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> | |
22 | #include <linux/file.h> | |
23 | #include <linux/writeback.h> | |
24 | #include <linux/blkdev.h> | |
25 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
26 | buffer_heads_over_limit */ | |
27 | #include <linux/mm_inline.h> | |
28 | #include <linux/pagevec.h> | |
29 | #include <linux/backing-dev.h> | |
30 | #include <linux/rmap.h> | |
31 | #include <linux/topology.h> | |
32 | #include <linux/cpu.h> | |
33 | #include <linux/cpuset.h> | |
34 | #include <linux/notifier.h> | |
35 | #include <linux/rwsem.h> | |
36 | ||
37 | #include <asm/tlbflush.h> | |
38 | #include <asm/div64.h> | |
39 | ||
40 | #include <linux/swapops.h> | |
41 | ||
0f8053a5 NP |
42 | #include "internal.h" |
43 | ||
1da177e4 LT |
44 | /* possible outcome of pageout() */ |
45 | typedef enum { | |
46 | /* failed to write page out, page is locked */ | |
47 | PAGE_KEEP, | |
48 | /* move page to the active list, page is locked */ | |
49 | PAGE_ACTIVATE, | |
50 | /* page has been sent to the disk successfully, page is unlocked */ | |
51 | PAGE_SUCCESS, | |
52 | /* page is clean and locked */ | |
53 | PAGE_CLEAN, | |
54 | } pageout_t; | |
55 | ||
56 | struct scan_control { | |
1da177e4 LT |
57 | /* Incremented by the number of inactive pages that were scanned */ |
58 | unsigned long nr_scanned; | |
59 | ||
1da177e4 LT |
60 | unsigned long nr_mapped; /* From page_state */ |
61 | ||
1da177e4 | 62 | /* This context's GFP mask */ |
6daa0e28 | 63 | gfp_t gfp_mask; |
1da177e4 LT |
64 | |
65 | int may_writepage; | |
66 | ||
f1fd1067 CL |
67 | /* Can pages be swapped as part of reclaim? */ |
68 | int may_swap; | |
69 | ||
1da177e4 LT |
70 | /* This context's SWAP_CLUSTER_MAX. If freeing memory for |
71 | * suspend, we effectively ignore SWAP_CLUSTER_MAX. | |
72 | * In this context, it doesn't matter that we scan the | |
73 | * whole list at once. */ | |
74 | int swap_cluster_max; | |
75 | }; | |
76 | ||
77 | /* | |
78 | * The list of shrinker callbacks used by to apply pressure to | |
79 | * ageable caches. | |
80 | */ | |
81 | struct shrinker { | |
82 | shrinker_t shrinker; | |
83 | struct list_head list; | |
84 | int seeks; /* seeks to recreate an obj */ | |
85 | long nr; /* objs pending delete */ | |
86 | }; | |
87 | ||
88 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) | |
89 | ||
90 | #ifdef ARCH_HAS_PREFETCH | |
91 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
92 | do { \ | |
93 | if ((_page)->lru.prev != _base) { \ | |
94 | struct page *prev; \ | |
95 | \ | |
96 | prev = lru_to_page(&(_page->lru)); \ | |
97 | prefetch(&prev->_field); \ | |
98 | } \ | |
99 | } while (0) | |
100 | #else | |
101 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
102 | #endif | |
103 | ||
104 | #ifdef ARCH_HAS_PREFETCHW | |
105 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
106 | do { \ | |
107 | if ((_page)->lru.prev != _base) { \ | |
108 | struct page *prev; \ | |
109 | \ | |
110 | prev = lru_to_page(&(_page->lru)); \ | |
111 | prefetchw(&prev->_field); \ | |
112 | } \ | |
113 | } while (0) | |
114 | #else | |
115 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
116 | #endif | |
117 | ||
118 | /* | |
119 | * From 0 .. 100. Higher means more swappy. | |
120 | */ | |
121 | int vm_swappiness = 60; | |
122 | static long total_memory; | |
123 | ||
124 | static LIST_HEAD(shrinker_list); | |
125 | static DECLARE_RWSEM(shrinker_rwsem); | |
126 | ||
127 | /* | |
128 | * Add a shrinker callback to be called from the vm | |
129 | */ | |
130 | struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker) | |
131 | { | |
132 | struct shrinker *shrinker; | |
133 | ||
134 | shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL); | |
135 | if (shrinker) { | |
136 | shrinker->shrinker = theshrinker; | |
137 | shrinker->seeks = seeks; | |
138 | shrinker->nr = 0; | |
139 | down_write(&shrinker_rwsem); | |
140 | list_add_tail(&shrinker->list, &shrinker_list); | |
141 | up_write(&shrinker_rwsem); | |
142 | } | |
143 | return shrinker; | |
144 | } | |
145 | EXPORT_SYMBOL(set_shrinker); | |
146 | ||
147 | /* | |
148 | * Remove one | |
149 | */ | |
150 | void remove_shrinker(struct shrinker *shrinker) | |
151 | { | |
152 | down_write(&shrinker_rwsem); | |
153 | list_del(&shrinker->list); | |
154 | up_write(&shrinker_rwsem); | |
155 | kfree(shrinker); | |
156 | } | |
157 | EXPORT_SYMBOL(remove_shrinker); | |
158 | ||
159 | #define SHRINK_BATCH 128 | |
160 | /* | |
161 | * Call the shrink functions to age shrinkable caches | |
162 | * | |
163 | * Here we assume it costs one seek to replace a lru page and that it also | |
164 | * takes a seek to recreate a cache object. With this in mind we age equal | |
165 | * percentages of the lru and ageable caches. This should balance the seeks | |
166 | * generated by these structures. | |
167 | * | |
168 | * If the vm encounted mapped pages on the LRU it increase the pressure on | |
169 | * slab to avoid swapping. | |
170 | * | |
171 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
172 | * | |
173 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
174 | * are eligible for the caller's allocation attempt. It is used for balancing | |
175 | * slab reclaim versus page reclaim. | |
b15e0905 | 176 | * |
177 | * Returns the number of slab objects which we shrunk. | |
1da177e4 | 178 | */ |
69e05944 AM |
179 | unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask, |
180 | unsigned long lru_pages) | |
1da177e4 LT |
181 | { |
182 | struct shrinker *shrinker; | |
69e05944 | 183 | unsigned long ret = 0; |
1da177e4 LT |
184 | |
185 | if (scanned == 0) | |
186 | scanned = SWAP_CLUSTER_MAX; | |
187 | ||
188 | if (!down_read_trylock(&shrinker_rwsem)) | |
b15e0905 | 189 | return 1; /* Assume we'll be able to shrink next time */ |
1da177e4 LT |
190 | |
191 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
192 | unsigned long long delta; | |
193 | unsigned long total_scan; | |
ea164d73 | 194 | unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask); |
1da177e4 LT |
195 | |
196 | delta = (4 * scanned) / shrinker->seeks; | |
ea164d73 | 197 | delta *= max_pass; |
1da177e4 LT |
198 | do_div(delta, lru_pages + 1); |
199 | shrinker->nr += delta; | |
ea164d73 AA |
200 | if (shrinker->nr < 0) { |
201 | printk(KERN_ERR "%s: nr=%ld\n", | |
202 | __FUNCTION__, shrinker->nr); | |
203 | shrinker->nr = max_pass; | |
204 | } | |
205 | ||
206 | /* | |
207 | * Avoid risking looping forever due to too large nr value: | |
208 | * never try to free more than twice the estimate number of | |
209 | * freeable entries. | |
210 | */ | |
211 | if (shrinker->nr > max_pass * 2) | |
212 | shrinker->nr = max_pass * 2; | |
1da177e4 LT |
213 | |
214 | total_scan = shrinker->nr; | |
215 | shrinker->nr = 0; | |
216 | ||
217 | while (total_scan >= SHRINK_BATCH) { | |
218 | long this_scan = SHRINK_BATCH; | |
219 | int shrink_ret; | |
b15e0905 | 220 | int nr_before; |
1da177e4 | 221 | |
b15e0905 | 222 | nr_before = (*shrinker->shrinker)(0, gfp_mask); |
1da177e4 LT |
223 | shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask); |
224 | if (shrink_ret == -1) | |
225 | break; | |
b15e0905 | 226 | if (shrink_ret < nr_before) |
227 | ret += nr_before - shrink_ret; | |
1da177e4 LT |
228 | mod_page_state(slabs_scanned, this_scan); |
229 | total_scan -= this_scan; | |
230 | ||
231 | cond_resched(); | |
232 | } | |
233 | ||
234 | shrinker->nr += total_scan; | |
235 | } | |
236 | up_read(&shrinker_rwsem); | |
b15e0905 | 237 | return ret; |
1da177e4 LT |
238 | } |
239 | ||
240 | /* Called without lock on whether page is mapped, so answer is unstable */ | |
241 | static inline int page_mapping_inuse(struct page *page) | |
242 | { | |
243 | struct address_space *mapping; | |
244 | ||
245 | /* Page is in somebody's page tables. */ | |
246 | if (page_mapped(page)) | |
247 | return 1; | |
248 | ||
249 | /* Be more reluctant to reclaim swapcache than pagecache */ | |
250 | if (PageSwapCache(page)) | |
251 | return 1; | |
252 | ||
253 | mapping = page_mapping(page); | |
254 | if (!mapping) | |
255 | return 0; | |
256 | ||
257 | /* File is mmap'd by somebody? */ | |
258 | return mapping_mapped(mapping); | |
259 | } | |
260 | ||
261 | static inline int is_page_cache_freeable(struct page *page) | |
262 | { | |
263 | return page_count(page) - !!PagePrivate(page) == 2; | |
264 | } | |
265 | ||
266 | static int may_write_to_queue(struct backing_dev_info *bdi) | |
267 | { | |
930d9152 | 268 | if (current->flags & PF_SWAPWRITE) |
1da177e4 LT |
269 | return 1; |
270 | if (!bdi_write_congested(bdi)) | |
271 | return 1; | |
272 | if (bdi == current->backing_dev_info) | |
273 | return 1; | |
274 | return 0; | |
275 | } | |
276 | ||
277 | /* | |
278 | * We detected a synchronous write error writing a page out. Probably | |
279 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
280 | * fsync(), msync() or close(). | |
281 | * | |
282 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
283 | * prevents it from being freed up. But we have a ref on the page and once | |
284 | * that page is locked, the mapping is pinned. | |
285 | * | |
