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> | |
5a0e3ad6 | 16 | #include <linux/gfp.h> |
1da177e4 LT |
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> | |
1da177e4 LT |
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> | |
3e7d3449 | 34 | #include <linux/compaction.h> |
1da177e4 LT |
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> |
929bea7c | 43 | #include <linux/oom.h> |
268bb0ce | 44 | #include <linux/prefetch.h> |
1da177e4 LT |
45 | |
46 | #include <asm/tlbflush.h> | |
47 | #include <asm/div64.h> | |
48 | ||
49 | #include <linux/swapops.h> | |
50 | ||
0f8053a5 NP |
51 | #include "internal.h" |
52 | ||
33906bc5 MG |
53 | #define CREATE_TRACE_POINTS |
54 | #include <trace/events/vmscan.h> | |
55 | ||
1da177e4 | 56 | struct scan_control { |
1da177e4 LT |
57 | /* Incremented by the number of inactive pages that were scanned */ |
58 | unsigned long nr_scanned; | |
59 | ||
a79311c1 RR |
60 | /* Number of pages freed so far during a call to shrink_zones() */ |
61 | unsigned long nr_reclaimed; | |
62 | ||
22fba335 KM |
63 | /* How many pages shrink_list() should reclaim */ |
64 | unsigned long nr_to_reclaim; | |
65 | ||
7b51755c KM |
66 | unsigned long hibernation_mode; |
67 | ||
1da177e4 | 68 | /* This context's GFP mask */ |
6daa0e28 | 69 | gfp_t gfp_mask; |
1da177e4 LT |
70 | |
71 | int may_writepage; | |
72 | ||
a6dc60f8 JW |
73 | /* Can mapped pages be reclaimed? */ |
74 | int may_unmap; | |
f1fd1067 | 75 | |
2e2e4259 KM |
76 | /* Can pages be swapped as part of reclaim? */ |
77 | int may_swap; | |
78 | ||
5ad333eb | 79 | int order; |
66e1707b | 80 | |
9e3b2f8c KK |
81 | /* Scan (total_size >> priority) pages at once */ |
82 | int priority; | |
83 | ||
f16015fb JW |
84 | /* |
85 | * The memory cgroup that hit its limit and as a result is the | |
86 | * primary target of this reclaim invocation. | |
87 | */ | |
88 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 89 | |
327c0e96 KH |
90 | /* |
91 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
92 | * are scanned. | |
93 | */ | |
94 | nodemask_t *nodemask; | |
1da177e4 LT |
95 | }; |
96 | ||
f16015fb JW |
97 | struct mem_cgroup_zone { |
98 | struct mem_cgroup *mem_cgroup; | |
99 | struct zone *zone; | |
100 | }; | |
101 | ||
1da177e4 LT |
102 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
103 | ||
104 | #ifdef ARCH_HAS_PREFETCH | |
105 | #define prefetch_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 | prefetch(&prev->_field); \ | |
112 | } \ | |
113 | } while (0) | |
114 | #else | |
115 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
116 | #endif | |
117 | ||
118 | #ifdef ARCH_HAS_PREFETCHW | |
119 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
120 | do { \ | |
121 | if ((_page)->lru.prev != _base) { \ | |
122 | struct page *prev; \ | |
123 | \ | |
124 | prev = lru_to_page(&(_page->lru)); \ | |
125 | prefetchw(&prev->_field); \ | |
126 | } \ | |
127 | } while (0) | |
128 | #else | |
129 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
130 | #endif | |
131 | ||
132 | /* | |
133 | * From 0 .. 100. Higher means more swappy. | |
134 | */ | |
135 | int vm_swappiness = 60; | |
bd1e22b8 | 136 | long vm_total_pages; /* The total number of pages which the VM controls */ |
1da177e4 LT |
137 | |
138 | static LIST_HEAD(shrinker_list); | |
139 | static DECLARE_RWSEM(shrinker_rwsem); | |
140 | ||
00f0b825 | 141 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
89b5fae5 JW |
142 | static bool global_reclaim(struct scan_control *sc) |
143 | { | |
f16015fb | 144 | return !sc->target_mem_cgroup; |
89b5fae5 | 145 | } |
91a45470 | 146 | #else |
89b5fae5 JW |
147 | static bool global_reclaim(struct scan_control *sc) |
148 | { | |
149 | return true; | |
150 | } | |
91a45470 KH |
151 | #endif |
152 | ||
f16015fb | 153 | static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz) |
6e901571 | 154 | { |
89abfab1 | 155 | return &mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup)->reclaim_stat; |
6e901571 KM |
156 | } |
157 | ||
f16015fb JW |
158 | static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz, |
159 | enum lru_list lru) | |
c9f299d9 | 160 | { |
c3c787e8 | 161 | if (!mem_cgroup_disabled()) |
f16015fb JW |
162 | return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup, |
163 | zone_to_nid(mz->zone), | |
164 | zone_idx(mz->zone), | |
165 | BIT(lru)); | |
a3d8e054 | 166 | |
f16015fb | 167 | return zone_page_state(mz->zone, NR_LRU_BASE + lru); |
c9f299d9 KM |
168 | } |
169 | ||
170 | ||
1da177e4 LT |
171 | /* |
172 | * Add a shrinker callback to be called from the vm | |
173 | */ | |
8e1f936b | 174 | void register_shrinker(struct shrinker *shrinker) |
1da177e4 | 175 | { |
83aeeada | 176 | atomic_long_set(&shrinker->nr_in_batch, 0); |
8e1f936b RR |
177 | down_write(&shrinker_rwsem); |
178 | list_add_tail(&shrinker->list, &shrinker_list); | |
179 | up_write(&shrinker_rwsem); | |
1da177e4 | 180 | } |
8e1f936b | 181 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
182 | |
183 | /* | |
184 | * Remove one | |
185 | */ | |
8e1f936b | 186 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
187 | { |
188 | down_write(&shrinker_rwsem); | |
189 | list_del(&shrinker->list); | |
190 | up_write(&shrinker_rwsem); | |
1da177e4 | 191 | } |
8e1f936b | 192 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 | 193 | |
1495f230 YH |
194 | static inline int do_shrinker_shrink(struct shrinker *shrinker, |
195 | struct shrink_control *sc, | |
196 | unsigned long nr_to_scan) | |
197 | { | |
198 | sc->nr_to_scan = nr_to_scan; | |
199 | return (*shrinker->shrink)(shrinker, sc); | |
200 | } | |
201 | ||
1da177e4 LT |
202 | #define SHRINK_BATCH 128 |
203 | /* | |
204 | * Call the shrink functions to age shrinkable caches | |
205 | * | |
206 | * Here we assume it costs one seek to replace a lru page and that it also | |
207 | * takes a seek to recreate a cache object. With this in mind we age equal | |
208 | * percentages of the lru and ageable caches. This should balance the seeks | |
209 | * generated by these structures. | |
210 | * | |
183ff22b | 211 | * If the vm encountered mapped pages on the LRU it increase the pressure on |
1da177e4 LT |
212 | * slab to avoid swapping. |
213 | * | |
214 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
215 | * | |
216 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
217 | * are eligible for the caller's allocation attempt. It is used for balancing | |
218 | * slab reclaim versus page reclaim. | |
b15e0905 | 219 | * |
220 | * Returns the number of slab objects which we shrunk. | |
1da177e4 | 221 | */ |
a09ed5e0 | 222 | unsigned long shrink_slab(struct shrink_control *shrink, |
1495f230 | 223 | unsigned long nr_pages_scanned, |
a09ed5e0 | 224 | unsigned long lru_pages) |
1da177e4 LT |
225 | { |
226 | struct shrinker *shrinker; | |
69e05944 | 227 | unsigned long ret = 0; |
1da177e4 | 228 | |
1495f230 YH |
229 | if (nr_pages_scanned == 0) |
230 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 231 | |
f06590bd MK |
232 | if (!down_read_trylock(&shrinker_rwsem)) { |
233 | /* Assume we'll be able to shrink next time */ | |
234 | ret = 1; | |
235 | goto out; | |
236 | } | |
1da177e4 LT |
237 | |
238 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
239 | unsigned long long delta; | |
635697c6 KK |
240 | long total_scan; |
241 | long max_pass; | |
09576073 | 242 | int shrink_ret = 0; |
acf92b48 DC |
243 | long nr; |
244 | long new_nr; | |
e9299f50 DC |
245 | long batch_size = shrinker->batch ? shrinker->batch |
246 | : SHRINK_BATCH; | |
1da177e4 | 247 | |
635697c6 KK |
248 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); |
249 | if (max_pass <= 0) | |
250 | continue; | |
251 | ||
acf92b48 DC |
252 | /* |
253 | * copy the current shrinker scan count into a local variable | |
254 | * and zero it so that other concurrent shrinker invocations | |
255 | * don't also do this scanning work. | |
256 | */ | |
83aeeada | 257 | nr = atomic_long_xchg(&shrinker->nr_in_batch, 0); |
acf92b48 DC |
258 | |
259 | total_scan = nr; | |
1495f230 | 260 | delta = (4 * nr_pages_scanned) / shrinker->seeks; |
ea164d73 | 261 | delta *= max_pass; |
1da177e4 | 262 | do_div(delta, lru_pages + 1); |
acf92b48 DC |
263 | total_scan += delta; |
264 | if (total_scan < 0) { | |
88c3bd70 DR |
265 | printk(KERN_ERR "shrink_slab: %pF negative objects to " |
266 | "delete nr=%ld\n", | |
acf92b48 DC |
267 | shrinker->shrink, total_scan); |
268 | total_scan = max_pass; | |
ea164d73 AA |
269 | } |
270 | ||
3567b59a DC |
271 | /* |
272 | * We need to avoid excessive windup on filesystem shrinkers | |
273 | * due to large numbers of GFP_NOFS allocations causing the | |
274 | * shrinkers to return -1 all the time. This results in a large | |
275 | * nr being built up so when a shrink that can do some work | |
276 | * comes along it empties the entire cache due to nr >>> | |
277 | * max_pass. This is bad for sustaining a working set in | |
278 | * memory. | |
279 | * | |
280 | * Hence only allow the shrinker to scan the entire cache when | |
281 | * a large delta change is calculated directly. | |
282 | */ | |
283 | if (delta < max_pass / 4) | |
284 | total_scan = min(total_scan, max_pass / 2); | |
285 | ||
ea164d73 AA |
286 | /* |
287 | * Avoid risking looping forever due to too large nr value: | |
288 | * never try to free more than twice the estimate number of | |
289 | * freeable entries. | |
290 | */ | |
acf92b48 DC |
291 | if (total_scan > max_pass * 2) |
292 | total_scan = max_pass * 2; | |
1da177e4 | 293 | |
acf92b48 | 294 | trace_mm_shrink_slab_start(shrinker, shrink, nr, |
09576073 DC |
295 | nr_pages_scanned, lru_pages, |
296 | max_pass, delta, total_scan); | |
297 | ||
e9299f50 | 298 | while (total_scan >= batch_size) { |
b15e0905 | 299 | int nr_before; |
1da177e4 | 300 | |
1495f230 YH |
301 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); |
302 | shrink_ret = do_shrinker_shrink(shrinker, shrink, | |
e9299f50 | 303 | batch_size); |
1da177e4 LT |
304 | if (shrink_ret == -1) |
305 | break; | |
b15e0905 | 306 | if (shrink_ret < nr_before) |
307 | ret += nr_before - shrink_ret; | |
e9299f50 DC |
308 | count_vm_events(SLABS_SCANNED, batch_size); |
309 | total_scan -= batch_size; | |
1da177e4 LT |
310 | |
311 | cond_resched(); | |
312 | } | |
313 | ||
acf92b48 DC |
314 | /* |
315 | * move the unused scan count back into the shrinker in a | |
316 | * manner that handles concurrent updates. If we exhausted the | |
317 | * scan, there is no need to do an update. | |
318 | */ | |
83aeeada KK |
319 | if (total_scan > 0) |
320 | new_nr = atomic_long_add_return(total_scan, | |
321 | &shrinker->nr_in_batch); | |
322 | else | |
323 | new_nr = atomic_long_read(&shrinker->nr_in_batch); | |
acf92b48 DC |
324 | |
325 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); | |
1da177e4 LT |
326 | } |
327 | up_read(&shrinker_rwsem); | |
f06590bd MK |
328 | out: |
329 | cond_resched(); | |
b15e0905 | 330 | return ret; |
1da177e4 LT |
331 | } |
332 | ||
1da177e4 LT |
333 | static inline int is_page_cache_freeable(struct page *page) |
334 | { | |
ceddc3a5 JW |
335 | /* |
336 | * A freeable page cache page is referenced only by the caller | |
337 | * that isolated the page, the page cache radix tree and | |
338 | * optional buffer heads at page->private. | |
339 | */ | |
edcf4748 | 340 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
341 | } |
342 | ||
7d3579e8 KM |
343 | static int may_write_to_queue(struct backing_dev_info *bdi, |
344 | struct scan_control *sc) | |
1da177e4 | 345 | { |
930d9152 | 346 | if (current->flags & PF_SWAPWRITE) |
1da177e4 LT |
347 | return 1; |
348 | if (!bdi_write_congested(bdi)) | |
349 | return 1; | |
350 | if (bdi == current->backing_dev_info) | |
351 | return 1; | |
352 | return 0; | |
353 | } | |
354 | ||
355 | /* | |
356 | * We detected a synchronous write error writing a page out. Probably | |
357 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
358 | * fsync(), msync() or close(). | |
359 | * | |
360 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
361 | * prevents it from being freed up. But we have a ref on the page and once | |
362 | * that page is locked, the mapping is pinned. | |
363 | * | |
364 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
365 | * __GFP_FS. | |
366 | */ | |
367 | static void handle_write_error(struct address_space *mapping, | |
368 | struct page *page, int error) | |
369 | { | |
7eaceacc | 370 | lock_page(page); |
3e9f45bd GC |
371 | if (page_mapping(page) == mapping) |
372 | mapping_set_error(mapping, error); | |
1da177e4 LT |
373 | unlock_page(page); |
374 | } | |
375 | ||
04e62a29 CL |
376 | /* possible outcome of pageout() */ |
377 | typedef enum { | |
378 | /* failed to write page out, page is locked */ | |
379 | PAGE_KEEP, | |
380 | /* move page to the active list, page is locked */ | |
381 | PAGE_ACTIVATE, | |
382 | /* page has been sent to the disk successfully, page is unlocked */ | |
383 | PAGE_SUCCESS, | |
384 | /* page is clean and locked */ | |
385 | PAGE_CLEAN, | |
386 | } pageout_t; | |
387 | ||
1da177e4 | 388 | /* |
1742f19f AM |
389 | * pageout is called by shrink_page_list() for each dirty page. |
390 | * Calls ->writepage(). | |
1da177e4 | 391 | */ |
c661b078 | 392 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 393 | struct scan_control *sc) |
1da177e4 LT |
394 | { |
395 | /* | |
396 | * If the page is dirty, only perform writeback if that write | |
397 | * will be non-blocking. To prevent this allocation from being | |
398 | * stalled by pagecache activity. But note that there may be | |
399 | * stalls if we need to run get_block(). We could test | |
400 | * PagePrivate for that. | |
401 | * | |
6aceb53b | 402 | * If this process is currently in __generic_file_aio_write() against |
1da177e4 LT |
403 | * this page's queue, we can perform writeback even if that |
404 | * will block. | |
405 | * | |
406 | * If the page is swapcache, write it back even if that would | |
407 | * block, for some throttling. This happens by accident, because | |
408 | * swap_backing_dev_info is bust: it doesn't reflect the | |
409 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
410 | */ |
411 | if (!is_page_cache_freeable(page)) | |
412 | return PAGE_KEEP; | |
413 | if (!mapping) { | |
414 | /* | |
415 | * Some data journaling orphaned pages can have | |
416 | * page->mapping == NULL while being dirty with clean buffers. | |
417 | */ | |
266cf658 | 418 | if (page_has_private(page)) { |
1da177e4 LT |
419 | if (try_to_free_buffers(page)) { |
420 | ClearPageDirty(page); | |
d40cee24 | 421 | printk("%s: orphaned page\n", __func__); |
1da177e4 LT |
422 | return PAGE_CLEAN; |
423 | } | |
424 | } | |
425 | return PAGE_KEEP; | |
426 | } | |
427 | if (mapping->a_ops->writepage == NULL) | |
428 | return PAGE_ACTIVATE; | |
0e093d99 | 429 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) |
1da177e4 LT |
430 | return PAGE_KEEP; |
431 | ||
432 | if (clear_page_dirty_for_io(page)) { | |
433 | int res; | |
434 | struct writeback_control wbc = { | |
435 | .sync_mode = WB_SYNC_NONE, | |
436 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
437 | .range_start = 0, |
438 | .range_end = LLONG_MAX, | |
1da177e4 LT |
439 | .for_reclaim = 1, |
440 | }; | |
441 | ||
442 | SetPageReclaim(page); | |
443 | res = mapping->a_ops->writepage(page, &wbc); | |
444 | if (res < 0) | |
445 | handle_write_error(mapping, page, res); | |
994fc28c | 446 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
447 | ClearPageReclaim(page); |
448 | return PAGE_ACTIVATE; | |
449 | } | |
c661b078 | 450 | |
1da177e4 LT |
451 | if (!PageWriteback(page)) { |
452 | /* synchronous write or broken a_ops? */ | |
453 | ClearPageReclaim(page); | |
454 | } | |
23b9da55 | 455 | trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); |
e129b5c2 | 456 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
457 | return PAGE_SUCCESS; |
458 | } | |
459 | ||
460 | return PAGE_CLEAN; | |
461 | } | |
462 | ||
a649fd92 | 463 | /* |
e286781d NP |
464 | * Same as remove_mapping, but if the page is removed from the mapping, it |
465 | * gets returned with a refcount of 0. | |
a649fd92 | 466 | */ |
e286781d | 467 | static int __remove_mapping(struct address_space *mapping, struct page *page) |
49d2e9cc | 468 | { |
28e4d965 NP |
469 | BUG_ON(!PageLocked(page)); |
470 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 471 | |
19fd6231 | 472 | spin_lock_irq(&mapping->tree_lock); |
49d2e9cc | 473 | /* |
0fd0e6b0 NP |
474 | * The non racy check for a busy page. |
475 | * | |
476 | * Must be careful with the order of the tests. When someone has | |
477 | * a ref to the page, it may be possible that they dirty it then | |
478 | * drop the reference. So if PageDirty is tested before page_count | |
479 | * here, then the following race may occur: | |
480 | * | |
481 | * get_user_pages(&page); | |
482 | * [user mapping goes away] | |
483 | * write_to(page); | |
484 | * !PageDirty(page) [good] | |
485 | * SetPageDirty(page); | |
