 | ||
4f98a2fe | 1142 | nr_active = clear_active_flags(&page_list, count); |
e9187bdc | 1143 | __count_vm_events(PGDEACTIVATE, nr_active); |
5ad333eb | 1144 | |
4f98a2fe RR |
1145 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, |
1146 | -count[LRU_ACTIVE_FILE]); | |
1147 | __mod_zone_page_state(zone, NR_INACTIVE_FILE, | |
1148 | -count[LRU_INACTIVE_FILE]); | |
1149 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, | |
1150 | -count[LRU_ACTIVE_ANON]); | |
1151 | __mod_zone_page_state(zone, NR_INACTIVE_ANON, | |
1152 | -count[LRU_INACTIVE_ANON]); | |
1153 | ||
a731286d KM |
1154 | nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON]; |
1155 | nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE]; | |
1156 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon); | |
1157 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file); | |
3e2f41f1 KM |
1158 | |
1159 | reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON]; | |
1160 | reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON]; | |
1161 | reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE]; | |
1162 | reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE]; | |
1163 | ||
1da177e4 LT |
1164 | spin_unlock_irq(&zone->lru_lock); |
1165 | ||
69e05944 | 1166 | nr_scanned += nr_scan; |
c661b078 AW |
1167 | nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC); |
1168 | ||
1169 | /* | |
1170 | * If we are direct reclaiming for contiguous pages and we do | |
1171 | * not reclaim everything in the list, try again and wait | |
1172 | * for IO to complete. This will stall high-order allocations | |
1173 | * but that should be acceptable to the caller | |
1174 | */ | |
1175 | if (nr_freed < nr_taken && !current_is_kswapd() && | |
78dc583d | 1176 | lumpy_reclaim) { |
8aa7e847 | 1177 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
c661b078 AW |
1178 | |
1179 | /* | |
1180 | * The attempt at page out may have made some | |
1181 | * of the pages active, mark them inactive again. | |
1182 | */ | |
4f98a2fe | 1183 | nr_active = clear_active_flags(&page_list, count); |
c661b078 AW |
1184 | count_vm_events(PGDEACTIVATE, nr_active); |
1185 | ||
1186 | nr_freed += shrink_page_list(&page_list, sc, | |
1187 | PAGEOUT_IO_SYNC); | |
1188 | } | |
1189 | ||
05ff5137 | 1190 | nr_reclaimed += nr_freed; |
b35ea17b | 1191 | |
a74609fa | 1192 | local_irq_disable(); |
b35ea17b | 1193 | if (current_is_kswapd()) |
f8891e5e | 1194 | __count_vm_events(KSWAPD_STEAL, nr_freed); |
918d3f90 | 1195 | __count_zone_vm_events(PGSTEAL, zone, nr_freed); |
a74609fa NP |
1196 | |
1197 | spin_lock(&zone->lru_lock); | |
1da177e4 LT |
1198 | /* |
1199 | * Put back any unfreeable pages. | |
1200 | */ | |
1201 | while (!list_empty(&page_list)) { | |
894bc310 | 1202 | int lru; |
1da177e4 | 1203 | page = lru_to_page(&page_list); |
725d704e | 1204 | VM_BUG_ON(PageLRU(page)); |
1da177e4 | 1205 | list_del(&page->lru); |
894bc310 LS |
1206 | if (unlikely(!page_evictable(page, NULL))) { |
1207 | spin_unlock_irq(&zone->lru_lock); | |
1208 | putback_lru_page(page); | |
1209 | spin_lock_irq(&zone->lru_lock); | |
1210 | continue; | |
1211 | } | |
1212 | SetPageLRU(page); | |
1213 | lru = page_lru(page); | |
1214 | add_page_to_lru_list(zone, page, lru); | |
74a1c48f | 1215 | if (is_active_lru(lru)) { |
b7c46d15 | 1216 | int file = is_file_lru(lru); |
6e901571 | 1217 | reclaim_stat->recent_rotated[file]++; |
4f98a2fe | 1218 | } |
1da177e4 LT |
1219 | if (!pagevec_add(&pvec, page)) { |
1220 | spin_unlock_irq(&zone->lru_lock); | |
1221 | __pagevec_release(&pvec); | |
1222 | spin_lock_irq(&zone->lru_lock); | |
1223 | } | |
1224 | } | |
a731286d KM |
1225 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon); |
1226 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file); | |
1227 | ||
69e05944 | 1228 | } while (nr_scanned < max_scan); |
b35ea17b | 1229 | |
1da177e4 | 1230 | done: |
b35ea17b | 1231 | spin_unlock_irq(&zone->lru_lock); |
1da177e4 | 1232 | pagevec_release(&pvec); |
05ff5137 | 1233 | return nr_reclaimed; |
1da177e4 LT |
1234 | } |
1235 | ||
3bb1a852 MB |
1236 | /* |
1237 | * We are about to scan this zone at a certain priority level. If that priority | |
1238 | * level is smaller (ie: more urgent) than the previous priority, then note | |
1239 | * that priority level within the zone. This is done so that when the next | |
1240 | * process comes in to scan this zone, it will immediately start out at this | |
1241 | * priority level rather than having to build up its own scanning priority. | |
1242 | * Here, this priority affects only the reclaim-mapped threshold. | |
1243 | */ | |
1244 | static inline void note_zone_scanning_priority(struct zone *zone, int priority) | |
1245 | { | |
1246 | if (priority < zone->prev_priority) | |
1247 | zone->prev_priority = priority; | |
1248 | } | |
1249 | ||
1da177e4 LT |
1250 | /* |
1251 | * This moves pages from the active list to the inactive list. | |
1252 | * | |
1253 | * We move them the other way if the page is referenced by one or more | |
1254 | * processes, from rmap. | |
1255 | * | |
1256 | * If the pages are mostly unmapped, the processing is fast and it is | |
1257 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1258 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1259 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1260 | * this, so instead we remove the pages from the LRU while processing them. | |
1261 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1262 | * nobody will play with that bit on a non-LRU page. | |
1263 | * | |
1264 | * The downside is that we have to touch page->_count against each page. | |
1265 | * But we had to alter page->flags anyway. | |
1266 | */ | |
1cfb419b | 1267 | |
3eb4140f WF |
1268 | static void move_active_pages_to_lru(struct zone *zone, |
1269 | struct list_head *list, | |
1270 | enum lru_list lru) | |
1271 | { | |
1272 | unsigned long pgmoved = 0; | |
1273 | struct pagevec pvec; | |
1274 | struct page *page; | |
1275 | ||
1276 | pagevec_init(&pvec, 1); | |
1277 | ||
1278 | while (!list_empty(list)) { | |
1279 | page = lru_to_page(list); | |
3eb4140f WF |
1280 | |
1281 | VM_BUG_ON(PageLRU(page)); | |
1282 | SetPageLRU(page); | |
1283 | ||
3eb4140f WF |
1284 | list_move(&page->lru, &zone->lru[lru].list); |
1285 | mem_cgroup_add_lru_list(page, lru); | |
1286 | pgmoved++; | |
1287 | ||
1288 | if (!pagevec_add(&pvec, page) || list_empty(list)) { | |
1289 | spin_unlock_irq(&zone->lru_lock); | |
1290 | if (buffer_heads_over_limit) | |
1291 | pagevec_strip(&pvec); | |
1292 | __pagevec_release(&pvec); | |
1293 | spin_lock_irq(&zone->lru_lock); | |
1294 | } | |
1295 | } | |
1296 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1297 | if (!is_active_lru(lru)) | |
1298 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1299 | } | |
1cfb419b | 1300 | |
1742f19f | 1301 | static void shrink_active_list(unsigned long nr_pages, struct zone *zone, |
4f98a2fe | 1302 | struct scan_control *sc, int priority, int file) |
1da177e4 | 1303 | { |
44c241f1 | 1304 | unsigned long nr_taken; |
69e05944 | 1305 | unsigned long pgscanned; |
6fe6b7e3 | 1306 | unsigned long vm_flags; |
1da177e4 | 1307 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1308 | LIST_HEAD(l_active); |
b69408e8 | 1309 | LIST_HEAD(l_inactive); |
1da177e4 | 1310 | struct page *page; |
6e901571 | 1311 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); |
44c241f1 | 1312 | unsigned long nr_rotated = 0; |
1da177e4 LT |
1313 | |
1314 | lru_add_drain(); | |
1315 | spin_lock_irq(&zone->lru_lock); | |
44c241f1 | 1316 | nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order, |
66e1707b | 1317 | ISOLATE_ACTIVE, zone, |
4f98a2fe | 1318 | sc->mem_cgroup, 1, file); |
1cfb419b KH |
1319 | /* |
1320 | * zone->pages_scanned is used for detect zone's oom | |
1321 | * mem_cgroup remembers nr_scan by itself. | |
1322 | */ | |
e72e2bd6 | 1323 | if (scanning_global_lru(sc)) { |
1cfb419b | 1324 | zone->pages_scanned += pgscanned; |
4f98a2fe | 1325 | } |
b7c46d15 | 1326 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1327 | |
3eb4140f | 1328 | __count_zone_vm_events(PGREFILL, zone, pgscanned); |
4f98a2fe | 1329 | if (file) |
44c241f1 | 1330 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken); |
4f98a2fe | 1331 | else |
44c241f1 | 1332 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken); |
a731286d | 1333 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
1da177e4 LT |
1334 | spin_unlock_irq(&zone->lru_lock); |
1335 | ||
1da177e4 LT |
1336 | while (!list_empty(&l_hold)) { |
1337 | cond_resched(); | |
1338 | page = lru_to_page(&l_hold); | |
1339 | list_del(&page->lru); | |
7e9cd484 | 1340 | |
894bc310 LS |
1341 | if (unlikely(!page_evictable(page, NULL))) { |
1342 | putback_lru_page(page); | |
1343 | continue; | |
1344 | } | |
1345 | ||
7e9cd484 RR |
1346 | /* page_referenced clears PageReferenced */ |
1347 | if (page_mapping_inuse(page) && | |
8cab4754 | 1348 | page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) { |
44c241f1 | 1349 | nr_rotated++; |
8cab4754 WF |
1350 | /* |