286 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
287 | * __GFP_FS. | |
288 | */ | |
289 | static void handle_write_error(struct address_space *mapping, | |
290 | struct page *page, int error) | |
291 | { | |
292 | lock_page(page); | |
293 | if (page_mapping(page) == mapping) { | |
294 | if (error == -ENOSPC) | |
295 | set_bit(AS_ENOSPC, &mapping->flags); | |
296 | else | |
297 | set_bit(AS_EIO, &mapping->flags); | |
298 | } | |
299 | unlock_page(page); | |
300 | } | |
301 | ||
302 | /* | |
1742f19f AM |
303 | * pageout is called by shrink_page_list() for each dirty page. |
304 | * Calls ->writepage(). | |
1da177e4 LT |
305 | */ |
306 | static pageout_t pageout(struct page *page, struct address_space *mapping) | |
307 | { | |
308 | /* | |
309 | * If the page is dirty, only perform writeback if that write | |
310 | * will be non-blocking. To prevent this allocation from being | |
311 | * stalled by pagecache activity. But note that there may be | |
312 | * stalls if we need to run get_block(). We could test | |
313 | * PagePrivate for that. | |
314 | * | |
315 | * If this process is currently in generic_file_write() against | |
316 | * this page's queue, we can perform writeback even if that | |
317 | * will block. | |
318 | * | |
319 | * If the page is swapcache, write it back even if that would | |
320 | * block, for some throttling. This happens by accident, because | |
321 | * swap_backing_dev_info is bust: it doesn't reflect the | |
322 | * congestion state of the swapdevs. Easy to fix, if needed. | |
323 | * See swapfile.c:page_queue_congested(). | |
324 | */ | |
325 | if (!is_page_cache_freeable(page)) | |
326 | return PAGE_KEEP; | |
327 | if (!mapping) { | |
328 | /* | |
329 | * Some data journaling orphaned pages can have | |
330 | * page->mapping == NULL while being dirty with clean buffers. | |
331 | */ | |
323aca6c | 332 | if (PagePrivate(page)) { |
1da177e4 LT |
333 | if (try_to_free_buffers(page)) { |
334 | ClearPageDirty(page); | |
335 | printk("%s: orphaned page\n", __FUNCTION__); | |
336 | return PAGE_CLEAN; | |
337 | } | |
338 | } | |
339 | return PAGE_KEEP; | |
340 | } | |
341 | if (mapping->a_ops->writepage == NULL) | |
342 | return PAGE_ACTIVATE; | |
343 | if (!may_write_to_queue(mapping->backing_dev_info)) | |
344 | return PAGE_KEEP; | |
345 | ||
346 | if (clear_page_dirty_for_io(page)) { | |
347 | int res; | |
348 | struct writeback_control wbc = { | |
349 | .sync_mode = WB_SYNC_NONE, | |
350 | .nr_to_write = SWAP_CLUSTER_MAX, | |
351 | .nonblocking = 1, | |
352 | .for_reclaim = 1, | |
353 | }; | |
354 | ||
355 | SetPageReclaim(page); | |
356 | res = mapping->a_ops->writepage(page, &wbc); | |
357 | if (res < 0) | |
358 | handle_write_error(mapping, page, res); | |
994fc28c | 359 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
360 | ClearPageReclaim(page); |
361 | return PAGE_ACTIVATE; | |
362 | } | |
363 | if (!PageWriteback(page)) { | |
364 | /* synchronous write or broken a_ops? */ | |
365 | ClearPageReclaim(page); | |
366 | } | |
367 | ||
368 | return PAGE_SUCCESS; | |
369 | } | |
370 | ||
371 | return PAGE_CLEAN; | |
372 | } | |
373 | ||
49d2e9cc CL |
374 | static int remove_mapping(struct address_space *mapping, struct page *page) |
375 | { | |
376 | if (!mapping) | |
377 | return 0; /* truncate got there first */ | |
378 | ||
379 | write_lock_irq(&mapping->tree_lock); | |
380 | ||
381 | /* | |
382 | * The non-racy check for busy page. It is critical to check | |
383 | * PageDirty _after_ making sure that the page is freeable and | |
384 | * not in use by anybody. (pagecache + us == 2) | |
385 | */ | |
386 | if (unlikely(page_count(page) != 2)) | |
387 | goto cannot_free; | |
388 | smp_rmb(); | |
389 | if (unlikely(PageDirty(page))) | |
390 | goto cannot_free; | |
391 | ||
392 | if (PageSwapCache(page)) { | |
393 | swp_entry_t swap = { .val = page_private(page) }; | |
394 | __delete_from_swap_cache(page); | |
395 | write_unlock_irq(&mapping->tree_lock); | |
396 | swap_free(swap); | |
397 | __put_page(page); /* The pagecache ref */ | |
398 | return 1; | |
399 | } | |
400 | ||
401 | __remove_from_page_cache(page); | |
402 | write_unlock_irq(&mapping->tree_lock); | |
403 | __put_page(page); | |
404 | return 1; | |
405 | ||
406 | cannot_free: | |
407 | write_unlock_irq(&mapping->tree_lock); | |
408 | return 0; | |
409 | } | |
410 | ||
1da177e4 | 411 | /* |
1742f19f | 412 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 413 | */ |
1742f19f AM |
414 | static unsigned long shrink_page_list(struct list_head *page_list, |
415 | struct scan_control *sc) | |
1da177e4 LT |
416 | { |
417 | LIST_HEAD(ret_pages); | |
418 | struct pagevec freed_pvec; | |
419 | int pgactivate = 0; | |
05ff5137 | 420 | unsigned long nr_reclaimed = 0; |
1da177e4 LT |
421 | |
422 | cond_resched(); | |
423 | ||
424 | pagevec_init(&freed_pvec, 1); | |
425 | while (!list_empty(page_list)) { | |
426 | struct address_space *mapping; | |
427 | struct page *page; | |
428 | int may_enter_fs; | |
429 | int referenced; | |
430 | ||
431 | cond_resched(); | |
432 | ||
433 | page = lru_to_page(page_list); | |
434 | list_del(&page->lru); | |
435 | ||
436 | if (TestSetPageLocked(page)) | |
437 | goto keep; | |
438 | ||
439 | BUG_ON(PageActive(page)); | |
440 | ||
441 | sc->nr_scanned++; | |
80e43426 CL |
442 | |
443 | if (!sc->may_swap && page_mapped(page)) | |
444 | goto keep_locked; | |
445 | ||
1da177e4 LT |
446 | /* Double the slab pressure for mapped and swapcache pages */ |
447 | if (page_mapped(page) || PageSwapCache(page)) | |
448 | sc->nr_scanned++; | |
449 | ||
450 | if (PageWriteback(page)) | |
451 | goto keep_locked; | |
452 | ||
f7b7fd8f | 453 | referenced = page_referenced(page, 1); |
1da177e4 LT |
454 | /* In active use or really unfreeable? Activate it. */ |
455 | if (referenced && page_mapping_inuse(page)) | |
456 | goto activate_locked; | |
457 | ||
458 | #ifdef CONFIG_SWAP | |
459 | /* | |
460 | * Anonymous process memory has backing store? | |
461 | * Try to allocate it some swap space here. | |
462 | */ | |
c340010e | 463 | if (PageAnon(page) && !PageSwapCache(page)) { |
f1fd1067 CL |
464 | if (!sc->may_swap) |
465 | goto keep_locked; | |
1480a540 | 466 | if (!add_to_swap(page, GFP_ATOMIC)) |
1da177e4 LT |
467 | goto activate_locked; |
468 | } | |
469 | #endif /* CONFIG_SWAP */ | |
470 | ||
471 | mapping = page_mapping(page); | |
472 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || | |
473 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
474 | ||
475 | /* | |
476 | * The page is mapped into the page tables of one or more | |
477 | * processes. Try to unmap it here. | |
478 | */ | |
479 | if (page_mapped(page) && mapping) { | |
aa3f18b3 CL |
480 | /* |
481 | * No unmapping if we do not swap | |
482 | */ | |
483 | if (!sc->may_swap) | |
484 | goto keep_locked; | |
485 | ||
a48d07af | 486 | switch (try_to_unmap(page, 0)) { |
1da177e4 LT |
487 | case SWAP_FAIL: |
488 | goto activate_locked; | |
489 | case SWAP_AGAIN: | |
490 | goto keep_locked; | |
491 | case SWAP_SUCCESS: | |
492 | ; /* try to free the page below */ | |
493 | } | |
494 | } | |
495 | ||
496 | if (PageDirty(page)) { | |
497 | if (referenced) | |
498 | goto keep_locked; | |
499 | if (!may_enter_fs) | |
500 | goto keep_locked; | |
52a8363e | 501 | if (!sc->may_writepage) |
1da177e4 LT |
502 | goto keep_locked; |
503 | ||
504 | /* Page is dirty, try to write it out here */ | |
505 | switch(pageout(page, mapping)) { | |
506 | case PAGE_KEEP: | |
507 | goto keep_locked; | |
508 | case PAGE_ACTIVATE: | |
509 | goto activate_locked; | |
510 | case PAGE_SUCCESS: | |
511 | if (PageWriteback(page) || PageDirty(page)) | |
512 | goto keep; | |
513 | /* | |
514 | * A synchronous write - probably a ramdisk. Go | |
515 | * ahead and try to reclaim the page. | |
516 | */ | |
517 | if (TestSetPageLocked(page)) | |
518 | goto keep; | |
519 | if (PageDirty(page) || PageWriteback(page)) | |
520 | goto keep_locked; | |
521 | mapping = page_mapping(page); | |
522 | case PAGE_CLEAN: | |
523 | ; /* try to free the page below */ | |
524 | } | |
525 | } | |
526 | ||
527 | /* | |
528 | * If the page has buffers, try to free the buffer mappings | |
529 | * associated with this page. If we succeed we try to free | |
530 | * the page as well. | |
531 | * | |
532 | * We do this even if the page is PageDirty(). | |
533 | * try_to_release_page() does not perform I/O, but it is | |
534 | * possible for a page to have PageDirty set, but it is actually | |
535 | * clean (all its buffers are clean). This happens if the | |
536 | * buffers were written out directly, with submit_bh(). ext3 | |
537 | * will do this, as well as the blockdev mapping. | |
538 | * try_to_release_page() will discover that cleanness and will | |
539 | * drop the buffers and mark the page clean - it can be freed. | |
540 | * | |
541 | * Rarely, pages can have buffers and no ->mapping. These are | |
542 | * the pages which were not successfully invalidated in | |
543 | * truncate_complete_page(). We try to drop those buffers here | |
544 | * and if that worked, and the page is no longer mapped into | |
545 | * process address space (page_count == 1) it can be freed. | |