486 | * put_page(page); | |
487 | * !page_count(page) [good, discard it] | |
488 | * | |
489 | * [oops, our write_to data is lost] | |
490 | * | |
491 | * Reversing the order of the tests ensures such a situation cannot | |
492 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
493 | * load is not satisfied before that of page->_count. | |
494 | * | |
495 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
496 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 497 | */ |
e286781d | 498 | if (!page_freeze_refs(page, 2)) |
49d2e9cc | 499 | goto cannot_free; |
e286781d NP |
500 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
501 | if (unlikely(PageDirty(page))) { | |
502 | page_unfreeze_refs(page, 2); | |
49d2e9cc | 503 | goto cannot_free; |
e286781d | 504 | } |
49d2e9cc CL |
505 | |
506 | if (PageSwapCache(page)) { | |
507 | swp_entry_t swap = { .val = page_private(page) }; | |
508 | __delete_from_swap_cache(page); | |
19fd6231 | 509 | spin_unlock_irq(&mapping->tree_lock); |
cb4b86ba | 510 | swapcache_free(swap, page); |
e286781d | 511 | } else { |
6072d13c LT |
512 | void (*freepage)(struct page *); |
513 | ||
514 | freepage = mapping->a_ops->freepage; | |
515 | ||
e64a782f | 516 | __delete_from_page_cache(page); |
19fd6231 | 517 | spin_unlock_irq(&mapping->tree_lock); |
e767e056 | 518 | mem_cgroup_uncharge_cache_page(page); |
6072d13c LT |
519 | |
520 | if (freepage != NULL) | |
521 | freepage(page); | |
49d2e9cc CL |
522 | } |
523 | ||
49d2e9cc CL |
524 | return 1; |
525 | ||
526 | cannot_free: | |
19fd6231 | 527 | spin_unlock_irq(&mapping->tree_lock); |
49d2e9cc CL |
528 | return 0; |
529 | } | |
530 | ||
e286781d NP |
531 | /* |
532 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
533 | * someone else has a ref on the page, abort and return 0. If it was | |
534 | * successfully detached, return 1. Assumes the caller has a single ref on | |
535 | * this page. | |
536 | */ | |
537 | int remove_mapping(struct address_space *mapping, struct page *page) | |
538 | { | |
539 | if (__remove_mapping(mapping, page)) { | |
540 | /* | |
541 | * Unfreezing the refcount with 1 rather than 2 effectively | |
542 | * drops the pagecache ref for us without requiring another | |
543 | * atomic operation. | |
544 | */ | |
545 | page_unfreeze_refs(page, 1); | |
546 | return 1; | |
547 | } | |
548 | return 0; | |
549 | } | |
550 | ||
894bc310 LS |
551 | /** |
552 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
553 | * @page: page to be put back to appropriate lru list | |
554 | * | |
555 | * Add previously isolated @page to appropriate LRU list. | |
556 | * Page may still be unevictable for other reasons. | |
557 | * | |
558 | * lru_lock must not be held, interrupts must be enabled. | |
559 | */ | |
894bc310 LS |
560 | void putback_lru_page(struct page *page) |
561 | { | |
562 | int lru; | |
563 | int active = !!TestClearPageActive(page); | |
bbfd28ee | 564 | int was_unevictable = PageUnevictable(page); |
894bc310 LS |
565 | |
566 | VM_BUG_ON(PageLRU(page)); | |
567 | ||
568 | redo: | |
569 | ClearPageUnevictable(page); | |
570 | ||
571 | if (page_evictable(page, NULL)) { | |
572 | /* | |
573 | * For evictable pages, we can use the cache. | |
574 | * In event of a race, worst case is we end up with an | |
575 | * unevictable page on [in]active list. | |
576 | * We know how to handle that. | |
577 | */ | |
401a8e1c | 578 | lru = active + page_lru_base_type(page); |
894bc310 LS |
579 | lru_cache_add_lru(page, lru); |
580 | } else { | |
581 | /* | |
582 | * Put unevictable pages directly on zone's unevictable | |
583 | * list. | |
584 | */ | |
585 | lru = LRU_UNEVICTABLE; | |
586 | add_page_to_unevictable_list(page); | |
6a7b9548 | 587 | /* |
21ee9f39 MK |
588 | * When racing with an mlock or AS_UNEVICTABLE clearing |
589 | * (page is unlocked) make sure that if the other thread | |
590 | * does not observe our setting of PG_lru and fails | |
24513264 | 591 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 592 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
593 | * the page back to the evictable list. |
594 | * | |
21ee9f39 | 595 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
596 | */ |
597 | smp_mb(); | |
894bc310 | 598 | } |
894bc310 LS |
599 | |
600 | /* | |
601 | * page's status can change while we move it among lru. If an evictable | |
602 | * page is on unevictable list, it never be freed. To avoid that, | |
603 | * check after we added it to the list, again. | |
604 | */ | |
605 | if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { | |
606 | if (!isolate_lru_page(page)) { | |
607 | put_page(page); | |
608 | goto redo; | |
609 | } | |
610 | /* This means someone else dropped this page from LRU | |
611 | * So, it will be freed or putback to LRU again. There is | |
612 | * nothing to do here. | |
613 | */ | |
614 | } | |
615 | ||
bbfd28ee LS |
616 | if (was_unevictable && lru != LRU_UNEVICTABLE) |
617 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
618 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) | |
619 | count_vm_event(UNEVICTABLE_PGCULLED); | |
620 | ||
894bc310 LS |
621 | put_page(page); /* drop ref from isolate */ |
622 | } | |
623 | ||
dfc8d636 JW |
624 | enum page_references { |
625 | PAGEREF_RECLAIM, | |
626 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 627 | PAGEREF_KEEP, |
dfc8d636 JW |
628 | PAGEREF_ACTIVATE, |
629 | }; | |
630 | ||
631 | static enum page_references page_check_references(struct page *page, | |
f16015fb | 632 | struct mem_cgroup_zone *mz, |
dfc8d636 JW |
633 | struct scan_control *sc) |
634 | { | |
64574746 | 635 | int referenced_ptes, referenced_page; |
dfc8d636 | 636 | unsigned long vm_flags; |
dfc8d636 | 637 | |
c3ac9a8a JW |
638 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
639 | &vm_flags); | |
64574746 | 640 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 641 | |
dfc8d636 JW |
642 | /* |
643 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
644 | * move the page to the unevictable list. | |
645 | */ | |
646 | if (vm_flags & VM_LOCKED) | |
647 | return PAGEREF_RECLAIM; | |
648 | ||
64574746 | 649 | if (referenced_ptes) { |
e4898273 | 650 | if (PageSwapBacked(page)) |
64574746 JW |
651 | return PAGEREF_ACTIVATE; |
652 | /* | |
653 | * All mapped pages start out with page table | |
654 | * references from the instantiating fault, so we need | |
655 | * to look twice if a mapped file page is used more | |
656 | * than once. | |
657 | * | |
658 | * Mark it and spare it for another trip around the | |
659 | * inactive list. Another page table reference will | |
660 | * lead to its activation. | |
661 | * | |
662 | * Note: the mark is set for activated pages as well | |
663 | * so that recently deactivated but used pages are | |
664 | * quickly recovered. | |
665 | */ | |
666 | SetPageReferenced(page); | |
667 | ||
34dbc67a | 668 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
669 | return PAGEREF_ACTIVATE; |
670 | ||
c909e993 KK |
671 | /* |
672 | * Activate file-backed executable pages after first usage. | |
673 | */ | |
674 | if (vm_flags & VM_EXEC) | |
675 | return PAGEREF_ACTIVATE; | |
676 | ||
64574746 JW |
677 | return PAGEREF_KEEP; |
678 | } | |
dfc8d636 JW |
679 | |
680 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 681 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
682 | return PAGEREF_RECLAIM_CLEAN; |
683 | ||
684 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
685 | } |
686 | ||
1da177e4 | 687 | /* |
1742f19f | 688 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 689 | */ |
1742f19f | 690 | static unsigned long shrink_page_list(struct list_head *page_list, |
f16015fb | 691 | struct mem_cgroup_zone *mz, |
f84f6e2b | 692 | struct scan_control *sc, |
92df3a72 MG |
693 | unsigned long *ret_nr_dirty, |
694 | unsigned long *ret_nr_writeback) | |
1da177e4 LT |
695 | { |
696 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 697 | LIST_HEAD(free_pages); |
1da177e4 | 698 | int pgactivate = 0; |
0e093d99 MG |
699 | unsigned long nr_dirty = 0; |
700 | unsigned long nr_congested = 0; | |
05ff5137 | 701 | unsigned long nr_reclaimed = 0; |
92df3a72 | 702 | unsigned long nr_writeback = 0; |
1da177e4 LT |
703 | |
704 | cond_resched(); | |
705 | ||
1da177e4 | 706 | while (!list_empty(page_list)) { |
dfc8d636 | 707 | enum page_references references; |
1da177e4 LT |
708 | struct address_space *mapping; |
709 | struct page *page; | |
710 | int may_enter_fs; | |
1da177e4 LT |
711 | |
712 | cond_resched(); | |
713 | ||
714 | page = lru_to_page(page_list); | |
715 | list_del(&page->lru); | |
716 | ||
529ae9aa | 717 | if (!trylock_page(page)) |
1da177e4 LT |
718 | goto keep; |
719 | ||
725d704e | 720 | VM_BUG_ON(PageActive(page)); |
f16015fb | 721 | VM_BUG_ON(page_zone(page) != mz->zone); |
1da177e4 LT |
722 | |
723 | sc->nr_scanned++; | |
80e43426 | 724 | |
b291f000 NP |
725 | if (unlikely(!page_evictable(page, NULL))) |
726 | goto cull_mlocked; | |
894bc310 | 727 | |
a6dc60f8 | 728 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
729 | goto keep_locked; |
730 | ||
1da177e4 LT |
731 | /* Double the slab pressure for mapped and swapcache pages */ |
732 | if (page_mapped(page) || PageSwapCache(page)) | |
733 | sc->nr_scanned++; | |
734 | ||
c661b078 AW |
735 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
736 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
737 | ||
738 | if (PageWriteback(page)) { | |
92df3a72 | 739 | nr_writeback++; |
41ac1999 MG |
740 | unlock_page(page); |
741 | goto keep; | |
c661b078 | 742 | } |
1da177e4 | 743 | |
f16015fb | 744 | references = page_check_references(page, mz, sc); |
dfc8d636 JW |
745 | switch (references) { |
746 | case PAGEREF_ACTIVATE: | |
1da177e4 | 747 | goto activate_locked; |
64574746 JW |
748 | case PAGEREF_KEEP: |
749 | goto keep_locked; | |
dfc8d636 JW |
750 | case PAGEREF_RECLAIM: |
751 | case PAGEREF_RECLAIM_CLEAN: | |
752 | ; /* try to reclaim the page below */ | |
753 | } | |
1da177e4 | 754 | |
1da177e4 LT |
755 | /* |
756 | * Anonymous process memory has backing store? | |
757 | * Try to allocate it some swap space here. | |
758 | */ | |
b291f000 | 759 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
760 | if (!(sc->gfp_mask & __GFP_IO)) |
761 | goto keep_locked; | |
ac47b003 | 762 | if (!add_to_swap(page)) |
1da177e4 | 763 | goto activate_locked; |
63eb6b93 | 764 | may_enter_fs = 1; |
b291f000 | 765 | } |
1da177e4 LT |
766 | |
767 | mapping = page_mapping(page); | |
1da177e4 LT |
768 | |
769 | /* | |
770 | * The page is mapped into the page tables of one or more | |
771 | * processes. Try to unmap it here. | |
772 | */ | |
773 | if (page_mapped(page) && mapping) { | |
14fa31b8 | 774 | switch (try_to_unmap(page, TTU_UNMAP)) { |
1da177e4 LT |
775 | case SWAP_FAIL: |
776 | goto activate_locked; | |
777 | case SWAP_AGAIN: | |
778 | goto keep_locked; | |
b291f000 NP |
779 | case SWAP_MLOCK: |
780 | goto cull_mlocked; | |
1da177e4 LT |
781 | case SWAP_SUCCESS: |
782 | ; /* try to free the page below */ | |
783 | } | |
784 | } | |
785 | ||
786 | if (PageDirty(page)) { | |
0e093d99 MG |
787 | nr_dirty++; |
788 | ||
ee72886d MG |
789 | /* |
790 | * Only kswapd can writeback filesystem pages to | |
f84f6e2b MG |
791 | * avoid risk of stack overflow but do not writeback |
792 | * unless under significant pressure. | |
ee72886d | 793 | */ |
f84f6e2b | 794 | if (page_is_file_cache(page) && |
9e3b2f8c KK |
795 | (!current_is_kswapd() || |
796 | sc->priority >= DEF_PRIORITY - 2)) { | |
49ea7eb6 MG |
797 | /* |
798 | * Immediately reclaim when written back. | |
799 | * Similar in principal to deactivate_page() | |
800 | * except we already have the page isolated | |
801 | * and know it's dirty | |
802 | */ | |
803 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
804 | SetPageReclaim(page); | |
805 | ||
ee72886d MG |
806 | goto keep_locked; |
807 | } | |
808 | ||
dfc8d636 | 809 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 810 | goto keep_locked; |
4dd4b920 | 811 | if (!may_enter_fs) |
1da177e4 | 812 | goto keep_locked; |
52a8363e | 813 | if (!sc->may_writepage) |
1da177e4 LT |
814 | goto keep_locked; |
815 | ||
816 | /* Page is dirty, try to write it out here */ | |
7d3579e8 | 817 | switch (pageout(page, mapping, sc)) { |
1da177e4 | 818 | case PAGE_KEEP: |
0e093d99 | 819 | nr_congested++; |
1da177e4 LT |
820 | goto keep_locked; |
821 | case PAGE_ACTIVATE: | |
822 | goto activate_locked; | |
823 | case PAGE_SUCCESS: | |
7d3579e8 | 824 | if (PageWriteback(page)) |
41ac1999 | 825 | goto keep; |
7d3579e8 | 826 | if (PageDirty(page)) |
1da177e4 | 827 | goto keep; |
7d3579e8 | 828 | |
1da177e4 LT |
829 | /* |
830 | * A synchronous write - probably a ramdisk. Go | |
831 | * ahead and try to reclaim the page. | |
832 | */ | |
529ae9aa | 833 | if (!trylock_page(page)) |
1da177e4 LT |
834 | goto keep; |
835 | if (PageDirty(page) || PageWriteback(page)) | |
836 | goto keep_locked; | |
837 | mapping = page_mapping(page); | |
838 | case PAGE_CLEAN: | |
839 | ; /* try to free the page below */ | |
840 | } | |
841 | } | |
842 | ||
843 | /* | |
844 | * If the page has buffers, try to free the buffer mappings | |
845 | * associated with this page. If we succeed we try to free | |
846 | * the page as well. | |
847 | * | |
848 | * We do this even if the page is PageDirty(). | |
849 | * try_to_release_page() does not perform I/O, but it is | |
850 | * possible for a page to have PageDirty set, but it is actually | |
851 | * clean (all its buffers are clean). This happens if the | |
852 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 853 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
854 | * try_to_release_page() will discover that cleanness and will |
855 | * drop the buffers and mark the page clean - it can be freed. | |
856 | * | |
857 | * Rarely, pages can have buffers and no ->mapping. These are | |
858 | * the pages which were not successfully invalidated in | |
859 | * truncate_complete_page(). We try to drop those buffers here | |
860 | * and if that worked, and the page is no longer mapped into | |
861 | * process address space (page_count == 1) it can be freed. | |
862 | * Otherwise, leave the page on the LRU so it is swappable. | |
863 | */ | |
266cf658 | 864 | if (page_has_private(page)) { |
1da177e4 LT |
865 | if (!try_to_release_page(page, sc->gfp_mask)) |
866 | goto activate_locked; | |
e286781d NP |
867 | if (!mapping && page_count(page) == 1) { |
868 | unlock_page(page); | |
869 | if (put_page_testzero(page)) | |
870 | goto free_it; | |
871 | else { | |
872 | /* | |
873 | * rare race with speculative reference. | |
874 | * the speculative reference will free | |
875 | * this page shortly, so we may | |
876 | * increment nr_reclaimed here (and | |
877 | * leave it off the LRU). | |
878 | */ | |
879 | nr_reclaimed++; | |
880 | continue; | |
881 | } | |
882 | } | |
1da177e4 LT |
883 | } |
884 | ||
e286781d | 885 | if (!mapping || !__remove_mapping(mapping, page)) |
49d2e9cc | 886 | goto keep_locked; |
1da177e4 | 887 | |
a978d6f5 NP |
888 | /* |
889 | * At this point, we have no other references and there is | |
890 | * no way to pick any more up (removed from LRU, removed | |
891 | * from pagecache). Can use non-atomic bitops now (and | |
892 | * we obviously don't have to worry about waking up a process | |
893 | * waiting on the page lock, because there are no references. | |
894 | */ | |
895 | __clear_page_locked(page); | |
e286781d | 896 | free_it: |
05ff5137 | 897 | nr_reclaimed++; |
abe4c3b5 MG |
898 | |
899 | /* | |
900 | * Is there need to periodically free_page_list? It would | |
901 | * appear not as the counts should be low | |
902 | */ | |
903 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
904 | continue; |
905 | ||
b291f000 | 906 | cull_mlocked: |
63d6c5ad HD |
907 | if (PageSwapCache(page)) |
908 | try_to_free_swap(page); | |
b291f000 NP |
909 | unlock_page(page); |
910 | putback_lru_page(page); | |
911 | continue; | |
912 | ||
1da177e4 | 913 | activate_locked: |
68a22394 RR |
914 | /* Not a candidate for swapping, so reclaim swap space. */ |
915 | if (PageSwapCache(page) && vm_swap_full()) | |
a2c43eed | 916 | try_to_free_swap(page); |
894bc310 | 917 | VM_BUG_ON(PageActive(page)); |
1da177e4 LT |
918 | SetPageActive(page); |
919 | pgactivate++; | |
920 | keep_locked: | |
921 | unlock_page(page); | |
922 | keep: | |
923 | list_add(&page->lru, &ret_pages); | |
b291f000 | 924 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); |
1da177e4 | 925 | } |
abe4c3b5 | 926 | |
0e093d99 MG |
927 | /* |
928 | * Tag a zone as congested if all the dirty pages encountered were | |
929 | * backed by a congested BDI. In this case, reclaimers should just | |
930 | * back off and wait for congestion to clear because further reclaim | |
931 | * will encounter the same problem | |
932 | */ | |
89b5fae5 | 933 | if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc)) |
f16015fb | 934 | zone_set_flag(mz->zone, ZONE_CONGESTED); |
0e093d99 | 935 | |
