]>
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> | |
70ddf637 | 22 | #include <linux/vmpressure.h> |
e129b5c2 | 23 | #include <linux/vmstat.h> |
1da177e4 LT |
24 | #include <linux/file.h> |
25 | #include <linux/writeback.h> | |
26 | #include <linux/blkdev.h> | |
27 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
28 | buffer_heads_over_limit */ | |
29 | #include <linux/mm_inline.h> | |
1da177e4 LT |
30 | #include <linux/backing-dev.h> |
31 | #include <linux/rmap.h> | |
32 | #include <linux/topology.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/cpuset.h> | |
3e7d3449 | 35 | #include <linux/compaction.h> |
1da177e4 LT |
36 | #include <linux/notifier.h> |
37 | #include <linux/rwsem.h> | |
248a0301 | 38 | #include <linux/delay.h> |
3218ae14 | 39 | #include <linux/kthread.h> |
7dfb7103 | 40 | #include <linux/freezer.h> |
66e1707b | 41 | #include <linux/memcontrol.h> |
873b4771 | 42 | #include <linux/delayacct.h> |
af936a16 | 43 | #include <linux/sysctl.h> |
929bea7c | 44 | #include <linux/oom.h> |
268bb0ce | 45 | #include <linux/prefetch.h> |
1da177e4 LT |
46 | |
47 | #include <asm/tlbflush.h> | |
48 | #include <asm/div64.h> | |
49 | ||
50 | #include <linux/swapops.h> | |
51 | ||
0f8053a5 NP |
52 | #include "internal.h" |
53 | ||
33906bc5 MG |
54 | #define CREATE_TRACE_POINTS |
55 | #include <trace/events/vmscan.h> | |
56 | ||
1da177e4 | 57 | struct scan_control { |
1da177e4 LT |
58 | /* Incremented by the number of inactive pages that were scanned */ |
59 | unsigned long nr_scanned; | |
60 | ||
a79311c1 RR |
61 | /* Number of pages freed so far during a call to shrink_zones() */ |
62 | unsigned long nr_reclaimed; | |
63 | ||
22fba335 KM |
64 | /* How many pages shrink_list() should reclaim */ |
65 | unsigned long nr_to_reclaim; | |
66 | ||
7b51755c KM |
67 | unsigned long hibernation_mode; |
68 | ||
1da177e4 | 69 | /* This context's GFP mask */ |
6daa0e28 | 70 | gfp_t gfp_mask; |
1da177e4 LT |
71 | |
72 | int may_writepage; | |
73 | ||
a6dc60f8 JW |
74 | /* Can mapped pages be reclaimed? */ |
75 | int may_unmap; | |
f1fd1067 | 76 | |
2e2e4259 KM |
77 | /* Can pages be swapped as part of reclaim? */ |
78 | int may_swap; | |
79 | ||
5ad333eb | 80 | int order; |
66e1707b | 81 | |
9e3b2f8c KK |
82 | /* Scan (total_size >> priority) pages at once */ |
83 | int priority; | |
84 | ||
f16015fb JW |
85 | /* |
86 | * The memory cgroup that hit its limit and as a result is the | |
87 | * primary target of this reclaim invocation. | |
88 | */ | |
89 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 90 | |
327c0e96 KH |
91 | /* |
92 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
93 | * are scanned. | |
94 | */ | |
95 | nodemask_t *nodemask; | |
1da177e4 LT |
96 | }; |
97 | ||
1da177e4 LT |
98 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
99 | ||
100 | #ifdef ARCH_HAS_PREFETCH | |
101 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
102 | do { \ | |
103 | if ((_page)->lru.prev != _base) { \ | |
104 | struct page *prev; \ | |
105 | \ | |
106 | prev = lru_to_page(&(_page->lru)); \ | |
107 | prefetch(&prev->_field); \ | |
108 | } \ | |
109 | } while (0) | |
110 | #else | |
111 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
112 | #endif | |
113 | ||
114 | #ifdef ARCH_HAS_PREFETCHW | |
115 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
116 | do { \ | |
117 | if ((_page)->lru.prev != _base) { \ | |
118 | struct page *prev; \ | |
119 | \ | |
120 | prev = lru_to_page(&(_page->lru)); \ | |
121 | prefetchw(&prev->_field); \ | |
122 | } \ | |
123 | } while (0) | |
124 | #else | |
125 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
126 | #endif | |
127 | ||
128 | /* | |
129 | * From 0 .. 100. Higher means more swappy. | |
130 | */ | |
131 | int vm_swappiness = 60; | |
b21e0b90 | 132 | unsigned long vm_total_pages; /* The total number of pages which the VM controls */ |
1da177e4 LT |
133 | |
134 | static LIST_HEAD(shrinker_list); | |
135 | static DECLARE_RWSEM(shrinker_rwsem); | |
136 | ||
c255a458 | 137 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
138 | static bool global_reclaim(struct scan_control *sc) |
139 | { | |
f16015fb | 140 | return !sc->target_mem_cgroup; |
89b5fae5 | 141 | } |
91a45470 | 142 | #else |
89b5fae5 JW |
143 | static bool global_reclaim(struct scan_control *sc) |
144 | { | |
145 | return true; | |
146 | } | |
91a45470 KH |
147 | #endif |
148 | ||
6e543d57 LD |
149 | unsigned long zone_reclaimable_pages(struct zone *zone) |
150 | { | |
151 | int nr; | |
152 | ||
153 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
154 | zone_page_state(zone, NR_INACTIVE_FILE); | |
155 | ||
156 | if (get_nr_swap_pages() > 0) | |
157 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
158 | zone_page_state(zone, NR_INACTIVE_ANON); | |
159 | ||
160 | return nr; | |
161 | } | |
162 | ||
163 | bool zone_reclaimable(struct zone *zone) | |
164 | { | |
165 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
166 | } | |
167 | ||
4d7dcca2 | 168 | static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) |
c9f299d9 | 169 | { |
c3c787e8 | 170 | if (!mem_cgroup_disabled()) |
4d7dcca2 | 171 | return mem_cgroup_get_lru_size(lruvec, lru); |
a3d8e054 | 172 | |
074291fe | 173 | return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); |
c9f299d9 KM |
174 | } |
175 | ||
1da177e4 LT |
176 | /* |
177 | * Add a shrinker callback to be called from the vm | |
178 | */ | |
8e1f936b | 179 | void register_shrinker(struct shrinker *shrinker) |
1da177e4 | 180 | { |
83aeeada | 181 | atomic_long_set(&shrinker->nr_in_batch, 0); |
8e1f936b RR |
182 | down_write(&shrinker_rwsem); |
183 | list_add_tail(&shrinker->list, &shrinker_list); | |
184 | up_write(&shrinker_rwsem); | |
1da177e4 | 185 | } |
8e1f936b | 186 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
187 | |
188 | /* | |
189 | * Remove one | |
190 | */ | |
8e1f936b | 191 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
192 | { |
193 | down_write(&shrinker_rwsem); | |
194 | list_del(&shrinker->list); | |
195 | up_write(&shrinker_rwsem); | |
1da177e4 | 196 | } |
8e1f936b | 197 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 | 198 | |
1495f230 YH |
199 | static inline int do_shrinker_shrink(struct shrinker *shrinker, |
200 | struct shrink_control *sc, | |
201 | unsigned long nr_to_scan) | |
202 | { | |
203 | sc->nr_to_scan = nr_to_scan; | |
204 | return (*shrinker->shrink)(shrinker, sc); | |
205 | } | |
206 | ||
1da177e4 LT |
207 | #define SHRINK_BATCH 128 |
208 | /* | |
209 | * Call the shrink functions to age shrinkable caches | |
210 | * | |
211 | * Here we assume it costs one seek to replace a lru page and that it also | |
212 | * takes a seek to recreate a cache object. With this in mind we age equal | |
213 | * percentages of the lru and ageable caches. This should balance the seeks | |
214 | * generated by these structures. | |
215 | * | |
183ff22b | 216 | * If the vm encountered mapped pages on the LRU it increase the pressure on |
1da177e4 LT |
217 | * slab to avoid swapping. |
218 | * | |
219 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
220 | * | |
221 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
222 | * are eligible for the caller's allocation attempt. It is used for balancing | |
223 | * slab reclaim versus page reclaim. | |
b15e0905 | 224 | * |
225 | * Returns the number of slab objects which we shrunk. | |
1da177e4 | 226 | */ |
a09ed5e0 | 227 | unsigned long shrink_slab(struct shrink_control *shrink, |
1495f230 | 228 | unsigned long nr_pages_scanned, |
a09ed5e0 | 229 | unsigned long lru_pages) |
1da177e4 LT |
230 | { |
231 | struct shrinker *shrinker; | |
69e05944 | 232 | unsigned long ret = 0; |
1da177e4 | 233 | |
1495f230 YH |
234 | if (nr_pages_scanned == 0) |
235 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 236 | |
f06590bd MK |
237 | if (!down_read_trylock(&shrinker_rwsem)) { |
238 | /* Assume we'll be able to shrink next time */ | |
239 | ret = 1; | |
240 | goto out; | |
241 | } | |
1da177e4 LT |
242 | |
243 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
244 | unsigned long long delta; | |
635697c6 KK |
245 | long total_scan; |
246 | long max_pass; | |
09576073 | 247 | int shrink_ret = 0; |
acf92b48 DC |
248 | long nr; |
249 | long new_nr; | |
e9299f50 DC |
250 | long batch_size = shrinker->batch ? shrinker->batch |
251 | : SHRINK_BATCH; | |
1da177e4 | 252 | |
635697c6 KK |
253 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); |
254 | if (max_pass <= 0) | |
255 | continue; | |
256 | ||
acf92b48 DC |
257 | /* |
258 | * copy the current shrinker scan count into a local variable | |
259 | * and zero it so that other concurrent shrinker invocations | |
260 | * don't also do this scanning work. | |
261 | */ | |
83aeeada | 262 | nr = atomic_long_xchg(&shrinker->nr_in_batch, 0); |
acf92b48 DC |
263 | |
264 | total_scan = nr; | |
1495f230 | 265 | delta = (4 * nr_pages_scanned) / shrinker->seeks; |
ea164d73 | 266 | delta *= max_pass; |
1da177e4 | 267 | do_div(delta, lru_pages + 1); |
acf92b48 DC |
268 | total_scan += delta; |
269 | if (total_scan < 0) { | |
88c3bd70 DR |
270 | printk(KERN_ERR "shrink_slab: %pF negative objects to " |
271 | "delete nr=%ld\n", | |
acf92b48 DC |
272 | shrinker->shrink, total_scan); |
273 | total_scan = max_pass; | |
ea164d73 AA |
274 | } |
275 | ||
3567b59a DC |
276 | /* |
277 | * We need to avoid excessive windup on filesystem shrinkers | |
278 | * due to large numbers of GFP_NOFS allocations causing the | |
279 | * shrinkers to return -1 all the time. This results in a large | |
280 | * nr being built up so when a shrink that can do some work | |
281 | * comes along it empties the entire cache due to nr >>> | |
282 | * max_pass. This is bad for sustaining a working set in | |
283 | * memory. | |
284 | * | |
285 | * Hence only allow the shrinker to scan the entire cache when | |
286 | * a large delta change is calculated directly. | |
287 | */ | |
288 | if (delta < max_pass / 4) | |
289 | total_scan = min(total_scan, max_pass / 2); | |
290 | ||
ea164d73 AA |
291 | /* |
292 | * Avoid risking looping forever due to too large nr value: | |
293 | * never try to free more than twice the estimate number of | |
294 | * freeable entries. | |
295 | */ | |
acf92b48 DC |
296 | if (total_scan > max_pass * 2) |
297 | total_scan = max_pass * 2; | |
1da177e4 | 298 | |
acf92b48 | 299 | trace_mm_shrink_slab_start(shrinker, shrink, nr, |
09576073 DC |
300 | nr_pages_scanned, lru_pages, |
301 | max_pass, delta, total_scan); | |
302 | ||
e9299f50 | 303 | while (total_scan >= batch_size) { |
b15e0905 | 304 | int nr_before; |
1da177e4 | 305 | |
1495f230 YH |
306 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); |
307 | shrink_ret = do_shrinker_shrink(shrinker, shrink, | |
e9299f50 | 308 | batch_size); |
1da177e4 LT |
309 | if (shrink_ret == -1) |
310 | break; | |
b15e0905 | 311 | if (shrink_ret < nr_before) |
312 | ret += nr_before - shrink_ret; | |
e9299f50 DC |
313 | count_vm_events(SLABS_SCANNED, batch_size); |
314 | total_scan -= batch_size; | |
1da177e4 LT |
315 | |
316 | cond_resched(); | |
317 | } | |
318 | ||
acf92b48 DC |
319 | /* |
320 | * move the unused scan count back into the shrinker in a | |
321 | * manner that handles concurrent updates. If we exhausted the | |
322 | * scan, there is no need to do an update. | |
323 | */ | |
83aeeada KK |
324 | if (total_scan > 0) |
325 | new_nr = atomic_long_add_return(total_scan, | |
326 | &shrinker->nr_in_batch); | |
327 | else | |
328 | new_nr = atomic_long_read(&shrinker->nr_in_batch); | |
acf92b48 DC |
329 | |
330 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); | |
1da177e4 LT |
331 | } |
332 | up_read(&shrinker_rwsem); | |
f06590bd MK |
333 | out: |
334 | cond_resched(); | |
b15e0905 | 335 | return ret; |
1da177e4 LT |
336 | } |
337 | ||
1da177e4 LT |
338 | static inline int is_page_cache_freeable(struct page *page) |
339 | { | |
ceddc3a5 JW |
340 | /* |
341 | * A freeable page cache page is referenced only by the caller | |
342 | * that isolated the page, the page cache radix tree and | |
343 | * optional buffer heads at page->private. | |
344 | */ | |
edcf4748 | 345 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
346 | } |
347 | ||
7d3579e8 KM |
348 | static int may_write_to_queue(struct backing_dev_info *bdi, |
349 | struct scan_control *sc) | |
1da177e4 | 350 | { |
930d9152 | 351 | if (current->flags & PF_SWAPWRITE) |
1da177e4 LT |
352 | return 1; |
353 | if (!bdi_write_congested(bdi)) | |
354 | return 1; | |
355 | if (bdi == current->backing_dev_info) | |
356 | return 1; | |
357 | return 0; | |
358 | } | |
359 | ||
360 | /* | |
361 | * We detected a synchronous write error writing a page out. Probably | |
362 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
363 | * fsync(), msync() or close(). | |
364 | * | |
365 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
366 | * prevents it from being freed up. But we have a ref on the page and once | |
367 | * that page is locked, the mapping is pinned. | |
368 | * | |
369 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
370 | * __GFP_FS. | |
371 | */ | |
372 | static void handle_write_error(struct address_space *mapping, | |
373 | struct page *page, int error) | |
374 | { | |
7eaceacc | 375 | lock_page(page); |
3e9f45bd GC |
376 | if (page_mapping(page) == mapping) |
377 | mapping_set_error(mapping, error); | |
1da177e4 LT |
378 | unlock_page(page); |
379 | } | |
380 | ||
04e62a29 CL |
381 | /* possible outcome of pageout() */ |
382 | typedef enum { | |
383 | /* failed to write page out, page is locked */ | |
384 | PAGE_KEEP, | |
385 | /* move page to the active list, page is locked */ | |
386 | PAGE_ACTIVATE, | |
387 | /* page has been sent to the disk successfully, page is unlocked */ | |
388 | PAGE_SUCCESS, | |
389 | /* page is clean and locked */ | |
390 | PAGE_CLEAN, | |
391 | } pageout_t; | |
392 | ||
1da177e4 | 393 | /* |
1742f19f AM |
394 | * pageout is called by shrink_page_list() for each dirty page. |
395 | * Calls ->writepage(). | |
1da177e4 | 396 | */ |
c661b078 | 397 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 398 | struct scan_control *sc) |
1da177e4 LT |
399 | { |
400 | /* | |
401 | * If the page is dirty, only perform writeback if that write | |
402 | * will be non-blocking. To prevent this allocation from being | |
403 | * stalled by pagecache activity. But note that there may be | |
404 | * stalls if we need to run get_block(). We could test | |
405 | * PagePrivate for that. | |
406 | * | |
6aceb53b | 407 | * If this process is currently in __generic_file_aio_write() against |
1da177e4 LT |
408 | * this page's queue, we can perform writeback even if that |
409 | * will block. | |
410 | * | |
411 | * If the page is swapcache, write it back even if that would | |
412 | * block, for some throttling. This happens by accident, because | |
413 | * swap_backing_dev_info is bust: it doesn't reflect the | |
414 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
415 | */ |
416 | if (!is_page_cache_freeable(page)) | |
417 | return PAGE_KEEP; | |
418 | if (!mapping) { | |
419 | /* | |
420 | * Some data journaling orphaned pages can have | |
421 | * page->mapping == NULL while being dirty with clean buffers. | |
422 | */ | |
266cf658 | 423 | if (page_has_private(page)) { |
1da177e4 LT |
424 | if (try_to_free_buffers(page)) { |
425 | ClearPageDirty(page); | |
d40cee24 | 426 | printk("%s: orphaned page\n", __func__); |
1da177e4 LT |
427 | return PAGE_CLEAN; |
428 | } | |
429 | } | |
430 | return PAGE_KEEP; | |
431 | } | |
432 | if (mapping->a_ops->writepage == NULL) | |
433 | return PAGE_ACTIVATE; | |
0e093d99 | 434 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) |
1da177e4 LT |
435 | return PAGE_KEEP; |
436 | ||
437 | if (clear_page_dirty_for_io(page)) { | |
438 | int res; | |
439 | struct writeback_control wbc = { | |
440 | .sync_mode = WB_SYNC_NONE, | |
441 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
442 | .range_start = 0, |
443 | .range_end = LLONG_MAX, | |
1da177e4 LT |
444 | .for_reclaim = 1, |
445 | }; | |
446 | ||
447 | SetPageReclaim(page); | |