546 | * Otherwise, leave the page on the LRU so it is swappable. | |
547 | */ | |
548 | if (PagePrivate(page)) { | |
549 | if (!try_to_release_page(page, sc->gfp_mask)) | |
550 | goto activate_locked; | |
551 | if (!mapping && page_count(page) == 1) | |
552 | goto free_it; | |
553 | } | |
554 | ||
49d2e9cc CL |
555 | if (!remove_mapping(mapping, page)) |
556 | goto keep_locked; | |
1da177e4 LT |
557 | |
558 | free_it: | |
559 | unlock_page(page); | |
05ff5137 | 560 | nr_reclaimed++; |
1da177e4 LT |
561 | if (!pagevec_add(&freed_pvec, page)) |
562 | __pagevec_release_nonlru(&freed_pvec); | |
563 | continue; | |
564 | ||
565 | activate_locked: | |
566 | SetPageActive(page); | |
567 | pgactivate++; | |
568 | keep_locked: | |
569 | unlock_page(page); | |
570 | keep: | |
571 | list_add(&page->lru, &ret_pages); | |
572 | BUG_ON(PageLRU(page)); | |
573 | } | |
574 | list_splice(&ret_pages, page_list); | |
575 | if (pagevec_count(&freed_pvec)) | |
576 | __pagevec_release_nonlru(&freed_pvec); | |
577 | mod_page_state(pgactivate, pgactivate); | |
05ff5137 | 578 | return nr_reclaimed; |
1da177e4 LT |
579 | } |
580 | ||
7cbe34cf | 581 | #ifdef CONFIG_MIGRATION |
8419c318 CL |
582 | static inline void move_to_lru(struct page *page) |
583 | { | |
584 | list_del(&page->lru); | |
585 | if (PageActive(page)) { | |
586 | /* | |
587 | * lru_cache_add_active checks that | |
588 | * the PG_active bit is off. | |
589 | */ | |
590 | ClearPageActive(page); | |
591 | lru_cache_add_active(page); | |
592 | } else { | |
593 | lru_cache_add(page); | |
594 | } | |
595 | put_page(page); | |
596 | } | |
597 | ||
598 | /* | |
053837fc | 599 | * Add isolated pages on the list back to the LRU. |
8419c318 CL |
600 | * |
601 | * returns the number of pages put back. | |
602 | */ | |
69e05944 | 603 | unsigned long putback_lru_pages(struct list_head *l) |
8419c318 CL |
604 | { |
605 | struct page *page; | |
606 | struct page *page2; | |
69e05944 | 607 | unsigned long count = 0; |
8419c318 CL |
608 | |
609 | list_for_each_entry_safe(page, page2, l, lru) { | |
610 | move_to_lru(page); | |
611 | count++; | |
612 | } | |
613 | return count; | |
614 | } | |
615 | ||
e965f963 CL |
616 | /* |
617 | * Non migratable page | |
618 | */ | |
619 | int fail_migrate_page(struct page *newpage, struct page *page) | |
620 | { | |
621 | return -EIO; | |
622 | } | |
623 | EXPORT_SYMBOL(fail_migrate_page); | |
624 | ||
49d2e9cc CL |
625 | /* |
626 | * swapout a single page | |
627 | * page is locked upon entry, unlocked on exit | |
49d2e9cc CL |
628 | */ |
629 | static int swap_page(struct page *page) | |
630 | { | |
631 | struct address_space *mapping = page_mapping(page); | |
632 | ||
633 | if (page_mapped(page) && mapping) | |
418aade4 | 634 | if (try_to_unmap(page, 1) != SWAP_SUCCESS) |
49d2e9cc CL |
635 | goto unlock_retry; |
636 | ||
637 | if (PageDirty(page)) { | |
638 | /* Page is dirty, try to write it out here */ | |
639 | switch(pageout(page, mapping)) { | |
640 | case PAGE_KEEP: | |
641 | case PAGE_ACTIVATE: | |
642 | goto unlock_retry; | |
643 | ||
644 | case PAGE_SUCCESS: | |
645 | goto retry; | |
646 | ||
647 | case PAGE_CLEAN: | |
648 | ; /* try to free the page below */ | |
649 | } | |
650 | } | |
651 | ||
652 | if (PagePrivate(page)) { | |
653 | if (!try_to_release_page(page, GFP_KERNEL) || | |
654 | (!mapping && page_count(page) == 1)) | |
655 | goto unlock_retry; | |
656 | } | |
657 | ||
658 | if (remove_mapping(mapping, page)) { | |
659 | /* Success */ | |
660 | unlock_page(page); | |
661 | return 0; | |
662 | } | |
663 | ||
664 | unlock_retry: | |
665 | unlock_page(page); | |
666 | ||
667 | retry: | |
d0d96328 | 668 | return -EAGAIN; |
49d2e9cc | 669 | } |
e965f963 | 670 | EXPORT_SYMBOL(swap_page); |
a48d07af CL |
671 | |
672 | /* | |
673 | * Page migration was first developed in the context of the memory hotplug | |
674 | * project. The main authors of the migration code are: | |
675 | * | |
676 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> | |
677 | * Hirokazu Takahashi <taka@valinux.co.jp> | |
678 | * Dave Hansen <haveblue@us.ibm.com> | |
679 | * Christoph Lameter <clameter@sgi.com> | |
680 | */ | |
681 | ||
682 | /* | |
683 | * Remove references for a page and establish the new page with the correct | |
684 | * basic settings to be able to stop accesses to the page. | |
685 | */ | |
e965f963 | 686 | int migrate_page_remove_references(struct page *newpage, |
a48d07af CL |
687 | struct page *page, int nr_refs) |
688 | { | |
689 | struct address_space *mapping = page_mapping(page); | |
690 | struct page **radix_pointer; | |
691 | ||
692 | /* | |
693 | * Avoid doing any of the following work if the page count | |
694 | * indicates that the page is in use or truncate has removed | |
695 | * the page. | |
696 | */ | |
697 | if (!mapping || page_mapcount(page) + nr_refs != page_count(page)) | |
4983da07 | 698 | return -EAGAIN; |
a48d07af CL |
699 | |
700 | /* | |
701 | * Establish swap ptes for anonymous pages or destroy pte | |
702 | * maps for files. | |
703 | * | |
704 | * In order to reestablish file backed mappings the fault handlers | |
705 | * will take the radix tree_lock which may then be used to stop | |
706 | * processses from accessing this page until the new page is ready. | |
707 | * | |
708 | * A process accessing via a swap pte (an anonymous page) will take a | |
709 | * page_lock on the old page which will block the process until the | |
710 | * migration attempt is complete. At that time the PageSwapCache bit | |
711 | * will be examined. If the page was migrated then the PageSwapCache | |
712 | * bit will be clear and the operation to retrieve the page will be | |
713 | * retried which will find the new page in the radix tree. Then a new | |
714 | * direct mapping may be generated based on the radix tree contents. | |
715 | * | |
716 | * If the page was not migrated then the PageSwapCache bit | |
717 | * is still set and the operation may continue. | |
718 | */ | |
4983da07 CL |
719 | if (try_to_unmap(page, 1) == SWAP_FAIL) |
720 | /* A vma has VM_LOCKED set -> Permanent failure */ | |
721 | return -EPERM; | |
a48d07af CL |
722 | |
723 | /* | |
724 | * Give up if we were unable to remove all mappings. | |
725 | */ | |
726 | if (page_mapcount(page)) | |
4983da07 | 727 | return -EAGAIN; |
a48d07af CL |
728 | |
729 | write_lock_irq(&mapping->tree_lock); | |
730 | ||
731 | radix_pointer = (struct page **)radix_tree_lookup_slot( | |
732 | &mapping->page_tree, | |
733 | page_index(page)); | |
734 | ||
735 | if (!page_mapping(page) || page_count(page) != nr_refs || | |
736 | *radix_pointer != page) { | |
737 | write_unlock_irq(&mapping->tree_lock); | |
4983da07 | 738 | return -EAGAIN; |
a48d07af CL |
739 | } |
740 | ||
741 | /* | |
742 | * Now we know that no one else is looking at the page. | |
743 | * | |
744 | * Certain minimal information about a page must be available | |
745 | * in order for other subsystems to properly handle the page if they | |
746 | * find it through the radix tree update before we are finished | |
747 | * copying the page. | |
748 | */ | |
749 | get_page(newpage); | |
750 | newpage->index = page->index; | |
751 | newpage->mapping = page->mapping; | |
752 | if (PageSwapCache(page)) { | |
753 | SetPageSwapCache(newpage); | |
754 | set_page_private(newpage, page_private(page)); | |
755 | } | |
756 | ||
757 | *radix_pointer = newpage; | |
758 | __put_page(page); | |
759 | write_unlock_irq(&mapping->tree_lock); | |
760 | ||
761 | return 0; | |
762 | } | |
e965f963 | 763 | EXPORT_SYMBOL(migrate_page_remove_references); |
a48d07af CL |
764 | |
765 | /* | |
766 | * Copy the page to its new location | |
767 | */ | |
768 | void migrate_page_copy(struct page *newpage, struct page *page) | |
769 | { | |
770 | copy_highpage(newpage, page); | |
771 | ||
772 | if (PageError(page)) | |
773 | SetPageError(newpage); | |
774 | if (PageReferenced(page)) | |
775 | SetPageReferenced(newpage); | |
776 | if (PageUptodate(page)) | |
777 | SetPageUptodate(newpage); | |
778 | if (PageActive(page)) | |
779 | SetPageActive(newpage); | |
780 | if (PageChecked(page)) | |
781 | SetPageChecked(newpage); | |
782 | if (PageMappedToDisk(page)) | |
783 | SetPageMappedToDisk(newpage); | |
784 | ||
785 | if (PageDirty(page)) { | |
786 | clear_page_dirty_for_io(page); | |
787 | set_page_dirty(newpage); | |
788 | } | |
789 | ||
790 | ClearPageSwapCache(page); | |
791 | ClearPageActive(page); | |
792 | ClearPagePrivate(page); | |
793 | set_page_private(page, 0); | |
794 | page->mapping = NULL; | |
795 | ||
796 | /* | |
797 | * If any waiters have accumulated on the new page then | |
798 | * wake them up. | |
799 | */ | |
800 | if (PageWriteback(newpage)) | |
801 | end_page_writeback(newpage); | |
802 | } | |
e965f963 | 803 | EXPORT_SYMBOL(migrate_page_copy); |
a48d07af CL |
804 | |
805 | /* | |
806 | * Common logic to directly migrate a single page suitable for | |
807 | * pages that do not use PagePrivate. | |
808 | * | |
809 | * Pages are locked upon entry and exit. | |
810 | */ | |
811 | int migrate_page(struct page *newpage, struct page *page) | |