cc59850e | 936 | free_hot_cold_page_list(&free_pages, 1); |
abe4c3b5 | 937 | |
1da177e4 | 938 | list_splice(&ret_pages, page_list); |
f8891e5e | 939 | count_vm_events(PGACTIVATE, pgactivate); |
92df3a72 MG |
940 | *ret_nr_dirty += nr_dirty; |
941 | *ret_nr_writeback += nr_writeback; | |
05ff5137 | 942 | return nr_reclaimed; |
1da177e4 LT |
943 | } |
944 | ||
5ad333eb AW |
945 | /* |
946 | * Attempt to remove the specified page from its LRU. Only take this page | |
947 | * if it is of the appropriate PageActive status. Pages which are being | |
948 | * freed elsewhere are also ignored. | |
949 | * | |
950 | * page: page to consider | |
951 | * mode: one of the LRU isolation modes defined above | |
952 | * | |
953 | * returns 0 on success, -ve errno on failure. | |
954 | */ | |
f3fd4a61 | 955 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
956 | { |
957 | int ret = -EINVAL; | |
958 | ||
959 | /* Only take pages on the LRU. */ | |
960 | if (!PageLRU(page)) | |
961 | return ret; | |
962 | ||
c53919ad | 963 | /* Do not give back unevictable pages for compaction */ |
894bc310 LS |
964 | if (PageUnevictable(page)) |
965 | return ret; | |
966 | ||
5ad333eb | 967 | ret = -EBUSY; |
08e552c6 | 968 | |
c8244935 MG |
969 | /* |
970 | * To minimise LRU disruption, the caller can indicate that it only | |
971 | * wants to isolate pages it will be able to operate on without | |
972 | * blocking - clean pages for the most part. | |
973 | * | |
974 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
975 | * is used by reclaim when it is cannot write to backing storage | |
976 | * | |
977 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
978 | * that it is possible to migrate without blocking | |
979 | */ | |
980 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
981 | /* All the caller can do on PageWriteback is block */ | |
982 | if (PageWriteback(page)) | |
983 | return ret; | |
984 | ||
985 | if (PageDirty(page)) { | |
986 | struct address_space *mapping; | |
987 | ||
988 | /* ISOLATE_CLEAN means only clean pages */ | |
989 | if (mode & ISOLATE_CLEAN) | |
990 | return ret; | |
991 | ||
992 | /* | |
993 | * Only pages without mappings or that have a | |
994 | * ->migratepage callback are possible to migrate | |
995 | * without blocking | |
996 | */ | |
997 | mapping = page_mapping(page); | |
998 | if (mapping && !mapping->a_ops->migratepage) | |
999 | return ret; | |
1000 | } | |
1001 | } | |
39deaf85 | 1002 | |
f80c0673 MK |
1003 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1004 | return ret; | |
1005 | ||
5ad333eb AW |
1006 | if (likely(get_page_unless_zero(page))) { |
1007 | /* | |
1008 | * Be careful not to clear PageLRU until after we're | |
1009 | * sure the page is not being freed elsewhere -- the | |
1010 | * page release code relies on it. | |
1011 | */ | |
1012 | ClearPageLRU(page); | |
1013 | ret = 0; | |
1014 | } | |
1015 | ||
1016 | return ret; | |
1017 | } | |
1018 | ||
1da177e4 LT |
1019 | /* |
1020 | * zone->lru_lock is heavily contended. Some of the functions that | |
1021 | * shrink the lists perform better by taking out a batch of pages | |
1022 | * and working on them outside the LRU lock. | |
1023 | * | |
1024 | * For pagecache intensive workloads, this function is the hottest | |
1025 | * spot in the kernel (apart from copy_*_user functions). | |
1026 | * | |
1027 | * Appropriate locks must be held before calling this function. | |
1028 | * | |
1029 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1030 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1031 | * @dst: The temp list to put pages on to. |
f626012d | 1032 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1033 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1034 | * @mode: One of the LRU isolation modes |
3cb99451 | 1035 | * @lru: LRU list id for isolating |
1da177e4 LT |
1036 | * |
1037 | * returns how many pages were moved onto *@dst. | |
1038 | */ | |
69e05944 | 1039 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1040 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1041 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1042 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1043 | { |
f626012d | 1044 | struct list_head *src; |
69e05944 | 1045 | unsigned long nr_taken = 0; |
c9b02d97 | 1046 | unsigned long scan; |
3cb99451 | 1047 | int file = is_file_lru(lru); |
f626012d | 1048 | |
f626012d | 1049 | src = &lruvec->lists[lru]; |
1da177e4 | 1050 | |
c9b02d97 | 1051 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
5ad333eb | 1052 | struct page *page; |
5ad333eb | 1053 | |
1da177e4 LT |
1054 | page = lru_to_page(src); |
1055 | prefetchw_prev_lru_page(page, src, flags); | |
1056 | ||
725d704e | 1057 | VM_BUG_ON(!PageLRU(page)); |
8d438f96 | 1058 | |
f3fd4a61 | 1059 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1060 | case 0: |
bbf808ed | 1061 | mem_cgroup_lru_del_list(page, lru); |
5ad333eb | 1062 | list_move(&page->lru, dst); |
2c888cfb | 1063 | nr_taken += hpage_nr_pages(page); |
5ad333eb AW |
1064 | break; |
1065 | ||
1066 | case -EBUSY: | |
1067 | /* else it is being freed elsewhere */ | |
1068 | list_move(&page->lru, src); | |
1069 | continue; | |
46453a6e | 1070 | |
5ad333eb AW |
1071 | default: |
1072 | BUG(); | |
1073 | } | |
1da177e4 LT |
1074 | } |
1075 | ||
f626012d | 1076 | *nr_scanned = scan; |
a8a94d15 | 1077 | |
fe2c2a10 | 1078 | trace_mm_vmscan_lru_isolate(sc->order, |
a8a94d15 MG |
1079 | nr_to_scan, scan, |
1080 | nr_taken, | |
ea4d349f | 1081 | mode, file); |
1da177e4 LT |
1082 | return nr_taken; |
1083 | } | |
1084 | ||
62695a84 NP |
1085 | /** |
1086 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1087 | * @page: page to isolate from its LRU list | |
1088 | * | |
1089 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1090 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1091 | * | |
1092 | * Returns 0 if the page was removed from an LRU list. | |
1093 | * Returns -EBUSY if the page was not on an LRU list. | |
1094 | * | |
1095 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1096 | * the active list, it will have PageActive set. If it was found on |
1097 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1098 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1099 | * |
1100 | * The vmstat statistic corresponding to the list on which the page was | |
1101 | * found will be decremented. | |
1102 | * | |
1103 | * Restrictions: | |
1104 | * (1) Must be called with an elevated refcount on the page. This is a | |
1105 | * fundamentnal difference from isolate_lru_pages (which is called | |
1106 | * without a stable reference). | |
1107 | * (2) the lru_lock must not be held. | |
1108 | * (3) interrupts must be enabled. | |
1109 | */ | |
1110 | int isolate_lru_page(struct page *page) | |
1111 | { | |
1112 | int ret = -EBUSY; | |
1113 | ||
0c917313 KK |
1114 | VM_BUG_ON(!page_count(page)); |
1115 | ||
62695a84 NP |
1116 | if (PageLRU(page)) { |
1117 | struct zone *zone = page_zone(page); | |
1118 | ||
1119 | spin_lock_irq(&zone->lru_lock); | |
0c917313 | 1120 | if (PageLRU(page)) { |
894bc310 | 1121 | int lru = page_lru(page); |
62695a84 | 1122 | ret = 0; |
0c917313 | 1123 | get_page(page); |
62695a84 | 1124 | ClearPageLRU(page); |
4f98a2fe | 1125 | |
4f98a2fe | 1126 | del_page_from_lru_list(zone, page, lru); |
62695a84 NP |
1127 | } |
1128 | spin_unlock_irq(&zone->lru_lock); | |
1129 | } | |
1130 | return ret; | |
1131 | } | |
1132 | ||
35cd7815 RR |
1133 | /* |
1134 | * Are there way too many processes in the direct reclaim path already? | |
1135 | */ | |
1136 | static int too_many_isolated(struct zone *zone, int file, | |
1137 | struct scan_control *sc) | |
1138 | { | |
1139 | unsigned long inactive, isolated; | |
1140 | ||
1141 | if (current_is_kswapd()) | |
1142 | return 0; | |
1143 | ||
89b5fae5 | 1144 | if (!global_reclaim(sc)) |
35cd7815 RR |
1145 | return 0; |
1146 | ||
1147 | if (file) { | |
1148 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1149 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1150 | } else { | |
1151 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1152 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1153 | } | |
1154 | ||
1155 | return isolated > inactive; | |
1156 | } | |
1157 | ||
66635629 | 1158 | static noinline_for_stack void |
3f79768f HD |
1159 | putback_inactive_pages(struct mem_cgroup_zone *mz, |
1160 | struct list_head *page_list) | |
66635629 | 1161 | { |
f16015fb | 1162 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
3f79768f HD |
1163 | struct zone *zone = mz->zone; |
1164 | LIST_HEAD(pages_to_free); | |
66635629 | 1165 | |
66635629 MG |
1166 | /* |
1167 | * Put back any unfreeable pages. | |
1168 | */ | |
66635629 | 1169 | while (!list_empty(page_list)) { |
3f79768f | 1170 | struct page *page = lru_to_page(page_list); |
66635629 | 1171 | int lru; |
3f79768f | 1172 | |
66635629 MG |
1173 | VM_BUG_ON(PageLRU(page)); |
1174 | list_del(&page->lru); | |
1175 | if (unlikely(!page_evictable(page, NULL))) { | |
1176 | spin_unlock_irq(&zone->lru_lock); | |
1177 | putback_lru_page(page); | |
1178 | spin_lock_irq(&zone->lru_lock); | |
1179 | continue; | |
1180 | } | |
7a608572 | 1181 | SetPageLRU(page); |
66635629 | 1182 | lru = page_lru(page); |
7a608572 | 1183 | add_page_to_lru_list(zone, page, lru); |
66635629 MG |
1184 | if (is_active_lru(lru)) { |
1185 | int file = is_file_lru(lru); | |
9992af10 RR |
1186 | int numpages = hpage_nr_pages(page); |
1187 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1188 | } |
2bcf8879 HD |
1189 | if (put_page_testzero(page)) { |
1190 | __ClearPageLRU(page); | |
1191 | __ClearPageActive(page); | |
1192 | del_page_from_lru_list(zone, page, lru); | |
1193 | ||
1194 | if (unlikely(PageCompound(page))) { | |
1195 | spin_unlock_irq(&zone->lru_lock); | |
1196 | (*get_compound_page_dtor(page))(page); | |
1197 | spin_lock_irq(&zone->lru_lock); | |
1198 | } else | |
1199 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1200 | } |
1201 | } | |
66635629 | 1202 | |
3f79768f HD |
1203 | /* |
1204 | * To save our caller's stack, now use input list for pages to free. | |
1205 | */ | |
1206 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1207 | } |
1208 | ||
f16015fb JW |
1209 | static noinline_for_stack void |
1210 | update_isolated_counts(struct mem_cgroup_zone *mz, | |
3f79768f | 1211 | struct list_head *page_list, |
f16015fb | 1212 | unsigned long *nr_anon, |
3f79768f | 1213 | unsigned long *nr_file) |
1489fa14 | 1214 | { |
f16015fb | 1215 | struct zone *zone = mz->zone; |
1489fa14 | 1216 | unsigned int count[NR_LRU_LISTS] = { 0, }; |
3f79768f HD |
1217 | unsigned long nr_active = 0; |
1218 | struct page *page; | |
1219 | int lru; | |
1220 | ||
1221 | /* | |
1222 | * Count pages and clear active flags | |
1223 | */ | |
1224 | list_for_each_entry(page, page_list, lru) { | |
1225 | int numpages = hpage_nr_pages(page); | |
1226 | lru = page_lru_base_type(page); | |
1227 | if (PageActive(page)) { | |
1228 | lru += LRU_ACTIVE; | |
1229 | ClearPageActive(page); | |
1230 | nr_active += numpages; | |
1231 | } | |
1232 | count[lru] += numpages; | |
1233 | } | |
1489fa14 | 1234 | |
d563c050 | 1235 | preempt_disable(); |
1489fa14 MG |
1236 | __count_vm_events(PGDEACTIVATE, nr_active); |
1237 | ||
1238 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, | |
1239 | -count[LRU_ACTIVE_FILE]); | |
1240 | __mod_zone_page_state(zone, NR_INACTIVE_FILE, | |
1241 | -count[LRU_INACTIVE_FILE]); | |
1242 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, | |
1243 | -count[LRU_ACTIVE_ANON]); | |
1244 | __mod_zone_page_state(zone, NR_INACTIVE_ANON, | |
1245 | -count[LRU_INACTIVE_ANON]); | |
1246 | ||
1247 | *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON]; | |
1248 | *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE]; | |
1489fa14 | 1249 | |
d563c050 HD |
1250 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon); |
1251 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file); | |
1252 | preempt_enable(); | |
1489fa14 MG |
1253 | } |
1254 | ||
1da177e4 | 1255 | /* |
1742f19f AM |
1256 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1257 | * of reclaimed pages | |
1da177e4 | 1258 | */ |
66635629 | 1259 | static noinline_for_stack unsigned long |
f16015fb | 1260 | shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz, |
9e3b2f8c | 1261 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1262 | { |
1263 | LIST_HEAD(page_list); | |
e247dbce | 1264 | unsigned long nr_scanned; |
05ff5137 | 1265 | unsigned long nr_reclaimed = 0; |
e247dbce | 1266 | unsigned long nr_taken; |
e247dbce KM |
1267 | unsigned long nr_anon; |
1268 | unsigned long nr_file; | |
92df3a72 MG |
1269 | unsigned long nr_dirty = 0; |
1270 | unsigned long nr_writeback = 0; | |
f3fd4a61 | 1271 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1272 | int file = is_file_lru(lru); |
f16015fb | 1273 | struct zone *zone = mz->zone; |
d563c050 | 1274 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
5dc35979 | 1275 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, mz->mem_cgroup); |
78dc583d | 1276 | |
35cd7815 | 1277 | while (unlikely(too_many_isolated(zone, file, sc))) { |
58355c78 | 1278 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1279 | |
1280 | /* We are about to die and free our memory. Return now. */ | |
1281 | if (fatal_signal_pending(current)) | |
1282 | return SWAP_CLUSTER_MAX; | |
1283 | } | |
1284 | ||
1da177e4 | 1285 | lru_add_drain(); |
f80c0673 MK |
1286 | |
1287 | if (!sc->may_unmap) | |
61317289 | 1288 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1289 | if (!sc->may_writepage) |
61317289 | 1290 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1291 | |
1da177e4 | 1292 | spin_lock_irq(&zone->lru_lock); |
b35ea17b | 1293 | |
5dc35979 KK |
1294 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1295 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1296 | if (global_reclaim(sc)) { |
e247dbce KM |
1297 | zone->pages_scanned += nr_scanned; |
1298 | if (current_is_kswapd()) | |
1299 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, | |
1300 | nr_scanned); | |
1301 | else | |
1302 | __count_zone_vm_events(PGSCAN_DIRECT, zone, | |
1303 | nr_scanned); | |
e247dbce | 1304 | } |
d563c050 | 1305 | spin_unlock_irq(&zone->lru_lock); |
b35ea17b | 1306 | |
d563c050 | 1307 | if (nr_taken == 0) |
66635629 | 1308 | return 0; |
5ad333eb | 1309 | |
3f79768f HD |
1310 | update_isolated_counts(mz, &page_list, &nr_anon, &nr_file); |
1311 | ||
9e3b2f8c | 1312 | nr_reclaimed = shrink_page_list(&page_list, mz, sc, |
92df3a72 | 1313 | &nr_dirty, &nr_writeback); |
c661b078 | 1314 | |
3f79768f HD |
1315 | spin_lock_irq(&zone->lru_lock); |
1316 | ||
d563c050 HD |
1317 | reclaim_stat->recent_scanned[0] += nr_anon; |
1318 | reclaim_stat->recent_scanned[1] += nr_file; | |
1319 | ||
904249aa YH |
1320 | if (global_reclaim(sc)) { |
1321 | if (current_is_kswapd()) | |
1322 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1323 | nr_reclaimed); | |
1324 | else | |
1325 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1326 | nr_reclaimed); | |
1327 | } | |
a74609fa | 1328 | |
3f79768f HD |
1329 | putback_inactive_pages(mz, &page_list); |
1330 | ||
1331 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon); | |
1332 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file); | |
1333 | ||
1334 | spin_unlock_irq(&zone->lru_lock); | |
1335 | ||
1336 | free_hot_cold_page_list(&page_list, 1); | |
e11da5b4 | 1337 | |
92df3a72 MG |
1338 | /* |
1339 | * If reclaim is isolating dirty pages under writeback, it implies | |
1340 | * that the long-lived page allocation rate is exceeding the page | |
1341 | * laundering rate. Either the global limits are not being effective | |
1342 | * at throttling processes due to the page distribution throughout | |
1343 | * zones or there is heavy usage of a slow backing device. The | |
1344 | * only option is to throttle from reclaim context which is not ideal | |
1345 | * as there is no guarantee the dirtying process is throttled in the | |
1346 | * same way balance_dirty_pages() manages. | |
1347 | * | |
1348 | * This scales the number of dirty pages that must be under writeback | |
1349 | * before throttling depending on priority. It is a simple backoff | |
1350 | * function that has the most effect in the range DEF_PRIORITY to | |
1351 | * DEF_PRIORITY-2 which is the priority reclaim is considered to be | |
1352 | * in trouble and reclaim is considered to be in trouble. | |
1353 | * | |
1354 | * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle | |
1355 | * DEF_PRIORITY-1 50% must be PageWriteback | |
1356 | * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble | |
1357 | * ... | |
1358 | * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any | |
1359 | * isolated page is PageWriteback | |
1360 | */ | |
9e3b2f8c KK |
1361 | if (nr_writeback && nr_writeback >= |
1362 | (nr_taken >> (DEF_PRIORITY - sc->priority))) | |
92df3a72 MG |