448 | res = mapping->a_ops->writepage(page, &wbc); | |
449 | if (res < 0) | |
450 | handle_write_error(mapping, page, res); | |
994fc28c | 451 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
452 | ClearPageReclaim(page); |
453 | return PAGE_ACTIVATE; | |
454 | } | |
c661b078 | 455 | |
1da177e4 LT |
456 | if (!PageWriteback(page)) { |
457 | /* synchronous write or broken a_ops? */ | |
458 | ClearPageReclaim(page); | |
459 | } | |
23b9da55 | 460 | trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); |
e129b5c2 | 461 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
462 | return PAGE_SUCCESS; |
463 | } | |
464 | ||
465 | return PAGE_CLEAN; | |
466 | } | |
467 | ||
a649fd92 | 468 | /* |
e286781d NP |
469 | * Same as remove_mapping, but if the page is removed from the mapping, it |
470 | * gets returned with a refcount of 0. | |
a649fd92 | 471 | */ |
e286781d | 472 | static int __remove_mapping(struct address_space *mapping, struct page *page) |
49d2e9cc | 473 | { |
28e4d965 NP |
474 | BUG_ON(!PageLocked(page)); |
475 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 476 | |
19fd6231 | 477 | spin_lock_irq(&mapping->tree_lock); |
49d2e9cc | 478 | /* |
0fd0e6b0 NP |
479 | * The non racy check for a busy page. |
480 | * | |
481 | * Must be careful with the order of the tests. When someone has | |
482 | * a ref to the page, it may be possible that they dirty it then | |
483 | * drop the reference. So if PageDirty is tested before page_count | |
484 | * here, then the following race may occur: | |
485 | * | |
486 | * get_user_pages(&page); | |
487 | * [user mapping goes away] | |
488 | * write_to(page); | |
489 | * !PageDirty(page) [good] | |
490 | * SetPageDirty(page); | |
491 | * put_page(page); | |
492 | * !page_count(page) [good, discard it] | |
493 | * | |
494 | * [oops, our write_to data is lost] | |
495 | * | |
496 | * Reversing the order of the tests ensures such a situation cannot | |
497 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
498 | * load is not satisfied before that of page->_count. | |
499 | * | |
500 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
501 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 502 | */ |
e286781d | 503 | if (!page_freeze_refs(page, 2)) |
49d2e9cc | 504 | goto cannot_free; |
e286781d NP |
505 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
506 | if (unlikely(PageDirty(page))) { | |
507 | page_unfreeze_refs(page, 2); | |
49d2e9cc | 508 | goto cannot_free; |
e286781d | 509 | } |
49d2e9cc CL |
510 | |
511 | if (PageSwapCache(page)) { | |
512 | swp_entry_t swap = { .val = page_private(page) }; | |
513 | __delete_from_swap_cache(page); | |
19fd6231 | 514 | spin_unlock_irq(&mapping->tree_lock); |
cb4b86ba | 515 | swapcache_free(swap, page); |
e286781d | 516 | } else { |
6072d13c LT |
517 | void (*freepage)(struct page *); |
518 | ||
519 | freepage = mapping->a_ops->freepage; | |
520 | ||
e64a782f | 521 | __delete_from_page_cache(page); |
19fd6231 | 522 | spin_unlock_irq(&mapping->tree_lock); |
e767e056 | 523 | mem_cgroup_uncharge_cache_page(page); |
6072d13c LT |
524 | |
525 | if (freepage != NULL) | |
526 | freepage(page); | |
49d2e9cc CL |
527 | } |
528 | ||
49d2e9cc CL |
529 | return 1; |
530 | ||
531 | cannot_free: | |
19fd6231 | 532 | spin_unlock_irq(&mapping->tree_lock); |
49d2e9cc CL |
533 | return 0; |
534 | } | |
535 | ||
e286781d NP |
536 | /* |
537 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
538 | * someone else has a ref on the page, abort and return 0. If it was | |
539 | * successfully detached, return 1. Assumes the caller has a single ref on | |
540 | * this page. | |
541 | */ | |
542 | int remove_mapping(struct address_space *mapping, struct page *page) | |
543 | { | |
544 | if (__remove_mapping(mapping, page)) { | |
545 | /* | |
546 | * Unfreezing the refcount with 1 rather than 2 effectively | |
547 | * drops the pagecache ref for us without requiring another | |
548 | * atomic operation. | |
549 | */ | |
550 | page_unfreeze_refs(page, 1); | |
551 | return 1; | |
552 | } | |
553 | return 0; | |
554 | } | |
555 | ||
894bc310 LS |
556 | /** |
557 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
558 | * @page: page to be put back to appropriate lru list | |
559 | * | |
560 | * Add previously isolated @page to appropriate LRU list. | |
561 | * Page may still be unevictable for other reasons. | |
562 | * | |
563 | * lru_lock must not be held, interrupts must be enabled. | |
564 | */ | |
894bc310 LS |
565 | void putback_lru_page(struct page *page) |
566 | { | |
0ec3b74c | 567 | bool is_unevictable; |
bbfd28ee | 568 | int was_unevictable = PageUnevictable(page); |
894bc310 LS |
569 | |
570 | VM_BUG_ON(PageLRU(page)); | |
571 | ||
572 | redo: | |
573 | ClearPageUnevictable(page); | |
574 | ||
39b5f29a | 575 | if (page_evictable(page)) { |
894bc310 LS |
576 | /* |
577 | * For evictable pages, we can use the cache. | |
578 | * In event of a race, worst case is we end up with an | |
579 | * unevictable page on [in]active list. | |
580 | * We know how to handle that. | |
581 | */ | |
0ec3b74c | 582 | is_unevictable = false; |
c53954a0 | 583 | lru_cache_add(page); |
894bc310 LS |
584 | } else { |
585 | /* | |
586 | * Put unevictable pages directly on zone's unevictable | |
587 | * list. | |
588 | */ | |
0ec3b74c | 589 | is_unevictable = true; |
894bc310 | 590 | add_page_to_unevictable_list(page); |
6a7b9548 | 591 | /* |
21ee9f39 MK |
592 | * When racing with an mlock or AS_UNEVICTABLE clearing |
593 | * (page is unlocked) make sure that if the other thread | |
594 | * does not observe our setting of PG_lru and fails | |
24513264 | 595 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 596 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
597 | * the page back to the evictable list. |
598 | * | |
21ee9f39 | 599 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
600 | */ |
601 | smp_mb(); | |
894bc310 | 602 | } |
894bc310 LS |
603 | |
604 | /* | |
605 | * page's status can change while we move it among lru. If an evictable | |
606 | * page is on unevictable list, it never be freed. To avoid that, | |
607 | * check after we added it to the list, again. | |
608 | */ | |
0ec3b74c | 609 | if (is_unevictable && page_evictable(page)) { |
894bc310 LS |
610 | if (!isolate_lru_page(page)) { |
611 | put_page(page); | |
612 | goto redo; | |
613 | } | |
614 | /* This means someone else dropped this page from LRU | |
615 | * So, it will be freed or putback to LRU again. There is | |
616 | * nothing to do here. | |
617 | */ | |
618 | } | |
619 | ||
0ec3b74c | 620 | if (was_unevictable && !is_unevictable) |
bbfd28ee | 621 | count_vm_event(UNEVICTABLE_PGRESCUED); |
0ec3b74c | 622 | else if (!was_unevictable && is_unevictable) |
bbfd28ee LS |
623 | count_vm_event(UNEVICTABLE_PGCULLED); |
624 | ||
894bc310 LS |
625 | put_page(page); /* drop ref from isolate */ |
626 | } | |
627 | ||
dfc8d636 JW |
628 | enum page_references { |
629 | PAGEREF_RECLAIM, | |
630 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 631 | PAGEREF_KEEP, |
dfc8d636 JW |
632 | PAGEREF_ACTIVATE, |
633 | }; | |
634 | ||
635 | static enum page_references page_check_references(struct page *page, | |
636 | struct scan_control *sc) | |
637 | { | |
64574746 | 638 | int referenced_ptes, referenced_page; |
dfc8d636 | 639 | unsigned long vm_flags; |
dfc8d636 | 640 | |
c3ac9a8a JW |
641 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
642 | &vm_flags); | |
64574746 | 643 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 644 | |
dfc8d636 JW |
645 | /* |
646 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
647 | * move the page to the unevictable list. | |
648 | */ | |
649 | if (vm_flags & VM_LOCKED) | |
650 | return PAGEREF_RECLAIM; | |
651 | ||
64574746 | 652 | if (referenced_ptes) { |
e4898273 | 653 | if (PageSwapBacked(page)) |
64574746 JW |
654 | return PAGEREF_ACTIVATE; |
655 | /* | |
656 | * All mapped pages start out with page table | |
657 | * references from the instantiating fault, so we need | |
658 | * to look twice if a mapped file page is used more | |
659 | * than once. | |
660 | * | |
661 | * Mark it and spare it for another trip around the | |
662 | * inactive list. Another page table reference will | |
663 | * lead to its activation. | |
664 | * | |
665 | * Note: the mark is set for activated pages as well | |
666 | * so that recently deactivated but used pages are | |
667 | * quickly recovered. | |
668 | */ | |
669 | SetPageReferenced(page); | |
670 | ||
34dbc67a | 671 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
672 | return PAGEREF_ACTIVATE; |
673 | ||
c909e993 KK |
674 | /* |
675 | * Activate file-backed executable pages after first usage. | |
676 | */ | |
677 | if (vm_flags & VM_EXEC) | |
678 | return PAGEREF_ACTIVATE; | |
679 | ||
64574746 JW |
680 | return PAGEREF_KEEP; |
681 | } | |
dfc8d636 JW |
682 | |
683 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 684 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
685 | return PAGEREF_RECLAIM_CLEAN; |
686 | ||
687 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
688 | } |
689 | ||
e2be15f6 MG |
690 | /* Check if a page is dirty or under writeback */ |
691 | static void page_check_dirty_writeback(struct page *page, | |
692 | bool *dirty, bool *writeback) | |
693 | { | |
b4597226 MG |
694 | struct address_space *mapping; |
695 | ||
e2be15f6 MG |
696 | /* |
697 | * Anonymous pages are not handled by flushers and must be written | |
698 | * from reclaim context. Do not stall reclaim based on them | |
699 | */ | |
700 | if (!page_is_file_cache(page)) { | |
701 | *dirty = false; | |
702 | *writeback = false; | |
703 | return; | |
704 | } | |
705 | ||
706 | /* By default assume that the page flags are accurate */ | |
707 | *dirty = PageDirty(page); | |
708 | *writeback = PageWriteback(page); | |
b4597226 MG |
709 | |
710 | /* Verify dirty/writeback state if the filesystem supports it */ | |
711 | if (!page_has_private(page)) | |
712 | return; | |
713 | ||
714 | mapping = page_mapping(page); | |
715 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
716 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
717 | } |
718 | ||
1da177e4 | 719 | /* |
1742f19f | 720 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 721 | */ |
1742f19f | 722 | static unsigned long shrink_page_list(struct list_head *page_list, |
6a18adb3 | 723 | struct zone *zone, |
f84f6e2b | 724 | struct scan_control *sc, |
02c6de8d | 725 | enum ttu_flags ttu_flags, |
8e950282 | 726 | unsigned long *ret_nr_dirty, |
d43006d5 | 727 | unsigned long *ret_nr_unqueued_dirty, |
8e950282 | 728 | unsigned long *ret_nr_congested, |
02c6de8d | 729 | unsigned long *ret_nr_writeback, |
b1a6f21e | 730 | unsigned long *ret_nr_immediate, |
02c6de8d | 731 | bool force_reclaim) |
1da177e4 LT |
732 | { |
733 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 734 | LIST_HEAD(free_pages); |
1da177e4 | 735 | int pgactivate = 0; |
d43006d5 | 736 | unsigned long nr_unqueued_dirty = 0; |
0e093d99 MG |
737 | unsigned long nr_dirty = 0; |
738 | unsigned long nr_congested = 0; | |
05ff5137 | 739 | unsigned long nr_reclaimed = 0; |
92df3a72 | 740 | unsigned long nr_writeback = 0; |
b1a6f21e | 741 | unsigned long nr_immediate = 0; |
1da177e4 LT |
742 | |
743 | cond_resched(); | |
744 | ||
69980e31 | 745 | mem_cgroup_uncharge_start(); |
1da177e4 LT |
746 | while (!list_empty(page_list)) { |
747 | struct address_space *mapping; | |
748 | struct page *page; | |
749 | int may_enter_fs; | |
02c6de8d | 750 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
e2be15f6 | 751 | bool dirty, writeback; |
1da177e4 LT |
752 | |
753 | cond_resched(); | |
754 | ||
755 | page = lru_to_page(page_list); | |
756 | list_del(&page->lru); | |
757 | ||
529ae9aa | 758 | if (!trylock_page(page)) |
1da177e4 LT |
759 | goto keep; |
760 | ||
725d704e | 761 | VM_BUG_ON(PageActive(page)); |
6a18adb3 | 762 | VM_BUG_ON(page_zone(page) != zone); |
1da177e4 LT |
763 | |
764 | sc->nr_scanned++; | |
80e43426 | 765 | |
39b5f29a | 766 | if (unlikely(!page_evictable(page))) |
b291f000 | 767 | goto cull_mlocked; |
894bc310 | 768 | |
a6dc60f8 | 769 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
770 | goto keep_locked; |
771 | ||
1da177e4 LT |
772 | /* Double the slab pressure for mapped and swapcache pages */ |
773 | if (page_mapped(page) || PageSwapCache(page)) | |
774 | sc->nr_scanned++; | |
775 | ||
c661b078 AW |
776 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
777 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
778 | ||
e2be15f6 MG |
779 | /* |
780 | * The number of dirty pages determines if a zone is marked | |
781 | * reclaim_congested which affects wait_iff_congested. kswapd | |
782 | * will stall and start writing pages if the tail of the LRU | |
783 | * is all dirty unqueued pages. | |
784 | */ | |
785 | page_check_dirty_writeback(page, &dirty, &writeback); | |
786 | if (dirty || writeback) | |
787 | nr_dirty++; | |
788 | ||
789 | if (dirty && !writeback) | |
790 | nr_unqueued_dirty++; | |
791 | ||
d04e8acd MG |
792 | /* |
793 | * Treat this page as congested if the underlying BDI is or if | |
794 | * pages are cycling through the LRU so quickly that the | |
795 | * pages marked for immediate reclaim are making it to the | |
796 | * end of the LRU a second time. | |
797 | */ | |
e2be15f6 | 798 | mapping = page_mapping(page); |
d04e8acd MG |
799 | if ((mapping && bdi_write_congested(mapping->backing_dev_info)) || |
800 | (writeback && PageReclaim(page))) | |
e2be15f6 MG |
801 | nr_congested++; |
802 | ||
283aba9f MG |
803 | /* |
804 | * If a page at the tail of the LRU is under writeback, there | |
805 | * are three cases to consider. | |
806 | * | |
807 | * 1) If reclaim is encountering an excessive number of pages | |
808 | * under writeback and this page is both under writeback and | |
809 | * PageReclaim then it indicates that pages are being queued | |
810 | * for IO but are being recycled through the LRU before the | |
811 | * IO can complete. Waiting on the page itself risks an | |
812 | * indefinite stall if it is impossible to writeback the | |
813 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
814 | * note that the LRU is being scanned too quickly and the |
815 | * caller can stall after page list has been processed. | |
283aba9f MG |
816 | * |
817 | * 2) Global reclaim encounters a page, memcg encounters a | |
818 | * page that is not marked for immediate reclaim or | |
819 | * the caller does not have __GFP_IO. In this case mark | |
820 | * the page for immediate reclaim and continue scanning. | |
821 | * | |
822 | * __GFP_IO is checked because a loop driver thread might | |
823 | * enter reclaim, and deadlock if it waits on a page for | |
824 | * which it is needed to do the write (loop masks off | |
825 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
826 | * would probably show more reasons. | |
827 | * | |
828 | * Don't require __GFP_FS, since we're not going into the | |
829 | * FS, just waiting on its writeback completion. Worryingly, | |
830 | * ext4 gfs2 and xfs allocate pages with | |
831 | * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing | |
832 | * may_enter_fs here is liable to OOM on them. | |
833 | * | |
834 | * 3) memcg encounters a page that is not already marked | |
835 | * PageReclaim. memcg does not have any dirty pages | |
836 | * throttling so we could easily OOM just because too many | |
837 | * pages are in writeback and there is nothing else to | |
838 | * reclaim. Wait for the writeback to complete. | |
839 | */ | |
c661b078 | 840 | if (PageWriteback(page)) { |
283aba9f MG |
841 | /* Case 1 above */ |
842 | if (current_is_kswapd() && | |
843 | PageReclaim(page) && | |
844 | zone_is_reclaim_writeback(zone)) { | |
b1a6f21e MG |
845 | nr_immediate++; |
846 | goto keep_locked; | |
283aba9f MG |
847 | |
848 | /* Case 2 above */ | |
849 | } else if (global_reclaim(sc) || | |
c3b94f44 HD |
850 | !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) { |
851 | /* | |
852 | * This is slightly racy - end_page_writeback() | |
853 | * might have just cleared PageReclaim, then | |