812 | { | |
4983da07 CL |
813 | int rc; |
814 | ||
a48d07af CL |
815 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ |
816 | ||
4983da07 CL |
817 | rc = migrate_page_remove_references(newpage, page, 2); |
818 | ||
819 | if (rc) | |
820 | return rc; | |
a48d07af CL |
821 | |
822 | migrate_page_copy(newpage, page); | |
823 | ||
a3351e52 CL |
824 | /* |
825 | * Remove auxiliary swap entries and replace | |
826 | * them with real ptes. | |
827 | * | |
828 | * Note that a real pte entry will allow processes that are not | |
829 | * waiting on the page lock to use the new page via the page tables | |
830 | * before the new page is unlocked. | |
831 | */ | |
832 | remove_from_swap(newpage); | |
a48d07af CL |
833 | return 0; |
834 | } | |
e965f963 | 835 | EXPORT_SYMBOL(migrate_page); |
a48d07af | 836 | |
49d2e9cc CL |
837 | /* |
838 | * migrate_pages | |
839 | * | |
840 | * Two lists are passed to this function. The first list | |
841 | * contains the pages isolated from the LRU to be migrated. | |
842 | * The second list contains new pages that the pages isolated | |
843 | * can be moved to. If the second list is NULL then all | |
844 | * pages are swapped out. | |
845 | * | |
846 | * The function returns after 10 attempts or if no pages | |
418aade4 | 847 | * are movable anymore because to has become empty |
49d2e9cc CL |
848 | * or no retryable pages exist anymore. |
849 | * | |
d0d96328 | 850 | * Return: Number of pages not migrated when "to" ran empty. |
49d2e9cc | 851 | */ |
69e05944 | 852 | unsigned long migrate_pages(struct list_head *from, struct list_head *to, |
d4984711 | 853 | struct list_head *moved, struct list_head *failed) |
49d2e9cc | 854 | { |
69e05944 AM |
855 | unsigned long retry; |
856 | unsigned long nr_failed = 0; | |
49d2e9cc CL |
857 | int pass = 0; |
858 | struct page *page; | |
859 | struct page *page2; | |
860 | int swapwrite = current->flags & PF_SWAPWRITE; | |
d0d96328 | 861 | int rc; |
49d2e9cc CL |
862 | |
863 | if (!swapwrite) | |
864 | current->flags |= PF_SWAPWRITE; | |
865 | ||
866 | redo: | |
867 | retry = 0; | |
868 | ||
d4984711 | 869 | list_for_each_entry_safe(page, page2, from, lru) { |
a48d07af CL |
870 | struct page *newpage = NULL; |
871 | struct address_space *mapping; | |
872 | ||
49d2e9cc CL |
873 | cond_resched(); |
874 | ||
d0d96328 CL |
875 | rc = 0; |
876 | if (page_count(page) == 1) | |
ee27497d | 877 | /* page was freed from under us. So we are done. */ |
d0d96328 CL |
878 | goto next; |
879 | ||
a48d07af CL |
880 | if (to && list_empty(to)) |
881 | break; | |
882 | ||
49d2e9cc CL |
883 | /* |
884 | * Skip locked pages during the first two passes to give the | |
7cbe34cf CL |
885 | * functions holding the lock time to release the page. Later we |
886 | * use lock_page() to have a higher chance of acquiring the | |
887 | * lock. | |
49d2e9cc | 888 | */ |
d0d96328 | 889 | rc = -EAGAIN; |
49d2e9cc CL |
890 | if (pass > 2) |
891 | lock_page(page); | |
892 | else | |
893 | if (TestSetPageLocked(page)) | |
d0d96328 | 894 | goto next; |
49d2e9cc CL |
895 | |
896 | /* | |
897 | * Only wait on writeback if we have already done a pass where | |
898 | * we we may have triggered writeouts for lots of pages. | |
899 | */ | |
7cbe34cf | 900 | if (pass > 0) { |
49d2e9cc | 901 | wait_on_page_writeback(page); |
7cbe34cf | 902 | } else { |
d0d96328 CL |
903 | if (PageWriteback(page)) |
904 | goto unlock_page; | |
7cbe34cf | 905 | } |
49d2e9cc | 906 | |
d0d96328 CL |
907 | /* |
908 | * Anonymous pages must have swap cache references otherwise | |
909 | * the information contained in the page maps cannot be | |
910 | * preserved. | |
911 | */ | |
49d2e9cc | 912 | if (PageAnon(page) && !PageSwapCache(page)) { |
1480a540 | 913 | if (!add_to_swap(page, GFP_KERNEL)) { |
d0d96328 CL |
914 | rc = -ENOMEM; |
915 | goto unlock_page; | |
49d2e9cc CL |
916 | } |
917 | } | |
49d2e9cc | 918 | |
a48d07af CL |
919 | if (!to) { |
920 | rc = swap_page(page); | |
921 | goto next; | |
922 | } | |
923 | ||
924 | newpage = lru_to_page(to); | |
925 | lock_page(newpage); | |
926 | ||
49d2e9cc | 927 | /* |
a48d07af | 928 | * Pages are properly locked and writeback is complete. |
49d2e9cc CL |
929 | * Try to migrate the page. |
930 | */ | |
a48d07af CL |
931 | mapping = page_mapping(page); |
932 | if (!mapping) | |
933 | goto unlock_both; | |
934 | ||
e965f963 | 935 | if (mapping->a_ops->migratepage) { |
418aade4 CL |
936 | /* |
937 | * Most pages have a mapping and most filesystems | |
938 | * should provide a migration function. Anonymous | |
939 | * pages are part of swap space which also has its | |
940 | * own migration function. This is the most common | |
941 | * path for page migration. | |
942 | */ | |
e965f963 CL |
943 | rc = mapping->a_ops->migratepage(newpage, page); |
944 | goto unlock_both; | |
945 | } | |
946 | ||
a48d07af | 947 | /* |
418aade4 CL |
948 | * Default handling if a filesystem does not provide |
949 | * a migration function. We can only migrate clean | |
950 | * pages so try to write out any dirty pages first. | |
a48d07af CL |
951 | */ |
952 | if (PageDirty(page)) { | |
953 | switch (pageout(page, mapping)) { | |
954 | case PAGE_KEEP: | |
955 | case PAGE_ACTIVATE: | |
956 | goto unlock_both; | |
957 | ||
958 | case PAGE_SUCCESS: | |
959 | unlock_page(newpage); | |
960 | goto next; | |
961 | ||
962 | case PAGE_CLEAN: | |
963 | ; /* try to migrate the page below */ | |
964 | } | |
965 | } | |
418aade4 | 966 | |
a48d07af | 967 | /* |
418aade4 CL |
968 | * Buffers are managed in a filesystem specific way. |
969 | * We must have no buffers or drop them. | |
a48d07af CL |
970 | */ |
971 | if (!page_has_buffers(page) || | |
972 | try_to_release_page(page, GFP_KERNEL)) { | |
973 | rc = migrate_page(newpage, page); | |
974 | goto unlock_both; | |
975 | } | |
976 | ||
977 | /* | |
978 | * On early passes with mapped pages simply | |
979 | * retry. There may be a lock held for some | |
980 | * buffers that may go away. Later | |
981 | * swap them out. | |
982 | */ | |
983 | if (pass > 4) { | |
418aade4 CL |
984 | /* |
985 | * Persistently unable to drop buffers..... As a | |
986 | * measure of last resort we fall back to | |
987 | * swap_page(). | |
988 | */ | |
a48d07af CL |
989 | unlock_page(newpage); |
990 | newpage = NULL; | |
991 | rc = swap_page(page); | |
992 | goto next; | |
993 | } | |
994 | ||
995 | unlock_both: | |
996 | unlock_page(newpage); | |
d0d96328 CL |
997 | |
998 | unlock_page: | |
999 | unlock_page(page); | |
1000 | ||
1001 | next: | |
1002 | if (rc == -EAGAIN) { | |
1003 | retry++; | |
1004 | } else if (rc) { | |
1005 | /* Permanent failure */ | |
1006 | list_move(&page->lru, failed); | |
1007 | nr_failed++; | |
1008 | } else { | |
a48d07af CL |
1009 | if (newpage) { |
1010 | /* Successful migration. Return page to LRU */ | |
1011 | move_to_lru(newpage); | |
1012 | } | |
d4984711 | 1013 | list_move(&page->lru, moved); |
d4984711 | 1014 | } |
49d2e9cc CL |
1015 | } |
1016 | if (retry && pass++ < 10) | |
1017 | goto redo; | |
1018 | ||
1019 | if (!swapwrite) | |
1020 | current->flags &= ~PF_SWAPWRITE; | |
1021 | ||
49d2e9cc CL |
1022 | return nr_failed + retry; |
1023 | } | |
8419c318 | 1024 | |
8419c318 CL |
1025 | /* |
1026 | * Isolate one page from the LRU lists and put it on the | |
053837fc | 1027 | * indicated list with elevated refcount. |
8419c318 CL |
1028 | * |
1029 | * Result: | |
1030 | * 0 = page not on LRU list | |
1031 | * 1 = page removed from LRU list and added to the specified list. | |
8419c318 CL |
1032 | */ |
1033 | int isolate_lru_page(struct page *page) | |
1034 | { | |
053837fc | 1035 | int ret = 0; |
8419c318 | 1036 | |
053837fc NP |
1037 | if (PageLRU(page)) { |
1038 | struct zone *zone = page_zone(page); | |
1039 | spin_lock_irq(&zone->lru_lock); | |
8d438f96 | 1040 | if (PageLRU(page)) { |
053837fc NP |
1041 | ret = 1; |
1042 | get_page(page); | |
8d438f96 | 1043 | ClearPageLRU(page); |
053837fc NP |
1044 | if (PageActive(page)) |
1045 | del_page_from_active_list(zone, page); | |
1046 | else | |
1047 | del_page_from_inactive_list(zone, page); | |
1048 | } | |
1049 | spin_unlock_irq(&zone->lru_lock); | |
8419c318 | 1050 | } |
053837fc NP |
1051 | |
1052 | return ret; | |
8419c318 | 1053 | } |
7cbe34cf | 1054 | #endif |
49d2e9cc | 1055 | |
1da177e4 LT |
1056 | /* |
1057 | * zone->lru_lock is heavily contended. Some of the functions that | |
1058 | * shrink the lists perform better by taking out a batch of pages | |
1059 | * and working on them outside the LRU lock. | |
1060 | * | |
1061 | * For pagecache intensive workloads, this function is the hottest | |
1062 | * spot in the kernel (apart from copy_*_user functions). | |
1063 | * | |
1064 | * Appropriate locks must be held before calling this function. | |
1065 | * | |
1066 | * @nr_to_scan: The number of pages to look through on the list. | |
1067 | * @src: The LRU list to pull pages off. | |
1068 | * @dst: The temp list to put pages on to. | |
1069 | * @scanned: The number of pages that were scanned. | |
1070 | * | |
1071 | * returns how many pages were moved onto *@dst. | |
1072 | */ | |
69e05944 AM |