1363 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); |
1364 | ||
e11da5b4 MG |
1365 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, |
1366 | zone_idx(zone), | |
1367 | nr_scanned, nr_reclaimed, | |
9e3b2f8c | 1368 | sc->priority, |
23b9da55 | 1369 | trace_shrink_flags(file)); |
05ff5137 | 1370 | return nr_reclaimed; |
1da177e4 LT |
1371 | } |
1372 | ||
1373 | /* | |
1374 | * This moves pages from the active list to the inactive list. | |
1375 | * | |
1376 | * We move them the other way if the page is referenced by one or more | |
1377 | * processes, from rmap. | |
1378 | * | |
1379 | * If the pages are mostly unmapped, the processing is fast and it is | |
1380 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1381 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1382 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1383 | * this, so instead we remove the pages from the LRU while processing them. | |
1384 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1385 | * nobody will play with that bit on a non-LRU page. | |
1386 | * | |
1387 | * The downside is that we have to touch page->_count against each page. | |
1388 | * But we had to alter page->flags anyway. | |
1389 | */ | |
1cfb419b | 1390 | |
3eb4140f WF |
1391 | static void move_active_pages_to_lru(struct zone *zone, |
1392 | struct list_head *list, | |
2bcf8879 | 1393 | struct list_head *pages_to_free, |
3eb4140f WF |
1394 | enum lru_list lru) |
1395 | { | |
1396 | unsigned long pgmoved = 0; | |
3eb4140f WF |
1397 | struct page *page; |
1398 | ||
3eb4140f | 1399 | while (!list_empty(list)) { |
925b7673 JW |
1400 | struct lruvec *lruvec; |
1401 | ||
3eb4140f | 1402 | page = lru_to_page(list); |
3eb4140f WF |
1403 | |
1404 | VM_BUG_ON(PageLRU(page)); | |
1405 | SetPageLRU(page); | |
1406 | ||
925b7673 JW |
1407 | lruvec = mem_cgroup_lru_add_list(zone, page, lru); |
1408 | list_move(&page->lru, &lruvec->lists[lru]); | |
2c888cfb | 1409 | pgmoved += hpage_nr_pages(page); |
3eb4140f | 1410 | |
2bcf8879 HD |
1411 | if (put_page_testzero(page)) { |
1412 | __ClearPageLRU(page); | |
1413 | __ClearPageActive(page); | |
1414 | del_page_from_lru_list(zone, page, lru); | |
1415 | ||
1416 | if (unlikely(PageCompound(page))) { | |
1417 | spin_unlock_irq(&zone->lru_lock); | |
1418 | (*get_compound_page_dtor(page))(page); | |
1419 | spin_lock_irq(&zone->lru_lock); | |
1420 | } else | |
1421 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1422 | } |
1423 | } | |
1424 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1425 | if (!is_active_lru(lru)) | |
1426 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1427 | } | |
1cfb419b | 1428 | |
f626012d | 1429 | static void shrink_active_list(unsigned long nr_to_scan, |
f16015fb JW |
1430 | struct mem_cgroup_zone *mz, |
1431 | struct scan_control *sc, | |
9e3b2f8c | 1432 | enum lru_list lru) |
1da177e4 | 1433 | { |
44c241f1 | 1434 | unsigned long nr_taken; |
f626012d | 1435 | unsigned long nr_scanned; |
6fe6b7e3 | 1436 | unsigned long vm_flags; |
1da177e4 | 1437 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1438 | LIST_HEAD(l_active); |
b69408e8 | 1439 | LIST_HEAD(l_inactive); |
1da177e4 | 1440 | struct page *page; |
f16015fb | 1441 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
44c241f1 | 1442 | unsigned long nr_rotated = 0; |
f3fd4a61 | 1443 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1444 | int file = is_file_lru(lru); |
f16015fb | 1445 | struct zone *zone = mz->zone; |
5dc35979 | 1446 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, mz->mem_cgroup); |
1da177e4 LT |
1447 | |
1448 | lru_add_drain(); | |
f80c0673 MK |
1449 | |
1450 | if (!sc->may_unmap) | |
61317289 | 1451 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1452 | if (!sc->may_writepage) |
61317289 | 1453 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1454 | |
1da177e4 | 1455 | spin_lock_irq(&zone->lru_lock); |
925b7673 | 1456 | |
5dc35979 KK |
1457 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1458 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1459 | if (global_reclaim(sc)) |
f626012d | 1460 | zone->pages_scanned += nr_scanned; |
89b5fae5 | 1461 | |
b7c46d15 | 1462 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1463 | |
f626012d | 1464 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); |
3cb99451 | 1465 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); |
a731286d | 1466 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
1da177e4 LT |
1467 | spin_unlock_irq(&zone->lru_lock); |
1468 | ||
1da177e4 LT |
1469 | while (!list_empty(&l_hold)) { |
1470 | cond_resched(); | |
1471 | page = lru_to_page(&l_hold); | |
1472 | list_del(&page->lru); | |
7e9cd484 | 1473 | |
894bc310 LS |
1474 | if (unlikely(!page_evictable(page, NULL))) { |
1475 | putback_lru_page(page); | |
1476 | continue; | |
1477 | } | |
1478 | ||
cc715d99 MG |
1479 | if (unlikely(buffer_heads_over_limit)) { |
1480 | if (page_has_private(page) && trylock_page(page)) { | |
1481 | if (page_has_private(page)) | |
1482 | try_to_release_page(page, 0); | |
1483 | unlock_page(page); | |
1484 | } | |
1485 | } | |
1486 | ||
c3ac9a8a JW |
1487 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1488 | &vm_flags)) { | |
9992af10 | 1489 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1490 | /* |
1491 | * Identify referenced, file-backed active pages and | |
1492 | * give them one more trip around the active list. So | |
1493 | * that executable code get better chances to stay in | |
1494 | * memory under moderate memory pressure. Anon pages | |
1495 | * are not likely to be evicted by use-once streaming | |
1496 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1497 | * so we ignore them here. | |
1498 | */ | |
41e20983 | 1499 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1500 | list_add(&page->lru, &l_active); |
1501 | continue; | |
1502 | } | |
1503 | } | |
7e9cd484 | 1504 | |
5205e56e | 1505 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1506 | list_add(&page->lru, &l_inactive); |
1507 | } | |
1508 | ||
b555749a | 1509 | /* |
8cab4754 | 1510 | * Move pages back to the lru list. |
b555749a | 1511 | */ |
2a1dc509 | 1512 | spin_lock_irq(&zone->lru_lock); |
556adecb | 1513 | /* |
8cab4754 WF |
1514 | * Count referenced pages from currently used mappings as rotated, |
1515 | * even though only some of them are actually re-activated. This | |
1516 | * helps balance scan pressure between file and anonymous pages in | |
1517 | * get_scan_ratio. | |
7e9cd484 | 1518 | */ |
b7c46d15 | 1519 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1520 | |
3cb99451 KK |
1521 | move_active_pages_to_lru(zone, &l_active, &l_hold, lru); |
1522 | move_active_pages_to_lru(zone, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
a731286d | 1523 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
f8891e5e | 1524 | spin_unlock_irq(&zone->lru_lock); |
2bcf8879 HD |
1525 | |
1526 | free_hot_cold_page_list(&l_hold, 1); | |
1da177e4 LT |
1527 | } |
1528 | ||
74e3f3c3 | 1529 | #ifdef CONFIG_SWAP |
14797e23 | 1530 | static int inactive_anon_is_low_global(struct zone *zone) |
f89eb90e KM |
1531 | { |
1532 | unsigned long active, inactive; | |
1533 | ||
1534 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1535 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1536 | ||
1537 | if (inactive * zone->inactive_ratio < active) | |
1538 | return 1; | |
1539 | ||
1540 | return 0; | |
1541 | } | |
1542 | ||
14797e23 KM |
1543 | /** |
1544 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
1545 | * @zone: zone to check | |
1546 | * @sc: scan control of this context | |
1547 | * | |
1548 | * Returns true if the zone does not have enough inactive anon pages, | |
1549 | * meaning some active anon pages need to be deactivated. | |
1550 | */ | |
f16015fb | 1551 | static int inactive_anon_is_low(struct mem_cgroup_zone *mz) |
14797e23 | 1552 | { |
74e3f3c3 MK |
1553 | /* |
1554 | * If we don't have swap space, anonymous page deactivation | |
1555 | * is pointless. | |
1556 | */ | |
1557 | if (!total_swap_pages) | |
1558 | return 0; | |
1559 | ||
c3c787e8 | 1560 | if (!mem_cgroup_disabled()) |
f16015fb JW |
1561 | return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup, |
1562 | mz->zone); | |
1563 | ||
1564 | return inactive_anon_is_low_global(mz->zone); | |
14797e23 | 1565 | } |
74e3f3c3 | 1566 | #else |
f16015fb | 1567 | static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz) |
74e3f3c3 MK |
1568 | { |
1569 | return 0; | |
1570 | } | |
1571 | #endif | |
14797e23 | 1572 | |
56e49d21 RR |
1573 | static int inactive_file_is_low_global(struct zone *zone) |
1574 | { | |
1575 | unsigned long active, inactive; | |
1576 | ||
1577 | active = zone_page_state(zone, NR_ACTIVE_FILE); | |
1578 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1579 | ||
1580 | return (active > inactive); | |
1581 | } | |
1582 | ||
1583 | /** | |
1584 | * inactive_file_is_low - check if file pages need to be deactivated | |
f16015fb | 1585 | * @mz: memory cgroup and zone to check |
56e49d21 RR |
1586 | * |
1587 | * When the system is doing streaming IO, memory pressure here | |
1588 | * ensures that active file pages get deactivated, until more | |
1589 | * than half of the file pages are on the inactive list. | |
1590 | * | |
1591 | * Once we get to that situation, protect the system's working | |
1592 | * set from being evicted by disabling active file page aging. | |
1593 | * | |
1594 | * This uses a different ratio than the anonymous pages, because | |
1595 | * the page cache uses a use-once replacement algorithm. | |
1596 | */ | |
f16015fb | 1597 | static int inactive_file_is_low(struct mem_cgroup_zone *mz) |
56e49d21 | 1598 | { |
c3c787e8 | 1599 | if (!mem_cgroup_disabled()) |
f16015fb JW |
1600 | return mem_cgroup_inactive_file_is_low(mz->mem_cgroup, |
1601 | mz->zone); | |
56e49d21 | 1602 | |
f16015fb | 1603 | return inactive_file_is_low_global(mz->zone); |
56e49d21 RR |
1604 | } |
1605 | ||
f16015fb | 1606 | static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file) |
b39415b2 RR |
1607 | { |
1608 | if (file) | |
f16015fb | 1609 | return inactive_file_is_low(mz); |
b39415b2 | 1610 | else |
f16015fb | 1611 | return inactive_anon_is_low(mz); |
b39415b2 RR |
1612 | } |
1613 | ||
4f98a2fe | 1614 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
f16015fb | 1615 | struct mem_cgroup_zone *mz, |
9e3b2f8c | 1616 | struct scan_control *sc) |
b69408e8 | 1617 | { |
4f98a2fe RR |
1618 | int file = is_file_lru(lru); |
1619 | ||
b39415b2 | 1620 | if (is_active_lru(lru)) { |
f16015fb | 1621 | if (inactive_list_is_low(mz, file)) |
9e3b2f8c | 1622 | shrink_active_list(nr_to_scan, mz, sc, lru); |
556adecb RR |
1623 | return 0; |
1624 | } | |
1625 | ||
9e3b2f8c | 1626 | return shrink_inactive_list(nr_to_scan, mz, sc, lru); |
4f98a2fe RR |
1627 | } |
1628 | ||
3d58ab5c | 1629 | static int vmscan_swappiness(struct scan_control *sc) |
1f4c025b | 1630 | { |
89b5fae5 | 1631 | if (global_reclaim(sc)) |
1f4c025b | 1632 | return vm_swappiness; |
3d58ab5c | 1633 | return mem_cgroup_swappiness(sc->target_mem_cgroup); |
1f4c025b KH |
1634 | } |
1635 | ||
4f98a2fe RR |
1636 | /* |
1637 | * Determine how aggressively the anon and file LRU lists should be | |
1638 | * scanned. The relative value of each set of LRU lists is determined | |
1639 | * by looking at the fraction of the pages scanned we did rotate back | |
1640 | * onto the active list instead of evict. | |
1641 | * | |
76a33fc3 | 1642 | * nr[0] = anon pages to scan; nr[1] = file pages to scan |
4f98a2fe | 1643 | */ |
f16015fb | 1644 | static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc, |
9e3b2f8c | 1645 | unsigned long *nr) |
4f98a2fe RR |
1646 | { |
1647 | unsigned long anon, file, free; | |
1648 | unsigned long anon_prio, file_prio; | |
1649 | unsigned long ap, fp; | |
f16015fb | 1650 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
76a33fc3 | 1651 | u64 fraction[2], denominator; |
4111304d | 1652 | enum lru_list lru; |
76a33fc3 | 1653 | int noswap = 0; |
a4d3e9e7 | 1654 | bool force_scan = false; |
246e87a9 | 1655 | |
f11c0ca5 JW |
1656 | /* |
1657 | * If the zone or memcg is small, nr[l] can be 0. This | |
1658 | * results in no scanning on this priority and a potential | |
1659 | * priority drop. Global direct reclaim can go to the next | |
1660 | * zone and tends to have no problems. Global kswapd is for | |
1661 | * zone balancing and it needs to scan a minimum amount. When | |
1662 | * reclaiming for a memcg, a priority drop can cause high | |
1663 | * latencies, so it's better to scan a minimum amount there as | |
1664 | * well. | |
1665 | */ | |
b95a2f2d | 1666 | if (current_is_kswapd() && mz->zone->all_unreclaimable) |
a4d3e9e7 | 1667 | force_scan = true; |
89b5fae5 | 1668 | if (!global_reclaim(sc)) |
a4d3e9e7 | 1669 | force_scan = true; |
76a33fc3 SL |
1670 | |
1671 | /* If we have no swap space, do not bother scanning anon pages. */ | |
1672 | if (!sc->may_swap || (nr_swap_pages <= 0)) { | |
1673 | noswap = 1; | |
1674 | fraction[0] = 0; | |
1675 | fraction[1] = 1; | |
1676 | denominator = 1; | |
1677 | goto out; | |
1678 | } | |
4f98a2fe | 1679 | |
f16015fb JW |
1680 | anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) + |
1681 | zone_nr_lru_pages(mz, LRU_INACTIVE_ANON); | |
1682 | file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) + | |
1683 | zone_nr_lru_pages(mz, LRU_INACTIVE_FILE); | |
a4d3e9e7 | 1684 | |
89b5fae5 | 1685 | if (global_reclaim(sc)) { |
f16015fb | 1686 | free = zone_page_state(mz->zone, NR_FREE_PAGES); |
eeee9a8c KM |
1687 | /* If we have very few page cache pages, |
1688 | force-scan anon pages. */ | |
f16015fb | 1689 | if (unlikely(file + free <= high_wmark_pages(mz->zone))) { |
76a33fc3 SL |
1690 | fraction[0] = 1; |
1691 | fraction[1] = 0; | |
1692 | denominator = 1; | |
1693 | goto out; | |
eeee9a8c | 1694 | } |
4f98a2fe RR |
1695 | } |
1696 | ||
58c37f6e KM |
1697 | /* |
1698 | * With swappiness at 100, anonymous and file have the same priority. | |
1699 | * This scanning priority is essentially the inverse of IO cost. | |
1700 | */ | |
3d58ab5c KK |
1701 | anon_prio = vmscan_swappiness(sc); |
1702 | file_prio = 200 - vmscan_swappiness(sc); | |
58c37f6e | 1703 | |
4f98a2fe RR |
1704 | /* |
1705 | * OK, so we have swap space and a fair amount of page cache | |
1706 | * pages. We use the recently rotated / recently scanned | |
1707 | * ratios to determine how valuable each cache is. | |
1708 | * | |
1709 | * Because workloads change over time (and to avoid overflow) | |
1710 | * we keep these statistics as a floating average, which ends | |
1711 | * up weighing recent references more than old ones. | |
1712 | * | |
1713 | * anon in [0], file in [1] | |
1714 | */ | |
f16015fb | 1715 | spin_lock_irq(&mz->zone->lru_lock); |
6e901571 | 1716 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
1717 | reclaim_stat->recent_scanned[0] /= 2; |
1718 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
1719 | } |
1720 | ||
6e901571 | 1721 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
1722 | reclaim_stat->recent_scanned[1] /= 2; |
1723 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
1724 | } |
1725 | ||
4f98a2fe | 1726 | /* |
00d8089c RR |
1727 | * The amount of pressure on anon vs file pages is inversely |
1728 | * proportional to the fraction of recently scanned pages on | |
1729 | * each list that were recently referenced and in active use. | |
4f98a2fe | 1730 | */ |
fe35004f | 1731 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 1732 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 1733 | |
fe35004f | 1734 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 1735 | fp /= reclaim_stat->recent_rotated[1] + 1; |
f16015fb | 1736 | spin_unlock_irq(&mz->zone->lru_lock); |
4f98a2fe | 1737 | |
76a33fc3 SL |
1738 | fraction[0] = ap; |
1739 | fraction[1] = fp; | |
1740 | denominator = ap + fp + 1; | |
1741 | out: | |
4111304d HD |
1742 | for_each_evictable_lru(lru) { |
1743 | int file = is_file_lru(lru); | |
76a33fc3 | 1744 | unsigned long scan; |
6e08a369 | 1745 | |
4111304d | 1746 | scan = zone_nr_lru_pages(mz, lru); |
9e3b2f8c KK |
1747 | if (sc->priority || noswap || !vmscan_swappiness(sc)) { |
1748 | scan >>= sc->priority; | |
f11c0ca5 JW |
1749 | if (!scan && force_scan) |
1750 | scan = SWAP_CLUSTER_MAX; | |
76a33fc3 SL |
1751 | scan = div64_u64(scan * fraction[file], denominator); |
1752 | } | |
4111304d | 1753 | nr[lru] = scan; |
76a33fc3 | 1754 | } |
6e08a369 | 1755 | } |
4f98a2fe | 1756 | |
23b9da55 | 1757 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 1758 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 MG |
1759 | { |
1760 | if (COMPACTION_BUILD && sc->order && | |
1761 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || | |
9e3b2f8c | 1762 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
1763 | return true; |
1764 | ||
1765 | return false; | |
1766 | } | |
1767 | ||
3e7d3449 | 1768 | /* |
23b9da55 MG |
1769 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