1351 | * Identify referenced, file-backed active pages and | |
1352 | * give them one more trip around the active list. So | |
1353 | * that executable code get better chances to stay in | |
1354 | * memory under moderate memory pressure. Anon pages | |
1355 | * are not likely to be evicted by use-once streaming | |
1356 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1357 | * so we ignore them here. | |
1358 | */ | |
1359 | if ((vm_flags & VM_EXEC) && !PageAnon(page)) { | |
1360 | list_add(&page->lru, &l_active); | |
1361 | continue; | |
1362 | } | |
1363 | } | |
7e9cd484 | 1364 | |
5205e56e | 1365 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1366 | list_add(&page->lru, &l_inactive); |
1367 | } | |
1368 | ||
b555749a | 1369 | /* |
8cab4754 | 1370 | * Move pages back to the lru list. |
b555749a | 1371 | */ |
2a1dc509 | 1372 | spin_lock_irq(&zone->lru_lock); |
556adecb | 1373 | /* |
8cab4754 WF |
1374 | * Count referenced pages from currently used mappings as rotated, |
1375 | * even though only some of them are actually re-activated. This | |
1376 | * helps balance scan pressure between file and anonymous pages in | |
1377 | * get_scan_ratio. | |
7e9cd484 | 1378 | */ |
b7c46d15 | 1379 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1380 | |
3eb4140f WF |
1381 | move_active_pages_to_lru(zone, &l_active, |
1382 | LRU_ACTIVE + file * LRU_FILE); | |
1383 | move_active_pages_to_lru(zone, &l_inactive, | |
1384 | LRU_BASE + file * LRU_FILE); | |
a731286d | 1385 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
f8891e5e | 1386 | spin_unlock_irq(&zone->lru_lock); |
1da177e4 LT |
1387 | } |
1388 | ||
14797e23 | 1389 | static int inactive_anon_is_low_global(struct zone *zone) |
f89eb90e KM |
1390 | { |
1391 | unsigned long active, inactive; | |
1392 | ||
1393 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1394 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1395 | ||
1396 | if (inactive * zone->inactive_ratio < active) | |
1397 | return 1; | |
1398 | ||
1399 | return 0; | |
1400 | } | |
1401 | ||
14797e23 KM |
1402 | /** |
1403 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
1404 | * @zone: zone to check | |
1405 | * @sc: scan control of this context | |
1406 | * | |
1407 | * Returns true if the zone does not have enough inactive anon pages, | |
1408 | * meaning some active anon pages need to be deactivated. | |
1409 | */ | |
1410 | static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc) | |
1411 | { | |
1412 | int low; | |
1413 | ||
e72e2bd6 | 1414 | if (scanning_global_lru(sc)) |
14797e23 KM |
1415 | low = inactive_anon_is_low_global(zone); |
1416 | else | |
c772be93 | 1417 | low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup); |
14797e23 KM |
1418 | return low; |
1419 | } | |
1420 | ||
56e49d21 RR |
1421 | static int inactive_file_is_low_global(struct zone *zone) |
1422 | { | |
1423 | unsigned long active, inactive; | |
1424 | ||
1425 | active = zone_page_state(zone, NR_ACTIVE_FILE); | |
1426 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1427 | ||
1428 | return (active > inactive); | |
1429 | } | |
1430 | ||
1431 | /** | |
1432 | * inactive_file_is_low - check if file pages need to be deactivated | |
1433 | * @zone: zone to check | |
1434 | * @sc: scan control of this context | |
1435 | * | |
1436 | * When the system is doing streaming IO, memory pressure here | |
1437 | * ensures that active file pages get deactivated, until more | |
1438 | * than half of the file pages are on the inactive list. | |
1439 | * | |
1440 | * Once we get to that situation, protect the system's working | |
1441 | * set from being evicted by disabling active file page aging. | |
1442 | * | |
1443 | * This uses a different ratio than the anonymous pages, because | |
1444 | * the page cache uses a use-once replacement algorithm. | |
1445 | */ | |
1446 | static int inactive_file_is_low(struct zone *zone, struct scan_control *sc) | |
1447 | { | |
1448 | int low; | |
1449 | ||
1450 | if (scanning_global_lru(sc)) | |
1451 | low = inactive_file_is_low_global(zone); | |
1452 | else | |
1453 | low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup); | |
1454 | return low; | |
1455 | } | |
1456 | ||
4f98a2fe | 1457 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
b69408e8 CL |
1458 | struct zone *zone, struct scan_control *sc, int priority) |
1459 | { | |
4f98a2fe RR |
1460 | int file = is_file_lru(lru); |
1461 | ||
56e49d21 | 1462 | if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) { |
556adecb RR |
1463 | shrink_active_list(nr_to_scan, zone, sc, priority, file); |
1464 | return 0; | |
1465 | } | |
1466 | ||
14797e23 | 1467 | if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) { |
4f98a2fe | 1468 | shrink_active_list(nr_to_scan, zone, sc, priority, file); |
b69408e8 CL |
1469 | return 0; |
1470 | } | |
33c120ed | 1471 | return shrink_inactive_list(nr_to_scan, zone, sc, priority, file); |
4f98a2fe RR |
1472 | } |
1473 | ||
1474 | /* | |
1475 | * Determine how aggressively the anon and file LRU lists should be | |
1476 | * scanned. The relative value of each set of LRU lists is determined | |
1477 | * by looking at the fraction of the pages scanned we did rotate back | |
1478 | * onto the active list instead of evict. | |
1479 | * | |
1480 | * percent[0] specifies how much pressure to put on ram/swap backed | |
1481 | * memory, while percent[1] determines pressure on the file LRUs. | |
1482 | */ | |
1483 | static void get_scan_ratio(struct zone *zone, struct scan_control *sc, | |
1484 | unsigned long *percent) | |
1485 | { | |
1486 | unsigned long anon, file, free; | |
1487 | unsigned long anon_prio, file_prio; | |
1488 | unsigned long ap, fp; | |
6e901571 | 1489 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); |
4f98a2fe | 1490 | |
0b217676 VL |
1491 | anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) + |
1492 | zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON); | |
1493 | file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) + | |
1494 | zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE); | |
b962716b | 1495 | |
e72e2bd6 | 1496 | if (scanning_global_lru(sc)) { |
eeee9a8c KM |
1497 | free = zone_page_state(zone, NR_FREE_PAGES); |
1498 | /* If we have very few page cache pages, | |
1499 | force-scan anon pages. */ | |
41858966 | 1500 | if (unlikely(file + free <= high_wmark_pages(zone))) { |
eeee9a8c KM |
1501 | percent[0] = 100; |
1502 | percent[1] = 0; | |
1503 | return; | |
1504 | } | |
4f98a2fe RR |
1505 | } |
1506 | ||
1507 | /* | |
1508 | * OK, so we have swap space and a fair amount of page cache | |
1509 | * pages. We use the recently rotated / recently scanned | |
1510 | * ratios to determine how valuable each cache is. | |
1511 | * | |
1512 | * Because workloads change over time (and to avoid overflow) | |
1513 | * we keep these statistics as a floating average, which ends | |
1514 | * up weighing recent references more than old ones. | |
1515 | * | |
1516 | * anon in [0], file in [1] | |
1517 | */ | |
6e901571 | 1518 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
4f98a2fe | 1519 | spin_lock_irq(&zone->lru_lock); |
6e901571 KM |
1520 | reclaim_stat->recent_scanned[0] /= 2; |
1521 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
1522 | spin_unlock_irq(&zone->lru_lock); |
1523 | } | |
1524 | ||
6e901571 | 1525 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
4f98a2fe | 1526 | spin_lock_irq(&zone->lru_lock); |
6e901571 KM |
1527 | reclaim_stat->recent_scanned[1] /= 2; |
1528 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
1529 | spin_unlock_irq(&zone->lru_lock); |
1530 | } | |
1531 | ||
1532 | /* | |
1533 | * With swappiness at 100, anonymous and file have the same priority. | |
1534 | * This scanning priority is essentially the inverse of IO cost. | |
1535 | */ | |
1536 | anon_prio = sc->swappiness; | |
1537 | file_prio = 200 - sc->swappiness; | |
1538 | ||
1539 | /* | |
00d8089c RR |
1540 | * The amount of pressure on anon vs file pages is inversely |
1541 | * proportional to the fraction of recently scanned pages on | |
1542 | * each list that were recently referenced and in active use. | |
4f98a2fe | 1543 | */ |
6e901571 KM |
1544 | ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1); |
1545 | ap /= reclaim_stat->recent_rotated[0] + 1; | |
4f98a2fe | 1546 | |
6e901571 KM |
1547 | fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1); |
1548 | fp /= reclaim_stat->recent_rotated[1] + 1; | |
4f98a2fe RR |
1549 | |
1550 | /* Normalize to percentages */ | |
1551 | percent[0] = 100 * ap / (ap + fp + 1); | |
1552 | percent[1] = 100 - percent[0]; | |
b69408e8 CL |
1553 | } |
1554 | ||
6e08a369 WF |
1555 | /* |
1556 | * Smallish @nr_to_scan's are deposited in @nr_saved_scan, | |
1557 | * until we collected @swap_cluster_max pages to scan. | |
1558 | */ | |
1559 | static unsigned long nr_scan_try_batch(unsigned long nr_to_scan, | |
1560 | unsigned long *nr_saved_scan, | |
1561 | unsigned long swap_cluster_max) | |
1562 | { | |
1563 | unsigned long nr; | |
1564 | ||
1565 | *nr_saved_scan += nr_to_scan; | |