854 | * setting PageReclaim here end up interpreted | |
855 | * as PageReadahead - but that does not matter | |
856 | * enough to care. What we do want is for this | |
857 | * page to have PageReclaim set next time memcg | |
858 | * reclaim reaches the tests above, so it will | |
859 | * then wait_on_page_writeback() to avoid OOM; | |
860 | * and it's also appropriate in global reclaim. | |
861 | */ | |
862 | SetPageReclaim(page); | |
e62e384e | 863 | nr_writeback++; |
283aba9f | 864 | |
c3b94f44 | 865 | goto keep_locked; |
283aba9f MG |
866 | |
867 | /* Case 3 above */ | |
868 | } else { | |
869 | wait_on_page_writeback(page); | |
e62e384e | 870 | } |
c661b078 | 871 | } |
1da177e4 | 872 | |
02c6de8d MK |
873 | if (!force_reclaim) |
874 | references = page_check_references(page, sc); | |
875 | ||
dfc8d636 JW |
876 | switch (references) { |
877 | case PAGEREF_ACTIVATE: | |
1da177e4 | 878 | goto activate_locked; |
64574746 JW |
879 | case PAGEREF_KEEP: |
880 | goto keep_locked; | |
dfc8d636 JW |
881 | case PAGEREF_RECLAIM: |
882 | case PAGEREF_RECLAIM_CLEAN: | |
883 | ; /* try to reclaim the page below */ | |
884 | } | |
1da177e4 | 885 | |
1da177e4 LT |
886 | /* |
887 | * Anonymous process memory has backing store? | |
888 | * Try to allocate it some swap space here. | |
889 | */ | |
b291f000 | 890 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
891 | if (!(sc->gfp_mask & __GFP_IO)) |
892 | goto keep_locked; | |
5bc7b8ac | 893 | if (!add_to_swap(page, page_list)) |
1da177e4 | 894 | goto activate_locked; |
63eb6b93 | 895 | may_enter_fs = 1; |
1da177e4 | 896 | |
e2be15f6 MG |
897 | /* Adding to swap updated mapping */ |
898 | mapping = page_mapping(page); | |
899 | } | |
1da177e4 LT |
900 | |
901 | /* | |
902 | * The page is mapped into the page tables of one or more | |
903 | * processes. Try to unmap it here. | |
904 | */ | |
905 | if (page_mapped(page) && mapping) { | |
02c6de8d | 906 | switch (try_to_unmap(page, ttu_flags)) { |
1da177e4 LT |
907 | case SWAP_FAIL: |
908 | goto activate_locked; | |
909 | case SWAP_AGAIN: | |
910 | goto keep_locked; | |
b291f000 NP |
911 | case SWAP_MLOCK: |
912 | goto cull_mlocked; | |
1da177e4 LT |
913 | case SWAP_SUCCESS: |
914 | ; /* try to free the page below */ | |
915 | } | |
916 | } | |
917 | ||
918 | if (PageDirty(page)) { | |
ee72886d MG |
919 | /* |
920 | * Only kswapd can writeback filesystem pages to | |
d43006d5 MG |
921 | * avoid risk of stack overflow but only writeback |
922 | * if many dirty pages have been encountered. | |
ee72886d | 923 | */ |
f84f6e2b | 924 | if (page_is_file_cache(page) && |
9e3b2f8c | 925 | (!current_is_kswapd() || |
d43006d5 | 926 | !zone_is_reclaim_dirty(zone))) { |
49ea7eb6 MG |
927 | /* |
928 | * Immediately reclaim when written back. | |
929 | * Similar in principal to deactivate_page() | |
930 | * except we already have the page isolated | |
931 | * and know it's dirty | |
932 | */ | |
933 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
934 | SetPageReclaim(page); | |
935 | ||
ee72886d MG |
936 | goto keep_locked; |
937 | } | |
938 | ||
dfc8d636 | 939 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 940 | goto keep_locked; |
4dd4b920 | 941 | if (!may_enter_fs) |
1da177e4 | 942 | goto keep_locked; |
52a8363e | 943 | if (!sc->may_writepage) |
1da177e4 LT |
944 | goto keep_locked; |
945 | ||
946 | /* Page is dirty, try to write it out here */ | |
7d3579e8 | 947 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
948 | case PAGE_KEEP: |
949 | goto keep_locked; | |
950 | case PAGE_ACTIVATE: | |
951 | goto activate_locked; | |
952 | case PAGE_SUCCESS: | |
7d3579e8 | 953 | if (PageWriteback(page)) |
41ac1999 | 954 | goto keep; |
7d3579e8 | 955 | if (PageDirty(page)) |
1da177e4 | 956 | goto keep; |
7d3579e8 | 957 | |
1da177e4 LT |
958 | /* |
959 | * A synchronous write - probably a ramdisk. Go | |
960 | * ahead and try to reclaim the page. | |
961 | */ | |
529ae9aa | 962 | if (!trylock_page(page)) |
1da177e4 LT |
963 | goto keep; |
964 | if (PageDirty(page) || PageWriteback(page)) | |
965 | goto keep_locked; | |
966 | mapping = page_mapping(page); | |
967 | case PAGE_CLEAN: | |
968 | ; /* try to free the page below */ | |
969 | } | |
970 | } | |
971 | ||
972 | /* | |
973 | * If the page has buffers, try to free the buffer mappings | |
974 | * associated with this page. If we succeed we try to free | |
975 | * the page as well. | |
976 | * | |
977 | * We do this even if the page is PageDirty(). | |
978 | * try_to_release_page() does not perform I/O, but it is | |
979 | * possible for a page to have PageDirty set, but it is actually | |
980 | * clean (all its buffers are clean). This happens if the | |
981 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 982 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
983 | * try_to_release_page() will discover that cleanness and will |
984 | * drop the buffers and mark the page clean - it can be freed. | |
985 | * | |
986 | * Rarely, pages can have buffers and no ->mapping. These are | |
987 | * the pages which were not successfully invalidated in | |
988 | * truncate_complete_page(). We try to drop those buffers here | |
989 | * and if that worked, and the page is no longer mapped into | |
990 | * process address space (page_count == 1) it can be freed. | |
991 | * Otherwise, leave the page on the LRU so it is swappable. | |
992 | */ | |
266cf658 | 993 | if (page_has_private(page)) { |
1da177e4 LT |
994 | if (!try_to_release_page(page, sc->gfp_mask)) |
995 | goto activate_locked; | |
e286781d NP |
996 | if (!mapping && page_count(page) == 1) { |
997 | unlock_page(page); | |
998 | if (put_page_testzero(page)) | |
999 | goto free_it; | |
1000 | else { | |
1001 | /* | |
1002 | * rare race with speculative reference. | |
1003 | * the speculative reference will free | |
1004 | * this page shortly, so we may | |
1005 | * increment nr_reclaimed here (and | |
1006 | * leave it off the LRU). | |
1007 | */ | |
1008 | nr_reclaimed++; | |
1009 | continue; | |
1010 | } | |
1011 | } | |
1da177e4 LT |
1012 | } |
1013 | ||
e286781d | 1014 | if (!mapping || !__remove_mapping(mapping, page)) |
49d2e9cc | 1015 | goto keep_locked; |
1da177e4 | 1016 | |
a978d6f5 NP |
1017 | /* |
1018 | * At this point, we have no other references and there is | |
1019 | * no way to pick any more up (removed from LRU, removed | |
1020 | * from pagecache). Can use non-atomic bitops now (and | |
1021 | * we obviously don't have to worry about waking up a process | |
1022 | * waiting on the page lock, because there are no references. | |
1023 | */ | |
1024 | __clear_page_locked(page); | |
e286781d | 1025 | free_it: |
05ff5137 | 1026 | nr_reclaimed++; |
abe4c3b5 MG |
1027 | |
1028 | /* | |
1029 | * Is there need to periodically free_page_list? It would | |
1030 | * appear not as the counts should be low | |
1031 | */ | |
1032 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1033 | continue; |
1034 | ||
b291f000 | 1035 | cull_mlocked: |
63d6c5ad HD |
1036 | if (PageSwapCache(page)) |
1037 | try_to_free_swap(page); | |
b291f000 NP |
1038 | unlock_page(page); |
1039 | putback_lru_page(page); | |
1040 | continue; | |
1041 | ||
1da177e4 | 1042 | activate_locked: |
68a22394 RR |
1043 | /* Not a candidate for swapping, so reclaim swap space. */ |
1044 | if (PageSwapCache(page) && vm_swap_full()) | |
a2c43eed | 1045 | try_to_free_swap(page); |
894bc310 | 1046 | VM_BUG_ON(PageActive(page)); |
1da177e4 LT |
1047 | SetPageActive(page); |
1048 | pgactivate++; | |
1049 | keep_locked: | |
1050 | unlock_page(page); | |
1051 | keep: | |
1052 | list_add(&page->lru, &ret_pages); | |
b291f000 | 1053 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); |
1da177e4 | 1054 | } |
abe4c3b5 | 1055 | |
cc59850e | 1056 | free_hot_cold_page_list(&free_pages, 1); |
abe4c3b5 | 1057 | |
1da177e4 | 1058 | list_splice(&ret_pages, page_list); |
f8891e5e | 1059 | count_vm_events(PGACTIVATE, pgactivate); |
69980e31 | 1060 | mem_cgroup_uncharge_end(); |
8e950282 MG |
1061 | *ret_nr_dirty += nr_dirty; |
1062 | *ret_nr_congested += nr_congested; | |
d43006d5 | 1063 | *ret_nr_unqueued_dirty += nr_unqueued_dirty; |
92df3a72 | 1064 | *ret_nr_writeback += nr_writeback; |
b1a6f21e | 1065 | *ret_nr_immediate += nr_immediate; |
05ff5137 | 1066 | return nr_reclaimed; |
1da177e4 LT |
1067 | } |
1068 | ||
02c6de8d MK |
1069 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1070 | struct list_head *page_list) | |
1071 | { | |
1072 | struct scan_control sc = { | |
1073 | .gfp_mask = GFP_KERNEL, | |
1074 | .priority = DEF_PRIORITY, | |
1075 | .may_unmap = 1, | |
1076 | }; | |
8e950282 | 1077 | unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; |
02c6de8d MK |
1078 | struct page *page, *next; |
1079 | LIST_HEAD(clean_pages); | |
1080 | ||
1081 | list_for_each_entry_safe(page, next, page_list, lru) { | |
1082 | if (page_is_file_cache(page) && !PageDirty(page)) { | |
1083 | ClearPageActive(page); | |
1084 | list_move(&page->lru, &clean_pages); | |
1085 | } | |
1086 | } | |
1087 | ||
1088 | ret = shrink_page_list(&clean_pages, zone, &sc, | |
8e950282 MG |
1089 | TTU_UNMAP|TTU_IGNORE_ACCESS, |
1090 | &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); | |
02c6de8d MK |
1091 | list_splice(&clean_pages, page_list); |
1092 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); | |
1093 | return ret; | |
1094 | } | |
1095 | ||
5ad333eb AW |
1096 | /* |
1097 | * Attempt to remove the specified page from its LRU. Only take this page | |
1098 | * if it is of the appropriate PageActive status. Pages which are being | |
1099 | * freed elsewhere are also ignored. | |
1100 | * | |
1101 | * page: page to consider | |
1102 | * mode: one of the LRU isolation modes defined above | |
1103 | * | |
1104 | * returns 0 on success, -ve errno on failure. | |
1105 | */ | |
f3fd4a61 | 1106 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1107 | { |
1108 | int ret = -EINVAL; | |
1109 | ||
1110 | /* Only take pages on the LRU. */ | |
1111 | if (!PageLRU(page)) | |
1112 | return ret; | |
1113 | ||
e46a2879 MK |
1114 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1115 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1116 | return ret; |
1117 | ||
5ad333eb | 1118 | ret = -EBUSY; |
08e552c6 | 1119 | |
c8244935 MG |
1120 | /* |
1121 | * To minimise LRU disruption, the caller can indicate that it only | |
1122 | * wants to isolate pages it will be able to operate on without | |
1123 | * blocking - clean pages for the most part. | |
1124 | * | |
1125 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1126 | * is used by reclaim when it is cannot write to backing storage | |
1127 | * | |
1128 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1129 | * that it is possible to migrate without blocking | |
1130 | */ | |
1131 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1132 | /* All the caller can do on PageWriteback is block */ | |
1133 | if (PageWriteback(page)) | |
1134 | return ret; | |
1135 | ||
1136 | if (PageDirty(page)) { | |
1137 | struct address_space *mapping; | |
1138 | ||
1139 | /* ISOLATE_CLEAN means only clean pages */ | |
1140 | if (mode & ISOLATE_CLEAN) | |
1141 | return ret; | |
1142 | ||
1143 | /* | |
1144 | * Only pages without mappings or that have a | |
1145 | * ->migratepage callback are possible to migrate | |
1146 | * without blocking | |
1147 | */ | |
1148 | mapping = page_mapping(page); | |
1149 | if (mapping && !mapping->a_ops->migratepage) | |
1150 | return ret; | |
1151 | } | |
1152 | } | |
39deaf85 | 1153 | |
f80c0673 MK |
1154 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1155 | return ret; | |
1156 | ||
5ad333eb AW |
1157 | if (likely(get_page_unless_zero(page))) { |
1158 | /* | |
1159 | * Be careful not to clear PageLRU until after we're | |
1160 | * sure the page is not being freed elsewhere -- the | |
1161 | * page release code relies on it. | |
1162 | */ | |
1163 | ClearPageLRU(page); | |
1164 | ret = 0; | |
1165 | } | |
1166 | ||
1167 | return ret; | |
1168 | } | |
1169 | ||
1da177e4 LT |
1170 | /* |
1171 | * zone->lru_lock is heavily contended. Some of the functions that | |
1172 | * shrink the lists perform better by taking out a batch of pages | |
1173 | * and working on them outside the LRU lock. | |
1174 | * | |
1175 | * For pagecache intensive workloads, this function is the hottest | |
1176 | * spot in the kernel (apart from copy_*_user functions). | |
1177 | * | |
1178 | * Appropriate locks must be held before calling this function. | |
1179 | * | |
1180 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1181 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1182 | * @dst: The temp list to put pages on to. |
f626012d | 1183 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1184 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1185 | * @mode: One of the LRU isolation modes |
3cb99451 | 1186 | * @lru: LRU list id for isolating |
1da177e4 LT |
1187 | * |
1188 | * returns how many pages were moved onto *@dst. | |
1189 | */ | |
69e05944 | 1190 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1191 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1192 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1193 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1194 | { |
75b00af7 | 1195 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1196 | unsigned long nr_taken = 0; |
c9b02d97 | 1197 | unsigned long scan; |
1da177e4 | 1198 | |
c9b02d97 | 1199 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
5ad333eb | 1200 | struct page *page; |
fa9add64 | 1201 | int nr_pages; |
5ad333eb | 1202 | |
1da177e4 LT |
1203 | page = lru_to_page(src); |
1204 | prefetchw_prev_lru_page(page, src, flags); | |
1205 | ||
725d704e | 1206 | VM_BUG_ON(!PageLRU(page)); |
8d438f96 | 1207 | |
f3fd4a61 | 1208 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1209 | case 0: |
fa9add64 HD |
1210 | nr_pages = hpage_nr_pages(page); |
1211 | mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); | |
5ad333eb | 1212 | list_move(&page->lru, dst); |
fa9add64 | 1213 | nr_taken += nr_pages; |
5ad333eb AW |
1214 | break; |
1215 | ||
1216 | case -EBUSY: | |
1217 | /* else it is being freed elsewhere */ | |
1218 | list_move(&page->lru, src); | |
1219 | continue; | |
46453a6e | 1220 | |
5ad333eb AW |
1221 | default: |
1222 | BUG(); | |
1223 | } | |
1da177e4 LT |
1224 | } |
1225 | ||
f626012d | 1226 | *nr_scanned = scan; |
75b00af7 HD |
1227 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, |
1228 | nr_taken, mode, is_file_lru(lru)); | |
1da177e4 LT |
1229 | return nr_taken; |
1230 | } | |
1231 | ||
62695a84 NP |
1232 | /** |
1233 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1234 | * @page: page to isolate from its LRU list | |
1235 | * | |
1236 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1237 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1238 | * | |
1239 | * Returns 0 if the page was removed from an LRU list. | |
1240 | * Returns -EBUSY if the page was not on an LRU list. | |
1241 | * | |
1242 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1243 | * the active list, it will have PageActive set. If it was found on |
1244 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1245 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1246 | * |
1247 | * The vmstat statistic corresponding to the list on which the page was | |
1248 | * found will be decremented. | |
1249 | * | |
1250 | * Restrictions: | |
1251 | * (1) Must be called with an elevated refcount on the page. This is a | |
1252 | * fundamentnal difference from isolate_lru_pages (which is called | |
1253 | * without a stable reference). | |
1254 | * (2) the lru_lock must not be held. | |
1255 | * (3) interrupts must be enabled. | |
1256 | */ | |
1257 | int isolate_lru_page(struct page *page) | |
1258 | { | |
1259 | int ret = -EBUSY; | |
1260 | ||
0c917313 KK |
1261 | VM_BUG_ON(!page_count(page)); |
1262 | ||
62695a84 NP |
1263 | if (PageLRU(page)) { |
1264 | struct zone *zone = page_zone(page); | |
fa9add64 | 1265 | struct lruvec *lruvec; |
62695a84 NP |
1266 | |
1267 | spin_lock_irq(&zone->lru_lock); | |
fa9add64 | 1268 | lruvec = mem_cgroup_page_lruvec(page, zone); |