1770 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
1771 | * true if more pages should be reclaimed such that when the page allocator | |
1772 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
1773 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 1774 | */ |
f16015fb | 1775 | static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz, |
3e7d3449 MG |
1776 | unsigned long nr_reclaimed, |
1777 | unsigned long nr_scanned, | |
1778 | struct scan_control *sc) | |
1779 | { | |
1780 | unsigned long pages_for_compaction; | |
1781 | unsigned long inactive_lru_pages; | |
1782 | ||
1783 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 1784 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
1785 | return false; |
1786 | ||
2876592f MG |
1787 | /* Consider stopping depending on scan and reclaim activity */ |
1788 | if (sc->gfp_mask & __GFP_REPEAT) { | |
1789 | /* | |
1790 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
1791 | * full LRU list has been scanned and we are still failing | |
1792 | * to reclaim pages. This full LRU scan is potentially | |
1793 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
1794 | */ | |
1795 | if (!nr_reclaimed && !nr_scanned) | |
1796 | return false; | |
1797 | } else { | |
1798 | /* | |
1799 | * For non-__GFP_REPEAT allocations which can presumably | |
1800 | * fail without consequence, stop if we failed to reclaim | |
1801 | * any pages from the last SWAP_CLUSTER_MAX number of | |
1802 | * pages that were scanned. This will return to the | |
1803 | * caller faster at the risk reclaim/compaction and | |
1804 | * the resulting allocation attempt fails | |
1805 | */ | |
1806 | if (!nr_reclaimed) | |
1807 | return false; | |
1808 | } | |
3e7d3449 MG |
1809 | |
1810 | /* | |
1811 | * If we have not reclaimed enough pages for compaction and the | |
1812 | * inactive lists are large enough, continue reclaiming | |
1813 | */ | |
1814 | pages_for_compaction = (2UL << sc->order); | |
f16015fb | 1815 | inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE); |
86cfd3a4 | 1816 | if (nr_swap_pages > 0) |
f16015fb | 1817 | inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON); |
3e7d3449 MG |
1818 | if (sc->nr_reclaimed < pages_for_compaction && |
1819 | inactive_lru_pages > pages_for_compaction) | |
1820 | return true; | |
1821 | ||
1822 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
f16015fb | 1823 | switch (compaction_suitable(mz->zone, sc->order)) { |
3e7d3449 MG |
1824 | case COMPACT_PARTIAL: |
1825 | case COMPACT_CONTINUE: | |
1826 | return false; | |
1827 | default: | |
1828 | return true; | |
1829 | } | |
1830 | } | |
1831 | ||
1da177e4 LT |
1832 | /* |
1833 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1834 | */ | |
9e3b2f8c | 1835 | static void shrink_mem_cgroup_zone(struct mem_cgroup_zone *mz, |
f16015fb | 1836 | struct scan_control *sc) |
1da177e4 | 1837 | { |
b69408e8 | 1838 | unsigned long nr[NR_LRU_LISTS]; |
8695949a | 1839 | unsigned long nr_to_scan; |
4111304d | 1840 | enum lru_list lru; |
f0fdc5e8 | 1841 | unsigned long nr_reclaimed, nr_scanned; |
22fba335 | 1842 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; |
3da367c3 | 1843 | struct blk_plug plug; |
e0f79b8f | 1844 | |
3e7d3449 MG |
1845 | restart: |
1846 | nr_reclaimed = 0; | |
f0fdc5e8 | 1847 | nr_scanned = sc->nr_scanned; |
9e3b2f8c | 1848 | get_scan_count(mz, sc, nr); |
1da177e4 | 1849 | |
3da367c3 | 1850 | blk_start_plug(&plug); |
556adecb RR |
1851 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || |
1852 | nr[LRU_INACTIVE_FILE]) { | |
4111304d HD |
1853 | for_each_evictable_lru(lru) { |
1854 | if (nr[lru]) { | |
ece74b2e | 1855 | nr_to_scan = min_t(unsigned long, |
4111304d HD |
1856 | nr[lru], SWAP_CLUSTER_MAX); |
1857 | nr[lru] -= nr_to_scan; | |
1da177e4 | 1858 | |
4111304d | 1859 | nr_reclaimed += shrink_list(lru, nr_to_scan, |
9e3b2f8c | 1860 | mz, sc); |
b69408e8 | 1861 | } |
1da177e4 | 1862 | } |
a79311c1 RR |
1863 | /* |
1864 | * On large memory systems, scan >> priority can become | |
1865 | * really large. This is fine for the starting priority; | |
1866 | * we want to put equal scanning pressure on each zone. | |
1867 | * However, if the VM has a harder time of freeing pages, | |
1868 | * with multiple processes reclaiming pages, the total | |
1869 | * freeing target can get unreasonably large. | |
1870 | */ | |
9e3b2f8c KK |
1871 | if (nr_reclaimed >= nr_to_reclaim && |
1872 | sc->priority < DEF_PRIORITY) | |
a79311c1 | 1873 | break; |
1da177e4 | 1874 | } |
3da367c3 | 1875 | blk_finish_plug(&plug); |
3e7d3449 | 1876 | sc->nr_reclaimed += nr_reclaimed; |
01dbe5c9 | 1877 | |
556adecb RR |
1878 | /* |
1879 | * Even if we did not try to evict anon pages at all, we want to | |
1880 | * rebalance the anon lru active/inactive ratio. | |
1881 | */ | |
f16015fb | 1882 | if (inactive_anon_is_low(mz)) |
3cb99451 | 1883 | shrink_active_list(SWAP_CLUSTER_MAX, mz, |
9e3b2f8c | 1884 | sc, LRU_ACTIVE_ANON); |
556adecb | 1885 | |
3e7d3449 | 1886 | /* reclaim/compaction might need reclaim to continue */ |
f16015fb | 1887 | if (should_continue_reclaim(mz, nr_reclaimed, |
9e3b2f8c | 1888 | sc->nr_scanned - nr_scanned, sc)) |
3e7d3449 MG |
1889 | goto restart; |
1890 | ||
232ea4d6 | 1891 | throttle_vm_writeout(sc->gfp_mask); |
1da177e4 LT |
1892 | } |
1893 | ||
9e3b2f8c | 1894 | static void shrink_zone(struct zone *zone, struct scan_control *sc) |
f16015fb | 1895 | { |
5660048c JW |
1896 | struct mem_cgroup *root = sc->target_mem_cgroup; |
1897 | struct mem_cgroup_reclaim_cookie reclaim = { | |
f16015fb | 1898 | .zone = zone, |
9e3b2f8c | 1899 | .priority = sc->priority, |
f16015fb | 1900 | }; |
5660048c JW |
1901 | struct mem_cgroup *memcg; |
1902 | ||
5660048c JW |
1903 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
1904 | do { | |
1905 | struct mem_cgroup_zone mz = { | |
1906 | .mem_cgroup = memcg, | |
1907 | .zone = zone, | |
1908 | }; | |
f16015fb | 1909 | |
9e3b2f8c | 1910 | shrink_mem_cgroup_zone(&mz, sc); |
5660048c JW |
1911 | /* |
1912 | * Limit reclaim has historically picked one memcg and | |
1913 | * scanned it with decreasing priority levels until | |
1914 | * nr_to_reclaim had been reclaimed. This priority | |
1915 | * cycle is thus over after a single memcg. | |
b95a2f2d JW |
1916 | * |
1917 | * Direct reclaim and kswapd, on the other hand, have | |
1918 | * to scan all memory cgroups to fulfill the overall | |
1919 | * scan target for the zone. | |
5660048c JW |
1920 | */ |
1921 | if (!global_reclaim(sc)) { | |
1922 | mem_cgroup_iter_break(root, memcg); | |
1923 | break; | |
1924 | } | |
1925 | memcg = mem_cgroup_iter(root, memcg, &reclaim); | |
1926 | } while (memcg); | |
f16015fb JW |
1927 | } |
1928 | ||
fe4b1b24 MG |
1929 | /* Returns true if compaction should go ahead for a high-order request */ |
1930 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
1931 | { | |
1932 | unsigned long balance_gap, watermark; | |
1933 | bool watermark_ok; | |
1934 | ||
1935 | /* Do not consider compaction for orders reclaim is meant to satisfy */ | |
1936 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
1937 | return false; | |
1938 | ||
1939 | /* | |
1940 | * Compaction takes time to run and there are potentially other | |
1941 | * callers using the pages just freed. Continue reclaiming until | |
1942 | * there is a buffer of free pages available to give compaction | |
1943 | * a reasonable chance of completing and allocating the page | |
1944 | */ | |
1945 | balance_gap = min(low_wmark_pages(zone), | |
1946 | (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
1947 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
1948 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order); | |
1949 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); | |
1950 | ||
1951 | /* | |
1952 | * If compaction is deferred, reclaim up to a point where | |
1953 | * compaction will have a chance of success when re-enabled | |
1954 | */ | |
aff62249 | 1955 | if (compaction_deferred(zone, sc->order)) |
fe4b1b24 MG |
1956 | return watermark_ok; |
1957 | ||
1958 | /* If compaction is not ready to start, keep reclaiming */ | |
1959 | if (!compaction_suitable(zone, sc->order)) | |
1960 | return false; | |
1961 | ||
1962 | return watermark_ok; | |
1963 | } | |
1964 | ||
1da177e4 LT |
1965 | /* |
1966 | * This is the direct reclaim path, for page-allocating processes. We only | |
1967 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
1968 | * request. | |
1969 | * | |
41858966 MG |
1970 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
1971 | * Because: | |
1da177e4 LT |
1972 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
1973 | * allocation or | |
41858966 MG |
1974 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
1975 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
1976 | * zone defense algorithm. | |
1da177e4 | 1977 | * |
1da177e4 LT |
1978 | * If a zone is deemed to be full of pinned pages then just give it a light |
1979 | * scan then give up on it. | |
e0c23279 MG |
1980 | * |
1981 | * This function returns true if a zone is being reclaimed for a costly | |
fe4b1b24 | 1982 | * high-order allocation and compaction is ready to begin. This indicates to |
0cee34fd MG |
1983 | * the caller that it should consider retrying the allocation instead of |
1984 | * further reclaim. | |
1da177e4 | 1985 | */ |
9e3b2f8c | 1986 | static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 1987 | { |
dd1a239f | 1988 | struct zoneref *z; |
54a6eb5c | 1989 | struct zone *zone; |
d149e3b2 YH |
1990 | unsigned long nr_soft_reclaimed; |
1991 | unsigned long nr_soft_scanned; | |
0cee34fd | 1992 | bool aborted_reclaim = false; |
1cfb419b | 1993 | |
cc715d99 MG |
1994 | /* |
1995 | * If the number of buffer_heads in the machine exceeds the maximum | |
1996 | * allowed level, force direct reclaim to scan the highmem zone as | |
1997 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
1998 | */ | |
1999 | if (buffer_heads_over_limit) | |
2000 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2001 | ||
d4debc66 MG |
2002 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
2003 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
f3fe6512 | 2004 | if (!populated_zone(zone)) |
1da177e4 | 2005 | continue; |
1cfb419b KH |
2006 | /* |
2007 | * Take care memory controller reclaiming has small influence | |
2008 | * to global LRU. | |
2009 | */ | |
89b5fae5 | 2010 | if (global_reclaim(sc)) { |
1cfb419b KH |
2011 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2012 | continue; | |
9e3b2f8c KK |
2013 | if (zone->all_unreclaimable && |
2014 | sc->priority != DEF_PRIORITY) | |
1cfb419b | 2015 | continue; /* Let kswapd poll it */ |
e0887c19 RR |
2016 | if (COMPACTION_BUILD) { |
2017 | /* | |
e0c23279 MG |
2018 | * If we already have plenty of memory free for |
2019 | * compaction in this zone, don't free any more. | |
2020 | * Even though compaction is invoked for any | |
2021 | * non-zero order, only frequent costly order | |
2022 | * reclamation is disruptive enough to become a | |
c7cfa37b CA |
2023 | * noticeable problem, like transparent huge |
2024 | * page allocations. | |
e0887c19 | 2025 | */ |
fe4b1b24 | 2026 | if (compaction_ready(zone, sc)) { |
0cee34fd | 2027 | aborted_reclaim = true; |
e0887c19 | 2028 | continue; |
e0c23279 | 2029 | } |
e0887c19 | 2030 | } |
ac34a1a3 KH |
2031 | /* |
2032 | * This steals pages from memory cgroups over softlimit | |
2033 | * and returns the number of reclaimed pages and | |
2034 | * scanned pages. This works for global memory pressure | |
2035 | * and balancing, not for a memcg's limit. | |
2036 | */ | |
2037 | nr_soft_scanned = 0; | |
2038 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2039 | sc->order, sc->gfp_mask, | |
2040 | &nr_soft_scanned); | |
2041 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2042 | sc->nr_scanned += nr_soft_scanned; | |
2043 | /* need some check for avoid more shrink_zone() */ | |
1cfb419b | 2044 | } |
408d8544 | 2045 | |
9e3b2f8c | 2046 | shrink_zone(zone, sc); |
1da177e4 | 2047 | } |
e0c23279 | 2048 | |
0cee34fd | 2049 | return aborted_reclaim; |
d1908362 MK |
2050 | } |
2051 | ||
2052 | static bool zone_reclaimable(struct zone *zone) | |
2053 | { | |
2054 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
2055 | } | |
2056 | ||
929bea7c | 2057 | /* All zones in zonelist are unreclaimable? */ |
d1908362 MK |
2058 | static bool all_unreclaimable(struct zonelist *zonelist, |
2059 | struct scan_control *sc) | |
2060 | { | |
2061 | struct zoneref *z; | |
2062 | struct zone *zone; | |
d1908362 MK |
2063 | |
2064 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2065 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2066 | if (!populated_zone(zone)) | |
2067 | continue; | |
2068 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2069 | continue; | |
929bea7c KM |
2070 | if (!zone->all_unreclaimable) |
2071 | return false; | |
d1908362 MK |
2072 | } |
2073 | ||
929bea7c | 2074 | return true; |
1da177e4 | 2075 | } |
4f98a2fe | 2076 | |
1da177e4 LT |
2077 | /* |
2078 | * This is the main entry point to direct page reclaim. | |
2079 | * | |
2080 | * If a full scan of the inactive list fails to free enough memory then we | |
2081 | * are "out of memory" and something needs to be killed. | |
2082 | * | |
2083 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2084 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2085 | * caller can't do much about. We kick the writeback threads and take explicit |
2086 | * naps in the hope that some of these pages can be written. But if the | |
2087 | * allocating task holds filesystem locks which prevent writeout this might not | |
2088 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2089 | * |
2090 | * returns: 0, if no pages reclaimed | |
2091 | * else, the number of pages reclaimed | |
1da177e4 | 2092 | */ |
dac1d27b | 2093 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
a09ed5e0 YH |
2094 | struct scan_control *sc, |
2095 | struct shrink_control *shrink) | |
1da177e4 | 2096 | { |
69e05944 | 2097 | unsigned long total_scanned = 0; |
1da177e4 | 2098 | struct reclaim_state *reclaim_state = current->reclaim_state; |
dd1a239f | 2099 | struct zoneref *z; |
54a6eb5c | 2100 | struct zone *zone; |
22fba335 | 2101 | unsigned long writeback_threshold; |
0cee34fd | 2102 | bool aborted_reclaim; |
1da177e4 | 2103 | |
873b4771 KK |
2104 | delayacct_freepages_start(); |
2105 | ||
89b5fae5 | 2106 | if (global_reclaim(sc)) |
1cfb419b | 2107 | count_vm_event(ALLOCSTALL); |
1da177e4 | 2108 | |
9e3b2f8c | 2109 | do { |
66e1707b | 2110 | sc->nr_scanned = 0; |
9e3b2f8c | 2111 | aborted_reclaim = shrink_zones(zonelist, sc); |
e0c23279 | 2112 | |
66e1707b BS |
2113 | /* |
2114 | * Don't shrink slabs when reclaiming memory from | |
2115 | * over limit cgroups | |
2116 | */ | |
89b5fae5 | 2117 | if (global_reclaim(sc)) { |
c6a8a8c5 | 2118 | unsigned long lru_pages = 0; |
d4debc66 MG |
2119 | for_each_zone_zonelist(zone, z, zonelist, |
2120 | gfp_zone(sc->gfp_mask)) { | |
c6a8a8c5 KM |
2121 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2122 | continue; | |
2123 | ||
2124 | lru_pages += zone_reclaimable_pages(zone); | |
2125 | } | |
2126 | ||
1495f230 | 2127 | shrink_slab(shrink, sc->nr_scanned, lru_pages); |
91a45470 | 2128 | if (reclaim_state) { |
a79311c1 | 2129 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
91a45470 KH |
2130 | reclaim_state->reclaimed_slab = 0; |
2131 | } | |
1da177e4 | 2132 | } |
66e1707b | 2133 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2134 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
1da177e4 | 2135 | goto out; |
1da177e4 LT |
2136 | |
2137 | /* | |
2138 | * Try to write back as many pages as we just scanned. This | |
2139 | * tends to cause slow streaming writers to write data to the | |
2140 | * disk smoothly, at the dirtying rate, which is nice. But | |
2141 | * that's undesirable in laptop mode, where we *want* lumpy | |
2142 | * writeout. So in laptop mode, write out the whole world. | |
2143 | */ | |
22fba335 KM |
2144 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2145 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2146 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2147 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2148 | sc->may_writepage = 1; |
1da177e4 LT |
2149 | } |
2150 | ||
2151 | /* Take a nap, wait for some writeback to complete */ | |
7b51755c | 2152 | if (!sc->hibernation_mode && sc->nr_scanned && |
9e3b2f8c | 2153 | sc->priority < DEF_PRIORITY - 2) { |
0e093d99 MG |
2154 | struct zone *preferred_zone; |
2155 | ||
2156 | first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), | |
f33261d7 DR |
2157 | &cpuset_current_mems_allowed, |
2158 | &preferred_zone); | |
0e093d99 MG |
2159 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); |
2160 | } | |
9e3b2f8c | 2161 | } while (--sc->priority >= 0); |
bb21c7ce | 2162 | |
1da177e4 | 2163 | out: |
873b4771 KK |