1073 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
1074 | struct list_head *src, struct list_head *dst, | |
1075 | unsigned long *scanned) | |
1da177e4 | 1076 | { |
69e05944 | 1077 | unsigned long nr_taken = 0; |
1da177e4 | 1078 | struct page *page; |
c9b02d97 | 1079 | unsigned long scan; |
1da177e4 | 1080 | |
c9b02d97 | 1081 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
7c8ee9a8 | 1082 | struct list_head *target; |
1da177e4 LT |
1083 | page = lru_to_page(src); |
1084 | prefetchw_prev_lru_page(page, src, flags); | |
1085 | ||
8d438f96 NP |
1086 | BUG_ON(!PageLRU(page)); |
1087 | ||
053837fc | 1088 | list_del(&page->lru); |
7c8ee9a8 NP |
1089 | target = src; |
1090 | if (likely(get_page_unless_zero(page))) { | |
053837fc | 1091 | /* |
7c8ee9a8 NP |
1092 | * Be careful not to clear PageLRU until after we're |
1093 | * sure the page is not being freed elsewhere -- the | |
1094 | * page release code relies on it. | |
053837fc | 1095 | */ |
7c8ee9a8 NP |
1096 | ClearPageLRU(page); |
1097 | target = dst; | |
1098 | nr_taken++; | |
1099 | } /* else it is being freed elsewhere */ | |
46453a6e | 1100 | |
7c8ee9a8 | 1101 | list_add(&page->lru, target); |
1da177e4 LT |
1102 | } |
1103 | ||
1104 | *scanned = scan; | |
1105 | return nr_taken; | |
1106 | } | |
1107 | ||
1108 | /* | |
1742f19f AM |
1109 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1110 | * of reclaimed pages | |
1da177e4 | 1111 | */ |
1742f19f AM |
1112 | static unsigned long shrink_inactive_list(unsigned long max_scan, |
1113 | struct zone *zone, struct scan_control *sc) | |
1da177e4 LT |
1114 | { |
1115 | LIST_HEAD(page_list); | |
1116 | struct pagevec pvec; | |
69e05944 | 1117 | unsigned long nr_scanned = 0; |
05ff5137 | 1118 | unsigned long nr_reclaimed = 0; |
1da177e4 LT |
1119 | |
1120 | pagevec_init(&pvec, 1); | |
1121 | ||
1122 | lru_add_drain(); | |
1123 | spin_lock_irq(&zone->lru_lock); | |
69e05944 | 1124 | do { |
1da177e4 | 1125 | struct page *page; |
69e05944 AM |
1126 | unsigned long nr_taken; |
1127 | unsigned long nr_scan; | |
1128 | unsigned long nr_freed; | |
1da177e4 LT |
1129 | |
1130 | nr_taken = isolate_lru_pages(sc->swap_cluster_max, | |
1131 | &zone->inactive_list, | |
1132 | &page_list, &nr_scan); | |
1133 | zone->nr_inactive -= nr_taken; | |
1134 | zone->pages_scanned += nr_scan; | |
1135 | spin_unlock_irq(&zone->lru_lock); | |
1136 | ||
69e05944 | 1137 | nr_scanned += nr_scan; |
1742f19f | 1138 | nr_freed = shrink_page_list(&page_list, sc); |
05ff5137 | 1139 | nr_reclaimed += nr_freed; |
a74609fa NP |
1140 | local_irq_disable(); |
1141 | if (current_is_kswapd()) { | |
1142 | __mod_page_state_zone(zone, pgscan_kswapd, nr_scan); | |
1143 | __mod_page_state(kswapd_steal, nr_freed); | |
1144 | } else | |
1145 | __mod_page_state_zone(zone, pgscan_direct, nr_scan); | |
1146 | __mod_page_state_zone(zone, pgsteal, nr_freed); | |
1147 | ||
fb8d14e1 WF |
1148 | if (nr_taken == 0) |
1149 | goto done; | |
1150 | ||
a74609fa | 1151 | spin_lock(&zone->lru_lock); |
1da177e4 LT |
1152 | /* |
1153 | * Put back any unfreeable pages. | |
1154 | */ | |
1155 | while (!list_empty(&page_list)) { | |
1156 | page = lru_to_page(&page_list); | |
8d438f96 NP |
1157 | BUG_ON(PageLRU(page)); |
1158 | SetPageLRU(page); | |
1da177e4 LT |
1159 | list_del(&page->lru); |
1160 | if (PageActive(page)) | |
1161 | add_page_to_active_list(zone, page); | |
1162 | else | |
1163 | add_page_to_inactive_list(zone, page); | |
1164 | if (!pagevec_add(&pvec, page)) { | |
1165 | spin_unlock_irq(&zone->lru_lock); | |
1166 | __pagevec_release(&pvec); | |
1167 | spin_lock_irq(&zone->lru_lock); | |
1168 | } | |
1169 | } | |
69e05944 | 1170 | } while (nr_scanned < max_scan); |
fb8d14e1 | 1171 | spin_unlock(&zone->lru_lock); |
1da177e4 | 1172 | done: |
fb8d14e1 | 1173 | local_irq_enable(); |
1da177e4 | 1174 | pagevec_release(&pvec); |
05ff5137 | 1175 | return nr_reclaimed; |
1da177e4 LT |
1176 | } |
1177 | ||
1178 | /* | |
1179 | * This moves pages from the active list to the inactive list. | |
1180 | * | |
1181 | * We move them the other way if the page is referenced by one or more | |
1182 | * processes, from rmap. | |
1183 | * | |
1184 | * If the pages are mostly unmapped, the processing is fast and it is | |
1185 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1186 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1187 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1188 | * this, so instead we remove the pages from the LRU while processing them. | |
1189 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1190 | * nobody will play with that bit on a non-LRU page. | |
1191 | * | |
1192 | * The downside is that we have to touch page->_count against each page. | |
1193 | * But we had to alter page->flags anyway. | |
1194 | */ | |
1742f19f AM |
1195 | static void shrink_active_list(unsigned long nr_pages, struct zone *zone, |
1196 | struct scan_control *sc) | |
1da177e4 | 1197 | { |
69e05944 | 1198 | unsigned long pgmoved; |
1da177e4 | 1199 | int pgdeactivate = 0; |
69e05944 | 1200 | unsigned long pgscanned; |
1da177e4 LT |
1201 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
1202 | LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */ | |
1203 | LIST_HEAD(l_active); /* Pages to go onto the active_list */ | |
1204 | struct page *page; | |
1205 | struct pagevec pvec; | |
1206 | int reclaim_mapped = 0; | |
2903fb16 CL |
1207 | |
1208 | if (unlikely(sc->may_swap)) { | |
1209 | long mapped_ratio; | |
1210 | long distress; | |
1211 | long swap_tendency; | |
1212 | ||
1213 | /* | |
1214 | * `distress' is a measure of how much trouble we're having | |
1215 | * reclaiming pages. 0 -> no problems. 100 -> great trouble. | |
1216 | */ | |
1217 | distress = 100 >> zone->prev_priority; | |
1218 | ||
1219 | /* | |
1220 | * The point of this algorithm is to decide when to start | |
1221 | * reclaiming mapped memory instead of just pagecache. Work out | |
1222 | * how much memory | |
1223 | * is mapped. | |
1224 | */ | |
1225 | mapped_ratio = (sc->nr_mapped * 100) / total_memory; | |
1226 | ||
1227 | /* | |
1228 | * Now decide how much we really want to unmap some pages. The | |
1229 | * mapped ratio is downgraded - just because there's a lot of | |
1230 | * mapped memory doesn't necessarily mean that page reclaim | |
1231 | * isn't succeeding. | |
1232 | * | |
1233 | * The distress ratio is important - we don't want to start | |
1234 | * going oom. | |
1235 | * | |
1236 | * A 100% value of vm_swappiness overrides this algorithm | |
1237 | * altogether. | |
1238 | */ | |
1239 | swap_tendency = mapped_ratio / 2 + distress + vm_swappiness; | |
1240 | ||
1241 | /* | |
1242 | * Now use this metric to decide whether to start moving mapped | |
1243 | * memory onto the inactive list. | |
1244 | */ | |
1245 | if (swap_tendency >= 100) | |
1246 | reclaim_mapped = 1; | |
1247 | } | |
1da177e4 LT |
1248 | |
1249 | lru_add_drain(); | |
1250 | spin_lock_irq(&zone->lru_lock); | |
1251 | pgmoved = isolate_lru_pages(nr_pages, &zone->active_list, | |
1252 | &l_hold, &pgscanned); | |
1253 | zone->pages_scanned += pgscanned; | |
1254 | zone->nr_active -= pgmoved; | |
1255 | spin_unlock_irq(&zone->lru_lock); | |
1256 | ||
1da177e4 LT |
1257 | while (!list_empty(&l_hold)) { |
1258 | cond_resched(); | |
1259 | page = lru_to_page(&l_hold); | |
1260 | list_del(&page->lru); | |
1261 | if (page_mapped(page)) { | |
1262 | if (!reclaim_mapped || | |
1263 | (total_swap_pages == 0 && PageAnon(page)) || | |
f7b7fd8f | 1264 | page_referenced(page, 0)) { |
1da177e4 LT |
1265 | list_add(&page->lru, &l_active); |
1266 | continue; | |
1267 | } | |
1268 | } | |
1269 | list_add(&page->lru, &l_inactive); | |
1270 | } | |
1271 | ||
1272 | pagevec_init(&pvec, 1); | |
1273 | pgmoved = 0; | |
1274 | spin_lock_irq(&zone->lru_lock); | |
1275 | while (!list_empty(&l_inactive)) { | |
1276 | page = lru_to_page(&l_inactive); | |
1277 | prefetchw_prev_lru_page(page, &l_inactive, flags); | |
8d438f96 NP |
1278 | BUG_ON(PageLRU(page)); |
1279 | SetPageLRU(page); | |
4c84cacf NP |
1280 | BUG_ON(!PageActive(page)); |
1281 | ClearPageActive(page); | |
1282 | ||
1da177e4 LT |
1283 | list_move(&page->lru, &zone->inactive_list); |
1284 | pgmoved++; | |
1285 | if (!pagevec_add(&pvec, page)) { | |
1286 | zone->nr_inactive += pgmoved; | |
1287 | spin_unlock_irq(&zone->lru_lock); | |
1288 | pgdeactivate += pgmoved; | |
1289 | pgmoved = 0; | |
1290 | if (buffer_heads_over_limit) | |
1291 | pagevec_strip(&pvec); | |
1292 | __pagevec_release(&pvec); | |
1293 | spin_lock_irq(&zone->lru_lock); | |
1294 | } | |
1295 | } | |
1296 | zone->nr_inactive += pgmoved; | |
1297 | pgdeactivate += pgmoved; | |
1298 | if (buffer_heads_over_limit) { | |
1299 | spin_unlock_irq(&zone->lru_lock); | |
1300 | pagevec_strip(&pvec); | |
1301 | spin_lock_irq(&zone->lru_lock); | |
1302 | } | |
1303 | ||
1304 | pgmoved = 0; | |
1305 | while (!list_empty(&l_active)) { | |
1306 | page = lru_to_page(&l_active); | |
1307 | prefetchw_prev_lru_page(page, &l_active, flags); | |
8d438f96 NP |
1308 | BUG_ON(PageLRU(page)); |
1309 | SetPageLRU(page); | |
1da177e4 LT |
1310 | BUG_ON(!PageActive(page)); |