0c917313 | 1269 | if (PageLRU(page)) { |
894bc310 | 1270 | int lru = page_lru(page); |
0c917313 | 1271 | get_page(page); |
62695a84 | 1272 | ClearPageLRU(page); |
fa9add64 HD |
1273 | del_page_from_lru_list(page, lruvec, lru); |
1274 | ret = 0; | |
62695a84 NP |
1275 | } |
1276 | spin_unlock_irq(&zone->lru_lock); | |
1277 | } | |
1278 | return ret; | |
1279 | } | |
1280 | ||
35cd7815 | 1281 | /* |
d37dd5dc FW |
1282 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1283 | * then get resheduled. When there are massive number of tasks doing page | |
1284 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1285 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1286 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 RR |
1287 | */ |
1288 | static int too_many_isolated(struct zone *zone, int file, | |
1289 | struct scan_control *sc) | |
1290 | { | |
1291 | unsigned long inactive, isolated; | |
1292 | ||
1293 | if (current_is_kswapd()) | |
1294 | return 0; | |
1295 | ||
89b5fae5 | 1296 | if (!global_reclaim(sc)) |
35cd7815 RR |
1297 | return 0; |
1298 | ||
1299 | if (file) { | |
1300 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1301 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1302 | } else { | |
1303 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1304 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1305 | } | |
1306 | ||
3cf23841 FW |
1307 | /* |
1308 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1309 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1310 | * deadlock. | |
1311 | */ | |
1312 | if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS) | |
1313 | inactive >>= 3; | |
1314 | ||
35cd7815 RR |
1315 | return isolated > inactive; |
1316 | } | |
1317 | ||
66635629 | 1318 | static noinline_for_stack void |
75b00af7 | 1319 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1320 | { |
27ac81d8 KK |
1321 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1322 | struct zone *zone = lruvec_zone(lruvec); | |
3f79768f | 1323 | LIST_HEAD(pages_to_free); |
66635629 | 1324 | |
66635629 MG |
1325 | /* |
1326 | * Put back any unfreeable pages. | |
1327 | */ | |
66635629 | 1328 | while (!list_empty(page_list)) { |
3f79768f | 1329 | struct page *page = lru_to_page(page_list); |
66635629 | 1330 | int lru; |
3f79768f | 1331 | |
66635629 MG |
1332 | VM_BUG_ON(PageLRU(page)); |
1333 | list_del(&page->lru); | |
39b5f29a | 1334 | if (unlikely(!page_evictable(page))) { |
66635629 MG |
1335 | spin_unlock_irq(&zone->lru_lock); |
1336 | putback_lru_page(page); | |
1337 | spin_lock_irq(&zone->lru_lock); | |
1338 | continue; | |
1339 | } | |
fa9add64 HD |
1340 | |
1341 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1342 | ||
7a608572 | 1343 | SetPageLRU(page); |
66635629 | 1344 | lru = page_lru(page); |
fa9add64 HD |
1345 | add_page_to_lru_list(page, lruvec, lru); |
1346 | ||
66635629 MG |
1347 | if (is_active_lru(lru)) { |
1348 | int file = is_file_lru(lru); | |
9992af10 RR |
1349 | int numpages = hpage_nr_pages(page); |
1350 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1351 | } |
2bcf8879 HD |
1352 | if (put_page_testzero(page)) { |
1353 | __ClearPageLRU(page); | |
1354 | __ClearPageActive(page); | |
fa9add64 | 1355 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1356 | |
1357 | if (unlikely(PageCompound(page))) { | |
1358 | spin_unlock_irq(&zone->lru_lock); | |
1359 | (*get_compound_page_dtor(page))(page); | |
1360 | spin_lock_irq(&zone->lru_lock); | |
1361 | } else | |
1362 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1363 | } |
1364 | } | |
66635629 | 1365 | |
3f79768f HD |
1366 | /* |
1367 | * To save our caller's stack, now use input list for pages to free. | |
1368 | */ | |
1369 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1370 | } |
1371 | ||
1da177e4 | 1372 | /* |
1742f19f AM |
1373 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1374 | * of reclaimed pages | |
1da177e4 | 1375 | */ |
66635629 | 1376 | static noinline_for_stack unsigned long |
1a93be0e | 1377 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1378 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1379 | { |
1380 | LIST_HEAD(page_list); | |
e247dbce | 1381 | unsigned long nr_scanned; |
05ff5137 | 1382 | unsigned long nr_reclaimed = 0; |
e247dbce | 1383 | unsigned long nr_taken; |
8e950282 MG |
1384 | unsigned long nr_dirty = 0; |
1385 | unsigned long nr_congested = 0; | |
e2be15f6 | 1386 | unsigned long nr_unqueued_dirty = 0; |
92df3a72 | 1387 | unsigned long nr_writeback = 0; |
b1a6f21e | 1388 | unsigned long nr_immediate = 0; |
f3fd4a61 | 1389 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1390 | int file = is_file_lru(lru); |
1a93be0e KK |
1391 | struct zone *zone = lruvec_zone(lruvec); |
1392 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
78dc583d | 1393 | |
35cd7815 | 1394 | while (unlikely(too_many_isolated(zone, file, sc))) { |
58355c78 | 1395 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1396 | |
1397 | /* We are about to die and free our memory. Return now. */ | |
1398 | if (fatal_signal_pending(current)) | |
1399 | return SWAP_CLUSTER_MAX; | |
1400 | } | |
1401 | ||
1da177e4 | 1402 | lru_add_drain(); |
f80c0673 MK |
1403 | |
1404 | if (!sc->may_unmap) | |
61317289 | 1405 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1406 | if (!sc->may_writepage) |
61317289 | 1407 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1408 | |
1da177e4 | 1409 | spin_lock_irq(&zone->lru_lock); |
b35ea17b | 1410 | |
5dc35979 KK |
1411 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1412 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc KK |
1413 | |
1414 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1415 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1416 | ||
89b5fae5 | 1417 | if (global_reclaim(sc)) { |
e247dbce KM |
1418 | zone->pages_scanned += nr_scanned; |
1419 | if (current_is_kswapd()) | |
75b00af7 | 1420 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); |
e247dbce | 1421 | else |
75b00af7 | 1422 | __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); |
e247dbce | 1423 | } |
d563c050 | 1424 | spin_unlock_irq(&zone->lru_lock); |
b35ea17b | 1425 | |
d563c050 | 1426 | if (nr_taken == 0) |
66635629 | 1427 | return 0; |
5ad333eb | 1428 | |
02c6de8d | 1429 | nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, |
8e950282 MG |
1430 | &nr_dirty, &nr_unqueued_dirty, &nr_congested, |
1431 | &nr_writeback, &nr_immediate, | |
1432 | false); | |
c661b078 | 1433 | |
3f79768f HD |
1434 | spin_lock_irq(&zone->lru_lock); |
1435 | ||
95d918fc | 1436 | reclaim_stat->recent_scanned[file] += nr_taken; |
d563c050 | 1437 | |
904249aa YH |
1438 | if (global_reclaim(sc)) { |
1439 | if (current_is_kswapd()) | |
1440 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1441 | nr_reclaimed); | |
1442 | else | |
1443 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1444 | nr_reclaimed); | |
1445 | } | |
a74609fa | 1446 | |
27ac81d8 | 1447 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1448 | |
95d918fc | 1449 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f HD |
1450 | |
1451 | spin_unlock_irq(&zone->lru_lock); | |
1452 | ||
1453 | free_hot_cold_page_list(&page_list, 1); | |
e11da5b4 | 1454 | |
92df3a72 MG |
1455 | /* |
1456 | * If reclaim is isolating dirty pages under writeback, it implies | |
1457 | * that the long-lived page allocation rate is exceeding the page | |
1458 | * laundering rate. Either the global limits are not being effective | |
1459 | * at throttling processes due to the page distribution throughout | |
1460 | * zones or there is heavy usage of a slow backing device. The | |
1461 | * only option is to throttle from reclaim context which is not ideal | |
1462 | * as there is no guarantee the dirtying process is throttled in the | |
1463 | * same way balance_dirty_pages() manages. | |
1464 | * | |
8e950282 MG |
1465 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1466 | * of pages under pages flagged for immediate reclaim and stall if any | |
1467 | * are encountered in the nr_immediate check below. | |
92df3a72 | 1468 | */ |
918fc718 | 1469 | if (nr_writeback && nr_writeback == nr_taken) |
283aba9f | 1470 | zone_set_flag(zone, ZONE_WRITEBACK); |
92df3a72 | 1471 | |
d43006d5 | 1472 | /* |
b1a6f21e MG |
1473 | * memcg will stall in page writeback so only consider forcibly |
1474 | * stalling for global reclaim | |
d43006d5 | 1475 | */ |
b1a6f21e | 1476 | if (global_reclaim(sc)) { |
8e950282 MG |
1477 | /* |
1478 | * Tag a zone as congested if all the dirty pages scanned were | |
1479 | * backed by a congested BDI and wait_iff_congested will stall. | |
1480 | */ | |
1481 | if (nr_dirty && nr_dirty == nr_congested) | |
1482 | zone_set_flag(zone, ZONE_CONGESTED); | |
1483 | ||
b1a6f21e MG |
1484 | /* |
1485 | * If dirty pages are scanned that are not queued for IO, it | |
1486 | * implies that flushers are not keeping up. In this case, flag | |
1487 | * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing | |
1488 | * pages from reclaim context. It will forcibly stall in the | |
1489 | * next check. | |
1490 | */ | |
1491 | if (nr_unqueued_dirty == nr_taken) | |
1492 | zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY); | |
1493 | ||
1494 | /* | |
1495 | * In addition, if kswapd scans pages marked marked for | |
1496 | * immediate reclaim and under writeback (nr_immediate), it | |
1497 | * implies that pages are cycling through the LRU faster than | |
1498 | * they are written so also forcibly stall. | |
1499 | */ | |
1500 | if (nr_unqueued_dirty == nr_taken || nr_immediate) | |
1501 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
e2be15f6 | 1502 | } |
d43006d5 | 1503 | |
8e950282 MG |
1504 | /* |
1505 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1506 | * is congested. Allow kswapd to continue until it starts encountering | |
1507 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1508 | */ | |
1509 | if (!sc->hibernation_mode && !current_is_kswapd()) | |
1510 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); | |
1511 | ||
e11da5b4 MG |
1512 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, |
1513 | zone_idx(zone), | |
1514 | nr_scanned, nr_reclaimed, | |
9e3b2f8c | 1515 | sc->priority, |
23b9da55 | 1516 | trace_shrink_flags(file)); |
05ff5137 | 1517 | return nr_reclaimed; |
1da177e4 LT |
1518 | } |
1519 | ||
1520 | /* | |
1521 | * This moves pages from the active list to the inactive list. | |
1522 | * | |
1523 | * We move them the other way if the page is referenced by one or more | |
1524 | * processes, from rmap. | |
1525 | * | |
1526 | * If the pages are mostly unmapped, the processing is fast and it is | |
1527 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1528 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1529 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1530 | * this, so instead we remove the pages from the LRU while processing them. | |
1531 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1532 | * nobody will play with that bit on a non-LRU page. | |
1533 | * | |
1534 | * The downside is that we have to touch page->_count against each page. | |
1535 | * But we had to alter page->flags anyway. | |
1536 | */ | |
1cfb419b | 1537 | |
fa9add64 | 1538 | static void move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1539 | struct list_head *list, |
2bcf8879 | 1540 | struct list_head *pages_to_free, |
3eb4140f WF |
1541 | enum lru_list lru) |
1542 | { | |
fa9add64 | 1543 | struct zone *zone = lruvec_zone(lruvec); |
3eb4140f | 1544 | unsigned long pgmoved = 0; |
3eb4140f | 1545 | struct page *page; |
fa9add64 | 1546 | int nr_pages; |
3eb4140f | 1547 | |
3eb4140f WF |
1548 | while (!list_empty(list)) { |
1549 | page = lru_to_page(list); | |
fa9add64 | 1550 | lruvec = mem_cgroup_page_lruvec(page, zone); |
3eb4140f WF |
1551 | |
1552 | VM_BUG_ON(PageLRU(page)); | |
1553 | SetPageLRU(page); | |
1554 | ||
fa9add64 HD |
1555 | nr_pages = hpage_nr_pages(page); |
1556 | mem_cgroup_update_lru_size(lruvec, lru, nr_pages); | |
925b7673 | 1557 | list_move(&page->lru, &lruvec->lists[lru]); |
fa9add64 | 1558 | pgmoved += nr_pages; |
3eb4140f | 1559 | |
2bcf8879 HD |
1560 | if (put_page_testzero(page)) { |
1561 | __ClearPageLRU(page); | |
1562 | __ClearPageActive(page); | |
fa9add64 | 1563 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1564 | |
1565 | if (unlikely(PageCompound(page))) { | |
1566 | spin_unlock_irq(&zone->lru_lock); | |
1567 | (*get_compound_page_dtor(page))(page); | |
1568 | spin_lock_irq(&zone->lru_lock); | |
1569 | } else | |
1570 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1571 | } |
1572 | } | |
1573 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1574 | if (!is_active_lru(lru)) | |
1575 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1576 | } | |
1cfb419b | 1577 | |
f626012d | 1578 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1579 | struct lruvec *lruvec, |
f16015fb | 1580 | struct scan_control *sc, |
9e3b2f8c | 1581 | enum lru_list lru) |
1da177e4 | 1582 | { |
44c241f1 | 1583 | unsigned long nr_taken; |
f626012d | 1584 | unsigned long nr_scanned; |
6fe6b7e3 | 1585 | unsigned long vm_flags; |
1da177e4 | 1586 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1587 | LIST_HEAD(l_active); |
b69408e8 | 1588 | LIST_HEAD(l_inactive); |
1da177e4 | 1589 | struct page *page; |
1a93be0e | 1590 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
44c241f1 | 1591 | unsigned long nr_rotated = 0; |
f3fd4a61 | 1592 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1593 | int file = is_file_lru(lru); |
1a93be0e | 1594 | struct zone *zone = lruvec_zone(lruvec); |
1da177e4 LT |
1595 | |
1596 | lru_add_drain(); | |
f80c0673 MK |
1597 | |
1598 | if (!sc->may_unmap) | |
61317289 | 1599 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1600 | if (!sc->may_writepage) |
61317289 | 1601 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1602 | |
1da177e4 | 1603 | spin_lock_irq(&zone->lru_lock); |
925b7673 | 1604 | |
5dc35979 KK |
1605 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1606 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1607 | if (global_reclaim(sc)) |
f626012d | 1608 | zone->pages_scanned += nr_scanned; |
89b5fae5 | 1609 | |
b7c46d15 | 1610 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1611 | |
f626012d | 1612 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); |
3cb99451 | 1613 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); |
a731286d | 1614 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
1da177e4 LT |
1615 | spin_unlock_irq(&zone->lru_lock); |
1616 | ||
1da177e4 LT |
1617 | while (!list_empty(&l_hold)) { |
1618 | cond_resched(); | |
1619 | page = lru_to_page(&l_hold); | |
1620 | list_del(&page->lru); | |
7e9cd484 | 1621 | |
39b5f29a | 1622 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
1623 | putback_lru_page(page); |
1624 | continue; | |
1625 | } | |
1626 | ||
cc715d99 MG |
1627 | if (unlikely(buffer_heads_over_limit)) { |
1628 | if (page_has_private(page) && trylock_page(page)) { | |
1629 | if (page_has_private(page)) | |
1630 | try_to_release_page(page, 0); | |
1631 | unlock_page(page); | |
1632 | } | |
1633 | } | |
1634 | ||
c3ac9a8a JW |
1635 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1636 | &vm_flags)) { | |
9992af10 | 1637 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1638 | /* |
1639 | * Identify referenced, file-backed active pages and | |
1640 | * give them one more trip around the active list. So | |
1641 | * that executable code get better chances to stay in | |
1642 | * memory under moderate memory pressure. Anon pages | |
1643 | * are not likely to be evicted by use-once streaming | |
1644 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1645 | * so we ignore them here. | |
1646 | */ | |
41e20983 | 1647 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1648 | list_add(&page->lru, &l_active); |
1649 | continue; | |
1650 | } | |
1651 | } | |
7e9cd484 | 1652 | |
5205e56e | 1653 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1654 | list_add(&page->lru, &l_inactive); |