1566 | nr = *nr_saved_scan; | |
1567 | ||
1568 | if (nr >= swap_cluster_max) | |
1569 | *nr_saved_scan = 0; | |
1570 | else | |
1571 | nr = 0; | |
1572 | ||
1573 | return nr; | |
1574 | } | |
4f98a2fe | 1575 | |
1da177e4 LT |
1576 | /* |
1577 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1578 | */ | |
a79311c1 | 1579 | static void shrink_zone(int priority, struct zone *zone, |
05ff5137 | 1580 | struct scan_control *sc) |
1da177e4 | 1581 | { |
b69408e8 | 1582 | unsigned long nr[NR_LRU_LISTS]; |
8695949a | 1583 | unsigned long nr_to_scan; |
4f98a2fe | 1584 | unsigned long percent[2]; /* anon @ 0; file @ 1 */ |
b69408e8 | 1585 | enum lru_list l; |
01dbe5c9 KM |
1586 | unsigned long nr_reclaimed = sc->nr_reclaimed; |
1587 | unsigned long swap_cluster_max = sc->swap_cluster_max; | |
f8629631 | 1588 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); |
9198e96c | 1589 | int noswap = 0; |
1da177e4 | 1590 | |
9198e96c DN |
1591 | /* If we have no swap space, do not bother scanning anon pages. */ |
1592 | if (!sc->may_swap || (nr_swap_pages <= 0)) { | |
1593 | noswap = 1; | |
1594 | percent[0] = 0; | |
1595 | percent[1] = 100; | |
1596 | } else | |
1597 | get_scan_ratio(zone, sc, percent); | |
4f98a2fe | 1598 | |
894bc310 | 1599 | for_each_evictable_lru(l) { |
9439c1c9 | 1600 | int file = is_file_lru(l); |
8713e012 | 1601 | unsigned long scan; |
e0f79b8f | 1602 | |
0b217676 | 1603 | scan = zone_nr_lru_pages(zone, sc, l); |
9198e96c | 1604 | if (priority || noswap) { |
9439c1c9 KM |
1605 | scan >>= priority; |
1606 | scan = (scan * percent[file]) / 100; | |
1607 | } | |
f8629631 WF |
1608 | nr[l] = nr_scan_try_batch(scan, |
1609 | &reclaim_stat->nr_saved_scan[l], | |
1610 | swap_cluster_max); | |
1cfb419b | 1611 | } |
1da177e4 | 1612 | |
556adecb RR |
1613 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || |
1614 | nr[LRU_INACTIVE_FILE]) { | |
894bc310 | 1615 | for_each_evictable_lru(l) { |
b69408e8 | 1616 | if (nr[l]) { |
01dbe5c9 | 1617 | nr_to_scan = min(nr[l], swap_cluster_max); |
b69408e8 | 1618 | nr[l] -= nr_to_scan; |
1da177e4 | 1619 | |
01dbe5c9 KM |
1620 | nr_reclaimed += shrink_list(l, nr_to_scan, |
1621 | zone, sc, priority); | |
b69408e8 | 1622 | } |
1da177e4 | 1623 | } |
a79311c1 RR |
1624 | /* |
1625 | * On large memory systems, scan >> priority can become | |
1626 | * really large. This is fine for the starting priority; | |
1627 | * we want to put equal scanning pressure on each zone. | |
1628 | * However, if the VM has a harder time of freeing pages, | |
1629 | * with multiple processes reclaiming pages, the total | |
1630 | * freeing target can get unreasonably large. | |
1631 | */ | |
01dbe5c9 | 1632 | if (nr_reclaimed > swap_cluster_max && |
a79311c1 RR |
1633 | priority < DEF_PRIORITY && !current_is_kswapd()) |
1634 | break; | |
1da177e4 LT |
1635 | } |
1636 | ||
01dbe5c9 KM |
1637 | sc->nr_reclaimed = nr_reclaimed; |
1638 | ||
556adecb RR |
1639 | /* |
1640 | * Even if we did not try to evict anon pages at all, we want to | |
1641 | * rebalance the anon lru active/inactive ratio. | |
1642 | */ | |
69c85481 | 1643 | if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0) |
556adecb RR |
1644 | shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0); |
1645 | ||
232ea4d6 | 1646 | throttle_vm_writeout(sc->gfp_mask); |
1da177e4 LT |
1647 | } |
1648 | ||
1649 | /* | |
1650 | * This is the direct reclaim path, for page-allocating processes. We only | |
1651 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
1652 | * request. | |
1653 | * | |
41858966 MG |
1654 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
1655 | * Because: | |
1da177e4 LT |
1656 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
1657 | * allocation or | |
41858966 MG |
1658 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
1659 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
1660 | * zone defense algorithm. | |
1da177e4 | 1661 | * |
1da177e4 LT |
1662 | * If a zone is deemed to be full of pinned pages then just give it a light |
1663 | * scan then give up on it. | |
1664 | */ | |
a79311c1 | 1665 | static void shrink_zones(int priority, struct zonelist *zonelist, |
05ff5137 | 1666 | struct scan_control *sc) |
1da177e4 | 1667 | { |
54a6eb5c | 1668 | enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask); |
dd1a239f | 1669 | struct zoneref *z; |
54a6eb5c | 1670 | struct zone *zone; |
1cfb419b | 1671 | |
408d8544 | 1672 | sc->all_unreclaimable = 1; |
327c0e96 KH |
1673 | for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, |
1674 | sc->nodemask) { | |
f3fe6512 | 1675 | if (!populated_zone(zone)) |
1da177e4 | 1676 | continue; |
1cfb419b KH |
1677 | /* |
1678 | * Take care memory controller reclaiming has small influence | |
1679 | * to global LRU. | |
1680 | */ | |
e72e2bd6 | 1681 | if (scanning_global_lru(sc)) { |
1cfb419b KH |
1682 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1683 | continue; | |
1684 | note_zone_scanning_priority(zone, priority); | |
1da177e4 | 1685 | |
1cfb419b KH |
1686 | if (zone_is_all_unreclaimable(zone) && |
1687 | priority != DEF_PRIORITY) | |
1688 | continue; /* Let kswapd poll it */ | |
1689 | sc->all_unreclaimable = 0; | |
1690 | } else { | |
1691 | /* | |
1692 | * Ignore cpuset limitation here. We just want to reduce | |
1693 | * # of used pages by us regardless of memory shortage. | |
1694 | */ | |
1695 | sc->all_unreclaimable = 0; | |
1696 | mem_cgroup_note_reclaim_priority(sc->mem_cgroup, | |
1697 | priority); | |
1698 | } | |
408d8544 | 1699 | |
a79311c1 | 1700 | shrink_zone(priority, zone, sc); |
1da177e4 LT |
1701 | } |
1702 | } | |
4f98a2fe | 1703 | |
1da177e4 LT |
1704 | /* |
1705 | * This is the main entry point to direct page reclaim. | |
1706 | * | |
1707 | * If a full scan of the inactive list fails to free enough memory then we | |
1708 | * are "out of memory" and something needs to be killed. | |
1709 | * | |
1710 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
1711 | * high - the zone may be full of dirty or under-writeback pages, which this | |
1712 | * caller can't do much about. We kick pdflush and take explicit naps in the | |
1713 | * hope that some of these pages can be written. But if the allocating task | |
1714 | * holds filesystem locks which prevent writeout this might not work, and the | |
1715 | * allocation attempt will fail. | |
a41f24ea NA |
1716 | * |
1717 | * returns: 0, if no pages reclaimed | |
1718 | * else, the number of pages reclaimed | |
1da177e4 | 1719 | */ |
dac1d27b | 1720 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
dd1a239f | 1721 | struct scan_control *sc) |
1da177e4 LT |
1722 | { |
1723 | int priority; | |
c700be3d | 1724 | unsigned long ret = 0; |
69e05944 | 1725 | unsigned long total_scanned = 0; |
1da177e4 | 1726 | struct reclaim_state *reclaim_state = current->reclaim_state; |
1da177e4 | 1727 | unsigned long lru_pages = 0; |
dd1a239f | 1728 | struct zoneref *z; |
54a6eb5c | 1729 | struct zone *zone; |
dd1a239f | 1730 | enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask); |
1da177e4 | 1731 | |
873b4771 KK |
1732 | delayacct_freepages_start(); |
1733 | ||
e72e2bd6 | 1734 | if (scanning_global_lru(sc)) |
1cfb419b KH |
1735 | count_vm_event(ALLOCSTALL); |
1736 | /* | |
1737 | * mem_cgroup will not do shrink_slab. | |
1738 | */ | |
e72e2bd6 | 1739 | if (scanning_global_lru(sc)) { |
54a6eb5c | 1740 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
1da177e4 | 1741 | |
1cfb419b KH |
1742 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1743 | continue; | |
1da177e4 | 1744 | |
adea02a1 | 1745 | lru_pages += zone_reclaimable_pages(zone); |
1cfb419b | 1746 | } |
1da177e4 LT |
1747 | } |
1748 | ||
1749 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { | |
66e1707b | 1750 | sc->nr_scanned = 0; |
f7b7fd8f RR |
1751 | if (!priority) |
1752 | disable_swap_token(); | |
a79311c1 | 1753 | shrink_zones(priority, zonelist, sc); |
66e1707b BS |
1754 | /* |
1755 | * Don't shrink slabs when reclaiming memory from | |
1756 | * over limit cgroups | |
1757 | */ | |
e72e2bd6 | 1758 | if (scanning_global_lru(sc)) { |
dd1a239f | 1759 | shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages); |
91a45470 | 1760 | if (reclaim_state) { |
a79311c1 | 1761 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
91a45470 KH |
1762 | reclaim_state->reclaimed_slab = 0; |
1763 | } | |
1da177e4 | 1764 | } |
66e1707b | 1765 | total_scanned += sc->nr_scanned; |
a79311c1 RR |
1766 | if (sc->nr_reclaimed >= sc->swap_cluster_max) { |
1767 | ret = sc->nr_reclaimed; | |
1da177e4 LT |
1768 | goto out; |
1769 | } | |
1770 | ||
1771 | /* | |
1772 | * Try to write back as many pages as we just scanned. This | |
1773 | * tends to cause slow streaming writers to write data to the | |
1774 | * disk smoothly, at the dirtying rate, which is nice. But | |
1775 | * that's undesirable in laptop mode, where we *want* lumpy | |
1776 | * writeout. So in laptop mode, write out the whole world. | |
1777 | */ | |
66e1707b BS |
1778 | if (total_scanned > sc->swap_cluster_max + |
1779 | sc->swap_cluster_max / 2) { | |