1311 | list_move(&page->lru, &zone->active_list); | |
1312 | pgmoved++; | |
1313 | if (!pagevec_add(&pvec, page)) { | |
1314 | zone->nr_active += pgmoved; | |
1315 | pgmoved = 0; | |
1316 | spin_unlock_irq(&zone->lru_lock); | |
1317 | __pagevec_release(&pvec); | |
1318 | spin_lock_irq(&zone->lru_lock); | |
1319 | } | |
1320 | } | |
1321 | zone->nr_active += pgmoved; | |
a74609fa NP |
1322 | spin_unlock(&zone->lru_lock); |
1323 | ||
1324 | __mod_page_state_zone(zone, pgrefill, pgscanned); | |
1325 | __mod_page_state(pgdeactivate, pgdeactivate); | |
1326 | local_irq_enable(); | |
1da177e4 | 1327 | |
a74609fa | 1328 | pagevec_release(&pvec); |
1da177e4 LT |
1329 | } |
1330 | ||
1331 | /* | |
1332 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1333 | */ | |
05ff5137 AM |
1334 | static unsigned long shrink_zone(int priority, struct zone *zone, |
1335 | struct scan_control *sc) | |
1da177e4 LT |
1336 | { |
1337 | unsigned long nr_active; | |
1338 | unsigned long nr_inactive; | |
8695949a | 1339 | unsigned long nr_to_scan; |
05ff5137 | 1340 | unsigned long nr_reclaimed = 0; |
1da177e4 | 1341 | |
53e9a615 MH |
1342 | atomic_inc(&zone->reclaim_in_progress); |
1343 | ||
1da177e4 LT |
1344 | /* |
1345 | * Add one to `nr_to_scan' just to make sure that the kernel will | |
1346 | * slowly sift through the active list. | |
1347 | */ | |
8695949a | 1348 | zone->nr_scan_active += (zone->nr_active >> priority) + 1; |
1da177e4 LT |
1349 | nr_active = zone->nr_scan_active; |
1350 | if (nr_active >= sc->swap_cluster_max) | |
1351 | zone->nr_scan_active = 0; | |
1352 | else | |
1353 | nr_active = 0; | |
1354 | ||
8695949a | 1355 | zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1; |
1da177e4 LT |
1356 | nr_inactive = zone->nr_scan_inactive; |
1357 | if (nr_inactive >= sc->swap_cluster_max) | |
1358 | zone->nr_scan_inactive = 0; | |
1359 | else | |
1360 | nr_inactive = 0; | |
1361 | ||
1da177e4 LT |
1362 | while (nr_active || nr_inactive) { |
1363 | if (nr_active) { | |
8695949a | 1364 | nr_to_scan = min(nr_active, |
1da177e4 | 1365 | (unsigned long)sc->swap_cluster_max); |
8695949a | 1366 | nr_active -= nr_to_scan; |
1742f19f | 1367 | shrink_active_list(nr_to_scan, zone, sc); |
1da177e4 LT |
1368 | } |
1369 | ||
1370 | if (nr_inactive) { | |
8695949a | 1371 | nr_to_scan = min(nr_inactive, |
1da177e4 | 1372 | (unsigned long)sc->swap_cluster_max); |
8695949a | 1373 | nr_inactive -= nr_to_scan; |
1742f19f AM |
1374 | nr_reclaimed += shrink_inactive_list(nr_to_scan, zone, |
1375 | sc); | |
1da177e4 LT |
1376 | } |
1377 | } | |
1378 | ||
1379 | throttle_vm_writeout(); | |
53e9a615 MH |
1380 | |
1381 | atomic_dec(&zone->reclaim_in_progress); | |
05ff5137 | 1382 | return nr_reclaimed; |
1da177e4 LT |
1383 | } |
1384 | ||
1385 | /* | |
1386 | * This is the direct reclaim path, for page-allocating processes. We only | |
1387 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
1388 | * request. | |
1389 | * | |
1390 | * We reclaim from a zone even if that zone is over pages_high. Because: | |
1391 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order | |
1392 | * allocation or | |
1393 | * b) The zones may be over pages_high but they must go *over* pages_high to | |
1394 | * satisfy the `incremental min' zone defense algorithm. | |
1395 | * | |
1396 | * Returns the number of reclaimed pages. | |
1397 | * | |
1398 | * If a zone is deemed to be full of pinned pages then just give it a light | |
1399 | * scan then give up on it. | |
1400 | */ | |
1742f19f | 1401 | static unsigned long shrink_zones(int priority, struct zone **zones, |
05ff5137 | 1402 | struct scan_control *sc) |
1da177e4 | 1403 | { |
05ff5137 | 1404 | unsigned long nr_reclaimed = 0; |
1da177e4 LT |
1405 | int i; |
1406 | ||
1407 | for (i = 0; zones[i] != NULL; i++) { | |
1408 | struct zone *zone = zones[i]; | |
1409 | ||
f3fe6512 | 1410 | if (!populated_zone(zone)) |
1da177e4 LT |
1411 | continue; |
1412 | ||
9bf2229f | 1413 | if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) |
1da177e4 LT |
1414 | continue; |
1415 | ||
8695949a CL |
1416 | zone->temp_priority = priority; |
1417 | if (zone->prev_priority > priority) | |
1418 | zone->prev_priority = priority; | |
1da177e4 | 1419 | |
8695949a | 1420 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
1da177e4 LT |
1421 | continue; /* Let kswapd poll it */ |
1422 | ||
05ff5137 | 1423 | nr_reclaimed += shrink_zone(priority, zone, sc); |
1da177e4 | 1424 | } |
05ff5137 | 1425 | return nr_reclaimed; |
1da177e4 LT |
1426 | } |
1427 | ||
1428 | /* | |
1429 | * This is the main entry point to direct page reclaim. | |
1430 | * | |
1431 | * If a full scan of the inactive list fails to free enough memory then we | |
1432 | * are "out of memory" and something needs to be killed. | |
1433 | * | |
1434 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
1435 | * high - the zone may be full of dirty or under-writeback pages, which this | |
1436 | * caller can't do much about. We kick pdflush and take explicit naps in the | |
1437 | * hope that some of these pages can be written. But if the allocating task | |
1438 | * holds filesystem locks which prevent writeout this might not work, and the | |
1439 | * allocation attempt will fail. | |
1440 | */ | |
69e05944 | 1441 | unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask) |
1da177e4 LT |
1442 | { |
1443 | int priority; | |
1444 | int ret = 0; | |
69e05944 | 1445 | unsigned long total_scanned = 0; |
05ff5137 | 1446 | unsigned long nr_reclaimed = 0; |
1da177e4 | 1447 | struct reclaim_state *reclaim_state = current->reclaim_state; |
1da177e4 LT |
1448 | unsigned long lru_pages = 0; |
1449 | int i; | |
179e9639 AM |
1450 | struct scan_control sc = { |
1451 | .gfp_mask = gfp_mask, | |
1452 | .may_writepage = !laptop_mode, | |
1453 | .swap_cluster_max = SWAP_CLUSTER_MAX, | |
1454 | .may_swap = 1, | |
1455 | }; | |
1da177e4 LT |
1456 | |
1457 | inc_page_state(allocstall); | |
1458 | ||
1459 | for (i = 0; zones[i] != NULL; i++) { | |
1460 | struct zone *zone = zones[i]; | |
1461 | ||
9bf2229f | 1462 | if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) |
1da177e4 LT |
1463 | continue; |
1464 | ||
1465 | zone->temp_priority = DEF_PRIORITY; | |
1466 | lru_pages += zone->nr_active + zone->nr_inactive; | |
1467 | } | |
1468 | ||
1469 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { | |
1470 | sc.nr_mapped = read_page_state(nr_mapped); | |
1471 | sc.nr_scanned = 0; | |
f7b7fd8f RR |
1472 | if (!priority) |
1473 | disable_swap_token(); | |
1742f19f | 1474 | nr_reclaimed += shrink_zones(priority, zones, &sc); |
1da177e4 LT |
1475 | shrink_slab(sc.nr_scanned, gfp_mask, lru_pages); |
1476 | if (reclaim_state) { | |
05ff5137 | 1477 | nr_reclaimed += reclaim_state->reclaimed_slab; |
1da177e4 LT |
1478 | reclaim_state->reclaimed_slab = 0; |
1479 | } | |
1480 | total_scanned += sc.nr_scanned; | |
05ff5137 | 1481 | if (nr_reclaimed >= sc.swap_cluster_max) { |
1da177e4 LT |
1482 | ret = 1; |
1483 | goto out; | |
1484 | } | |
1485 | ||
1486 | /* | |
1487 | * Try to write back as many pages as we just scanned. This | |
1488 | * tends to cause slow streaming writers to write data to the | |
1489 | * disk smoothly, at the dirtying rate, which is nice. But | |
1490 | * that's undesirable in laptop mode, where we *want* lumpy | |
1491 | * writeout. So in laptop mode, write out the whole world. | |
1492 | */ | |
179e9639 AM |
1493 | if (total_scanned > sc.swap_cluster_max + |
1494 | sc.swap_cluster_max / 2) { | |
687a21ce | 1495 | wakeup_pdflush(laptop_mode ? 0 : total_scanned); |
1da177e4 LT |
1496 | sc.may_writepage = 1; |
1497 | } | |
1498 | ||
1499 | /* Take a nap, wait for some writeback to complete */ | |
1500 | if (sc.nr_scanned && priority < DEF_PRIORITY - 2) | |
1501 | blk_congestion_wait(WRITE, HZ/10); | |
1502 | } | |
1503 | out: | |
1504 | for (i = 0; zones[i] != 0; i++) { | |
1505 | struct zone *zone = zones[i]; | |
1506 | ||
9bf2229f | 1507 | if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) |
1da177e4 LT |
1508 | continue; |
1509 | ||
1510 | zone->prev_priority = zone->temp_priority; | |
1511 | } | |
1512 | return ret; | |
1513 | } | |
1514 | ||
1515 | /* | |
1516 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
1517 | * they are all at pages_high. | |
1518 | * | |
1519 | * If `nr_pages' is non-zero then it is the number of pages which are to be | |
1520 | * reclaimed, regardless of the zone occupancies. This is a software suspend | |
1521 | * special. | |
1522 | * | |
1523 | * Returns the number of pages which were actually freed. | |
1524 | * | |
1525 | * There is special handling here for zones which are full of pinned pages. | |
1526 | * This can happen if the pages are all mlocked, or if they are all used by | |
1527 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
1528 | * What we do is to detect the case where all pages in the zone have been | |
1529 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
1530 | * dead and from now on, only perform a short scan. Basically we're polling | |
1531 | * the zone for when the problem goes away. | |
1532 | * | |
1533 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