2164 | delayacct_freepages_end(); |
2165 | ||
bb21c7ce KM |
2166 | if (sc->nr_reclaimed) |
2167 | return sc->nr_reclaimed; | |
2168 | ||
929bea7c KM |
2169 | /* |
2170 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2171 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2172 | * check. | |
2173 | */ | |
2174 | if (oom_killer_disabled) | |
2175 | return 0; | |
2176 | ||
0cee34fd MG |
2177 | /* Aborted reclaim to try compaction? don't OOM, then */ |
2178 | if (aborted_reclaim) | |
7335084d MG |
2179 | return 1; |
2180 | ||
bb21c7ce | 2181 | /* top priority shrink_zones still had more to do? don't OOM, then */ |
89b5fae5 | 2182 | if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc)) |
bb21c7ce KM |
2183 | return 1; |
2184 | ||
2185 | return 0; | |
1da177e4 LT |
2186 | } |
2187 | ||
dac1d27b | 2188 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2189 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2190 | { |
33906bc5 | 2191 | unsigned long nr_reclaimed; |
66e1707b BS |
2192 | struct scan_control sc = { |
2193 | .gfp_mask = gfp_mask, | |
2194 | .may_writepage = !laptop_mode, | |
22fba335 | 2195 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
a6dc60f8 | 2196 | .may_unmap = 1, |
2e2e4259 | 2197 | .may_swap = 1, |
66e1707b | 2198 | .order = order, |
9e3b2f8c | 2199 | .priority = DEF_PRIORITY, |
f16015fb | 2200 | .target_mem_cgroup = NULL, |
327c0e96 | 2201 | .nodemask = nodemask, |
66e1707b | 2202 | }; |
a09ed5e0 YH |
2203 | struct shrink_control shrink = { |
2204 | .gfp_mask = sc.gfp_mask, | |
2205 | }; | |
66e1707b | 2206 | |
33906bc5 MG |
2207 | trace_mm_vmscan_direct_reclaim_begin(order, |
2208 | sc.may_writepage, | |
2209 | gfp_mask); | |
2210 | ||
a09ed5e0 | 2211 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
33906bc5 MG |
2212 | |
2213 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2214 | ||
2215 | return nr_reclaimed; | |
66e1707b BS |
2216 | } |
2217 | ||
00f0b825 | 2218 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
66e1707b | 2219 | |
72835c86 | 2220 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, |
4e416953 | 2221 | gfp_t gfp_mask, bool noswap, |
0ae5e89c YH |
2222 | struct zone *zone, |
2223 | unsigned long *nr_scanned) | |
4e416953 BS |
2224 | { |
2225 | struct scan_control sc = { | |
0ae5e89c | 2226 | .nr_scanned = 0, |
b8f5c566 | 2227 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
4e416953 BS |
2228 | .may_writepage = !laptop_mode, |
2229 | .may_unmap = 1, | |
2230 | .may_swap = !noswap, | |
4e416953 | 2231 | .order = 0, |
9e3b2f8c | 2232 | .priority = 0, |
72835c86 | 2233 | .target_mem_cgroup = memcg, |
4e416953 | 2234 | }; |
5660048c | 2235 | struct mem_cgroup_zone mz = { |
72835c86 | 2236 | .mem_cgroup = memcg, |
5660048c JW |
2237 | .zone = zone, |
2238 | }; | |
0ae5e89c | 2239 | |
4e416953 BS |
2240 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2241 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 2242 | |
9e3b2f8c | 2243 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e KM |
2244 | sc.may_writepage, |
2245 | sc.gfp_mask); | |
2246 | ||
4e416953 BS |
2247 | /* |
2248 | * NOTE: Although we can get the priority field, using it | |
2249 | * here is not a good idea, since it limits the pages we can scan. | |
2250 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2251 | * will pick up pages from other mem cgroup's as well. We hack | |
2252 | * the priority and make it zero. | |
2253 | */ | |
9e3b2f8c | 2254 | shrink_mem_cgroup_zone(&mz, &sc); |
bdce6d9e KM |
2255 | |
2256 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2257 | ||
0ae5e89c | 2258 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2259 | return sc.nr_reclaimed; |
2260 | } | |
2261 | ||
72835c86 | 2262 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
a7885eb8 | 2263 | gfp_t gfp_mask, |
185efc0f | 2264 | bool noswap) |
66e1707b | 2265 | { |
4e416953 | 2266 | struct zonelist *zonelist; |
bdce6d9e | 2267 | unsigned long nr_reclaimed; |
889976db | 2268 | int nid; |
66e1707b | 2269 | struct scan_control sc = { |
66e1707b | 2270 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2271 | .may_unmap = 1, |
2e2e4259 | 2272 | .may_swap = !noswap, |
22fba335 | 2273 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
66e1707b | 2274 | .order = 0, |
9e3b2f8c | 2275 | .priority = DEF_PRIORITY, |
72835c86 | 2276 | .target_mem_cgroup = memcg, |
327c0e96 | 2277 | .nodemask = NULL, /* we don't care the placement */ |
a09ed5e0 YH |
2278 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2279 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2280 | }; | |
2281 | struct shrink_control shrink = { | |
2282 | .gfp_mask = sc.gfp_mask, | |
66e1707b | 2283 | }; |
66e1707b | 2284 | |
889976db YH |
2285 | /* |
2286 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2287 | * take care of from where we get pages. So the node where we start the | |
2288 | * scan does not need to be the current node. | |
2289 | */ | |
72835c86 | 2290 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2291 | |
2292 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2293 | |
2294 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2295 | sc.may_writepage, | |
2296 | sc.gfp_mask); | |
2297 | ||
a09ed5e0 | 2298 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
bdce6d9e KM |
2299 | |
2300 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2301 | ||
2302 | return nr_reclaimed; | |
66e1707b BS |
2303 | } |
2304 | #endif | |
2305 | ||
9e3b2f8c | 2306 | static void age_active_anon(struct zone *zone, struct scan_control *sc) |
f16015fb | 2307 | { |
b95a2f2d | 2308 | struct mem_cgroup *memcg; |
f16015fb | 2309 | |
b95a2f2d JW |
2310 | if (!total_swap_pages) |
2311 | return; | |
2312 | ||
2313 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2314 | do { | |
2315 | struct mem_cgroup_zone mz = { | |
2316 | .mem_cgroup = memcg, | |
2317 | .zone = zone, | |
2318 | }; | |
2319 | ||
2320 | if (inactive_anon_is_low(&mz)) | |
2321 | shrink_active_list(SWAP_CLUSTER_MAX, &mz, | |
9e3b2f8c | 2322 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
2323 | |
2324 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2325 | } while (memcg); | |
f16015fb JW |
2326 | } |
2327 | ||
1741c877 MG |
2328 | /* |
2329 | * pgdat_balanced is used when checking if a node is balanced for high-order | |
2330 | * allocations. Only zones that meet watermarks and are in a zone allowed | |
2331 | * by the callers classzone_idx are added to balanced_pages. The total of | |
2332 | * balanced pages must be at least 25% of the zones allowed by classzone_idx | |
2333 | * for the node to be considered balanced. Forcing all zones to be balanced | |
2334 | * for high orders can cause excessive reclaim when there are imbalanced zones. | |
2335 | * The choice of 25% is due to | |
2336 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2337 | * reasonable sized machine | |
2338 | * o On all other machines, the top zone must be at least a reasonable | |
25985edc | 2339 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
1741c877 MG |
2340 | * would need to be at least 256M for it to be balance a whole node. |
2341 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2342 | * to balance a node on its own. These seemed like reasonable ratios. | |
2343 | */ | |
2344 | static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages, | |
2345 | int classzone_idx) | |
2346 | { | |
2347 | unsigned long present_pages = 0; | |
2348 | int i; | |
2349 | ||
2350 | for (i = 0; i <= classzone_idx; i++) | |
2351 | present_pages += pgdat->node_zones[i].present_pages; | |
2352 | ||
4746efde SL |
2353 | /* A special case here: if zone has no page, we think it's balanced */ |
2354 | return balanced_pages >= (present_pages >> 2); | |
1741c877 MG |
2355 | } |
2356 | ||
f50de2d3 | 2357 | /* is kswapd sleeping prematurely? */ |
dc83edd9 MG |
2358 | static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining, |
2359 | int classzone_idx) | |
f50de2d3 | 2360 | { |
bb3ab596 | 2361 | int i; |
1741c877 MG |
2362 | unsigned long balanced = 0; |
2363 | bool all_zones_ok = true; | |
f50de2d3 MG |
2364 | |
2365 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ | |
2366 | if (remaining) | |
dc83edd9 | 2367 | return true; |
f50de2d3 | 2368 | |
0abdee2b | 2369 | /* Check the watermark levels */ |
08951e54 | 2370 | for (i = 0; i <= classzone_idx; i++) { |
bb3ab596 KM |
2371 | struct zone *zone = pgdat->node_zones + i; |
2372 | ||
2373 | if (!populated_zone(zone)) | |
2374 | continue; | |
2375 | ||
355b09c4 MG |
2376 | /* |
2377 | * balance_pgdat() skips over all_unreclaimable after | |
2378 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2379 | * they must be considered balanced here as well if kswapd | |
2380 | * is to sleep | |
2381 | */ | |
2382 | if (zone->all_unreclaimable) { | |
2383 | balanced += zone->present_pages; | |
de3fab39 | 2384 | continue; |
355b09c4 | 2385 | } |
de3fab39 | 2386 | |
88f5acf8 | 2387 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone), |
da175d06 | 2388 | i, 0)) |
1741c877 MG |
2389 | all_zones_ok = false; |
2390 | else | |
2391 | balanced += zone->present_pages; | |
bb3ab596 | 2392 | } |
f50de2d3 | 2393 | |
1741c877 MG |
2394 | /* |
2395 | * For high-order requests, the balanced zones must contain at least | |
2396 | * 25% of the nodes pages for kswapd to sleep. For order-0, all zones | |
2397 | * must be balanced | |
2398 | */ | |
2399 | if (order) | |
afc7e326 | 2400 | return !pgdat_balanced(pgdat, balanced, classzone_idx); |
1741c877 MG |
2401 | else |
2402 | return !all_zones_ok; | |
f50de2d3 MG |
2403 | } |
2404 | ||
1da177e4 LT |
2405 | /* |
2406 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 2407 | * they are all at high_wmark_pages(zone). |
1da177e4 | 2408 | * |
0abdee2b | 2409 | * Returns the final order kswapd was reclaiming at |
1da177e4 LT |
2410 | * |
2411 | * There is special handling here for zones which are full of pinned pages. | |
2412 | * This can happen if the pages are all mlocked, or if they are all used by | |
2413 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
2414 | * What we do is to detect the case where all pages in the zone have been | |
2415 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
2416 | * dead and from now on, only perform a short scan. Basically we're polling | |
2417 | * the zone for when the problem goes away. | |
2418 | * | |
2419 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
2420 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
2421 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
2422 | * lower zones regardless of the number of free pages in the lower zones. This | |
2423 | * interoperates with the page allocator fallback scheme to ensure that aging | |
2424 | * of pages is balanced across the zones. | |
1da177e4 | 2425 | */ |
99504748 | 2426 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, |
dc83edd9 | 2427 | int *classzone_idx) |
1da177e4 | 2428 | { |
1da177e4 | 2429 | int all_zones_ok; |
1741c877 | 2430 | unsigned long balanced; |
1da177e4 | 2431 | int i; |
99504748 | 2432 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
69e05944 | 2433 | unsigned long total_scanned; |
1da177e4 | 2434 | struct reclaim_state *reclaim_state = current->reclaim_state; |
0ae5e89c YH |
2435 | unsigned long nr_soft_reclaimed; |
2436 | unsigned long nr_soft_scanned; | |
179e9639 AM |
2437 | struct scan_control sc = { |
2438 | .gfp_mask = GFP_KERNEL, | |
a6dc60f8 | 2439 | .may_unmap = 1, |
2e2e4259 | 2440 | .may_swap = 1, |
22fba335 KM |
2441 | /* |
2442 | * kswapd doesn't want to be bailed out while reclaim. because | |
2443 | * we want to put equal scanning pressure on each zone. | |
2444 | */ | |
2445 | .nr_to_reclaim = ULONG_MAX, | |
5ad333eb | 2446 | .order = order, |
f16015fb | 2447 | .target_mem_cgroup = NULL, |
179e9639 | 2448 | }; |
a09ed5e0 YH |
2449 | struct shrink_control shrink = { |
2450 | .gfp_mask = sc.gfp_mask, | |
2451 | }; | |
1da177e4 LT |
2452 | loop_again: |
2453 | total_scanned = 0; | |
9e3b2f8c | 2454 | sc.priority = DEF_PRIORITY; |
a79311c1 | 2455 | sc.nr_reclaimed = 0; |
c0bbbc73 | 2456 | sc.may_writepage = !laptop_mode; |
f8891e5e | 2457 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 2458 | |
9e3b2f8c | 2459 | do { |
1da177e4 | 2460 | unsigned long lru_pages = 0; |
bb3ab596 | 2461 | int has_under_min_watermark_zone = 0; |
1da177e4 LT |
2462 | |
2463 | all_zones_ok = 1; | |
1741c877 | 2464 | balanced = 0; |
1da177e4 | 2465 | |
d6277db4 RW |
2466 | /* |
2467 | * Scan in the highmem->dma direction for the highest | |
2468 | * zone which needs scanning | |
2469 | */ | |
2470 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
2471 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 2472 | |
d6277db4 RW |
2473 | if (!populated_zone(zone)) |
2474 | continue; | |
1da177e4 | 2475 | |
9e3b2f8c KK |
2476 | if (zone->all_unreclaimable && |
2477 | sc.priority != DEF_PRIORITY) | |
d6277db4 | 2478 | continue; |
1da177e4 | 2479 | |
556adecb RR |
2480 | /* |
2481 | * Do some background aging of the anon list, to give | |
2482 | * pages a chance to be referenced before reclaiming. | |
2483 | */ | |
9e3b2f8c | 2484 | age_active_anon(zone, &sc); |
556adecb | 2485 | |
cc715d99 MG |
2486 | /* |
2487 | * If the number of buffer_heads in the machine | |
2488 | * exceeds the maximum allowed level and this node | |
2489 | * has a highmem zone, force kswapd to reclaim from | |
2490 | * it to relieve lowmem pressure. | |
2491 | */ | |
2492 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
2493 | end_zone = i; | |
2494 | break; | |
2495 | } | |
2496 | ||
88f5acf8 | 2497 | if (!zone_watermark_ok_safe(zone, order, |
41858966 | 2498 | high_wmark_pages(zone), 0, 0)) { |
d6277db4 | 2499 | end_zone = i; |
e1dbeda6 | 2500 | break; |
439423f6 SL |
2501 | } else { |
2502 | /* If balanced, clear the congested flag */ | |
2503 | zone_clear_flag(zone, ZONE_CONGESTED); | |
1da177e4 | 2504 | } |
1da177e4 | 2505 | } |
e1dbeda6 AM |
2506 | if (i < 0) |
2507 | goto out; | |
2508 | ||
1da177e4 LT |
2509 | for (i = 0; i <= end_zone; i++) { |
2510 | struct zone *zone = pgdat->node_zones + i; | |
2511 | ||
adea02a1 | 2512 | lru_pages += zone_reclaimable_pages(zone); |
1da177e4 LT |
2513 | } |
2514 | ||
2515 | /* | |
2516 | * Now scan the zone in the dma->highmem direction, stopping | |
2517 | * at the last zone which needs scanning. | |
2518 | * | |
2519 | * We do this because the page allocator works in the opposite | |
2520 | * direction. This prevents the page allocator from allocating | |
2521 | * pages behind kswapd's direction of progress, which would | |
2522 | * cause too much scanning of the lower zones. | |
2523 | */ | |
2524 | for (i = 0; i <= end_zone; i++) { | |
2525 | struct zone *zone = pgdat->node_zones + i; | |
fe2c2a10 | 2526 | int nr_slab, testorder; |
8afdcece | 2527 | unsigned long balance_gap; |
1da177e4 | 2528 | |
f3fe6512 | 2529 | if (!populated_zone(zone)) |
1da177e4 LT |
2530 | continue; |
2531 | ||
9e3b2f8c KK |
2532 | if (zone->all_unreclaimable && |
2533 | sc.priority != DEF_PRIORITY) | |
1da177e4 LT |
2534 | continue; |
2535 | ||
1da177e4 | 2536 | sc.nr_scanned = 0; |
4e416953 | 2537 | |
0ae5e89c | 2538 | nr_soft_scanned = 0; |
4e416953 BS |
2539 | /* |
2540 | * Call soft limit reclaim before calling shrink_zone. | |
4e416953 | 2541 | */ |
0ae5e89c YH |
2542 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, |
2543 | order, sc.gfp_mask, | |
2544 | &nr_soft_scanned); | |
2545 | sc.nr_reclaimed += nr_soft_reclaimed; | |
2546 | total_scanned += nr_soft_scanned; | |
00918b6a | 2547 | |
32a4330d | 2548 | /* |
8afdcece MG |
2549 | * We put equal pressure on every zone, unless |
2550 | * one zone has way too many pages free | |
2551 | * already. The "too many pages" is defined | |
2552 | * as the high wmark plus a "gap" where the | |
2553 | * gap is either the low watermark or 1% | |
2554 | * of the zone, whichever is smaller. | |
32a4330d | 2555 | */ |
8afdcece MG |
2556 | balance_gap = min(low_wmark_pages(zone), |
2557 | (zone->present_pages + | |
2558 | KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2559 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
fe2c2a10 RR |
2560 | /* |
2561 | * Kswapd reclaims only single pages with compaction | |
2562 | * enabled. Trying too hard to reclaim until contiguous | |
2563 | * free pages have become available can hurt performance | |
2564 | * by evicting too much useful data from memory. | |
2565 | * Do not reclaim more than needed for compaction. | |
2566 | */ | |
2567 | testorder = order; | |
2568 | if (COMPACTION_BUILD && order && | |
2569 | compaction_suitable(zone, order) != | |
2570 | COMPACT_SKIPPED) | |
2571 | testorder = 0; | |
2572 | ||
cc715d99 | 2573 | if ((buffer_heads_over_limit && is_highmem_idx(i)) || |
643ac9fc | 2574 | !zone_watermark_ok_safe(zone, testorder, |
8afdcece | 2575 | high_wmark_pages(zone) + balance_gap, |
d7868dae | 2576 | end_zone, 0)) { |
9e3b2f8c | 2577 | shrink_zone(zone, &sc); |