03ba3782 | 1780 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned); |
66e1707b | 1781 | sc->may_writepage = 1; |
1da177e4 LT |
1782 | } |
1783 | ||
1784 | /* Take a nap, wait for some writeback to complete */ | |
4dd4b920 | 1785 | if (sc->nr_scanned && priority < DEF_PRIORITY - 2) |
8aa7e847 | 1786 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
1da177e4 | 1787 | } |
87547ee9 | 1788 | /* top priority shrink_zones still had more to do? don't OOM, then */ |
e72e2bd6 | 1789 | if (!sc->all_unreclaimable && scanning_global_lru(sc)) |
a79311c1 | 1790 | ret = sc->nr_reclaimed; |
1da177e4 | 1791 | out: |
3bb1a852 MB |
1792 | /* |
1793 | * Now that we've scanned all the zones at this priority level, note | |
1794 | * that level within the zone so that the next thread which performs | |
1795 | * scanning of this zone will immediately start out at this priority | |
1796 | * level. This affects only the decision whether or not to bring | |
1797 | * mapped pages onto the inactive list. | |
1798 | */ | |
1799 | if (priority < 0) | |
1800 | priority = 0; | |
1da177e4 | 1801 | |
e72e2bd6 | 1802 | if (scanning_global_lru(sc)) { |
54a6eb5c | 1803 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
1cfb419b KH |
1804 | |
1805 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
1806 | continue; | |
1807 | ||
1808 | zone->prev_priority = priority; | |
1809 | } | |
1810 | } else | |
1811 | mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority); | |
1da177e4 | 1812 | |
873b4771 KK |
1813 | delayacct_freepages_end(); |
1814 | ||
1da177e4 LT |
1815 | return ret; |
1816 | } | |
1817 | ||
dac1d27b | 1818 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 1819 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b BS |
1820 | { |
1821 | struct scan_control sc = { | |
1822 | .gfp_mask = gfp_mask, | |
1823 | .may_writepage = !laptop_mode, | |
1824 | .swap_cluster_max = SWAP_CLUSTER_MAX, | |
a6dc60f8 | 1825 | .may_unmap = 1, |
2e2e4259 | 1826 | .may_swap = 1, |
66e1707b BS |
1827 | .swappiness = vm_swappiness, |
1828 | .order = order, | |
1829 | .mem_cgroup = NULL, | |
1830 | .isolate_pages = isolate_pages_global, | |
327c0e96 | 1831 | .nodemask = nodemask, |
66e1707b BS |
1832 | }; |
1833 | ||
dd1a239f | 1834 | return do_try_to_free_pages(zonelist, &sc); |
66e1707b BS |
1835 | } |
1836 | ||
00f0b825 | 1837 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
66e1707b | 1838 | |
4e416953 BS |
1839 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem, |
1840 | gfp_t gfp_mask, bool noswap, | |
1841 | unsigned int swappiness, | |
1842 | struct zone *zone, int nid) | |
1843 | { | |
1844 | struct scan_control sc = { | |
1845 | .may_writepage = !laptop_mode, | |
1846 | .may_unmap = 1, | |
1847 | .may_swap = !noswap, | |
1848 | .swap_cluster_max = SWAP_CLUSTER_MAX, | |
1849 | .swappiness = swappiness, | |
1850 | .order = 0, | |
1851 | .mem_cgroup = mem, | |
1852 | .isolate_pages = mem_cgroup_isolate_pages, | |
1853 | }; | |
1854 | nodemask_t nm = nodemask_of_node(nid); | |
1855 | ||
1856 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | | |
1857 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
1858 | sc.nodemask = &nm; | |
1859 | sc.nr_reclaimed = 0; | |
1860 | sc.nr_scanned = 0; | |
1861 | /* | |
1862 | * NOTE: Although we can get the priority field, using it | |
1863 | * here is not a good idea, since it limits the pages we can scan. | |
1864 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
1865 | * will pick up pages from other mem cgroup's as well. We hack | |
1866 | * the priority and make it zero. | |
1867 | */ | |
1868 | shrink_zone(0, zone, &sc); | |
1869 | return sc.nr_reclaimed; | |
1870 | } | |
1871 | ||
e1a1cd59 | 1872 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont, |
a7885eb8 KM |
1873 | gfp_t gfp_mask, |
1874 | bool noswap, | |
1875 | unsigned int swappiness) | |
66e1707b | 1876 | { |
4e416953 | 1877 | struct zonelist *zonelist; |
66e1707b | 1878 | struct scan_control sc = { |
66e1707b | 1879 | .may_writepage = !laptop_mode, |
a6dc60f8 | 1880 | .may_unmap = 1, |
2e2e4259 | 1881 | .may_swap = !noswap, |
66e1707b | 1882 | .swap_cluster_max = SWAP_CLUSTER_MAX, |
a7885eb8 | 1883 | .swappiness = swappiness, |
66e1707b BS |
1884 | .order = 0, |
1885 | .mem_cgroup = mem_cont, | |
1886 | .isolate_pages = mem_cgroup_isolate_pages, | |
327c0e96 | 1887 | .nodemask = NULL, /* we don't care the placement */ |
66e1707b | 1888 | }; |
66e1707b | 1889 | |
dd1a239f MG |
1890 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
1891 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
1892 | zonelist = NODE_DATA(numa_node_id())->node_zonelists; | |
1893 | return do_try_to_free_pages(zonelist, &sc); | |
66e1707b BS |
1894 | } |
1895 | #endif | |
1896 | ||
1da177e4 LT |
1897 | /* |
1898 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 1899 | * they are all at high_wmark_pages(zone). |
1da177e4 | 1900 | * |
1da177e4 LT |
1901 | * Returns the number of pages which were actually freed. |
1902 | * | |
1903 | * There is special handling here for zones which are full of pinned pages. | |
1904 | * This can happen if the pages are all mlocked, or if they are all used by | |
1905 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
1906 | * What we do is to detect the case where all pages in the zone have been | |
1907 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
1908 | * dead and from now on, only perform a short scan. Basically we're polling | |
1909 | * the zone for when the problem goes away. | |
1910 | * | |
1911 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
1912 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
1913 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
1914 | * lower zones regardless of the number of free pages in the lower zones. This | |
1915 | * interoperates with the page allocator fallback scheme to ensure that aging | |
1916 | * of pages is balanced across the zones. | |
1da177e4 | 1917 | */ |
d6277db4 | 1918 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order) |
1da177e4 | 1919 | { |
1da177e4 LT |
1920 | int all_zones_ok; |
1921 | int priority; | |
1922 | int i; | |
69e05944 | 1923 | unsigned long total_scanned; |
1da177e4 | 1924 | struct reclaim_state *reclaim_state = current->reclaim_state; |
179e9639 AM |
1925 | struct scan_control sc = { |
1926 | .gfp_mask = GFP_KERNEL, | |
a6dc60f8 | 1927 | .may_unmap = 1, |
2e2e4259 | 1928 | .may_swap = 1, |
d6277db4 RW |
1929 | .swap_cluster_max = SWAP_CLUSTER_MAX, |
1930 | .swappiness = vm_swappiness, | |
5ad333eb | 1931 | .order = order, |
66e1707b BS |
1932 | .mem_cgroup = NULL, |
1933 | .isolate_pages = isolate_pages_global, | |
179e9639 | 1934 | }; |
3bb1a852 MB |
1935 | /* |
1936 | * temp_priority is used to remember the scanning priority at which | |
41858966 MG |
1937 | * this zone was successfully refilled to |
1938 | * free_pages == high_wmark_pages(zone). | |
3bb1a852 MB |
1939 | */ |
1940 | int temp_priority[MAX_NR_ZONES]; | |
1da177e4 LT |
1941 | |
1942 | loop_again: | |
1943 | total_scanned = 0; | |
a79311c1 | 1944 | sc.nr_reclaimed = 0; |
c0bbbc73 | 1945 | sc.may_writepage = !laptop_mode; |
f8891e5e | 1946 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 1947 | |
3bb1a852 MB |
1948 | for (i = 0; i < pgdat->nr_zones; i++) |
1949 | temp_priority[i] = DEF_PRIORITY; | |
1da177e4 LT |
1950 | |
1951 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { | |
1952 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ | |
1953 | unsigned long lru_pages = 0; | |
1954 | ||
f7b7fd8f RR |
1955 | /* The swap token gets in the way of swapout... */ |
1956 | if (!priority) | |
1957 | disable_swap_token(); | |
1958 | ||
1da177e4 LT |
1959 | all_zones_ok = 1; |
1960 | ||
d6277db4 RW |
1961 | /* |
1962 | * Scan in the highmem->dma direction for the highest | |
1963 | * zone which needs scanning | |
1964 | */ | |
1965 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
1966 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 1967 | |
d6277db4 RW |
1968 | if (!populated_zone(zone)) |
1969 | continue; | |
1da177e4 | 1970 | |
e815af95 DR |
1971 | if (zone_is_all_unreclaimable(zone) && |
1972 | priority != DEF_PRIORITY) | |
d6277db4 | 1973 | continue; |
1da177e4 | 1974 | |
556adecb RR |
1975 | /* |
1976 | * Do some background aging of the anon list, to give | |
1977 | * pages a chance to be referenced before reclaiming. | |
1978 | */ | |
14797e23 | 1979 | if (inactive_anon_is_low(zone, &sc)) |
556adecb RR |
1980 | shrink_active_list(SWAP_CLUSTER_MAX, zone, |
1981 | &sc, priority, 0); | |
1982 | ||
41858966 MG |
1983 | if (!zone_watermark_ok(zone, order, |
1984 | high_wmark_pages(zone), 0, 0)) { | |
d6277db4 | 1985 | end_zone = i; |
e1dbeda6 | 1986 | break; |
1da177e4 | 1987 | } |
1da177e4 | 1988 | } |
e1dbeda6 AM |
1989 | if (i < 0) |
1990 | goto out; | |
1991 | ||
1da177e4 LT |
1992 | for (i = 0; i <= end_zone; i++) { |
1993 | struct zone *zone = pgdat->node_zones + i; | |
1994 | ||
adea02a1 | 1995 | lru_pages += zone_reclaimable_pages(zone); |
1da177e4 LT |
1996 | } |
1997 | ||
1998 | /* | |