1534 | * zones which have free_pages > pages_high, but once a zone is found to have | |
1535 | * free_pages <= pages_high, we scan that zone and the lower zones regardless | |
1536 | * of the number of free pages in the lower zones. This interoperates with | |
1537 | * the page allocator fallback scheme to ensure that aging of pages is balanced | |
1538 | * across the zones. | |
1539 | */ | |
69e05944 AM |
1540 | static unsigned long balance_pgdat(pg_data_t *pgdat, unsigned long nr_pages, |
1541 | int order) | |
1da177e4 | 1542 | { |
69e05944 | 1543 | unsigned long to_free = nr_pages; |
1da177e4 LT |
1544 | int all_zones_ok; |
1545 | int priority; | |
1546 | int i; | |
69e05944 | 1547 | unsigned long total_scanned; |
05ff5137 | 1548 | unsigned long nr_reclaimed; |
1da177e4 | 1549 | struct reclaim_state *reclaim_state = current->reclaim_state; |
179e9639 AM |
1550 | struct scan_control sc = { |
1551 | .gfp_mask = GFP_KERNEL, | |
1552 | .may_swap = 1, | |
1553 | .swap_cluster_max = nr_pages ? nr_pages : SWAP_CLUSTER_MAX, | |
1554 | }; | |
1da177e4 LT |
1555 | |
1556 | loop_again: | |
1557 | total_scanned = 0; | |
05ff5137 | 1558 | nr_reclaimed = 0; |
179e9639 | 1559 | sc.may_writepage = !laptop_mode, |
1da177e4 LT |
1560 | sc.nr_mapped = read_page_state(nr_mapped); |
1561 | ||
1562 | inc_page_state(pageoutrun); | |
1563 | ||
1564 | for (i = 0; i < pgdat->nr_zones; i++) { | |
1565 | struct zone *zone = pgdat->node_zones + i; | |
1566 | ||
1567 | zone->temp_priority = DEF_PRIORITY; | |
1568 | } | |
1569 | ||
1570 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { | |
1571 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ | |
1572 | unsigned long lru_pages = 0; | |
1573 | ||
f7b7fd8f RR |
1574 | /* The swap token gets in the way of swapout... */ |
1575 | if (!priority) | |
1576 | disable_swap_token(); | |
1577 | ||
1da177e4 LT |
1578 | all_zones_ok = 1; |
1579 | ||
1580 | if (nr_pages == 0) { | |
1581 | /* | |
1582 | * Scan in the highmem->dma direction for the highest | |
1583 | * zone which needs scanning | |
1584 | */ | |
1585 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
1586 | struct zone *zone = pgdat->node_zones + i; | |
1587 | ||
f3fe6512 | 1588 | if (!populated_zone(zone)) |
1da177e4 LT |
1589 | continue; |
1590 | ||
1591 | if (zone->all_unreclaimable && | |
1592 | priority != DEF_PRIORITY) | |
1593 | continue; | |
1594 | ||
1595 | if (!zone_watermark_ok(zone, order, | |
7fb1d9fc | 1596 | zone->pages_high, 0, 0)) { |
1da177e4 LT |
1597 | end_zone = i; |
1598 | goto scan; | |
1599 | } | |
1600 | } | |
1601 | goto out; | |
1602 | } else { | |
1603 | end_zone = pgdat->nr_zones - 1; | |
1604 | } | |
1605 | scan: | |
1606 | for (i = 0; i <= end_zone; i++) { | |
1607 | struct zone *zone = pgdat->node_zones + i; | |
1608 | ||
1609 | lru_pages += zone->nr_active + zone->nr_inactive; | |
1610 | } | |
1611 | ||
1612 | /* | |
1613 | * Now scan the zone in the dma->highmem direction, stopping | |
1614 | * at the last zone which needs scanning. | |
1615 | * | |
1616 | * We do this because the page allocator works in the opposite | |
1617 | * direction. This prevents the page allocator from allocating | |
1618 | * pages behind kswapd's direction of progress, which would | |
1619 | * cause too much scanning of the lower zones. | |
1620 | */ | |
1621 | for (i = 0; i <= end_zone; i++) { | |
1622 | struct zone *zone = pgdat->node_zones + i; | |
b15e0905 | 1623 | int nr_slab; |
1da177e4 | 1624 | |
f3fe6512 | 1625 | if (!populated_zone(zone)) |
1da177e4 LT |
1626 | continue; |
1627 | ||
1628 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) | |
1629 | continue; | |
1630 | ||
1631 | if (nr_pages == 0) { /* Not software suspend */ | |
1632 | if (!zone_watermark_ok(zone, order, | |
7fb1d9fc | 1633 | zone->pages_high, end_zone, 0)) |
1da177e4 LT |
1634 | all_zones_ok = 0; |
1635 | } | |
1636 | zone->temp_priority = priority; | |
1637 | if (zone->prev_priority > priority) | |
1638 | zone->prev_priority = priority; | |
1639 | sc.nr_scanned = 0; | |
05ff5137 | 1640 | nr_reclaimed += shrink_zone(priority, zone, &sc); |
1da177e4 | 1641 | reclaim_state->reclaimed_slab = 0; |
b15e0905 | 1642 | nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, |
1643 | lru_pages); | |
05ff5137 | 1644 | nr_reclaimed += reclaim_state->reclaimed_slab; |
1da177e4 LT |
1645 | total_scanned += sc.nr_scanned; |
1646 | if (zone->all_unreclaimable) | |
1647 | continue; | |
b15e0905 | 1648 | if (nr_slab == 0 && zone->pages_scanned >= |
1649 | (zone->nr_active + zone->nr_inactive) * 4) | |
1da177e4 LT |
1650 | zone->all_unreclaimable = 1; |
1651 | /* | |
1652 | * If we've done a decent amount of scanning and | |
1653 | * the reclaim ratio is low, start doing writepage | |
1654 | * even in laptop mode | |
1655 | */ | |
1656 | if (total_scanned > SWAP_CLUSTER_MAX * 2 && | |
05ff5137 | 1657 | total_scanned > nr_reclaimed + nr_reclaimed / 2) |
1da177e4 LT |
1658 | sc.may_writepage = 1; |
1659 | } | |
05ff5137 | 1660 | if (nr_pages && to_free > nr_reclaimed) |
1da177e4 LT |
1661 | continue; /* swsusp: need to do more work */ |
1662 | if (all_zones_ok) | |
1663 | break; /* kswapd: all done */ | |
1664 | /* | |
1665 | * OK, kswapd is getting into trouble. Take a nap, then take | |
1666 | * another pass across the zones. | |
1667 | */ | |
1668 | if (total_scanned && priority < DEF_PRIORITY - 2) | |
1669 | blk_congestion_wait(WRITE, HZ/10); | |
1670 | ||
1671 | /* | |
1672 | * We do this so kswapd doesn't build up large priorities for | |
1673 | * example when it is freeing in parallel with allocators. It | |
1674 | * matches the direct reclaim path behaviour in terms of impact | |
1675 | * on zone->*_priority. | |
1676 | */ | |
05ff5137 | 1677 | if ((nr_reclaimed >= SWAP_CLUSTER_MAX) && !nr_pages) |
1da177e4 LT |
1678 | break; |
1679 | } | |
1680 | out: | |
1681 | for (i = 0; i < pgdat->nr_zones; i++) { | |
1682 | struct zone *zone = pgdat->node_zones + i; | |
1683 | ||
1684 | zone->prev_priority = zone->temp_priority; | |
1685 | } | |
1686 | if (!all_zones_ok) { | |
1687 | cond_resched(); | |
1688 | goto loop_again; | |
1689 | } | |
1690 | ||
05ff5137 | 1691 | return nr_reclaimed; |
1da177e4 LT |
1692 | } |
1693 | ||
1694 | /* | |
1695 | * The background pageout daemon, started as a kernel thread | |
1696 | * from the init process. | |
1697 | * | |
1698 | * This basically trickles out pages so that we have _some_ | |
1699 | * free memory available even if there is no other activity | |
1700 | * that frees anything up. This is needed for things like routing | |
1701 | * etc, where we otherwise might have all activity going on in | |
1702 | * asynchronous contexts that cannot page things out. | |
1703 | * | |
1704 | * If there are applications that are active memory-allocators | |
1705 | * (most normal use), this basically shouldn't matter. | |
1706 | */ | |
1707 | static int kswapd(void *p) | |
1708 | { | |
1709 | unsigned long order; | |
1710 | pg_data_t *pgdat = (pg_data_t*)p; | |
1711 | struct task_struct *tsk = current; | |
1712 | DEFINE_WAIT(wait); | |
1713 | struct reclaim_state reclaim_state = { | |
1714 | .reclaimed_slab = 0, | |
1715 | }; | |
1716 | cpumask_t cpumask; | |
1717 | ||
1718 | daemonize("kswapd%d", pgdat->node_id); | |
1719 | cpumask = node_to_cpumask(pgdat->node_id); | |
1720 | if (!cpus_empty(cpumask)) | |
1721 | set_cpus_allowed(tsk, cpumask); | |
1722 | current->reclaim_state = &reclaim_state; | |
1723 | ||
1724 | /* | |
1725 | * Tell the memory management that we're a "memory allocator", | |
1726 | * and that if we need more memory we should get access to it | |
1727 | * regardless (see "__alloc_pages()"). "kswapd" should | |
1728 | * never get caught in the normal page freeing logic. | |
1729 | * | |
1730 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
1731 | * you need a small amount of memory in order to be able to | |
1732 | * page out something else, and this flag essentially protects | |
1733 | * us from recursively trying to free more memory as we're | |
1734 | * trying to free the first piece of memory in the first place). | |
1735 | */ | |
930d9152 | 1736 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
1da177e4 LT |
1737 | |
1738 | order = 0; | |
1739 | for ( ; ; ) { | |
1740 | unsigned long new_order; | |
3e1d1d28 CL |
1741 | |
1742 | try_to_freeze(); | |
1da177e4 LT |
1743 | |
1744 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
1745 | new_order = pgdat->kswapd_max_order; | |
1746 | pgdat->kswapd_max_order = 0; | |
1747 | if (order < new_order) { | |
1748 | /* | |
1749 | * Don't sleep if someone wants a larger 'order' | |
1750 | * allocation | |
1751 | */ | |
1752 | order = new_order; | |
1753 | } else { | |
1754 | schedule(); | |
1755 | order = pgdat->kswapd_max_order; | |
1756 | } | |
1757 | finish_wait(&pgdat->kswapd_wait, &wait); | |
1758 | ||
1759 | balance_pgdat(pgdat, 0, order); | |
1760 | } | |
1761 | return 0; | |
1762 | } | |
1763 | ||
1764 | /* | |
1765 | * A zone is low on free memory, so wake its kswapd task to service it. | |
1766 | */ | |
1767 | void wakeup_kswapd(struct zone *zone, int order) | |