5a03b051 | 2578 | |
d7868dae MG |
2579 | reclaim_state->reclaimed_slab = 0; |
2580 | nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages); | |
2581 | sc.nr_reclaimed += reclaim_state->reclaimed_slab; | |
2582 | total_scanned += sc.nr_scanned; | |
2583 | ||
2584 | if (nr_slab == 0 && !zone_reclaimable(zone)) | |
2585 | zone->all_unreclaimable = 1; | |
2586 | } | |
2587 | ||
1da177e4 LT |
2588 | /* |
2589 | * If we've done a decent amount of scanning and | |
2590 | * the reclaim ratio is low, start doing writepage | |
2591 | * even in laptop mode | |
2592 | */ | |
2593 | if (total_scanned > SWAP_CLUSTER_MAX * 2 && | |
a79311c1 | 2594 | total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) |
1da177e4 | 2595 | sc.may_writepage = 1; |
bb3ab596 | 2596 | |
215ddd66 MG |
2597 | if (zone->all_unreclaimable) { |
2598 | if (end_zone && end_zone == i) | |
2599 | end_zone--; | |
d7868dae | 2600 | continue; |
215ddd66 | 2601 | } |
d7868dae | 2602 | |
fe2c2a10 | 2603 | if (!zone_watermark_ok_safe(zone, testorder, |
45973d74 MK |
2604 | high_wmark_pages(zone), end_zone, 0)) { |
2605 | all_zones_ok = 0; | |
2606 | /* | |
2607 | * We are still under min water mark. This | |
2608 | * means that we have a GFP_ATOMIC allocation | |
2609 | * failure risk. Hurry up! | |
2610 | */ | |
88f5acf8 | 2611 | if (!zone_watermark_ok_safe(zone, order, |
45973d74 MK |
2612 | min_wmark_pages(zone), end_zone, 0)) |
2613 | has_under_min_watermark_zone = 1; | |
0e093d99 MG |
2614 | } else { |
2615 | /* | |
2616 | * If a zone reaches its high watermark, | |
2617 | * consider it to be no longer congested. It's | |
2618 | * possible there are dirty pages backed by | |
2619 | * congested BDIs but as pressure is relieved, | |
2620 | * spectulatively avoid congestion waits | |
2621 | */ | |
2622 | zone_clear_flag(zone, ZONE_CONGESTED); | |
dc83edd9 | 2623 | if (i <= *classzone_idx) |
1741c877 | 2624 | balanced += zone->present_pages; |
45973d74 | 2625 | } |
bb3ab596 | 2626 | |
1da177e4 | 2627 | } |
dc83edd9 | 2628 | if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx))) |
1da177e4 LT |
2629 | break; /* kswapd: all done */ |
2630 | /* | |
2631 | * OK, kswapd is getting into trouble. Take a nap, then take | |
2632 | * another pass across the zones. | |
2633 | */ | |
9e3b2f8c | 2634 | if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) { |
bb3ab596 KM |
2635 | if (has_under_min_watermark_zone) |
2636 | count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT); | |
2637 | else | |
2638 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2639 | } | |
1da177e4 LT |
2640 | |
2641 | /* | |
2642 | * We do this so kswapd doesn't build up large priorities for | |
2643 | * example when it is freeing in parallel with allocators. It | |
2644 | * matches the direct reclaim path behaviour in terms of impact | |
2645 | * on zone->*_priority. | |
2646 | */ | |
a79311c1 | 2647 | if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) |
1da177e4 | 2648 | break; |
9e3b2f8c | 2649 | } while (--sc.priority >= 0); |
1da177e4 | 2650 | out: |
99504748 MG |
2651 | |
2652 | /* | |
2653 | * order-0: All zones must meet high watermark for a balanced node | |
1741c877 MG |
2654 | * high-order: Balanced zones must make up at least 25% of the node |
2655 | * for the node to be balanced | |
99504748 | 2656 | */ |
dc83edd9 | 2657 | if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) { |
1da177e4 | 2658 | cond_resched(); |
8357376d RW |
2659 | |
2660 | try_to_freeze(); | |
2661 | ||
73ce02e9 KM |
2662 | /* |
2663 | * Fragmentation may mean that the system cannot be | |
2664 | * rebalanced for high-order allocations in all zones. | |
2665 | * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, | |
2666 | * it means the zones have been fully scanned and are still | |
2667 | * not balanced. For high-order allocations, there is | |
2668 | * little point trying all over again as kswapd may | |
2669 | * infinite loop. | |
2670 | * | |
2671 | * Instead, recheck all watermarks at order-0 as they | |
2672 | * are the most important. If watermarks are ok, kswapd will go | |
2673 | * back to sleep. High-order users can still perform direct | |
2674 | * reclaim if they wish. | |
2675 | */ | |
2676 | if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) | |
2677 | order = sc.order = 0; | |
2678 | ||
1da177e4 LT |
2679 | goto loop_again; |
2680 | } | |
2681 | ||
99504748 MG |
2682 | /* |
2683 | * If kswapd was reclaiming at a higher order, it has the option of | |
2684 | * sleeping without all zones being balanced. Before it does, it must | |
2685 | * ensure that the watermarks for order-0 on *all* zones are met and | |
2686 | * that the congestion flags are cleared. The congestion flag must | |
2687 | * be cleared as kswapd is the only mechanism that clears the flag | |
2688 | * and it is potentially going to sleep here. | |
2689 | */ | |
2690 | if (order) { | |
7be62de9 RR |
2691 | int zones_need_compaction = 1; |
2692 | ||
99504748 MG |
2693 | for (i = 0; i <= end_zone; i++) { |
2694 | struct zone *zone = pgdat->node_zones + i; | |
2695 | ||
2696 | if (!populated_zone(zone)) | |
2697 | continue; | |
2698 | ||
9e3b2f8c KK |
2699 | if (zone->all_unreclaimable && |
2700 | sc.priority != DEF_PRIORITY) | |
99504748 MG |
2701 | continue; |
2702 | ||
fe2c2a10 | 2703 | /* Would compaction fail due to lack of free memory? */ |
496b919b RR |
2704 | if (COMPACTION_BUILD && |
2705 | compaction_suitable(zone, order) == COMPACT_SKIPPED) | |
fe2c2a10 RR |
2706 | goto loop_again; |
2707 | ||
99504748 MG |
2708 | /* Confirm the zone is balanced for order-0 */ |
2709 | if (!zone_watermark_ok(zone, 0, | |
2710 | high_wmark_pages(zone), 0, 0)) { | |
2711 | order = sc.order = 0; | |
2712 | goto loop_again; | |
2713 | } | |
2714 | ||
7be62de9 RR |
2715 | /* Check if the memory needs to be defragmented. */ |
2716 | if (zone_watermark_ok(zone, order, | |
2717 | low_wmark_pages(zone), *classzone_idx, 0)) | |
2718 | zones_need_compaction = 0; | |
2719 | ||
99504748 MG |
2720 | /* If balanced, clear the congested flag */ |
2721 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2722 | } | |
7be62de9 RR |
2723 | |
2724 | if (zones_need_compaction) | |
2725 | compact_pgdat(pgdat, order); | |
99504748 MG |
2726 | } |
2727 | ||
0abdee2b MG |
2728 | /* |
2729 | * Return the order we were reclaiming at so sleeping_prematurely() | |
2730 | * makes a decision on the order we were last reclaiming at. However, | |
2731 | * if another caller entered the allocator slow path while kswapd | |
2732 | * was awake, order will remain at the higher level | |
2733 | */ | |
dc83edd9 | 2734 | *classzone_idx = end_zone; |
0abdee2b | 2735 | return order; |
1da177e4 LT |
2736 | } |
2737 | ||
dc83edd9 | 2738 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f0bc0a60 KM |
2739 | { |
2740 | long remaining = 0; | |
2741 | DEFINE_WAIT(wait); | |
2742 | ||
2743 | if (freezing(current) || kthread_should_stop()) | |
2744 | return; | |
2745 | ||
2746 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2747 | ||
2748 | /* Try to sleep for a short interval */ | |
dc83edd9 | 2749 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
2750 | remaining = schedule_timeout(HZ/10); |
2751 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2752 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2753 | } | |
2754 | ||
2755 | /* | |
2756 | * After a short sleep, check if it was a premature sleep. If not, then | |
2757 | * go fully to sleep until explicitly woken up. | |
2758 | */ | |
dc83edd9 | 2759 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
2760 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
2761 | ||
2762 | /* | |
2763 | * vmstat counters are not perfectly accurate and the estimated | |
2764 | * value for counters such as NR_FREE_PAGES can deviate from the | |
2765 | * true value by nr_online_cpus * threshold. To avoid the zone | |
2766 | * watermarks being breached while under pressure, we reduce the | |
2767 | * per-cpu vmstat threshold while kswapd is awake and restore | |
2768 | * them before going back to sleep. | |
2769 | */ | |
2770 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
2771 | schedule(); | |
2772 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); | |
2773 | } else { | |
2774 | if (remaining) | |
2775 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
2776 | else | |
2777 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
2778 | } | |
2779 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2780 | } | |
2781 | ||
1da177e4 LT |
2782 | /* |
2783 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 2784 | * from the init process. |
1da177e4 LT |
2785 | * |
2786 | * This basically trickles out pages so that we have _some_ | |
2787 | * free memory available even if there is no other activity | |
2788 | * that frees anything up. This is needed for things like routing | |
2789 | * etc, where we otherwise might have all activity going on in | |
2790 | * asynchronous contexts that cannot page things out. | |
2791 | * | |
2792 | * If there are applications that are active memory-allocators | |
2793 | * (most normal use), this basically shouldn't matter. | |
2794 | */ | |
2795 | static int kswapd(void *p) | |
2796 | { | |
215ddd66 | 2797 | unsigned long order, new_order; |
d2ebd0f6 | 2798 | unsigned balanced_order; |
215ddd66 | 2799 | int classzone_idx, new_classzone_idx; |
d2ebd0f6 | 2800 | int balanced_classzone_idx; |
1da177e4 LT |
2801 | pg_data_t *pgdat = (pg_data_t*)p; |
2802 | struct task_struct *tsk = current; | |
f0bc0a60 | 2803 | |
1da177e4 LT |
2804 | struct reclaim_state reclaim_state = { |
2805 | .reclaimed_slab = 0, | |
2806 | }; | |
a70f7302 | 2807 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 2808 | |
cf40bd16 NP |
2809 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
2810 | ||
174596a0 | 2811 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 2812 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
2813 | current->reclaim_state = &reclaim_state; |
2814 | ||
2815 | /* | |
2816 | * Tell the memory management that we're a "memory allocator", | |
2817 | * and that if we need more memory we should get access to it | |
2818 | * regardless (see "__alloc_pages()"). "kswapd" should | |
2819 | * never get caught in the normal page freeing logic. | |
2820 | * | |
2821 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
2822 | * you need a small amount of memory in order to be able to | |
2823 | * page out something else, and this flag essentially protects | |
2824 | * us from recursively trying to free more memory as we're | |
2825 | * trying to free the first piece of memory in the first place). | |
2826 | */ | |
930d9152 | 2827 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 2828 | set_freezable(); |
1da177e4 | 2829 | |
215ddd66 | 2830 | order = new_order = 0; |
d2ebd0f6 | 2831 | balanced_order = 0; |
215ddd66 | 2832 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; |
d2ebd0f6 | 2833 | balanced_classzone_idx = classzone_idx; |
1da177e4 | 2834 | for ( ; ; ) { |
8fe23e05 | 2835 | int ret; |
3e1d1d28 | 2836 | |
215ddd66 MG |
2837 | /* |
2838 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
2839 | * new request of a similar or harder type will succeed soon | |
2840 | * so consider going to sleep on the basis we reclaimed at | |
2841 | */ | |
d2ebd0f6 AS |
2842 | if (balanced_classzone_idx >= new_classzone_idx && |
2843 | balanced_order == new_order) { | |
215ddd66 MG |
2844 | new_order = pgdat->kswapd_max_order; |
2845 | new_classzone_idx = pgdat->classzone_idx; | |
2846 | pgdat->kswapd_max_order = 0; | |
2847 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
2848 | } | |
2849 | ||
99504748 | 2850 | if (order < new_order || classzone_idx > new_classzone_idx) { |
1da177e4 LT |
2851 | /* |
2852 | * Don't sleep if someone wants a larger 'order' | |
99504748 | 2853 | * allocation or has tigher zone constraints |
1da177e4 LT |
2854 | */ |
2855 | order = new_order; | |
99504748 | 2856 | classzone_idx = new_classzone_idx; |
1da177e4 | 2857 | } else { |
d2ebd0f6 AS |
2858 | kswapd_try_to_sleep(pgdat, balanced_order, |
2859 | balanced_classzone_idx); | |
1da177e4 | 2860 | order = pgdat->kswapd_max_order; |
99504748 | 2861 | classzone_idx = pgdat->classzone_idx; |
f0dfcde0 AS |
2862 | new_order = order; |
2863 | new_classzone_idx = classzone_idx; | |
4d40502e | 2864 | pgdat->kswapd_max_order = 0; |
215ddd66 | 2865 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
1da177e4 | 2866 | } |
1da177e4 | 2867 | |
8fe23e05 DR |
2868 | ret = try_to_freeze(); |
2869 | if (kthread_should_stop()) | |
2870 | break; | |
2871 | ||
2872 | /* | |
2873 | * We can speed up thawing tasks if we don't call balance_pgdat | |
2874 | * after returning from the refrigerator | |
2875 | */ | |
33906bc5 MG |
2876 | if (!ret) { |
2877 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
d2ebd0f6 AS |
2878 | balanced_classzone_idx = classzone_idx; |
2879 | balanced_order = balance_pgdat(pgdat, order, | |
2880 | &balanced_classzone_idx); | |
33906bc5 | 2881 | } |
1da177e4 LT |
2882 | } |
2883 | return 0; | |
2884 | } | |
2885 | ||
2886 | /* | |
2887 | * A zone is low on free memory, so wake its kswapd task to service it. | |
2888 | */ | |
99504748 | 2889 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
2890 | { |
2891 | pg_data_t *pgdat; | |
2892 | ||
f3fe6512 | 2893 | if (!populated_zone(zone)) |
1da177e4 LT |
2894 | return; |
2895 | ||
88f5acf8 | 2896 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1da177e4 | 2897 | return; |
88f5acf8 | 2898 | pgdat = zone->zone_pgdat; |
99504748 | 2899 | if (pgdat->kswapd_max_order < order) { |
1da177e4 | 2900 | pgdat->kswapd_max_order = order; |
99504748 MG |
2901 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
2902 | } | |
8d0986e2 | 2903 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 2904 | return; |
88f5acf8 MG |
2905 | if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) |
2906 | return; | |
2907 | ||
2908 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 2909 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
2910 | } |
2911 | ||
adea02a1 WF |
2912 | /* |
2913 | * The reclaimable count would be mostly accurate. | |
2914 | * The less reclaimable pages may be | |
2915 | * - mlocked pages, which will be moved to unevictable list when encountered | |
2916 | * - mapped pages, which may require several travels to be reclaimed | |
2917 | * - dirty pages, which is not "instantly" reclaimable | |
2918 | */ | |
2919 | unsigned long global_reclaimable_pages(void) | |
4f98a2fe | 2920 | { |
adea02a1 WF |
2921 | int nr; |
2922 | ||
2923 | nr = global_page_state(NR_ACTIVE_FILE) + | |
2924 | global_page_state(NR_INACTIVE_FILE); | |
2925 | ||
2926 | if (nr_swap_pages > 0) | |
2927 | nr += global_page_state(NR_ACTIVE_ANON) + | |
2928 | global_page_state(NR_INACTIVE_ANON); | |
2929 | ||
2930 | return nr; | |
2931 | } | |
2932 | ||
2933 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
2934 | { | |
2935 | int nr; | |
2936 | ||
2937 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
2938 | zone_page_state(zone, NR_INACTIVE_FILE); | |
2939 | ||
2940 | if (nr_swap_pages > 0) | |
2941 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
2942 | zone_page_state(zone, NR_INACTIVE_ANON); | |
2943 | ||
2944 | return nr; | |
4f98a2fe RR |
2945 | } |
2946 | ||
c6f37f12 | 2947 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 2948 | /* |
7b51755c | 2949 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
2950 | * freed pages. |
2951 | * | |
2952 | * Rather than trying to age LRUs the aim is to preserve the overall | |
2953 | * LRU order by reclaiming preferentially | |
2954 | * inactive > active > active referenced > active mapped | |
1da177e4 | 2955 | */ |
7b51755c | 2956 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 2957 | { |
d6277db4 | 2958 | struct reclaim_state reclaim_state; |
d6277db4 | 2959 | struct scan_control sc = { |
7b51755c KM |
2960 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
2961 | .may_swap = 1, | |
2962 | .may_unmap = 1, | |
d6277db4 | 2963 | .may_writepage = 1, |
7b51755c KM |
2964 | .nr_to_reclaim = nr_to_reclaim, |
2965 | .hibernation_mode = 1, | |
7b51755c | 2966 | .order = 0, |
9e3b2f8c | 2967 | .priority = DEF_PRIORITY, |
1da177e4 | 2968 | }; |
a09ed5e0 YH |
2969 | struct shrink_control shrink = { |
2970 | .gfp_mask = sc.gfp_mask, | |
2971 | }; | |
2972 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
7b51755c KM |
2973 | struct task_struct *p = current; |
2974 | unsigned long nr_reclaimed; | |
1da177e4 | 2975 | |
7b51755c KM |
2976 | p->flags |= PF_MEMALLOC; |
2977 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
2978 | reclaim_state.reclaimed_slab = 0; | |
2979 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 2980 | |
a09ed5e0 | 2981 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
d979677c | 2982 | |
7b51755c KM |
2983 | p->reclaim_state = NULL; |
2984 | lockdep_clear_current_reclaim_state(); | |