1655 | } | |
1656 | ||
b555749a | 1657 | /* |
8cab4754 | 1658 | * Move pages back to the lru list. |
b555749a | 1659 | */ |
2a1dc509 | 1660 | spin_lock_irq(&zone->lru_lock); |
556adecb | 1661 | /* |
8cab4754 WF |
1662 | * Count referenced pages from currently used mappings as rotated, |
1663 | * even though only some of them are actually re-activated. This | |
1664 | * helps balance scan pressure between file and anonymous pages in | |
1665 | * get_scan_ratio. | |
7e9cd484 | 1666 | */ |
b7c46d15 | 1667 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1668 | |
fa9add64 HD |
1669 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
1670 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
a731286d | 1671 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
f8891e5e | 1672 | spin_unlock_irq(&zone->lru_lock); |
2bcf8879 HD |
1673 | |
1674 | free_hot_cold_page_list(&l_hold, 1); | |
1da177e4 LT |
1675 | } |
1676 | ||
74e3f3c3 | 1677 | #ifdef CONFIG_SWAP |
14797e23 | 1678 | static int inactive_anon_is_low_global(struct zone *zone) |
f89eb90e KM |
1679 | { |
1680 | unsigned long active, inactive; | |
1681 | ||
1682 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1683 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1684 | ||
1685 | if (inactive * zone->inactive_ratio < active) | |
1686 | return 1; | |
1687 | ||
1688 | return 0; | |
1689 | } | |
1690 | ||
14797e23 KM |
1691 | /** |
1692 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
c56d5c7d | 1693 | * @lruvec: LRU vector to check |
14797e23 KM |
1694 | * |
1695 | * Returns true if the zone does not have enough inactive anon pages, | |
1696 | * meaning some active anon pages need to be deactivated. | |
1697 | */ | |
c56d5c7d | 1698 | static int inactive_anon_is_low(struct lruvec *lruvec) |
14797e23 | 1699 | { |
74e3f3c3 MK |
1700 | /* |
1701 | * If we don't have swap space, anonymous page deactivation | |
1702 | * is pointless. | |
1703 | */ | |
1704 | if (!total_swap_pages) | |
1705 | return 0; | |
1706 | ||
c3c787e8 | 1707 | if (!mem_cgroup_disabled()) |
c56d5c7d | 1708 | return mem_cgroup_inactive_anon_is_low(lruvec); |
f16015fb | 1709 | |
c56d5c7d | 1710 | return inactive_anon_is_low_global(lruvec_zone(lruvec)); |
14797e23 | 1711 | } |
74e3f3c3 | 1712 | #else |
c56d5c7d | 1713 | static inline int inactive_anon_is_low(struct lruvec *lruvec) |
74e3f3c3 MK |
1714 | { |
1715 | return 0; | |
1716 | } | |
1717 | #endif | |
14797e23 | 1718 | |
56e49d21 RR |
1719 | /** |
1720 | * inactive_file_is_low - check if file pages need to be deactivated | |
c56d5c7d | 1721 | * @lruvec: LRU vector to check |
56e49d21 RR |
1722 | * |
1723 | * When the system is doing streaming IO, memory pressure here | |
1724 | * ensures that active file pages get deactivated, until more | |
1725 | * than half of the file pages are on the inactive list. | |
1726 | * | |
1727 | * Once we get to that situation, protect the system's working | |
1728 | * set from being evicted by disabling active file page aging. | |
1729 | * | |
1730 | * This uses a different ratio than the anonymous pages, because | |
1731 | * the page cache uses a use-once replacement algorithm. | |
1732 | */ | |
c56d5c7d | 1733 | static int inactive_file_is_low(struct lruvec *lruvec) |
56e49d21 | 1734 | { |
e3790144 JW |
1735 | unsigned long inactive; |
1736 | unsigned long active; | |
1737 | ||
1738 | inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
1739 | active = get_lru_size(lruvec, LRU_ACTIVE_FILE); | |
56e49d21 | 1740 | |
e3790144 | 1741 | return active > inactive; |
56e49d21 RR |
1742 | } |
1743 | ||
75b00af7 | 1744 | static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) |
b39415b2 | 1745 | { |
75b00af7 | 1746 | if (is_file_lru(lru)) |
c56d5c7d | 1747 | return inactive_file_is_low(lruvec); |
b39415b2 | 1748 | else |
c56d5c7d | 1749 | return inactive_anon_is_low(lruvec); |
b39415b2 RR |
1750 | } |
1751 | ||
4f98a2fe | 1752 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
1a93be0e | 1753 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 1754 | { |
b39415b2 | 1755 | if (is_active_lru(lru)) { |
75b00af7 | 1756 | if (inactive_list_is_low(lruvec, lru)) |
1a93be0e | 1757 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
1758 | return 0; |
1759 | } | |
1760 | ||
1a93be0e | 1761 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
1762 | } |
1763 | ||
3d58ab5c | 1764 | static int vmscan_swappiness(struct scan_control *sc) |
1f4c025b | 1765 | { |
89b5fae5 | 1766 | if (global_reclaim(sc)) |
1f4c025b | 1767 | return vm_swappiness; |
3d58ab5c | 1768 | return mem_cgroup_swappiness(sc->target_mem_cgroup); |
1f4c025b KH |
1769 | } |
1770 | ||
9a265114 JW |
1771 | enum scan_balance { |
1772 | SCAN_EQUAL, | |
1773 | SCAN_FRACT, | |
1774 | SCAN_ANON, | |
1775 | SCAN_FILE, | |
1776 | }; | |
1777 | ||
4f98a2fe RR |
1778 | /* |
1779 | * Determine how aggressively the anon and file LRU lists should be | |
1780 | * scanned. The relative value of each set of LRU lists is determined | |
1781 | * by looking at the fraction of the pages scanned we did rotate back | |
1782 | * onto the active list instead of evict. | |
1783 | * | |
be7bd59d WL |
1784 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
1785 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 1786 | */ |
90126375 | 1787 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
9e3b2f8c | 1788 | unsigned long *nr) |
4f98a2fe | 1789 | { |
9a265114 JW |
1790 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1791 | u64 fraction[2]; | |
1792 | u64 denominator = 0; /* gcc */ | |
1793 | struct zone *zone = lruvec_zone(lruvec); | |
4f98a2fe | 1794 | unsigned long anon_prio, file_prio; |
9a265114 JW |
1795 | enum scan_balance scan_balance; |
1796 | unsigned long anon, file, free; | |
1797 | bool force_scan = false; | |
4f98a2fe | 1798 | unsigned long ap, fp; |
4111304d | 1799 | enum lru_list lru; |
246e87a9 | 1800 | |
f11c0ca5 JW |
1801 | /* |
1802 | * If the zone or memcg is small, nr[l] can be 0. This | |
1803 | * results in no scanning on this priority and a potential | |
1804 | * priority drop. Global direct reclaim can go to the next | |
1805 | * zone and tends to have no problems. Global kswapd is for | |
1806 | * zone balancing and it needs to scan a minimum amount. When | |
1807 | * reclaiming for a memcg, a priority drop can cause high | |
1808 | * latencies, so it's better to scan a minimum amount there as | |
1809 | * well. | |
1810 | */ | |
6e543d57 | 1811 | if (current_is_kswapd() && !zone_reclaimable(zone)) |
a4d3e9e7 | 1812 | force_scan = true; |
89b5fae5 | 1813 | if (!global_reclaim(sc)) |
a4d3e9e7 | 1814 | force_scan = true; |
76a33fc3 SL |
1815 | |
1816 | /* If we have no swap space, do not bother scanning anon pages. */ | |
ec8acf20 | 1817 | if (!sc->may_swap || (get_nr_swap_pages() <= 0)) { |
9a265114 | 1818 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
1819 | goto out; |
1820 | } | |
4f98a2fe | 1821 | |
10316b31 JW |
1822 | /* |
1823 | * Global reclaim will swap to prevent OOM even with no | |
1824 | * swappiness, but memcg users want to use this knob to | |
1825 | * disable swapping for individual groups completely when | |
1826 | * using the memory controller's swap limit feature would be | |
1827 | * too expensive. | |
1828 | */ | |
1829 | if (!global_reclaim(sc) && !vmscan_swappiness(sc)) { | |
9a265114 | 1830 | scan_balance = SCAN_FILE; |
10316b31 JW |
1831 | goto out; |
1832 | } | |
1833 | ||
1834 | /* | |
1835 | * Do not apply any pressure balancing cleverness when the | |
1836 | * system is close to OOM, scan both anon and file equally | |
1837 | * (unless the swappiness setting disagrees with swapping). | |
1838 | */ | |
1839 | if (!sc->priority && vmscan_swappiness(sc)) { | |
9a265114 | 1840 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
1841 | goto out; |
1842 | } | |
1843 | ||
4d7dcca2 HD |
1844 | anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + |
1845 | get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
1846 | file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
1847 | get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
a4d3e9e7 | 1848 | |
11d16c25 JW |
1849 | /* |
1850 | * If it's foreseeable that reclaiming the file cache won't be | |
1851 | * enough to get the zone back into a desirable shape, we have | |
1852 | * to swap. Better start now and leave the - probably heavily | |
1853 | * thrashing - remaining file pages alone. | |
1854 | */ | |
89b5fae5 | 1855 | if (global_reclaim(sc)) { |
11d16c25 | 1856 | free = zone_page_state(zone, NR_FREE_PAGES); |
90126375 | 1857 | if (unlikely(file + free <= high_wmark_pages(zone))) { |
9a265114 | 1858 | scan_balance = SCAN_ANON; |
76a33fc3 | 1859 | goto out; |
eeee9a8c | 1860 | } |
4f98a2fe RR |
1861 | } |
1862 | ||
7c5bd705 JW |
1863 | /* |
1864 | * There is enough inactive page cache, do not reclaim | |
1865 | * anything from the anonymous working set right now. | |
1866 | */ | |
1867 | if (!inactive_file_is_low(lruvec)) { | |
9a265114 | 1868 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
1869 | goto out; |
1870 | } | |
1871 | ||
9a265114 JW |
1872 | scan_balance = SCAN_FRACT; |
1873 | ||
58c37f6e KM |
1874 | /* |
1875 | * With swappiness at 100, anonymous and file have the same priority. | |
1876 | * This scanning priority is essentially the inverse of IO cost. | |
1877 | */ | |
3d58ab5c | 1878 | anon_prio = vmscan_swappiness(sc); |
75b00af7 | 1879 | file_prio = 200 - anon_prio; |
58c37f6e | 1880 | |
4f98a2fe RR |
1881 | /* |
1882 | * OK, so we have swap space and a fair amount of page cache | |
1883 | * pages. We use the recently rotated / recently scanned | |
1884 | * ratios to determine how valuable each cache is. | |
1885 | * | |
1886 | * Because workloads change over time (and to avoid overflow) | |
1887 | * we keep these statistics as a floating average, which ends | |
1888 | * up weighing recent references more than old ones. | |
1889 | * | |
1890 | * anon in [0], file in [1] | |
1891 | */ | |
90126375 | 1892 | spin_lock_irq(&zone->lru_lock); |
6e901571 | 1893 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
1894 | reclaim_stat->recent_scanned[0] /= 2; |
1895 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
1896 | } |
1897 | ||
6e901571 | 1898 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
1899 | reclaim_stat->recent_scanned[1] /= 2; |
1900 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
1901 | } |
1902 | ||
4f98a2fe | 1903 | /* |
00d8089c RR |
1904 | * The amount of pressure on anon vs file pages is inversely |
1905 | * proportional to the fraction of recently scanned pages on | |
1906 | * each list that were recently referenced and in active use. | |
4f98a2fe | 1907 | */ |
fe35004f | 1908 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 1909 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 1910 | |
fe35004f | 1911 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 1912 | fp /= reclaim_stat->recent_rotated[1] + 1; |
90126375 | 1913 | spin_unlock_irq(&zone->lru_lock); |
4f98a2fe | 1914 | |
76a33fc3 SL |
1915 | fraction[0] = ap; |
1916 | fraction[1] = fp; | |
1917 | denominator = ap + fp + 1; | |
1918 | out: | |
4111304d HD |
1919 | for_each_evictable_lru(lru) { |
1920 | int file = is_file_lru(lru); | |
d778df51 | 1921 | unsigned long size; |
76a33fc3 | 1922 | unsigned long scan; |
6e08a369 | 1923 | |
d778df51 | 1924 | size = get_lru_size(lruvec, lru); |
10316b31 | 1925 | scan = size >> sc->priority; |
9a265114 | 1926 | |
10316b31 JW |
1927 | if (!scan && force_scan) |
1928 | scan = min(size, SWAP_CLUSTER_MAX); | |
9a265114 JW |
1929 | |
1930 | switch (scan_balance) { | |
1931 | case SCAN_EQUAL: | |
1932 | /* Scan lists relative to size */ | |
1933 | break; | |
1934 | case SCAN_FRACT: | |
1935 | /* | |
1936 | * Scan types proportional to swappiness and | |
1937 | * their relative recent reclaim efficiency. | |
1938 | */ | |
1939 | scan = div64_u64(scan * fraction[file], denominator); | |
1940 | break; | |
1941 | case SCAN_FILE: | |
1942 | case SCAN_ANON: | |
1943 | /* Scan one type exclusively */ | |
1944 | if ((scan_balance == SCAN_FILE) != file) | |
1945 | scan = 0; | |
1946 | break; | |
1947 | default: | |
1948 | /* Look ma, no brain */ | |
1949 | BUG(); | |
1950 | } | |
4111304d | 1951 | nr[lru] = scan; |
76a33fc3 | 1952 | } |
6e08a369 | 1953 | } |
4f98a2fe | 1954 | |
9b4f98cd JW |
1955 | /* |
1956 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1957 | */ | |
1958 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) | |
1959 | { | |
1960 | unsigned long nr[NR_LRU_LISTS]; | |
e82e0561 | 1961 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
1962 | unsigned long nr_to_scan; |
1963 | enum lru_list lru; | |
1964 | unsigned long nr_reclaimed = 0; | |
1965 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
1966 | struct blk_plug plug; | |
e82e0561 | 1967 | bool scan_adjusted = false; |
9b4f98cd JW |
1968 | |
1969 | get_scan_count(lruvec, sc, nr); | |
1970 | ||
e82e0561 MG |
1971 | /* Record the original scan target for proportional adjustments later */ |
1972 | memcpy(targets, nr, sizeof(nr)); | |
1973 | ||
9b4f98cd JW |
1974 | blk_start_plug(&plug); |
1975 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
1976 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
1977 | unsigned long nr_anon, nr_file, percentage; |
1978 | unsigned long nr_scanned; | |
1979 | ||
9b4f98cd JW |
1980 | for_each_evictable_lru(lru) { |
1981 | if (nr[lru]) { | |
1982 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
1983 | nr[lru] -= nr_to_scan; | |
1984 | ||
1985 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
1986 | lruvec, sc); | |
1987 | } | |
1988 | } | |
e82e0561 MG |
1989 | |
1990 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) | |
1991 | continue; | |
1992 | ||
9b4f98cd | 1993 | /* |
e82e0561 MG |
1994 | * For global direct reclaim, reclaim only the number of pages |
1995 | * requested. Less care is taken to scan proportionally as it | |
1996 | * is more important to minimise direct reclaim stall latency | |
1997 | * than it is to properly age the LRU lists. | |
9b4f98cd | 1998 | */ |
e82e0561 | 1999 | if (global_reclaim(sc) && !current_is_kswapd()) |
9b4f98cd | 2000 | break; |
e82e0561 MG |
2001 | |
2002 | /* | |
2003 | * For kswapd and memcg, reclaim at least the number of pages | |
2004 | * requested. Ensure that the anon and file LRUs shrink | |
2005 | * proportionally what was requested by get_scan_count(). We | |
2006 | * stop reclaiming one LRU and reduce the amount scanning | |
2007 | * proportional to the original scan target. | |
2008 | */ | |
2009 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2010 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2011 | ||
2012 | if (nr_file > nr_anon) { | |
2013 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2014 | targets[LRU_ACTIVE_ANON] + 1; | |
2015 | lru = LRU_BASE; | |
2016 | percentage = nr_anon * 100 / scan_target; | |
2017 | } else { | |
2018 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2019 | targets[LRU_ACTIVE_FILE] + 1; | |
2020 | lru = LRU_FILE; | |
2021 | percentage = nr_file * 100 / scan_target; | |
2022 | } | |
2023 | ||
2024 | /* Stop scanning the smaller of the LRU */ | |
2025 | nr[lru] = 0; | |
2026 | nr[lru + LRU_ACTIVE] = 0; | |
2027 | ||
2028 | /* | |
2029 | * Recalculate the other LRU scan count based on its original | |
2030 | * scan target and the percentage scanning already complete | |
2031 | */ | |
2032 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2033 | nr_scanned = targets[lru] - nr[lru]; | |
2034 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2035 | nr[lru] -= min(nr[lru], nr_scanned); | |
2036 | ||
2037 | lru += LRU_ACTIVE; | |
2038 | nr_scanned = targets[lru] - nr[lru]; | |
2039 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2040 | nr[lru] -= min(nr[lru], nr_scanned); | |
2041 | ||
2042 | scan_adjusted = true; | |
9b4f98cd JW |
2043 | } |
2044 | blk_finish_plug(&plug); | |
2045 | sc->nr_reclaimed += nr_reclaimed; | |
2046 | ||
2047 | /* | |
2048 | * Even if we did not try to evict anon pages at all, we want to | |