1999 | * Now scan the zone in the dma->highmem direction, stopping | |
2000 | * at the last zone which needs scanning. | |
2001 | * | |
2002 | * We do this because the page allocator works in the opposite | |
2003 | * direction. This prevents the page allocator from allocating | |
2004 | * pages behind kswapd's direction of progress, which would | |
2005 | * cause too much scanning of the lower zones. | |
2006 | */ | |
2007 | for (i = 0; i <= end_zone; i++) { | |
2008 | struct zone *zone = pgdat->node_zones + i; | |
b15e0905 | 2009 | int nr_slab; |
4e416953 | 2010 | int nid, zid; |
1da177e4 | 2011 | |
f3fe6512 | 2012 | if (!populated_zone(zone)) |
1da177e4 LT |
2013 | continue; |
2014 | ||
e815af95 DR |
2015 | if (zone_is_all_unreclaimable(zone) && |
2016 | priority != DEF_PRIORITY) | |
1da177e4 LT |
2017 | continue; |
2018 | ||
41858966 MG |
2019 | if (!zone_watermark_ok(zone, order, |
2020 | high_wmark_pages(zone), end_zone, 0)) | |
d6277db4 | 2021 | all_zones_ok = 0; |
3bb1a852 | 2022 | temp_priority[i] = priority; |
1da177e4 | 2023 | sc.nr_scanned = 0; |
3bb1a852 | 2024 | note_zone_scanning_priority(zone, priority); |
4e416953 BS |
2025 | |
2026 | nid = pgdat->node_id; | |
2027 | zid = zone_idx(zone); | |
2028 | /* | |
2029 | * Call soft limit reclaim before calling shrink_zone. | |
2030 | * For now we ignore the return value | |
2031 | */ | |
2032 | mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask, | |
2033 | nid, zid); | |
32a4330d RR |
2034 | /* |
2035 | * We put equal pressure on every zone, unless one | |
2036 | * zone has way too many pages free already. | |
2037 | */ | |
41858966 MG |
2038 | if (!zone_watermark_ok(zone, order, |
2039 | 8*high_wmark_pages(zone), end_zone, 0)) | |
a79311c1 | 2040 | shrink_zone(priority, zone, &sc); |
1da177e4 | 2041 | reclaim_state->reclaimed_slab = 0; |
b15e0905 | 2042 | nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, |
2043 | lru_pages); | |
a79311c1 | 2044 | sc.nr_reclaimed += reclaim_state->reclaimed_slab; |
1da177e4 | 2045 | total_scanned += sc.nr_scanned; |
e815af95 | 2046 | if (zone_is_all_unreclaimable(zone)) |
1da177e4 | 2047 | continue; |
b15e0905 | 2048 | if (nr_slab == 0 && zone->pages_scanned >= |
adea02a1 | 2049 | (zone_reclaimable_pages(zone) * 6)) |
e815af95 DR |
2050 | zone_set_flag(zone, |
2051 | ZONE_ALL_UNRECLAIMABLE); | |
1da177e4 LT |
2052 | /* |
2053 | * If we've done a decent amount of scanning and | |
2054 | * the reclaim ratio is low, start doing writepage | |
2055 | * even in laptop mode | |
2056 | */ | |
2057 | if (total_scanned > SWAP_CLUSTER_MAX * 2 && | |
a79311c1 | 2058 | total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) |
1da177e4 LT |
2059 | sc.may_writepage = 1; |
2060 | } | |
1da177e4 LT |
2061 | if (all_zones_ok) |
2062 | break; /* kswapd: all done */ | |
2063 | /* | |
2064 | * OK, kswapd is getting into trouble. Take a nap, then take | |
2065 | * another pass across the zones. | |
2066 | */ | |
4dd4b920 | 2067 | if (total_scanned && priority < DEF_PRIORITY - 2) |
8aa7e847 | 2068 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
1da177e4 LT |
2069 | |
2070 | /* | |
2071 | * We do this so kswapd doesn't build up large priorities for | |
2072 | * example when it is freeing in parallel with allocators. It | |
2073 | * matches the direct reclaim path behaviour in terms of impact | |
2074 | * on zone->*_priority. | |
2075 | */ | |
a79311c1 | 2076 | if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) |
1da177e4 LT |
2077 | break; |
2078 | } | |
2079 | out: | |
3bb1a852 MB |
2080 | /* |
2081 | * Note within each zone the priority level at which this zone was | |
2082 | * brought into a happy state. So that the next thread which scans this | |
2083 | * zone will start out at that priority level. | |
2084 | */ | |
1da177e4 LT |
2085 | for (i = 0; i < pgdat->nr_zones; i++) { |
2086 | struct zone *zone = pgdat->node_zones + i; | |
2087 | ||
3bb1a852 | 2088 | zone->prev_priority = temp_priority[i]; |
1da177e4 LT |
2089 | } |
2090 | if (!all_zones_ok) { | |
2091 | cond_resched(); | |
8357376d RW |
2092 | |
2093 | try_to_freeze(); | |
2094 | ||
73ce02e9 KM |
2095 | /* |
2096 | * Fragmentation may mean that the system cannot be | |
2097 | * rebalanced for high-order allocations in all zones. | |
2098 | * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, | |
2099 | * it means the zones have been fully scanned and are still | |
2100 | * not balanced. For high-order allocations, there is | |
2101 | * little point trying all over again as kswapd may | |
2102 | * infinite loop. | |
2103 | * | |
2104 | * Instead, recheck all watermarks at order-0 as they | |
2105 | * are the most important. If watermarks are ok, kswapd will go | |
2106 | * back to sleep. High-order users can still perform direct | |
2107 | * reclaim if they wish. | |
2108 | */ | |
2109 | if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) | |
2110 | order = sc.order = 0; | |
2111 | ||
1da177e4 LT |
2112 | goto loop_again; |
2113 | } | |
2114 | ||
a79311c1 | 2115 | return sc.nr_reclaimed; |
1da177e4 LT |
2116 | } |
2117 | ||
2118 | /* | |
2119 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 2120 | * from the init process. |
1da177e4 LT |
2121 | * |
2122 | * This basically trickles out pages so that we have _some_ | |
2123 | * free memory available even if there is no other activity | |
2124 | * that frees anything up. This is needed for things like routing | |
2125 | * etc, where we otherwise might have all activity going on in | |
2126 | * asynchronous contexts that cannot page things out. | |
2127 | * | |
2128 | * If there are applications that are active memory-allocators | |
2129 | * (most normal use), this basically shouldn't matter. | |
2130 | */ | |
2131 | static int kswapd(void *p) | |
2132 | { | |
2133 | unsigned long order; | |
2134 | pg_data_t *pgdat = (pg_data_t*)p; | |
2135 | struct task_struct *tsk = current; | |
2136 | DEFINE_WAIT(wait); | |
2137 | struct reclaim_state reclaim_state = { | |
2138 | .reclaimed_slab = 0, | |
2139 | }; | |
a70f7302 | 2140 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 2141 | |
cf40bd16 NP |
2142 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
2143 | ||
174596a0 | 2144 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 2145 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
2146 | current->reclaim_state = &reclaim_state; |
2147 | ||
2148 | /* | |
2149 | * Tell the memory management that we're a "memory allocator", | |
2150 | * and that if we need more memory we should get access to it | |
2151 | * regardless (see "__alloc_pages()"). "kswapd" should | |
2152 | * never get caught in the normal page freeing logic. | |
2153 | * | |
2154 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
2155 | * you need a small amount of memory in order to be able to | |
2156 | * page out something else, and this flag essentially protects | |
2157 | * us from recursively trying to free more memory as we're | |
2158 | * trying to free the first piece of memory in the first place). | |
2159 | */ | |
930d9152 | 2160 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 2161 | set_freezable(); |
1da177e4 LT |
2162 | |
2163 | order = 0; | |
2164 | for ( ; ; ) { | |
2165 | unsigned long new_order; | |
3e1d1d28 | 2166 | |
1da177e4 LT |
2167 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); |
2168 | new_order = pgdat->kswapd_max_order; | |
2169 | pgdat->kswapd_max_order = 0; | |
2170 | if (order < new_order) { | |
2171 | /* | |
2172 | * Don't sleep if someone wants a larger 'order' | |
2173 | * allocation | |
2174 | */ | |
2175 | order = new_order; | |
2176 | } else { | |
b1296cc4 RW |
2177 | if (!freezing(current)) |
2178 | schedule(); | |
2179 | ||
1da177e4 LT |
2180 | order = pgdat->kswapd_max_order; |
2181 | } | |
2182 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2183 | ||
b1296cc4 RW |
2184 | if (!try_to_freeze()) { |
2185 | /* We can speed up thawing tasks if we don't call | |
2186 | * balance_pgdat after returning from the refrigerator | |
2187 | */ | |
2188 | balance_pgdat(pgdat, order); | |
2189 | } | |
1da177e4 LT |
2190 | } |
2191 | return 0; | |
2192 | } | |
2193 | ||
2194 | /* | |
2195 | * A zone is low on free memory, so wake its kswapd task to service it. | |
2196 | */ | |
2197 | void wakeup_kswapd(struct zone *zone, int order) | |
2198 | { | |
2199 | pg_data_t *pgdat; | |
2200 | ||
f3fe6512 | 2201 | if (!populated_zone(zone)) |
1da177e4 LT |
2202 | return; |
2203 | ||
2204 | pgdat = zone->zone_pgdat; | |
41858966 | 2205 | if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0)) |
1da177e4 LT |
2206 | return; |
2207 | if (pgdat->kswapd_max_order < order) | |
2208 | pgdat->kswapd_max_order = order; | |
02a0e53d | 2209 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1da177e4 | 2210 | return; |
8d0986e2 | 2211 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 2212 | return; |
8d0986e2 | 2213 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
2214 | } |
2215 | ||
adea02a1 WF |
2216 | /* |
2217 | * The reclaimable count would be mostly accurate. | |