1768 | { | |
1769 | pg_data_t *pgdat; | |
1770 | ||
f3fe6512 | 1771 | if (!populated_zone(zone)) |
1da177e4 LT |
1772 | return; |
1773 | ||
1774 | pgdat = zone->zone_pgdat; | |
7fb1d9fc | 1775 | if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0)) |
1da177e4 LT |
1776 | return; |
1777 | if (pgdat->kswapd_max_order < order) | |
1778 | pgdat->kswapd_max_order = order; | |
9bf2229f | 1779 | if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) |
1da177e4 | 1780 | return; |
8d0986e2 | 1781 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 1782 | return; |
8d0986e2 | 1783 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
1784 | } |
1785 | ||
1786 | #ifdef CONFIG_PM | |
1787 | /* | |
1788 | * Try to free `nr_pages' of memory, system-wide. Returns the number of freed | |
1789 | * pages. | |
1790 | */ | |
69e05944 | 1791 | unsigned long shrink_all_memory(unsigned long nr_pages) |
1da177e4 LT |
1792 | { |
1793 | pg_data_t *pgdat; | |
69e05944 AM |
1794 | unsigned long nr_to_free = nr_pages; |
1795 | unsigned long ret = 0; | |
1da177e4 LT |
1796 | struct reclaim_state reclaim_state = { |
1797 | .reclaimed_slab = 0, | |
1798 | }; | |
1799 | ||
1800 | current->reclaim_state = &reclaim_state; | |
1801 | for_each_pgdat(pgdat) { | |
69e05944 AM |
1802 | unsigned long freed; |
1803 | ||
1da177e4 LT |
1804 | freed = balance_pgdat(pgdat, nr_to_free, 0); |
1805 | ret += freed; | |
1806 | nr_to_free -= freed; | |
69e05944 | 1807 | if ((long)nr_to_free <= 0) |
1da177e4 LT |
1808 | break; |
1809 | } | |
1810 | current->reclaim_state = NULL; | |
1811 | return ret; | |
1812 | } | |
1813 | #endif | |
1814 | ||
1815 | #ifdef CONFIG_HOTPLUG_CPU | |
1816 | /* It's optimal to keep kswapds on the same CPUs as their memory, but | |
1817 | not required for correctness. So if the last cpu in a node goes | |
1818 | away, we get changed to run anywhere: as the first one comes back, | |
1819 | restore their cpu bindings. */ | |
1820 | static int __devinit cpu_callback(struct notifier_block *nfb, | |
69e05944 | 1821 | unsigned long action, void *hcpu) |
1da177e4 LT |
1822 | { |
1823 | pg_data_t *pgdat; | |
1824 | cpumask_t mask; | |
1825 | ||
1826 | if (action == CPU_ONLINE) { | |
1827 | for_each_pgdat(pgdat) { | |
1828 | mask = node_to_cpumask(pgdat->node_id); | |
1829 | if (any_online_cpu(mask) != NR_CPUS) | |
1830 | /* One of our CPUs online: restore mask */ | |
1831 | set_cpus_allowed(pgdat->kswapd, mask); | |
1832 | } | |
1833 | } | |
1834 | return NOTIFY_OK; | |
1835 | } | |
1836 | #endif /* CONFIG_HOTPLUG_CPU */ | |
1837 | ||
1838 | static int __init kswapd_init(void) | |
1839 | { | |
1840 | pg_data_t *pgdat; | |
69e05944 | 1841 | |
1da177e4 | 1842 | swap_setup(); |
69e05944 AM |
1843 | for_each_pgdat(pgdat) { |
1844 | pid_t pid; | |
1845 | ||
1846 | pid = kernel_thread(kswapd, pgdat, CLONE_KERNEL); | |
1847 | BUG_ON(pid < 0); | |
1848 | pgdat->kswapd = find_task_by_pid(pid); | |
1849 | } | |
1da177e4 LT |
1850 | total_memory = nr_free_pagecache_pages(); |
1851 | hotcpu_notifier(cpu_callback, 0); | |
1852 | return 0; | |
1853 | } | |
1854 | ||
1855 | module_init(kswapd_init) | |
9eeff239 CL |
1856 | |
1857 | #ifdef CONFIG_NUMA | |
1858 | /* | |
1859 | * Zone reclaim mode | |
1860 | * | |
1861 | * If non-zero call zone_reclaim when the number of free pages falls below | |
1862 | * the watermarks. | |
1863 | * | |
1864 | * In the future we may add flags to the mode. However, the page allocator | |
1865 | * should only have to check that zone_reclaim_mode != 0 before calling | |
1866 | * zone_reclaim(). | |
1867 | */ | |
1868 | int zone_reclaim_mode __read_mostly; | |
1869 | ||
1b2ffb78 CL |
1870 | #define RECLAIM_OFF 0 |
1871 | #define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */ | |
1872 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ | |
1873 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
2a16e3f4 | 1874 | #define RECLAIM_SLAB (1<<3) /* Do a global slab shrink if the zone is out of memory */ |
1b2ffb78 | 1875 | |
9eeff239 CL |
1876 | /* |
1877 | * Mininum time between zone reclaim scans | |
1878 | */ | |
2a11ff06 | 1879 | int zone_reclaim_interval __read_mostly = 30*HZ; |
a92f7126 CL |
1880 | |
1881 | /* | |
1882 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
1883 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
1884 | * a zone. | |
1885 | */ | |
1886 | #define ZONE_RECLAIM_PRIORITY 4 | |
1887 | ||
9eeff239 CL |
1888 | /* |
1889 | * Try to free up some pages from this zone through reclaim. | |
1890 | */ | |
179e9639 | 1891 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 1892 | { |
7fb2d46d | 1893 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 1894 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
1895 | struct task_struct *p = current; |
1896 | struct reclaim_state reclaim_state; | |
8695949a | 1897 | int priority; |
05ff5137 | 1898 | unsigned long nr_reclaimed = 0; |
179e9639 AM |
1899 | struct scan_control sc = { |
1900 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
1901 | .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP), | |
1902 | .nr_mapped = read_page_state(nr_mapped), | |
69e05944 AM |
1903 | .swap_cluster_max = max_t(unsigned long, nr_pages, |
1904 | SWAP_CLUSTER_MAX), | |
179e9639 AM |
1905 | .gfp_mask = gfp_mask, |
1906 | }; | |
9eeff239 CL |
1907 | |
1908 | disable_swap_token(); | |
9eeff239 | 1909 | cond_resched(); |
d4f7796e CL |
1910 | /* |
1911 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
1912 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
1913 | * and RECLAIM_SWAP. | |
1914 | */ | |
1915 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
9eeff239 CL |
1916 | reclaim_state.reclaimed_slab = 0; |
1917 | p->reclaim_state = &reclaim_state; | |
c84db23c | 1918 | |
a92f7126 CL |
1919 | /* |
1920 | * Free memory by calling shrink zone with increasing priorities | |
1921 | * until we have enough memory freed. | |
1922 | */ | |
8695949a | 1923 | priority = ZONE_RECLAIM_PRIORITY; |
a92f7126 | 1924 | do { |
05ff5137 | 1925 | nr_reclaimed += shrink_zone(priority, zone, &sc); |
8695949a | 1926 | priority--; |
05ff5137 | 1927 | } while (priority >= 0 && nr_reclaimed < nr_pages); |
c84db23c | 1928 | |
05ff5137 | 1929 | if (nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) { |
2a16e3f4 | 1930 | /* |
7fb2d46d CL |
1931 | * shrink_slab() does not currently allow us to determine how |
1932 | * many pages were freed in this zone. So we just shake the slab | |
1933 | * a bit and then go off node for this particular allocation | |
1934 | * despite possibly having freed enough memory to allocate in | |
1935 | * this zone. If we freed local memory then the next | |
1936 | * allocations will be local again. | |
2a16e3f4 CL |
1937 | * |
1938 | * shrink_slab will free memory on all zones and may take | |
1939 | * a long time. | |
1940 | */ | |
1941 | shrink_slab(sc.nr_scanned, gfp_mask, order); | |
2a16e3f4 CL |
1942 | } |
1943 | ||
9eeff239 | 1944 | p->reclaim_state = NULL; |
d4f7796e | 1945 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
9eeff239 | 1946 | |
7fb2d46d CL |
1947 | if (nr_reclaimed == 0) { |
1948 | /* | |
1949 | * We were unable to reclaim enough pages to stay on node. We | |
1950 | * now allow off node accesses for a certain time period before | |
1951 | * trying again to reclaim pages from the local zone. | |
1952 | */ | |
9eeff239 | 1953 | zone->last_unsuccessful_zone_reclaim = jiffies; |
7fb2d46d | 1954 | } |
9eeff239 | 1955 | |
05ff5137 | 1956 | return nr_reclaimed >= nr_pages; |
9eeff239 | 1957 | } |
179e9639 AM |
1958 | |
1959 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
1960 | { | |
1961 | cpumask_t mask; | |
1962 | int node_id; | |
1963 | ||
1964 | /* | |
1965 | * Do not reclaim if there was a recent unsuccessful attempt at zone | |
1966 | * reclaim. In that case we let allocations go off node for the | |
1967 | * zone_reclaim_interval. Otherwise we would scan for each off-node | |
1968 | * page allocation. | |
1969 | */ | |
1970 | if (time_before(jiffies, | |
1971 | zone->last_unsuccessful_zone_reclaim + zone_reclaim_interval)) | |
1972 | return 0; | |
1973 | ||
1974 | /* | |
1975 | * Avoid concurrent zone reclaims, do not reclaim in a zone that does | |
1976 | * not have reclaimable pages and if we should not delay the allocation | |
1977 | * then do not scan. | |
1978 | */ | |
1979 | if (!(gfp_mask & __GFP_WAIT) || | |
1980 | zone->all_unreclaimable || | |
1981 | atomic_read(&zone->reclaim_in_progress) > 0 || | |
1982 | (current->flags & PF_MEMALLOC)) | |
1983 | return 0; | |
1984 | ||
1985 | /* | |
1986 | * Only run zone reclaim on the local zone or on zones that do not | |
1987 | * have associated processors. This will favor the local processor | |
1988 | * over remote processors and spread off node memory allocations | |
1989 | * as wide as possible. | |
1990 | */ | |
1991 | node_id = zone->zone_pgdat->node_id; | |
1992 | mask = node_to_cpumask(node_id); | |
1993 | if (!cpus_empty(mask) && node_id != numa_node_id()) | |
1994 | return 0; | |
1995 | return __zone_reclaim(zone, gfp_mask, order); | |
1996 | } | |
9eeff239 | 1997 | #endif |