2049 | * rebalance the anon lru active/inactive ratio. | |
2050 | */ | |
2051 | if (inactive_anon_is_low(lruvec)) | |
2052 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
2053 | sc, LRU_ACTIVE_ANON); | |
2054 | ||
2055 | throttle_vm_writeout(sc->gfp_mask); | |
2056 | } | |
2057 | ||
23b9da55 | 2058 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2059 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2060 | { |
d84da3f9 | 2061 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2062 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2063 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2064 | return true; |
2065 | ||
2066 | return false; | |
2067 | } | |
2068 | ||
3e7d3449 | 2069 | /* |
23b9da55 MG |
2070 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2071 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2072 | * true if more pages should be reclaimed such that when the page allocator | |
2073 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2074 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2075 | */ |
9b4f98cd | 2076 | static inline bool should_continue_reclaim(struct zone *zone, |
3e7d3449 MG |
2077 | unsigned long nr_reclaimed, |
2078 | unsigned long nr_scanned, | |
2079 | struct scan_control *sc) | |
2080 | { | |
2081 | unsigned long pages_for_compaction; | |
2082 | unsigned long inactive_lru_pages; | |
2083 | ||
2084 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2085 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2086 | return false; |
2087 | ||
2876592f MG |
2088 | /* Consider stopping depending on scan and reclaim activity */ |
2089 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2090 | /* | |
2091 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2092 | * full LRU list has been scanned and we are still failing | |
2093 | * to reclaim pages. This full LRU scan is potentially | |
2094 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2095 | */ | |
2096 | if (!nr_reclaimed && !nr_scanned) | |
2097 | return false; | |
2098 | } else { | |
2099 | /* | |
2100 | * For non-__GFP_REPEAT allocations which can presumably | |
2101 | * fail without consequence, stop if we failed to reclaim | |
2102 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2103 | * pages that were scanned. This will return to the | |
2104 | * caller faster at the risk reclaim/compaction and | |
2105 | * the resulting allocation attempt fails | |
2106 | */ | |
2107 | if (!nr_reclaimed) | |
2108 | return false; | |
2109 | } | |
3e7d3449 MG |
2110 | |
2111 | /* | |
2112 | * If we have not reclaimed enough pages for compaction and the | |
2113 | * inactive lists are large enough, continue reclaiming | |
2114 | */ | |
2115 | pages_for_compaction = (2UL << sc->order); | |
9b4f98cd | 2116 | inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); |
ec8acf20 | 2117 | if (get_nr_swap_pages() > 0) |
9b4f98cd | 2118 | inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); |
3e7d3449 MG |
2119 | if (sc->nr_reclaimed < pages_for_compaction && |
2120 | inactive_lru_pages > pages_for_compaction) | |
2121 | return true; | |
2122 | ||
2123 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
9b4f98cd | 2124 | switch (compaction_suitable(zone, sc->order)) { |
3e7d3449 MG |
2125 | case COMPACT_PARTIAL: |
2126 | case COMPACT_CONTINUE: | |
2127 | return false; | |
2128 | default: | |
2129 | return true; | |
2130 | } | |
2131 | } | |
2132 | ||
9b4f98cd | 2133 | static void shrink_zone(struct zone *zone, struct scan_control *sc) |
1da177e4 | 2134 | { |
f0fdc5e8 | 2135 | unsigned long nr_reclaimed, nr_scanned; |
1da177e4 | 2136 | |
9b4f98cd JW |
2137 | do { |
2138 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2139 | struct mem_cgroup_reclaim_cookie reclaim = { | |
2140 | .zone = zone, | |
2141 | .priority = sc->priority, | |
2142 | }; | |
2143 | struct mem_cgroup *memcg; | |
3e7d3449 | 2144 | |
9b4f98cd JW |
2145 | nr_reclaimed = sc->nr_reclaimed; |
2146 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2147 | |
9b4f98cd JW |
2148 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2149 | do { | |
2150 | struct lruvec *lruvec; | |
5660048c | 2151 | |
9b4f98cd | 2152 | lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
f9be23d6 | 2153 | |
9b4f98cd | 2154 | shrink_lruvec(lruvec, sc); |
f16015fb | 2155 | |
9b4f98cd | 2156 | /* |
a394cb8e MH |
2157 | * Direct reclaim and kswapd have to scan all memory |
2158 | * cgroups to fulfill the overall scan target for the | |
9b4f98cd | 2159 | * zone. |
a394cb8e MH |
2160 | * |
2161 | * Limit reclaim, on the other hand, only cares about | |
2162 | * nr_to_reclaim pages to be reclaimed and it will | |
2163 | * retry with decreasing priority if one round over the | |
2164 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2165 | */ |
a394cb8e MH |
2166 | if (!global_reclaim(sc) && |
2167 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2168 | mem_cgroup_iter_break(root, memcg); |
2169 | break; | |
2170 | } | |
2171 | memcg = mem_cgroup_iter(root, memcg, &reclaim); | |
2172 | } while (memcg); | |
70ddf637 AV |
2173 | |
2174 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, | |
2175 | sc->nr_scanned - nr_scanned, | |
2176 | sc->nr_reclaimed - nr_reclaimed); | |
2177 | ||
9b4f98cd JW |
2178 | } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, |
2179 | sc->nr_scanned - nr_scanned, sc)); | |
f16015fb JW |
2180 | } |
2181 | ||
fe4b1b24 MG |
2182 | /* Returns true if compaction should go ahead for a high-order request */ |
2183 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
2184 | { | |
2185 | unsigned long balance_gap, watermark; | |
2186 | bool watermark_ok; | |
2187 | ||
2188 | /* Do not consider compaction for orders reclaim is meant to satisfy */ | |
2189 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
2190 | return false; | |
2191 | ||
2192 | /* | |
2193 | * Compaction takes time to run and there are potentially other | |
2194 | * callers using the pages just freed. Continue reclaiming until | |
2195 | * there is a buffer of free pages available to give compaction | |
2196 | * a reasonable chance of completing and allocating the page | |
2197 | */ | |
2198 | balance_gap = min(low_wmark_pages(zone), | |
b40da049 | 2199 | (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / |
fe4b1b24 MG |
2200 | KSWAPD_ZONE_BALANCE_GAP_RATIO); |
2201 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order); | |
2202 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); | |
2203 | ||
2204 | /* | |
2205 | * If compaction is deferred, reclaim up to a point where | |
2206 | * compaction will have a chance of success when re-enabled | |
2207 | */ | |
aff62249 | 2208 | if (compaction_deferred(zone, sc->order)) |
fe4b1b24 MG |
2209 | return watermark_ok; |
2210 | ||
2211 | /* If compaction is not ready to start, keep reclaiming */ | |
2212 | if (!compaction_suitable(zone, sc->order)) | |
2213 | return false; | |
2214 | ||
2215 | return watermark_ok; | |
2216 | } | |
2217 | ||
1da177e4 LT |
2218 | /* |
2219 | * This is the direct reclaim path, for page-allocating processes. We only | |
2220 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2221 | * request. | |
2222 | * | |
41858966 MG |
2223 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
2224 | * Because: | |
1da177e4 LT |
2225 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
2226 | * allocation or | |
41858966 MG |
2227 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
2228 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2229 | * zone defense algorithm. | |
1da177e4 | 2230 | * |
1da177e4 LT |
2231 | * If a zone is deemed to be full of pinned pages then just give it a light |
2232 | * scan then give up on it. | |
e0c23279 MG |
2233 | * |
2234 | * This function returns true if a zone is being reclaimed for a costly | |
fe4b1b24 | 2235 | * high-order allocation and compaction is ready to begin. This indicates to |
0cee34fd MG |
2236 | * the caller that it should consider retrying the allocation instead of |
2237 | * further reclaim. | |
1da177e4 | 2238 | */ |
9e3b2f8c | 2239 | static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2240 | { |
dd1a239f | 2241 | struct zoneref *z; |
54a6eb5c | 2242 | struct zone *zone; |
d149e3b2 YH |
2243 | unsigned long nr_soft_reclaimed; |
2244 | unsigned long nr_soft_scanned; | |
0cee34fd | 2245 | bool aborted_reclaim = false; |
1cfb419b | 2246 | |
cc715d99 MG |
2247 | /* |
2248 | * If the number of buffer_heads in the machine exceeds the maximum | |
2249 | * allowed level, force direct reclaim to scan the highmem zone as | |
2250 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2251 | */ | |
2252 | if (buffer_heads_over_limit) | |
2253 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2254 | ||
d4debc66 MG |
2255 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
2256 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
f3fe6512 | 2257 | if (!populated_zone(zone)) |
1da177e4 | 2258 | continue; |
1cfb419b KH |
2259 | /* |
2260 | * Take care memory controller reclaiming has small influence | |
2261 | * to global LRU. | |
2262 | */ | |
89b5fae5 | 2263 | if (global_reclaim(sc)) { |
1cfb419b KH |
2264 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2265 | continue; | |
6e543d57 LD |
2266 | if (sc->priority != DEF_PRIORITY && |
2267 | !zone_reclaimable(zone)) | |
1cfb419b | 2268 | continue; /* Let kswapd poll it */ |
d84da3f9 | 2269 | if (IS_ENABLED(CONFIG_COMPACTION)) { |
e0887c19 | 2270 | /* |
e0c23279 MG |
2271 | * If we already have plenty of memory free for |
2272 | * compaction in this zone, don't free any more. | |
2273 | * Even though compaction is invoked for any | |
2274 | * non-zero order, only frequent costly order | |
2275 | * reclamation is disruptive enough to become a | |
c7cfa37b CA |
2276 | * noticeable problem, like transparent huge |
2277 | * page allocations. | |
e0887c19 | 2278 | */ |
fe4b1b24 | 2279 | if (compaction_ready(zone, sc)) { |
0cee34fd | 2280 | aborted_reclaim = true; |
e0887c19 | 2281 | continue; |
e0c23279 | 2282 | } |
e0887c19 | 2283 | } |
ac34a1a3 KH |
2284 | /* |
2285 | * This steals pages from memory cgroups over softlimit | |
2286 | * and returns the number of reclaimed pages and | |
2287 | * scanned pages. This works for global memory pressure | |
2288 | * and balancing, not for a memcg's limit. | |
2289 | */ | |
2290 | nr_soft_scanned = 0; | |
2291 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2292 | sc->order, sc->gfp_mask, | |
2293 | &nr_soft_scanned); | |
2294 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2295 | sc->nr_scanned += nr_soft_scanned; | |
2296 | /* need some check for avoid more shrink_zone() */ | |
1cfb419b | 2297 | } |
408d8544 | 2298 | |
9e3b2f8c | 2299 | shrink_zone(zone, sc); |
1da177e4 | 2300 | } |
e0c23279 | 2301 | |
0cee34fd | 2302 | return aborted_reclaim; |
d1908362 MK |
2303 | } |
2304 | ||
929bea7c | 2305 | /* All zones in zonelist are unreclaimable? */ |
d1908362 MK |
2306 | static bool all_unreclaimable(struct zonelist *zonelist, |
2307 | struct scan_control *sc) | |
2308 | { | |
2309 | struct zoneref *z; | |
2310 | struct zone *zone; | |
d1908362 MK |
2311 | |
2312 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2313 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2314 | if (!populated_zone(zone)) | |
2315 | continue; | |
2316 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2317 | continue; | |
6e543d57 | 2318 | if (zone_reclaimable(zone)) |
929bea7c | 2319 | return false; |
d1908362 MK |
2320 | } |
2321 | ||
929bea7c | 2322 | return true; |
1da177e4 | 2323 | } |
4f98a2fe | 2324 | |
1da177e4 LT |
2325 | /* |
2326 | * This is the main entry point to direct page reclaim. | |
2327 | * | |
2328 | * If a full scan of the inactive list fails to free enough memory then we | |
2329 | * are "out of memory" and something needs to be killed. | |
2330 | * | |
2331 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2332 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2333 | * caller can't do much about. We kick the writeback threads and take explicit |
2334 | * naps in the hope that some of these pages can be written. But if the | |
2335 | * allocating task holds filesystem locks which prevent writeout this might not | |
2336 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2337 | * |
2338 | * returns: 0, if no pages reclaimed | |
2339 | * else, the number of pages reclaimed | |
1da177e4 | 2340 | */ |
dac1d27b | 2341 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
a09ed5e0 YH |
2342 | struct scan_control *sc, |
2343 | struct shrink_control *shrink) | |
1da177e4 | 2344 | { |
69e05944 | 2345 | unsigned long total_scanned = 0; |
1da177e4 | 2346 | struct reclaim_state *reclaim_state = current->reclaim_state; |
dd1a239f | 2347 | struct zoneref *z; |
54a6eb5c | 2348 | struct zone *zone; |
22fba335 | 2349 | unsigned long writeback_threshold; |
0cee34fd | 2350 | bool aborted_reclaim; |
1da177e4 | 2351 | |
873b4771 KK |
2352 | delayacct_freepages_start(); |
2353 | ||
89b5fae5 | 2354 | if (global_reclaim(sc)) |
1cfb419b | 2355 | count_vm_event(ALLOCSTALL); |
1da177e4 | 2356 | |
9e3b2f8c | 2357 | do { |
70ddf637 AV |
2358 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2359 | sc->priority); | |
66e1707b | 2360 | sc->nr_scanned = 0; |
9e3b2f8c | 2361 | aborted_reclaim = shrink_zones(zonelist, sc); |
e0c23279 | 2362 | |
66e1707b | 2363 | /* |
5a1c9cbc MG |
2364 | * Don't shrink slabs when reclaiming memory from over limit |
2365 | * cgroups but do shrink slab at least once when aborting | |
2366 | * reclaim for compaction to avoid unevenly scanning file/anon | |
2367 | * LRU pages over slab pages. | |
66e1707b | 2368 | */ |
89b5fae5 | 2369 | if (global_reclaim(sc)) { |
c6a8a8c5 | 2370 | unsigned long lru_pages = 0; |
d4debc66 MG |
2371 | for_each_zone_zonelist(zone, z, zonelist, |
2372 | gfp_zone(sc->gfp_mask)) { | |
c6a8a8c5 KM |
2373 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2374 | continue; | |
2375 | ||
2376 | lru_pages += zone_reclaimable_pages(zone); | |
2377 | } | |
2378 | ||
1495f230 | 2379 | shrink_slab(shrink, sc->nr_scanned, lru_pages); |
91a45470 | 2380 | if (reclaim_state) { |
a79311c1 | 2381 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
91a45470 KH |
2382 | reclaim_state->reclaimed_slab = 0; |
2383 | } | |
1da177e4 | 2384 | } |
66e1707b | 2385 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2386 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
1da177e4 | 2387 | goto out; |
1da177e4 | 2388 | |
0e50ce3b MK |
2389 | /* |
2390 | * If we're getting trouble reclaiming, start doing | |
2391 | * writepage even in laptop mode. | |
2392 | */ | |
2393 | if (sc->priority < DEF_PRIORITY - 2) | |
2394 | sc->may_writepage = 1; | |
2395 | ||
1da177e4 LT |
2396 | /* |
2397 | * Try to write back as many pages as we just scanned. This | |
2398 | * tends to cause slow streaming writers to write data to the | |
2399 | * disk smoothly, at the dirtying rate, which is nice. But | |
2400 | * that's undesirable in laptop mode, where we *want* lumpy | |
2401 | * writeout. So in laptop mode, write out the whole world. | |
2402 | */ | |
22fba335 KM |
2403 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2404 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2405 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2406 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2407 | sc->may_writepage = 1; |
1da177e4 | 2408 | } |
5a1c9cbc | 2409 | } while (--sc->priority >= 0 && !aborted_reclaim); |
bb21c7ce | 2410 | |
1da177e4 | 2411 | out: |
873b4771 KK |
2412 | delayacct_freepages_end(); |
2413 | ||
bb21c7ce KM |
2414 | if (sc->nr_reclaimed) |
2415 | return sc->nr_reclaimed; | |
2416 | ||
929bea7c KM |
2417 | /* |
2418 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2419 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2420 | * check. | |
2421 | */ | |
2422 | if (oom_killer_disabled) | |
2423 | return 0; | |
2424 | ||
0cee34fd MG |
2425 | /* Aborted reclaim to try compaction? don't OOM, then */ |
2426 | if (aborted_reclaim) | |
7335084d MG |
2427 | return 1; |
2428 | ||