2985 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 2986 | |
7b51755c | 2987 | return nr_reclaimed; |
1da177e4 | 2988 | } |
c6f37f12 | 2989 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 2990 | |
1da177e4 LT |
2991 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
2992 | not required for correctness. So if the last cpu in a node goes | |
2993 | away, we get changed to run anywhere: as the first one comes back, | |
2994 | restore their cpu bindings. */ | |
9c7b216d | 2995 | static int __devinit cpu_callback(struct notifier_block *nfb, |
69e05944 | 2996 | unsigned long action, void *hcpu) |
1da177e4 | 2997 | { |
58c0a4a7 | 2998 | int nid; |
1da177e4 | 2999 | |
8bb78442 | 3000 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
58c0a4a7 | 3001 | for_each_node_state(nid, N_HIGH_MEMORY) { |
c5f59f08 | 3002 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3003 | const struct cpumask *mask; |
3004 | ||
3005 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3006 | |
3e597945 | 3007 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3008 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3009 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3010 | } |
3011 | } | |
3012 | return NOTIFY_OK; | |
3013 | } | |
1da177e4 | 3014 | |
3218ae14 YG |
3015 | /* |
3016 | * This kswapd start function will be called by init and node-hot-add. | |
3017 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3018 | */ | |
3019 | int kswapd_run(int nid) | |
3020 | { | |
3021 | pg_data_t *pgdat = NODE_DATA(nid); | |
3022 | int ret = 0; | |
3023 | ||
3024 | if (pgdat->kswapd) | |
3025 | return 0; | |
3026 | ||
3027 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3028 | if (IS_ERR(pgdat->kswapd)) { | |
3029 | /* failure at boot is fatal */ | |
3030 | BUG_ON(system_state == SYSTEM_BOOTING); | |
3031 | printk("Failed to start kswapd on node %d\n",nid); | |
3032 | ret = -1; | |
3033 | } | |
3034 | return ret; | |
3035 | } | |
3036 | ||
8fe23e05 DR |
3037 | /* |
3038 | * Called by memory hotplug when all memory in a node is offlined. | |
3039 | */ | |
3040 | void kswapd_stop(int nid) | |
3041 | { | |
3042 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3043 | ||
3044 | if (kswapd) | |
3045 | kthread_stop(kswapd); | |
3046 | } | |
3047 | ||
1da177e4 LT |
3048 | static int __init kswapd_init(void) |
3049 | { | |
3218ae14 | 3050 | int nid; |
69e05944 | 3051 | |
1da177e4 | 3052 | swap_setup(); |
9422ffba | 3053 | for_each_node_state(nid, N_HIGH_MEMORY) |
3218ae14 | 3054 | kswapd_run(nid); |
1da177e4 LT |
3055 | hotcpu_notifier(cpu_callback, 0); |
3056 | return 0; | |
3057 | } | |
3058 | ||
3059 | module_init(kswapd_init) | |
9eeff239 CL |
3060 | |
3061 | #ifdef CONFIG_NUMA | |
3062 | /* | |
3063 | * Zone reclaim mode | |
3064 | * | |
3065 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3066 | * the watermarks. | |
9eeff239 CL |
3067 | */ |
3068 | int zone_reclaim_mode __read_mostly; | |
3069 | ||
1b2ffb78 | 3070 | #define RECLAIM_OFF 0 |
7d03431c | 3071 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 CL |
3072 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
3073 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3074 | ||
a92f7126 CL |
3075 | /* |
3076 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3077 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3078 | * a zone. | |
3079 | */ | |
3080 | #define ZONE_RECLAIM_PRIORITY 4 | |
3081 | ||
9614634f CL |
3082 | /* |
3083 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3084 | * occur. | |
3085 | */ | |
3086 | int sysctl_min_unmapped_ratio = 1; | |
3087 | ||
0ff38490 CL |
3088 | /* |
3089 | * If the number of slab pages in a zone grows beyond this percentage then | |
3090 | * slab reclaim needs to occur. | |
3091 | */ | |
3092 | int sysctl_min_slab_ratio = 5; | |
3093 | ||
90afa5de MG |
3094 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
3095 | { | |
3096 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3097 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3098 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3099 | ||
3100 | /* | |
3101 | * It's possible for there to be more file mapped pages than | |
3102 | * accounted for by the pages on the file LRU lists because | |
3103 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3104 | */ | |
3105 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3106 | } | |
3107 | ||
3108 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3109 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3110 | { | |
3111 | long nr_pagecache_reclaimable; | |
3112 | long delta = 0; | |
3113 | ||
3114 | /* | |
3115 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3116 | * potentially reclaimable. Otherwise, we have to worry about | |
3117 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3118 | * a better estimate | |
3119 | */ | |
3120 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3121 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3122 | else | |
3123 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3124 | ||
3125 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3126 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3127 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3128 | ||
3129 | /* Watch for any possible underflows due to delta */ | |
3130 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3131 | delta = nr_pagecache_reclaimable; | |
3132 | ||
3133 | return nr_pagecache_reclaimable - delta; | |
3134 | } | |
3135 | ||
9eeff239 CL |
3136 | /* |
3137 | * Try to free up some pages from this zone through reclaim. | |
3138 | */ | |
179e9639 | 3139 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3140 | { |
7fb2d46d | 3141 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3142 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3143 | struct task_struct *p = current; |
3144 | struct reclaim_state reclaim_state; | |
179e9639 AM |
3145 | struct scan_control sc = { |
3146 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
a6dc60f8 | 3147 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), |
2e2e4259 | 3148 | .may_swap = 1, |
22fba335 KM |
3149 | .nr_to_reclaim = max_t(unsigned long, nr_pages, |
3150 | SWAP_CLUSTER_MAX), | |
179e9639 | 3151 | .gfp_mask = gfp_mask, |
bd2f6199 | 3152 | .order = order, |
9e3b2f8c | 3153 | .priority = ZONE_RECLAIM_PRIORITY, |
179e9639 | 3154 | }; |
a09ed5e0 YH |
3155 | struct shrink_control shrink = { |
3156 | .gfp_mask = sc.gfp_mask, | |
3157 | }; | |
15748048 | 3158 | unsigned long nr_slab_pages0, nr_slab_pages1; |
9eeff239 | 3159 | |
9eeff239 | 3160 | cond_resched(); |
d4f7796e CL |
3161 | /* |
3162 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3163 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3164 | * and RECLAIM_SWAP. | |
3165 | */ | |
3166 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3167 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3168 | reclaim_state.reclaimed_slab = 0; |
3169 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3170 | |
90afa5de | 3171 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
3172 | /* |
3173 | * Free memory by calling shrink zone with increasing | |
3174 | * priorities until we have enough memory freed. | |
3175 | */ | |
0ff38490 | 3176 | do { |
9e3b2f8c KK |
3177 | shrink_zone(zone, &sc); |
3178 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); | |
0ff38490 | 3179 | } |
c84db23c | 3180 | |
15748048 KM |
3181 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3182 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
2a16e3f4 | 3183 | /* |
7fb2d46d | 3184 | * shrink_slab() does not currently allow us to determine how |
0ff38490 CL |
3185 | * many pages were freed in this zone. So we take the current |
3186 | * number of slab pages and shake the slab until it is reduced | |
3187 | * by the same nr_pages that we used for reclaiming unmapped | |
3188 | * pages. | |
2a16e3f4 | 3189 | * |
0ff38490 CL |
3190 | * Note that shrink_slab will free memory on all zones and may |
3191 | * take a long time. | |
2a16e3f4 | 3192 | */ |
4dc4b3d9 KM |
3193 | for (;;) { |
3194 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3195 | ||
3196 | /* No reclaimable slab or very low memory pressure */ | |
1495f230 | 3197 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) |
4dc4b3d9 KM |
3198 | break; |
3199 | ||
3200 | /* Freed enough memory */ | |
3201 | nr_slab_pages1 = zone_page_state(zone, | |
3202 | NR_SLAB_RECLAIMABLE); | |
3203 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3204 | break; | |
3205 | } | |
83e33a47 CL |
3206 | |
3207 | /* | |
3208 | * Update nr_reclaimed by the number of slab pages we | |
3209 | * reclaimed from this zone. | |
3210 | */ | |
15748048 KM |
3211 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3212 | if (nr_slab_pages1 < nr_slab_pages0) | |
3213 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
2a16e3f4 CL |
3214 | } |
3215 | ||
9eeff239 | 3216 | p->reclaim_state = NULL; |
d4f7796e | 3217 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3218 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3219 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3220 | } |
179e9639 AM |
3221 | |
3222 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3223 | { | |
179e9639 | 3224 | int node_id; |
d773ed6b | 3225 | int ret; |
179e9639 AM |
3226 | |
3227 | /* | |
0ff38490 CL |
3228 | * Zone reclaim reclaims unmapped file backed pages and |
3229 | * slab pages if we are over the defined limits. | |
34aa1330 | 3230 | * |
9614634f CL |
3231 | * A small portion of unmapped file backed pages is needed for |
3232 | * file I/O otherwise pages read by file I/O will be immediately | |
3233 | * thrown out if the zone is overallocated. So we do not reclaim | |
3234 | * if less than a specified percentage of the zone is used by | |
3235 | * unmapped file backed pages. | |
179e9639 | 3236 | */ |
90afa5de MG |
3237 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
3238 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 3239 | return ZONE_RECLAIM_FULL; |
179e9639 | 3240 | |
93e4a89a | 3241 | if (zone->all_unreclaimable) |
fa5e084e | 3242 | return ZONE_RECLAIM_FULL; |
d773ed6b | 3243 | |
179e9639 | 3244 | /* |
d773ed6b | 3245 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3246 | */ |
d773ed6b | 3247 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 3248 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
3249 | |
3250 | /* | |
3251 | * Only run zone reclaim on the local zone or on zones that do not | |
3252 | * have associated processors. This will favor the local processor | |
3253 | * over remote processors and spread off node memory allocations | |
3254 | * as wide as possible. | |
3255 | */ | |
89fa3024 | 3256 | node_id = zone_to_nid(zone); |
37c0708d | 3257 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 3258 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b DR |
3259 | |
3260 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
fa5e084e MG |
3261 | return ZONE_RECLAIM_NOSCAN; |
3262 | ||
d773ed6b DR |
3263 | ret = __zone_reclaim(zone, gfp_mask, order); |
3264 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3265 | ||
24cf7251 MG |
3266 | if (!ret) |
3267 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3268 | ||
d773ed6b | 3269 | return ret; |
179e9639 | 3270 | } |
9eeff239 | 3271 | #endif |
894bc310 | 3272 | |
894bc310 LS |
3273 | /* |
3274 | * page_evictable - test whether a page is evictable | |
3275 | * @page: the page to test | |
3276 | * @vma: the VMA in which the page is or will be mapped, may be NULL | |
3277 | * | |
3278 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
b291f000 NP |
3279 | * lists vs unevictable list. The vma argument is !NULL when called from the |
3280 | * fault path to determine how to instantate a new page. | |
894bc310 LS |
3281 | * |
3282 | * Reasons page might not be evictable: | |
ba9ddf49 | 3283 | * (1) page's mapping marked unevictable |
b291f000 | 3284 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3285 | * |
894bc310 LS |
3286 | */ |
3287 | int page_evictable(struct page *page, struct vm_area_struct *vma) | |
3288 | { | |
3289 | ||
ba9ddf49 LS |
3290 | if (mapping_unevictable(page_mapping(page))) |
3291 | return 0; | |
3292 | ||
096a7cf4 | 3293 | if (PageMlocked(page) || (vma && mlocked_vma_newpage(vma, page))) |
b291f000 | 3294 | return 0; |
894bc310 LS |
3295 | |
3296 | return 1; | |
3297 | } | |
89e004ea | 3298 | |
85046579 | 3299 | #ifdef CONFIG_SHMEM |
89e004ea | 3300 | /** |
24513264 HD |
3301 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3302 | * @pages: array of pages to check | |
3303 | * @nr_pages: number of pages to check | |
89e004ea | 3304 | * |
24513264 | 3305 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3306 | * |
3307 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3308 | */ |
24513264 | 3309 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3310 | { |
925b7673 | 3311 | struct lruvec *lruvec; |
24513264 HD |
3312 | struct zone *zone = NULL; |
3313 | int pgscanned = 0; | |
3314 | int pgrescued = 0; | |
3315 | int i; | |
89e004ea | 3316 | |
24513264 HD |
3317 | for (i = 0; i < nr_pages; i++) { |
3318 | struct page *page = pages[i]; | |
3319 | struct zone *pagezone; | |
89e004ea | 3320 | |
24513264 HD |
3321 | pgscanned++; |
3322 | pagezone = page_zone(page); | |
3323 | if (pagezone != zone) { | |
3324 | if (zone) | |
3325 | spin_unlock_irq(&zone->lru_lock); | |
3326 | zone = pagezone; | |
3327 | spin_lock_irq(&zone->lru_lock); | |
3328 | } | |
89e004ea | 3329 | |
24513264 HD |
3330 | if (!PageLRU(page) || !PageUnevictable(page)) |
3331 | continue; | |
89e004ea | 3332 | |
24513264 HD |
3333 | if (page_evictable(page, NULL)) { |
3334 | enum lru_list lru = page_lru_base_type(page); | |
3335 | ||
3336 | VM_BUG_ON(PageActive(page)); | |
3337 | ClearPageUnevictable(page); | |
3338 | __dec_zone_state(zone, NR_UNEVICTABLE); | |
3339 | lruvec = mem_cgroup_lru_move_lists(zone, page, | |
3340 | LRU_UNEVICTABLE, lru); | |
3341 | list_move(&page->lru, &lruvec->lists[lru]); | |
3342 | __inc_zone_state(zone, NR_INACTIVE_ANON + lru); | |
3343 | pgrescued++; | |
89e004ea | 3344 | } |
24513264 | 3345 | } |
89e004ea | 3346 | |
24513264 HD |
3347 | if (zone) { |
3348 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3349 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3350 | spin_unlock_irq(&zone->lru_lock); | |
89e004ea | 3351 | } |
89e004ea | 3352 | } |
85046579 | 3353 | #endif /* CONFIG_SHMEM */ |
af936a16 | 3354 | |
264e56d8 | 3355 | static void warn_scan_unevictable_pages(void) |
af936a16 | 3356 | { |
264e56d8 | 3357 | printk_once(KERN_WARNING |
25bd91bd | 3358 | "%s: The scan_unevictable_pages sysctl/node-interface has been " |
264e56d8 | 3359 | "disabled for lack of a legitimate use case. If you have " |
25bd91bd KM |
3360 | "one, please send an email to linux-mm@kvack.org.\n", |
3361 | current->comm); | |
af936a16 LS |
3362 | } |
3363 | ||
3364 | /* | |
3365 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3366 | * all nodes' unevictable lists for evictable pages | |
3367 | */ | |
3368 | unsigned long scan_unevictable_pages; | |
3369 | ||
3370 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
8d65af78 | 3371 | void __user *buffer, |
af936a16 LS |
3372 | size_t *length, loff_t *ppos) |
3373 | { | |
264e56d8 | 3374 | warn_scan_unevictable_pages(); |
8d65af78 | 3375 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
af936a16 LS |
3376 | scan_unevictable_pages = 0; |
3377 | return 0; | |
3378 | } | |
3379 | ||
e4455abb | 3380 | #ifdef CONFIG_NUMA |
af936a16 LS |
3381 | /* |
3382 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3383 | * a specified node's per zone unevictable lists for evictable pages. | |
3384 | */ | |
3385 | ||
10fbcf4c KS |
3386 | static ssize_t read_scan_unevictable_node(struct device *dev, |
3387 | struct device_attribute *attr, | |
af936a16 LS |
3388 | char *buf) |
3389 | { | |
264e56d8 | 3390 | warn_scan_unevictable_pages(); |
af936a16 LS |
3391 | return sprintf(buf, "0\n"); /* always zero; should fit... */ |
3392 | } | |
3393 | ||
10fbcf4c KS |
3394 | static ssize_t write_scan_unevictable_node(struct device *dev, |
3395 | struct device_attribute *attr, | |
af936a16 LS |
3396 | const char *buf, size_t count) |
3397 | { | |
264e56d8 | 3398 | warn_scan_unevictable_pages(); |
af936a16 LS |
3399 | return 1; |
3400 | } | |
3401 | ||
3402 | ||
10fbcf4c | 3403 | static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, |
af936a16 LS |
3404 | read_scan_unevictable_node, |
3405 | write_scan_unevictable_node); | |
3406 | ||
3407 | int scan_unevictable_register_node(struct node *node) | |
3408 | { | |
10fbcf4c | 3409 | return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 LS |
3410 | } |
3411 | ||
3412 | void scan_unevictable_unregister_node(struct node *node) | |
3413 | { | |
10fbcf4c | 3414 | device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 | 3415 | } |
e4455abb | 3416 | #endif |