2218 | * The less reclaimable pages may be | |
2219 | * - mlocked pages, which will be moved to unevictable list when encountered | |
2220 | * - mapped pages, which may require several travels to be reclaimed | |
2221 | * - dirty pages, which is not "instantly" reclaimable | |
2222 | */ | |
2223 | unsigned long global_reclaimable_pages(void) | |
4f98a2fe | 2224 | { |
adea02a1 WF |
2225 | int nr; |
2226 | ||
2227 | nr = global_page_state(NR_ACTIVE_FILE) + | |
2228 | global_page_state(NR_INACTIVE_FILE); | |
2229 | ||
2230 | if (nr_swap_pages > 0) | |
2231 | nr += global_page_state(NR_ACTIVE_ANON) + | |
2232 | global_page_state(NR_INACTIVE_ANON); | |
2233 | ||
2234 | return nr; | |
2235 | } | |
2236 | ||
2237 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
2238 | { | |
2239 | int nr; | |
2240 | ||
2241 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
2242 | zone_page_state(zone, NR_INACTIVE_FILE); | |
2243 | ||
2244 | if (nr_swap_pages > 0) | |
2245 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
2246 | zone_page_state(zone, NR_INACTIVE_ANON); | |
2247 | ||
2248 | return nr; | |
4f98a2fe RR |
2249 | } |
2250 | ||
c6f37f12 | 2251 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 2252 | /* |
d6277db4 | 2253 | * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages |
d979677c | 2254 | * from LRU lists system-wide, for given pass and priority. |
d6277db4 RW |
2255 | * |
2256 | * For pass > 3 we also try to shrink the LRU lists that contain a few pages | |
2257 | */ | |
d979677c | 2258 | static void shrink_all_zones(unsigned long nr_pages, int prio, |
e07aa05b | 2259 | int pass, struct scan_control *sc) |
d6277db4 RW |
2260 | { |
2261 | struct zone *zone; | |
d979677c | 2262 | unsigned long nr_reclaimed = 0; |
f8629631 | 2263 | struct zone_reclaim_stat *reclaim_stat; |
d6277db4 | 2264 | |
ee99c71c | 2265 | for_each_populated_zone(zone) { |
0cb57258 | 2266 | enum lru_list l; |
d6277db4 | 2267 | |
e815af95 | 2268 | if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY) |
d6277db4 RW |
2269 | continue; |
2270 | ||
894bc310 | 2271 | for_each_evictable_lru(l) { |
0cb57258 JW |
2272 | enum zone_stat_item ls = NR_LRU_BASE + l; |
2273 | unsigned long lru_pages = zone_page_state(zone, ls); | |
2274 | ||
894bc310 | 2275 | /* For pass = 0, we don't shrink the active list */ |
0cb57258 JW |
2276 | if (pass == 0 && (l == LRU_ACTIVE_ANON || |
2277 | l == LRU_ACTIVE_FILE)) | |
b69408e8 CL |
2278 | continue; |
2279 | ||
f8629631 WF |
2280 | reclaim_stat = get_reclaim_stat(zone, sc); |
2281 | reclaim_stat->nr_saved_scan[l] += | |
2282 | (lru_pages >> prio) + 1; | |
2283 | if (reclaim_stat->nr_saved_scan[l] | |
2284 | >= nr_pages || pass > 3) { | |
0cb57258 JW |
2285 | unsigned long nr_to_scan; |
2286 | ||
f8629631 | 2287 | reclaim_stat->nr_saved_scan[l] = 0; |
0cb57258 | 2288 | nr_to_scan = min(nr_pages, lru_pages); |
d979677c | 2289 | nr_reclaimed += shrink_list(l, nr_to_scan, zone, |
b69408e8 | 2290 | sc, prio); |
d979677c | 2291 | if (nr_reclaimed >= nr_pages) { |
a21e2553 | 2292 | sc->nr_reclaimed += nr_reclaimed; |
d979677c MK |
2293 | return; |
2294 | } | |
d6277db4 RW |
2295 | } |
2296 | } | |
d6277db4 | 2297 | } |
a21e2553 | 2298 | sc->nr_reclaimed += nr_reclaimed; |
d6277db4 RW |
2299 | } |
2300 | ||
2301 | /* | |
2302 | * Try to free `nr_pages' of memory, system-wide, and return the number of | |
2303 | * freed pages. | |
2304 | * | |
2305 | * Rather than trying to age LRUs the aim is to preserve the overall | |
2306 | * LRU order by reclaiming preferentially | |
2307 | * inactive > active > active referenced > active mapped | |
1da177e4 | 2308 | */ |
69e05944 | 2309 | unsigned long shrink_all_memory(unsigned long nr_pages) |
1da177e4 | 2310 | { |
d6277db4 | 2311 | unsigned long lru_pages, nr_slab; |
d6277db4 RW |
2312 | int pass; |
2313 | struct reclaim_state reclaim_state; | |
d6277db4 RW |
2314 | struct scan_control sc = { |
2315 | .gfp_mask = GFP_KERNEL, | |
a6dc60f8 | 2316 | .may_unmap = 0, |
d6277db4 | 2317 | .may_writepage = 1, |
66e1707b | 2318 | .isolate_pages = isolate_pages_global, |
a21e2553 | 2319 | .nr_reclaimed = 0, |
1da177e4 LT |
2320 | }; |
2321 | ||
2322 | current->reclaim_state = &reclaim_state; | |
69e05944 | 2323 | |
adea02a1 | 2324 | lru_pages = global_reclaimable_pages(); |
972d1a7b | 2325 | nr_slab = global_page_state(NR_SLAB_RECLAIMABLE); |
d6277db4 RW |
2326 | /* If slab caches are huge, it's better to hit them first */ |
2327 | while (nr_slab >= lru_pages) { | |
2328 | reclaim_state.reclaimed_slab = 0; | |
2329 | shrink_slab(nr_pages, sc.gfp_mask, lru_pages); | |
2330 | if (!reclaim_state.reclaimed_slab) | |
1da177e4 | 2331 | break; |
d6277db4 | 2332 | |
d979677c MK |
2333 | sc.nr_reclaimed += reclaim_state.reclaimed_slab; |
2334 | if (sc.nr_reclaimed >= nr_pages) | |
d6277db4 RW |
2335 | goto out; |
2336 | ||
2337 | nr_slab -= reclaim_state.reclaimed_slab; | |
1da177e4 | 2338 | } |
d6277db4 RW |
2339 | |
2340 | /* | |
2341 | * We try to shrink LRUs in 5 passes: | |
2342 | * 0 = Reclaim from inactive_list only | |
2343 | * 1 = Reclaim from active list but don't reclaim mapped | |
2344 | * 2 = 2nd pass of type 1 | |
2345 | * 3 = Reclaim mapped (normal reclaim) | |
2346 | * 4 = 2nd pass of type 3 | |
2347 | */ | |
2348 | for (pass = 0; pass < 5; pass++) { | |
2349 | int prio; | |
2350 | ||
d6277db4 | 2351 | /* Force reclaiming mapped pages in the passes #3 and #4 */ |
3049103d | 2352 | if (pass > 2) |
a6dc60f8 | 2353 | sc.may_unmap = 1; |
d6277db4 RW |
2354 | |
2355 | for (prio = DEF_PRIORITY; prio >= 0; prio--) { | |
d979677c | 2356 | unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed; |
d6277db4 | 2357 | |
d6277db4 | 2358 | sc.nr_scanned = 0; |
9786bf84 | 2359 | sc.swap_cluster_max = nr_to_scan; |
d979677c MK |
2360 | shrink_all_zones(nr_to_scan, prio, pass, &sc); |
2361 | if (sc.nr_reclaimed >= nr_pages) | |
d6277db4 RW |
2362 | goto out; |
2363 | ||
2364 | reclaim_state.reclaimed_slab = 0; | |
76395d37 | 2365 | shrink_slab(sc.nr_scanned, sc.gfp_mask, |
adea02a1 | 2366 | global_reclaimable_pages()); |
d979677c MK |
2367 | sc.nr_reclaimed += reclaim_state.reclaimed_slab; |
2368 | if (sc.nr_reclaimed >= nr_pages) | |
d6277db4 RW |
2369 | goto out; |
2370 | ||
2371 | if (sc.nr_scanned && prio < DEF_PRIORITY - 2) | |
8aa7e847 | 2372 | congestion_wait(BLK_RW_ASYNC, HZ / 10); |
d6277db4 | 2373 | } |
248a0301 | 2374 | } |
d6277db4 RW |
2375 | |
2376 | /* | |
d979677c MK |
2377 | * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be |
2378 | * something in slab caches | |
d6277db4 | 2379 | */ |
d979677c | 2380 | if (!sc.nr_reclaimed) { |
d6277db4 RW |
2381 | do { |
2382 | reclaim_state.reclaimed_slab = 0; | |
adea02a1 WF |
2383 | shrink_slab(nr_pages, sc.gfp_mask, |
2384 | global_reclaimable_pages()); | |
d979677c MK |
2385 | sc.nr_reclaimed += reclaim_state.reclaimed_slab; |
2386 | } while (sc.nr_reclaimed < nr_pages && | |
2387 | reclaim_state.reclaimed_slab > 0); | |
76395d37 | 2388 | } |
d6277db4 | 2389 | |
d979677c | 2390 | |
d6277db4 | 2391 | out: |
1da177e4 | 2392 | current->reclaim_state = NULL; |
d6277db4 | 2393 | |
d979677c | 2394 | return sc.nr_reclaimed; |
1da177e4 | 2395 | } |
c6f37f12 | 2396 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 2397 | |
1da177e4 LT |
2398 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
2399 | not required for correctness. So if the last cpu in a node goes | |
2400 | away, we get changed to run anywhere: as the first one comes back, | |
2401 | restore their cpu bindings. */ | |
9c7b216d | 2402 | static int __devinit cpu_callback(struct notifier_block *nfb, |
69e05944 | 2403 | unsigned long action, void *hcpu) |
1da177e4 | 2404 | { |
58c0a4a7 | 2405 | int nid; |
1da177e4 | 2406 | |
8bb78442 | 2407 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
58c0a4a7 | 2408 | for_each_node_state(nid, N_HIGH_MEMORY) { |
c5f59f08 | 2409 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
2410 | const struct cpumask *mask; |
2411 | ||
2412 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 2413 | |
3e597945 | 2414 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 2415 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 2416 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
2417 | } |
2418 | } | |
2419 | return NOTIFY_OK; | |
2420 | } | |
1da177e4 | 2421 | |
3218ae14 YG |
2422 | /* |
2423 | * This kswapd start function will be called by init and node-hot-add. | |
2424 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
2425 | */ | |
2426 | int kswapd_run(int nid) | |
2427 | { | |
2428 | pg_data_t *pgdat = NODE_DATA(nid); | |
2429 | int ret = 0; | |
2430 | ||
2431 | if (pgdat->kswapd) | |
2432 | return 0; | |
2433 | ||
2434 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
2435 | if (IS_ERR(pgdat->kswapd)) { | |
2436 | /* failure at boot is fatal */ | |
2437 | BUG_ON(system_state == SYSTEM_BOOTING); | |
2438 | printk("Failed to start kswapd on node %d\n",nid); | |
2439 | ret = -1; | |
2440 | } | |
2441 | return ret; | |
2442 | } | |
2443 | ||
1da177e4 LT |
2444 | static int __init kswapd_init(void) |