bb21c7ce | 2429 | /* top priority shrink_zones still had more to do? don't OOM, then */ |
89b5fae5 | 2430 | if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc)) |
bb21c7ce KM |
2431 | return 1; |
2432 | ||
2433 | return 0; | |
1da177e4 LT |
2434 | } |
2435 | ||
5515061d MG |
2436 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) |
2437 | { | |
2438 | struct zone *zone; | |
2439 | unsigned long pfmemalloc_reserve = 0; | |
2440 | unsigned long free_pages = 0; | |
2441 | int i; | |
2442 | bool wmark_ok; | |
2443 | ||
2444 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2445 | zone = &pgdat->node_zones[i]; | |
2446 | pfmemalloc_reserve += min_wmark_pages(zone); | |
2447 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2448 | } | |
2449 | ||
2450 | wmark_ok = free_pages > pfmemalloc_reserve / 2; | |
2451 | ||
2452 | /* kswapd must be awake if processes are being throttled */ | |
2453 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2454 | pgdat->classzone_idx = min(pgdat->classzone_idx, | |
2455 | (enum zone_type)ZONE_NORMAL); | |
2456 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2457 | } | |
2458 | ||
2459 | return wmark_ok; | |
2460 | } | |
2461 | ||
2462 | /* | |
2463 | * Throttle direct reclaimers if backing storage is backed by the network | |
2464 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2465 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2466 | * when the low watermark is reached. |
2467 | * | |
2468 | * Returns true if a fatal signal was delivered during throttling. If this | |
2469 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2470 | */ |
50694c28 | 2471 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2472 | nodemask_t *nodemask) |
2473 | { | |
2474 | struct zone *zone; | |
2475 | int high_zoneidx = gfp_zone(gfp_mask); | |
2476 | pg_data_t *pgdat; | |
2477 | ||
2478 | /* | |
2479 | * Kernel threads should not be throttled as they may be indirectly | |
2480 | * responsible for cleaning pages necessary for reclaim to make forward | |
2481 | * progress. kjournald for example may enter direct reclaim while | |
2482 | * committing a transaction where throttling it could forcing other | |
2483 | * processes to block on log_wait_commit(). | |
2484 | */ | |
2485 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2486 | goto out; |
2487 | ||
2488 | /* | |
2489 | * If a fatal signal is pending, this process should not throttle. | |
2490 | * It should return quickly so it can exit and free its memory | |
2491 | */ | |
2492 | if (fatal_signal_pending(current)) | |
2493 | goto out; | |
5515061d MG |
2494 | |
2495 | /* Check if the pfmemalloc reserves are ok */ | |
2496 | first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone); | |
2497 | pgdat = zone->zone_pgdat; | |
2498 | if (pfmemalloc_watermark_ok(pgdat)) | |
50694c28 | 2499 | goto out; |
5515061d | 2500 | |
68243e76 MG |
2501 | /* Account for the throttling */ |
2502 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2503 | ||
5515061d MG |
2504 | /* |
2505 | * If the caller cannot enter the filesystem, it's possible that it | |
2506 | * is due to the caller holding an FS lock or performing a journal | |
2507 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2508 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2509 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2510 | * second before continuing. | |
2511 | */ | |
2512 | if (!(gfp_mask & __GFP_FS)) { | |
2513 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2514 | pfmemalloc_watermark_ok(pgdat), HZ); | |
50694c28 MG |
2515 | |
2516 | goto check_pending; | |
5515061d MG |
2517 | } |
2518 | ||
2519 | /* Throttle until kswapd wakes the process */ | |
2520 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2521 | pfmemalloc_watermark_ok(pgdat)); | |
50694c28 MG |
2522 | |
2523 | check_pending: | |
2524 | if (fatal_signal_pending(current)) | |
2525 | return true; | |
2526 | ||
2527 | out: | |
2528 | return false; | |
5515061d MG |
2529 | } |
2530 | ||
dac1d27b | 2531 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2532 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2533 | { |
33906bc5 | 2534 | unsigned long nr_reclaimed; |
66e1707b | 2535 | struct scan_control sc = { |
21caf2fc | 2536 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
66e1707b | 2537 | .may_writepage = !laptop_mode, |
22fba335 | 2538 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
a6dc60f8 | 2539 | .may_unmap = 1, |
2e2e4259 | 2540 | .may_swap = 1, |
66e1707b | 2541 | .order = order, |
9e3b2f8c | 2542 | .priority = DEF_PRIORITY, |
f16015fb | 2543 | .target_mem_cgroup = NULL, |
327c0e96 | 2544 | .nodemask = nodemask, |
66e1707b | 2545 | }; |
a09ed5e0 YH |
2546 | struct shrink_control shrink = { |
2547 | .gfp_mask = sc.gfp_mask, | |
2548 | }; | |
66e1707b | 2549 | |
5515061d | 2550 | /* |
50694c28 MG |
2551 | * Do not enter reclaim if fatal signal was delivered while throttled. |
2552 | * 1 is returned so that the page allocator does not OOM kill at this | |
2553 | * point. | |
5515061d | 2554 | */ |
50694c28 | 2555 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) |
5515061d MG |
2556 | return 1; |
2557 | ||
33906bc5 MG |
2558 | trace_mm_vmscan_direct_reclaim_begin(order, |
2559 | sc.may_writepage, | |
2560 | gfp_mask); | |
2561 | ||
a09ed5e0 | 2562 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
33906bc5 MG |
2563 | |
2564 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2565 | ||
2566 | return nr_reclaimed; | |
66e1707b BS |
2567 | } |
2568 | ||
c255a458 | 2569 | #ifdef CONFIG_MEMCG |
66e1707b | 2570 | |
72835c86 | 2571 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, |
4e416953 | 2572 | gfp_t gfp_mask, bool noswap, |
0ae5e89c YH |
2573 | struct zone *zone, |
2574 | unsigned long *nr_scanned) | |
4e416953 BS |
2575 | { |
2576 | struct scan_control sc = { | |
0ae5e89c | 2577 | .nr_scanned = 0, |
b8f5c566 | 2578 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
4e416953 BS |
2579 | .may_writepage = !laptop_mode, |
2580 | .may_unmap = 1, | |
2581 | .may_swap = !noswap, | |
4e416953 | 2582 | .order = 0, |
9e3b2f8c | 2583 | .priority = 0, |
72835c86 | 2584 | .target_mem_cgroup = memcg, |
4e416953 | 2585 | }; |
f9be23d6 | 2586 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
0ae5e89c | 2587 | |
4e416953 BS |
2588 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2589 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 2590 | |
9e3b2f8c | 2591 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e KM |
2592 | sc.may_writepage, |
2593 | sc.gfp_mask); | |
2594 | ||
4e416953 BS |
2595 | /* |
2596 | * NOTE: Although we can get the priority field, using it | |
2597 | * here is not a good idea, since it limits the pages we can scan. | |
2598 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2599 | * will pick up pages from other mem cgroup's as well. We hack | |
2600 | * the priority and make it zero. | |
2601 | */ | |
f9be23d6 | 2602 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
2603 | |
2604 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2605 | ||
0ae5e89c | 2606 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2607 | return sc.nr_reclaimed; |
2608 | } | |
2609 | ||
72835c86 | 2610 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
a7885eb8 | 2611 | gfp_t gfp_mask, |
185efc0f | 2612 | bool noswap) |
66e1707b | 2613 | { |
4e416953 | 2614 | struct zonelist *zonelist; |
bdce6d9e | 2615 | unsigned long nr_reclaimed; |
889976db | 2616 | int nid; |
66e1707b | 2617 | struct scan_control sc = { |
66e1707b | 2618 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2619 | .may_unmap = 1, |
2e2e4259 | 2620 | .may_swap = !noswap, |
22fba335 | 2621 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
66e1707b | 2622 | .order = 0, |
9e3b2f8c | 2623 | .priority = DEF_PRIORITY, |
72835c86 | 2624 | .target_mem_cgroup = memcg, |
327c0e96 | 2625 | .nodemask = NULL, /* we don't care the placement */ |
a09ed5e0 YH |
2626 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2627 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2628 | }; | |
2629 | struct shrink_control shrink = { | |
2630 | .gfp_mask = sc.gfp_mask, | |
66e1707b | 2631 | }; |
66e1707b | 2632 | |
889976db YH |
2633 | /* |
2634 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2635 | * take care of from where we get pages. So the node where we start the | |
2636 | * scan does not need to be the current node. | |
2637 | */ | |
72835c86 | 2638 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2639 | |
2640 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2641 | |
2642 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2643 | sc.may_writepage, | |
2644 | sc.gfp_mask); | |
2645 | ||
a09ed5e0 | 2646 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
bdce6d9e KM |
2647 | |
2648 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2649 | ||
2650 | return nr_reclaimed; | |
66e1707b BS |
2651 | } |
2652 | #endif | |
2653 | ||
9e3b2f8c | 2654 | static void age_active_anon(struct zone *zone, struct scan_control *sc) |
f16015fb | 2655 | { |
b95a2f2d | 2656 | struct mem_cgroup *memcg; |
f16015fb | 2657 | |
b95a2f2d JW |
2658 | if (!total_swap_pages) |
2659 | return; | |
2660 | ||
2661 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2662 | do { | |
c56d5c7d | 2663 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
b95a2f2d | 2664 | |
c56d5c7d | 2665 | if (inactive_anon_is_low(lruvec)) |
1a93be0e | 2666 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 2667 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
2668 | |
2669 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2670 | } while (memcg); | |
f16015fb JW |
2671 | } |
2672 | ||
60cefed4 JW |
2673 | static bool zone_balanced(struct zone *zone, int order, |
2674 | unsigned long balance_gap, int classzone_idx) | |
2675 | { | |
2676 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + | |
2677 | balance_gap, classzone_idx, 0)) | |
2678 | return false; | |
2679 | ||
d84da3f9 KS |
2680 | if (IS_ENABLED(CONFIG_COMPACTION) && order && |
2681 | !compaction_suitable(zone, order)) | |
60cefed4 JW |
2682 | return false; |
2683 | ||
2684 | return true; | |
2685 | } | |
2686 | ||
1741c877 | 2687 | /* |
4ae0a48b ZC |
2688 | * pgdat_balanced() is used when checking if a node is balanced. |
2689 | * | |
2690 | * For order-0, all zones must be balanced! | |
2691 | * | |
2692 | * For high-order allocations only zones that meet watermarks and are in a | |
2693 | * zone allowed by the callers classzone_idx are added to balanced_pages. The | |
2694 | * total of balanced pages must be at least 25% of the zones allowed by | |
2695 | * classzone_idx for the node to be considered balanced. Forcing all zones to | |
2696 | * be balanced for high orders can cause excessive reclaim when there are | |
2697 | * imbalanced zones. | |
1741c877 MG |
2698 | * The choice of 25% is due to |
2699 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2700 | * reasonable sized machine | |
2701 | * o On all other machines, the top zone must be at least a reasonable | |
25985edc | 2702 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
1741c877 MG |
2703 | * would need to be at least 256M for it to be balance a whole node. |
2704 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2705 | * to balance a node on its own. These seemed like reasonable ratios. | |
2706 | */ | |
4ae0a48b | 2707 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) |
1741c877 | 2708 | { |
b40da049 | 2709 | unsigned long managed_pages = 0; |
4ae0a48b | 2710 | unsigned long balanced_pages = 0; |
1741c877 MG |
2711 | int i; |
2712 | ||
4ae0a48b ZC |
2713 | /* Check the watermark levels */ |
2714 | for (i = 0; i <= classzone_idx; i++) { | |
2715 | struct zone *zone = pgdat->node_zones + i; | |
1741c877 | 2716 | |
4ae0a48b ZC |
2717 | if (!populated_zone(zone)) |
2718 | continue; | |
2719 | ||
b40da049 | 2720 | managed_pages += zone->managed_pages; |
4ae0a48b ZC |
2721 | |
2722 | /* | |
2723 | * A special case here: | |
2724 | * | |
2725 | * balance_pgdat() skips over all_unreclaimable after | |
2726 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2727 | * they must be considered balanced here as well! | |
2728 | */ | |
6e543d57 | 2729 | if (!zone_reclaimable(zone)) { |
b40da049 | 2730 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
2731 | continue; |
2732 | } | |
2733 | ||
2734 | if (zone_balanced(zone, order, 0, i)) | |
b40da049 | 2735 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
2736 | else if (!order) |
2737 | return false; | |
2738 | } | |
2739 | ||
2740 | if (order) | |
b40da049 | 2741 | return balanced_pages >= (managed_pages >> 2); |
4ae0a48b ZC |
2742 | else |
2743 | return true; | |
1741c877 MG |
2744 | } |
2745 | ||
5515061d MG |
2746 | /* |
2747 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
2748 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
2749 | * | |
2750 | * Returns true if kswapd is ready to sleep | |
2751 | */ | |
2752 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, | |
dc83edd9 | 2753 | int classzone_idx) |
f50de2d3 | 2754 | { |
f50de2d3 MG |
2755 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ |
2756 | if (remaining) | |
5515061d MG |
2757 | return false; |
2758 | ||
2759 | /* | |
2760 | * There is a potential race between when kswapd checks its watermarks | |
2761 | * and a process gets throttled. There is also a potential race if | |
2762 | * processes get throttled, kswapd wakes, a large process exits therby | |
2763 | * balancing the zones that causes kswapd to miss a wakeup. If kswapd | |
2764 | * is going to sleep, no process should be sleeping on pfmemalloc_wait | |
2765 | * so wake them now if necessary. If necessary, processes will wake | |
2766 | * kswapd and get throttled again | |
2767 | */ | |
2768 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) { | |
2769 | wake_up(&pgdat->pfmemalloc_wait); | |
2770 | return false; | |
2771 | } | |
f50de2d3 | 2772 | |
4ae0a48b | 2773 | return pgdat_balanced(pgdat, order, classzone_idx); |
f50de2d3 MG |
2774 | } |
2775 | ||
75485363 MG |
2776 | /* |
2777 | * kswapd shrinks the zone by the number of pages required to reach | |
2778 | * the high watermark. | |
b8e83b94 MG |
2779 | * |
2780 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
2781 | * reclaim or if the lack of progress was due to pages under writeback. |
2782 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 2783 | */ |
b8e83b94 | 2784 | static bool kswapd_shrink_zone(struct zone *zone, |
7c954f6d | 2785 | int classzone_idx, |
75485363 | 2786 | struct scan_control *sc, |
2ab44f43 MG |
2787 | unsigned long lru_pages, |
2788 | unsigned long *nr_attempted) | |
75485363 | 2789 | { |
7c954f6d MG |
2790 | int testorder = sc->order; |
2791 | unsigned long balance_gap; | |
75485363 MG |
2792 | struct reclaim_state *reclaim_state = current->reclaim_state; |
2793 | struct shrink_control shrink = { | |
2794 | .gfp_mask = sc->gfp_mask, | |
2795 | }; | |
7c954f6d | 2796 | bool lowmem_pressure; |
75485363 MG |
2797 | |
2798 | /* Reclaim above the high watermark. */ | |
2799 | sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone)); | |
7c954f6d MG |
2800 | |
2801 | /* | |
2802 | * Kswapd reclaims only single pages with compaction enabled. Trying | |
2803 | * too hard to reclaim until contiguous free pages have become | |
2804 | * available can hurt performance by evicting too much useful data | |
2805 | * from memory. Do not reclaim more than needed for compaction. | |
2806 | */ | |
2807 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && | |
2808 | compaction_suitable(zone, sc->order) != | |
2809 | COMPACT_SKIPPED) | |
2810 | testorder = 0; | |
2811 | ||
2812 | /* | |
2813 | * We put equal pressure on every zone, unless one zone has way too | |
2814 | * many pages free already. The "too many pages" is defined as the | |
2815 | * high wmark plus a "gap" where the gap is either the low | |
2816 | * watermark or 1% of the zone, whichever is smaller. | |
2817 | */ | |
2818 | balance_gap = min(low_wmark_pages(zone), | |