2445 | { | |
3218ae14 | 2446 | int nid; |
69e05944 | 2447 | |
1da177e4 | 2448 | swap_setup(); |
9422ffba | 2449 | for_each_node_state(nid, N_HIGH_MEMORY) |
3218ae14 | 2450 | kswapd_run(nid); |
1da177e4 LT |
2451 | hotcpu_notifier(cpu_callback, 0); |
2452 | return 0; | |
2453 | } | |
2454 | ||
2455 | module_init(kswapd_init) | |
9eeff239 CL |
2456 | |
2457 | #ifdef CONFIG_NUMA | |
2458 | /* | |
2459 | * Zone reclaim mode | |
2460 | * | |
2461 | * If non-zero call zone_reclaim when the number of free pages falls below | |
2462 | * the watermarks. | |
9eeff239 CL |
2463 | */ |
2464 | int zone_reclaim_mode __read_mostly; | |
2465 | ||
1b2ffb78 | 2466 | #define RECLAIM_OFF 0 |
7d03431c | 2467 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 CL |
2468 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
2469 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
2470 | ||
a92f7126 CL |
2471 | /* |
2472 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
2473 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
2474 | * a zone. | |
2475 | */ | |
2476 | #define ZONE_RECLAIM_PRIORITY 4 | |
2477 | ||
9614634f CL |
2478 | /* |
2479 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
2480 | * occur. | |
2481 | */ | |
2482 | int sysctl_min_unmapped_ratio = 1; | |
2483 | ||
0ff38490 CL |
2484 | /* |
2485 | * If the number of slab pages in a zone grows beyond this percentage then | |
2486 | * slab reclaim needs to occur. | |
2487 | */ | |
2488 | int sysctl_min_slab_ratio = 5; | |
2489 | ||
90afa5de MG |
2490 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
2491 | { | |
2492 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
2493 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
2494 | zone_page_state(zone, NR_ACTIVE_FILE); | |
2495 | ||
2496 | /* | |
2497 | * It's possible for there to be more file mapped pages than | |
2498 | * accounted for by the pages on the file LRU lists because | |
2499 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
2500 | */ | |
2501 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
2502 | } | |
2503 | ||
2504 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
2505 | static long zone_pagecache_reclaimable(struct zone *zone) | |
2506 | { | |
2507 | long nr_pagecache_reclaimable; | |
2508 | long delta = 0; | |
2509 | ||
2510 | /* | |
2511 | * If RECLAIM_SWAP is set, then all file pages are considered | |
2512 | * potentially reclaimable. Otherwise, we have to worry about | |
2513 | * pages like swapcache and zone_unmapped_file_pages() provides | |
2514 | * a better estimate | |
2515 | */ | |
2516 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
2517 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
2518 | else | |
2519 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
2520 | ||
2521 | /* If we can't clean pages, remove dirty pages from consideration */ | |
2522 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
2523 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
2524 | ||
2525 | /* Watch for any possible underflows due to delta */ | |
2526 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
2527 | delta = nr_pagecache_reclaimable; | |
2528 | ||
2529 | return nr_pagecache_reclaimable - delta; | |
2530 | } | |
2531 | ||
9eeff239 CL |
2532 | /* |
2533 | * Try to free up some pages from this zone through reclaim. | |
2534 | */ | |
179e9639 | 2535 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 2536 | { |
7fb2d46d | 2537 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 2538 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
2539 | struct task_struct *p = current; |
2540 | struct reclaim_state reclaim_state; | |
8695949a | 2541 | int priority; |
179e9639 AM |
2542 | struct scan_control sc = { |
2543 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
a6dc60f8 | 2544 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), |
2e2e4259 | 2545 | .may_swap = 1, |
69e05944 AM |
2546 | .swap_cluster_max = max_t(unsigned long, nr_pages, |
2547 | SWAP_CLUSTER_MAX), | |
179e9639 | 2548 | .gfp_mask = gfp_mask, |
d6277db4 | 2549 | .swappiness = vm_swappiness, |
bd2f6199 | 2550 | .order = order, |
66e1707b | 2551 | .isolate_pages = isolate_pages_global, |
179e9639 | 2552 | }; |
83e33a47 | 2553 | unsigned long slab_reclaimable; |
9eeff239 CL |
2554 | |
2555 | disable_swap_token(); | |
9eeff239 | 2556 | cond_resched(); |
d4f7796e CL |
2557 | /* |
2558 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
2559 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
2560 | * and RECLAIM_SWAP. | |
2561 | */ | |
2562 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
9eeff239 CL |
2563 | reclaim_state.reclaimed_slab = 0; |
2564 | p->reclaim_state = &reclaim_state; | |
c84db23c | 2565 | |
90afa5de | 2566 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
2567 | /* |
2568 | * Free memory by calling shrink zone with increasing | |
2569 | * priorities until we have enough memory freed. | |
2570 | */ | |
2571 | priority = ZONE_RECLAIM_PRIORITY; | |
2572 | do { | |
3bb1a852 | 2573 | note_zone_scanning_priority(zone, priority); |
a79311c1 | 2574 | shrink_zone(priority, zone, &sc); |
0ff38490 | 2575 | priority--; |
a79311c1 | 2576 | } while (priority >= 0 && sc.nr_reclaimed < nr_pages); |
0ff38490 | 2577 | } |
c84db23c | 2578 | |
83e33a47 CL |
2579 | slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
2580 | if (slab_reclaimable > zone->min_slab_pages) { | |
2a16e3f4 | 2581 | /* |
7fb2d46d | 2582 | * shrink_slab() does not currently allow us to determine how |
0ff38490 CL |
2583 | * many pages were freed in this zone. So we take the current |
2584 | * number of slab pages and shake the slab until it is reduced | |
2585 | * by the same nr_pages that we used for reclaiming unmapped | |
2586 | * pages. | |
2a16e3f4 | 2587 | * |
0ff38490 CL |
2588 | * Note that shrink_slab will free memory on all zones and may |
2589 | * take a long time. | |
2a16e3f4 | 2590 | */ |
0ff38490 | 2591 | while (shrink_slab(sc.nr_scanned, gfp_mask, order) && |
83e33a47 CL |
2592 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) > |
2593 | slab_reclaimable - nr_pages) | |
0ff38490 | 2594 | ; |
83e33a47 CL |
2595 | |
2596 | /* | |
2597 | * Update nr_reclaimed by the number of slab pages we | |
2598 | * reclaimed from this zone. | |
2599 | */ | |
a79311c1 | 2600 | sc.nr_reclaimed += slab_reclaimable - |
83e33a47 | 2601 | zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
2a16e3f4 CL |
2602 | } |
2603 | ||
9eeff239 | 2604 | p->reclaim_state = NULL; |
d4f7796e | 2605 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
a79311c1 | 2606 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 2607 | } |
179e9639 AM |
2608 | |
2609 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
2610 | { | |
179e9639 | 2611 | int node_id; |
d773ed6b | 2612 | int ret; |
179e9639 AM |
2613 | |
2614 | /* | |
0ff38490 CL |
2615 | * Zone reclaim reclaims unmapped file backed pages and |
2616 | * slab pages if we are over the defined limits. | |
34aa1330 | 2617 | * |
9614634f CL |
2618 | * A small portion of unmapped file backed pages is needed for |
2619 | * file I/O otherwise pages read by file I/O will be immediately | |
2620 | * thrown out if the zone is overallocated. So we do not reclaim | |
2621 | * if less than a specified percentage of the zone is used by | |
2622 | * unmapped file backed pages. | |
179e9639 | 2623 | */ |
90afa5de MG |
2624 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
2625 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 2626 | return ZONE_RECLAIM_FULL; |
179e9639 | 2627 | |
d773ed6b | 2628 | if (zone_is_all_unreclaimable(zone)) |
fa5e084e | 2629 | return ZONE_RECLAIM_FULL; |
d773ed6b | 2630 | |
179e9639 | 2631 | /* |
d773ed6b | 2632 | * Do not scan if the allocation should not be delayed. |
179e9639 | 2633 | */ |
d773ed6b | 2634 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 2635 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
2636 | |
2637 | /* | |
2638 | * Only run zone reclaim on the local zone or on zones that do not | |
2639 | * have associated processors. This will favor the local processor | |
2640 | * over remote processors and spread off node memory allocations | |
2641 | * as wide as possible. | |
2642 | */ | |
89fa3024 | 2643 | node_id = zone_to_nid(zone); |
37c0708d | 2644 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 2645 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b DR |
2646 | |
2647 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
fa5e084e MG |
2648 | return ZONE_RECLAIM_NOSCAN; |
2649 | ||
d773ed6b DR |
2650 | ret = __zone_reclaim(zone, gfp_mask, order); |
2651 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
2652 | ||
24cf7251 MG |
2653 | if (!ret) |
2654 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
2655 | ||
d773ed6b | 2656 | return ret; |
179e9639 | 2657 | } |
9eeff239 | 2658 | #endif |
894bc310 | 2659 | |
894bc310 LS |
2660 | /* |
2661 | * page_evictable - test whether a page is evictable | |
2662 | * @page: the page to test | |
2663 | * @vma: the VMA in which the page is or will be mapped, may be NULL | |
2664 | * | |
2665 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
b291f000 NP |
2666 | * lists vs unevictable list. The vma argument is !NULL when called from the |
2667 | * fault path to determine how to instantate a new page. | |
894bc310 LS |
2668 | * |
2669 | * Reasons page might not be evictable: | |
ba9ddf49 | 2670 | * (1) page's mapping marked unevictable |
b291f000 | 2671 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 2672 | * |
894bc310 LS |
2673 | */ |
2674 | int page_evictable(struct page *page, struct vm_area_struct *vma) | |
2675 | { | |
2676 | ||
ba9ddf49 LS |
2677 | if (mapping_unevictable(page_mapping(page))) |
2678 | return 0; | |
2679 | ||
b291f000 NP |
2680 | if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page))) |
2681 | return 0; | |
894bc310 LS |
2682 | |
2683 | return 1; | |
2684 | } | |
89e004ea LS |
2685 | |
2686 | /** | |
2687 | * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list | |
2688 | * @page: page to check evictability and move to appropriate lru list | |
2689 | * @zone: zone page is in | |
2690 | * | |
2691 | * Checks a page for evictability and moves the page to the appropriate | |
2692 | * zone lru list. | |
2693 | * | |
2694 | * Restrictions: zone->lru_lock must be held, page must be on LRU and must | |
2695 | * have PageUnevictable set. | |
2696 | */ | |
2697 | static void check_move_unevictable_page(struct page *page, struct zone *zone) | |
2698 | { | |
2699 | VM_BUG_ON(PageActive(page)); | |
2700 | ||
2701 | retry: | |
2702 | ClearPageUnevictable(page); | |
2703 | if (page_evictable(page, NULL)) { | |
401a8e1c | 2704 | enum lru_list l = page_lru_base_type(page); |
af936a16 | 2705 | |
89e004ea LS |
2706 | __dec_zone_state(zone, NR_UNEVICTABLE); |
2707 | list_move(&page->lru, &zone->lru[l].list); | |
08e552c6 | 2708 | mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l); |
89e004ea LS |
2709 | __inc_zone_state(zone, NR_INACTIVE_ANON + l); |
2710 | __count_vm_event(UNEVICTABLE_PGRESCUED); | |
2711 | } else { | |
2712 | /* | |
2713 | * rotate unevictable list | |
2714 | */ | |
2715 | SetPageUnevictable(page); | |
2716 | list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list); | |
08e552c6 | 2717 | mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE); |
89e004ea LS |
2718 | if (page_evictable(page, NULL)) |
2719 | goto retry; | |
2720 | } | |
2721 | } | |
2722 | ||
2723 | /** | |
2724 | * scan_mapping_unevictable_pages - scan an address space for evictable pages | |
2725 | * @mapping: struct address_space to scan for evictable pages | |
2726 | * | |
2727 | * Scan all pages in mapping. Check unevictable pages for | |
2728 | * evictability and move them to the appropriate zone lru list. | |
2729 | */ | |
2730 | void scan_mapping_unevictable_pages(struct address_space *mapping) | |
2731 | { | |
2732 | pgoff_t next = 0; | |
2733 | pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >> | |
2734 | PAGE_CACHE_SHIFT; | |
2735 | struct zone *zone; | |
2736 | struct pagevec pvec; | |
2737 | ||
2738 | if (mapping->nrpages == 0) | |
2739 | return; | |
2740 | ||
2741 | pagevec_init(&pvec, 0); | |
2742 | while (next < end && | |
2743 | pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { | |
2744 | int i; | |
2745 | int pg_scanned = 0; | |
2746 | ||
2747 | zone = NULL; | |
2748 | ||
2749 | for (i = 0; i < pagevec_count(&pvec); i++) { | |
2750 | struct page *page = pvec.pages[i]; | |
2751 | pgoff_t page_index = page->index; | |
2752 | struct zone *pagezone = page_zone(page); | |
2753 | ||
2754 | pg_scanned++; | |
2755 | if (page_index > next) | |
2756 | next = page_index; | |
2757 | next++; | |
2758 | ||
2759 | if (pagezone != zone) { | |
2760 | if (zone) | |
2761 | spin_unlock_irq(&zone->lru_lock); | |
2762 | zone = pagezone; | |
2763 | spin_lock_irq(&zone->lru_lock); | |
2764 | } | |
2765 | ||
2766 | if (PageLRU(page) && PageUnevictable(page)) | |
2767 | check_move_unevictable_page(page, zone); | |
2768 | } | |
2769 | if (zone) | |
2770 | spin_unlock_irq(&zone->lru_lock); | |
2771 | pagevec_release(&pvec); | |
2772 | ||
2773 | count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned); | |
2774 | } | |
2775 | ||
2776 | } | |
af936a16 LS |
2777 | |
2778 | /** | |
2779 | * scan_zone_unevictable_pages - check unevictable list for evictable pages | |
2780 | * @zone - zone of which to scan the unevictable list | |
2781 | * | |
2782 | * Scan @zone's unevictable LRU lists to check for pages that have become | |
2783 | * evictable. Move those that have to @zone's inactive list where they | |
2784 | * become candidates for reclaim, unless shrink_inactive_zone() decides | |
2785 | * to reactivate them. Pages that are still unevictable are rotated | |
2786 | * back onto @zone's unevictable list. | |
2787 | */ | |
2788 | #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */ | |
14b90b22 | 2789 | static void scan_zone_unevictable_pages(struct zone *zone) |
af936a16 LS |
2790 | { |
2791 | struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list; | |
2792 | unsigned long scan; | |
2793 | unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE); | |
2794 | ||
2795 | while (nr_to_scan > 0) { | |
2796 | unsigned long batch_size = min(nr_to_scan, | |
2797 | SCAN_UNEVICTABLE_BATCH_SIZE); | |
2798 | ||
2799 | spin_lock_irq(&zone->lru_lock); | |
2800 | for (scan = 0; scan < batch_size; scan++) { | |
2801 | struct page *page = lru_to_page(l_unevictable); | |
2802 | ||
2803 | if (!trylock_page(page)) | |
2804 | continue; | |
2805 | ||
2806 | prefetchw_prev_lru_page(page, l_unevictable, flags); | |
2807 | ||
2808 | if (likely(PageLRU(page) && PageUnevictable(page))) | |
2809 | check_move_unevictable_page(page, zone); | |
2810 | ||
2811 | unlock_page(page); | |
2812 | } | |
2813 | spin_unlock_irq(&zone->lru_lock); | |
2814 | ||
2815 | nr_to_scan -= batch_size; | |
2816 | } | |
2817 | } | |
2818 | ||
2819 | ||
2820 | /** | |
2821 | * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages | |
2822 | * | |
2823 | * A really big hammer: scan all zones' unevictable LRU lists to check for | |
2824 | * pages that have become evictable. Move those back to the zones' | |
2825 | * inactive list where they become candidates for reclaim. | |
2826 | * This occurs when, e.g., we have unswappable pages on the unevictable lists, | |
2827 | * and we add swap to the system. As such, it runs in the context of a task | |
2828 | * that has possibly/probably made some previously unevictable pages | |
2829 | * evictable. | |
2830 | */ | |
ff30153b | 2831 | static void scan_all_zones_unevictable_pages(void) |
af936a16 LS |
2832 | { |
2833 | struct zone *zone; | |
2834 | ||
2835 | for_each_zone(zone) { | |
2836 | scan_zone_unevictable_pages(zone); | |
2837 | } | |
2838 | } | |
2839 | ||
2840 | /* | |
2841 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
2842 | * all nodes' unevictable lists for evictable pages | |
2843 | */ | |
2844 | unsigned long scan_unevictable_pages; | |
2845 | ||
2846 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
2847 | struct file *file, void __user *buffer, | |
2848 | size_t *length, loff_t *ppos) | |
2849 | { | |
2850 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); | |
2851 | ||
2852 | if (write && *(unsigned long *)table->data) | |
2853 | scan_all_zones_unevictable_pages(); | |
2854 | ||
2855 | scan_unevictable_pages = 0; | |
2856 | return 0; | |
2857 | } | |
2858 | ||
2859 | /* | |
2860 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
2861 | * a specified node's per zone unevictable lists for evictable pages. | |
2862 | */ | |
2863 | ||
2864 | static ssize_t read_scan_unevictable_node(struct sys_device *dev, | |
2865 | struct sysdev_attribute *attr, | |
2866 | char *buf) | |
2867 | { | |
2868 | return sprintf(buf, "0\n"); /* always zero; should fit... */ | |
2869 | } | |
2870 | ||
2871 | static ssize_t write_scan_unevictable_node(struct sys_device *dev, | |
2872 | struct sysdev_attribute *attr, | |
2873 | const char *buf, size_t count) | |
2874 | { | |
2875 | struct zone *node_zones = NODE_DATA(dev->id)->node_zones; | |
2876 | struct zone *zone; | |
2877 | unsigned long res; | |
2878 | unsigned long req = strict_strtoul(buf, 10, &res); | |
2879 | ||
2880 | if (!req) | |
2881 | return 1; /* zero is no-op */ | |
2882 | ||
2883 | for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { | |
2884 | if (!populated_zone(zone)) | |
2885 | continue; | |
2886 | scan_zone_unevictable_pages(zone); | |
2887 | } | |
2888 | return 1; | |
2889 | } | |
2890 | ||
2891 | ||
2892 | static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, | |
2893 | read_scan_unevictable_node, | |
2894 | write_scan_unevictable_node); | |
2895 | ||
2896 | int scan_unevictable_register_node(struct node *node) | |
2897 | { | |
2898 | return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages); | |
2899 | } | |
2900 | ||
2901 | void scan_unevictable_unregister_node(struct node *node) | |
2902 | { | |
2903 | sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages); | |
2904 | } | |
2905 |