2819 | (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2820 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
2821 | ||
2822 | /* | |
2823 | * If there is no low memory pressure or the zone is balanced then no | |
2824 | * reclaim is necessary | |
2825 | */ | |
2826 | lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone)); | |
2827 | if (!lowmem_pressure && zone_balanced(zone, testorder, | |
2828 | balance_gap, classzone_idx)) | |
2829 | return true; | |
2830 | ||
75485363 MG |
2831 | shrink_zone(zone, sc); |
2832 | ||
2833 | reclaim_state->reclaimed_slab = 0; | |
6e543d57 | 2834 | shrink_slab(&shrink, sc->nr_scanned, lru_pages); |
75485363 MG |
2835 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
2836 | ||
2ab44f43 MG |
2837 | /* Account for the number of pages attempted to reclaim */ |
2838 | *nr_attempted += sc->nr_to_reclaim; | |
2839 | ||
283aba9f MG |
2840 | zone_clear_flag(zone, ZONE_WRITEBACK); |
2841 | ||
7c954f6d MG |
2842 | /* |
2843 | * If a zone reaches its high watermark, consider it to be no longer | |
2844 | * congested. It's possible there are dirty pages backed by congested | |
2845 | * BDIs but as pressure is relieved, speculatively avoid congestion | |
2846 | * waits. | |
2847 | */ | |
6e543d57 | 2848 | if (zone_reclaimable(zone) && |
7c954f6d MG |
2849 | zone_balanced(zone, testorder, 0, classzone_idx)) { |
2850 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2851 | zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY); | |
2852 | } | |
2853 | ||
b8e83b94 | 2854 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
2855 | } |
2856 | ||
1da177e4 LT |
2857 | /* |
2858 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 2859 | * they are all at high_wmark_pages(zone). |
1da177e4 | 2860 | * |
0abdee2b | 2861 | * Returns the final order kswapd was reclaiming at |
1da177e4 LT |
2862 | * |
2863 | * There is special handling here for zones which are full of pinned pages. | |
2864 | * This can happen if the pages are all mlocked, or if they are all used by | |
2865 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
2866 | * What we do is to detect the case where all pages in the zone have been | |
2867 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
2868 | * dead and from now on, only perform a short scan. Basically we're polling | |
2869 | * the zone for when the problem goes away. | |
2870 | * | |
2871 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
2872 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
2873 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
2874 | * lower zones regardless of the number of free pages in the lower zones. This | |
2875 | * interoperates with the page allocator fallback scheme to ensure that aging | |
2876 | * of pages is balanced across the zones. | |
1da177e4 | 2877 | */ |
99504748 | 2878 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, |
dc83edd9 | 2879 | int *classzone_idx) |
1da177e4 | 2880 | { |
1da177e4 | 2881 | int i; |
99504748 | 2882 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
0ae5e89c YH |
2883 | unsigned long nr_soft_reclaimed; |
2884 | unsigned long nr_soft_scanned; | |
179e9639 AM |
2885 | struct scan_control sc = { |
2886 | .gfp_mask = GFP_KERNEL, | |
b8e83b94 | 2887 | .priority = DEF_PRIORITY, |
a6dc60f8 | 2888 | .may_unmap = 1, |
2e2e4259 | 2889 | .may_swap = 1, |
b8e83b94 | 2890 | .may_writepage = !laptop_mode, |
5ad333eb | 2891 | .order = order, |
f16015fb | 2892 | .target_mem_cgroup = NULL, |
179e9639 | 2893 | }; |
f8891e5e | 2894 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 2895 | |
9e3b2f8c | 2896 | do { |
1da177e4 | 2897 | unsigned long lru_pages = 0; |
2ab44f43 | 2898 | unsigned long nr_attempted = 0; |
b8e83b94 | 2899 | bool raise_priority = true; |
2ab44f43 | 2900 | bool pgdat_needs_compaction = (order > 0); |
b8e83b94 MG |
2901 | |
2902 | sc.nr_reclaimed = 0; | |
1da177e4 | 2903 | |
d6277db4 RW |
2904 | /* |
2905 | * Scan in the highmem->dma direction for the highest | |
2906 | * zone which needs scanning | |
2907 | */ | |
2908 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
2909 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 2910 | |
d6277db4 RW |
2911 | if (!populated_zone(zone)) |
2912 | continue; | |
1da177e4 | 2913 | |
6e543d57 LD |
2914 | if (sc.priority != DEF_PRIORITY && |
2915 | !zone_reclaimable(zone)) | |
d6277db4 | 2916 | continue; |
1da177e4 | 2917 | |
556adecb RR |
2918 | /* |
2919 | * Do some background aging of the anon list, to give | |
2920 | * pages a chance to be referenced before reclaiming. | |
2921 | */ | |
9e3b2f8c | 2922 | age_active_anon(zone, &sc); |
556adecb | 2923 | |
cc715d99 MG |
2924 | /* |
2925 | * If the number of buffer_heads in the machine | |
2926 | * exceeds the maximum allowed level and this node | |
2927 | * has a highmem zone, force kswapd to reclaim from | |
2928 | * it to relieve lowmem pressure. | |
2929 | */ | |
2930 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
2931 | end_zone = i; | |
2932 | break; | |
2933 | } | |
2934 | ||
60cefed4 | 2935 | if (!zone_balanced(zone, order, 0, 0)) { |
d6277db4 | 2936 | end_zone = i; |
e1dbeda6 | 2937 | break; |
439423f6 | 2938 | } else { |
d43006d5 MG |
2939 | /* |
2940 | * If balanced, clear the dirty and congested | |
2941 | * flags | |
2942 | */ | |
439423f6 | 2943 | zone_clear_flag(zone, ZONE_CONGESTED); |
d43006d5 | 2944 | zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY); |
1da177e4 | 2945 | } |
1da177e4 | 2946 | } |
dafcb73e | 2947 | |
b8e83b94 | 2948 | if (i < 0) |
e1dbeda6 AM |
2949 | goto out; |
2950 | ||
1da177e4 LT |
2951 | for (i = 0; i <= end_zone; i++) { |
2952 | struct zone *zone = pgdat->node_zones + i; | |
2953 | ||
2ab44f43 MG |
2954 | if (!populated_zone(zone)) |
2955 | continue; | |
2956 | ||
adea02a1 | 2957 | lru_pages += zone_reclaimable_pages(zone); |
2ab44f43 MG |
2958 | |
2959 | /* | |
2960 | * If any zone is currently balanced then kswapd will | |
2961 | * not call compaction as it is expected that the | |
2962 | * necessary pages are already available. | |
2963 | */ | |
2964 | if (pgdat_needs_compaction && | |
2965 | zone_watermark_ok(zone, order, | |
2966 | low_wmark_pages(zone), | |
2967 | *classzone_idx, 0)) | |
2968 | pgdat_needs_compaction = false; | |
1da177e4 LT |
2969 | } |
2970 | ||
b7ea3c41 MG |
2971 | /* |
2972 | * If we're getting trouble reclaiming, start doing writepage | |
2973 | * even in laptop mode. | |
2974 | */ | |
2975 | if (sc.priority < DEF_PRIORITY - 2) | |
2976 | sc.may_writepage = 1; | |
2977 | ||
1da177e4 LT |
2978 | /* |
2979 | * Now scan the zone in the dma->highmem direction, stopping | |
2980 | * at the last zone which needs scanning. | |
2981 | * | |
2982 | * We do this because the page allocator works in the opposite | |
2983 | * direction. This prevents the page allocator from allocating | |
2984 | * pages behind kswapd's direction of progress, which would | |
2985 | * cause too much scanning of the lower zones. | |
2986 | */ | |
2987 | for (i = 0; i <= end_zone; i++) { | |
2988 | struct zone *zone = pgdat->node_zones + i; | |
2989 | ||
f3fe6512 | 2990 | if (!populated_zone(zone)) |
1da177e4 LT |
2991 | continue; |
2992 | ||
6e543d57 LD |
2993 | if (sc.priority != DEF_PRIORITY && |
2994 | !zone_reclaimable(zone)) | |
1da177e4 LT |
2995 | continue; |
2996 | ||
1da177e4 | 2997 | sc.nr_scanned = 0; |
4e416953 | 2998 | |
0ae5e89c | 2999 | nr_soft_scanned = 0; |
4e416953 BS |
3000 | /* |
3001 | * Call soft limit reclaim before calling shrink_zone. | |
4e416953 | 3002 | */ |
0ae5e89c YH |
3003 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, |
3004 | order, sc.gfp_mask, | |
3005 | &nr_soft_scanned); | |
3006 | sc.nr_reclaimed += nr_soft_reclaimed; | |
00918b6a | 3007 | |
32a4330d | 3008 | /* |
7c954f6d MG |
3009 | * There should be no need to raise the scanning |
3010 | * priority if enough pages are already being scanned | |
3011 | * that that high watermark would be met at 100% | |
3012 | * efficiency. | |
fe2c2a10 | 3013 | */ |
7c954f6d MG |
3014 | if (kswapd_shrink_zone(zone, end_zone, &sc, |
3015 | lru_pages, &nr_attempted)) | |
3016 | raise_priority = false; | |
1da177e4 | 3017 | } |
5515061d MG |
3018 | |
3019 | /* | |
3020 | * If the low watermark is met there is no need for processes | |
3021 | * to be throttled on pfmemalloc_wait as they should not be | |
3022 | * able to safely make forward progress. Wake them | |
3023 | */ | |
3024 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
3025 | pfmemalloc_watermark_ok(pgdat)) | |
3026 | wake_up(&pgdat->pfmemalloc_wait); | |
3027 | ||
1da177e4 | 3028 | /* |
b8e83b94 MG |
3029 | * Fragmentation may mean that the system cannot be rebalanced |
3030 | * for high-order allocations in all zones. If twice the | |
3031 | * allocation size has been reclaimed and the zones are still | |
3032 | * not balanced then recheck the watermarks at order-0 to | |
3033 | * prevent kswapd reclaiming excessively. Assume that a | |
3034 | * process requested a high-order can direct reclaim/compact. | |
1da177e4 | 3035 | */ |
b8e83b94 MG |
3036 | if (order && sc.nr_reclaimed >= 2UL << order) |
3037 | order = sc.order = 0; | |
8357376d | 3038 | |
b8e83b94 MG |
3039 | /* Check if kswapd should be suspending */ |
3040 | if (try_to_freeze() || kthread_should_stop()) | |
3041 | break; | |
8357376d | 3042 | |
2ab44f43 MG |
3043 | /* |
3044 | * Compact if necessary and kswapd is reclaiming at least the | |
3045 | * high watermark number of pages as requsted | |
3046 | */ | |
3047 | if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted) | |
3048 | compact_pgdat(pgdat, order); | |
3049 | ||
73ce02e9 | 3050 | /* |
b8e83b94 MG |
3051 | * Raise priority if scanning rate is too low or there was no |
3052 | * progress in reclaiming pages | |
73ce02e9 | 3053 | */ |
b8e83b94 MG |
3054 | if (raise_priority || !sc.nr_reclaimed) |
3055 | sc.priority--; | |
9aa41348 | 3056 | } while (sc.priority >= 1 && |
b8e83b94 | 3057 | !pgdat_balanced(pgdat, order, *classzone_idx)); |
1da177e4 | 3058 | |
b8e83b94 | 3059 | out: |
0abdee2b | 3060 | /* |
5515061d | 3061 | * Return the order we were reclaiming at so prepare_kswapd_sleep() |
0abdee2b MG |
3062 | * makes a decision on the order we were last reclaiming at. However, |
3063 | * if another caller entered the allocator slow path while kswapd | |
3064 | * was awake, order will remain at the higher level | |
3065 | */ | |
dc83edd9 | 3066 | *classzone_idx = end_zone; |
0abdee2b | 3067 | return order; |
1da177e4 LT |
3068 | } |
3069 | ||
dc83edd9 | 3070 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f0bc0a60 KM |
3071 | { |
3072 | long remaining = 0; | |
3073 | DEFINE_WAIT(wait); | |
3074 | ||
3075 | if (freezing(current) || kthread_should_stop()) | |
3076 | return; | |
3077 | ||
3078 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3079 | ||
3080 | /* Try to sleep for a short interval */ | |
5515061d | 3081 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
3082 | remaining = schedule_timeout(HZ/10); |
3083 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3084 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3085 | } | |
3086 | ||
3087 | /* | |
3088 | * After a short sleep, check if it was a premature sleep. If not, then | |
3089 | * go fully to sleep until explicitly woken up. | |
3090 | */ | |
5515061d | 3091 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
3092 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3093 | ||
3094 | /* | |
3095 | * vmstat counters are not perfectly accurate and the estimated | |
3096 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3097 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3098 | * watermarks being breached while under pressure, we reduce the | |
3099 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3100 | * them before going back to sleep. | |
3101 | */ | |
3102 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c | 3103 | |
62997027 MG |
3104 | /* |
3105 | * Compaction records what page blocks it recently failed to | |
3106 | * isolate pages from and skips them in the future scanning. | |
3107 | * When kswapd is going to sleep, it is reasonable to assume | |
3108 | * that pages and compaction may succeed so reset the cache. | |
3109 | */ | |
3110 | reset_isolation_suitable(pgdat); | |
3111 | ||
1c7e7f6c AK |
3112 | if (!kthread_should_stop()) |
3113 | schedule(); | |
3114 | ||
f0bc0a60 KM |
3115 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3116 | } else { | |
3117 | if (remaining) | |
3118 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3119 | else | |
3120 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3121 | } | |
3122 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3123 | } | |
3124 | ||
1da177e4 LT |
3125 | /* |
3126 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3127 | * from the init process. |
1da177e4 LT |
3128 | * |
3129 | * This basically trickles out pages so that we have _some_ | |
3130 | * free memory available even if there is no other activity | |
3131 | * that frees anything up. This is needed for things like routing | |
3132 | * etc, where we otherwise might have all activity going on in | |
3133 | * asynchronous contexts that cannot page things out. | |
3134 | * | |
3135 | * If there are applications that are active memory-allocators | |
3136 | * (most normal use), this basically shouldn't matter. | |
3137 | */ | |
3138 | static int kswapd(void *p) | |
3139 | { | |
215ddd66 | 3140 | unsigned long order, new_order; |
d2ebd0f6 | 3141 | unsigned balanced_order; |
215ddd66 | 3142 | int classzone_idx, new_classzone_idx; |
d2ebd0f6 | 3143 | int balanced_classzone_idx; |
1da177e4 LT |
3144 | pg_data_t *pgdat = (pg_data_t*)p; |
3145 | struct task_struct *tsk = current; | |
f0bc0a60 | 3146 | |
1da177e4 LT |
3147 | struct reclaim_state reclaim_state = { |
3148 | .reclaimed_slab = 0, | |
3149 | }; | |
a70f7302 | 3150 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3151 | |
cf40bd16 NP |
3152 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3153 | ||
174596a0 | 3154 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3155 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3156 | current->reclaim_state = &reclaim_state; |
3157 | ||
3158 | /* | |
3159 | * Tell the memory management that we're a "memory allocator", | |
3160 | * and that if we need more memory we should get access to it | |
3161 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3162 | * never get caught in the normal page freeing logic. | |
3163 | * | |
3164 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3165 | * you need a small amount of memory in order to be able to | |
3166 | * page out something else, and this flag essentially protects | |
3167 | * us from recursively trying to free more memory as we're | |
3168 | * trying to free the first piece of memory in the first place). | |
3169 | */ | |
930d9152 | 3170 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3171 | set_freezable(); |
1da177e4 | 3172 | |
215ddd66 | 3173 | order = new_order = 0; |
d2ebd0f6 | 3174 | balanced_order = 0; |
215ddd66 | 3175 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; |
d2ebd0f6 | 3176 | balanced_classzone_idx = classzone_idx; |
1da177e4 | 3177 | for ( ; ; ) { |
6f6313d4 | 3178 | bool ret; |
3e1d1d28 | 3179 | |
215ddd66 MG |
3180 | /* |
3181 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
3182 | * new request of a similar or harder type will succeed soon | |
3183 | * so consider going to sleep on the basis we reclaimed at | |
3184 | */ | |
d2ebd0f6 AS |
3185 | if (balanced_classzone_idx >= new_classzone_idx && |
3186 | balanced_order == new_order) { | |
215ddd66 MG |
3187 | new_order = pgdat->kswapd_max_order; |
3188 | new_classzone_idx = pgdat->classzone_idx; | |
3189 | pgdat->kswapd_max_order = 0; | |
3190 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3191 | } | |
3192 | ||
99504748 | 3193 | if (order < new_order || classzone_idx > new_classzone_idx) { |
1da177e4 LT |
3194 | /* |
3195 | * Don't sleep if someone wants a larger 'order' | |
99504748 | 3196 | * allocation or has tigher zone constraints |
1da177e4 LT |
3197 | */ |
3198 | order = new_order; | |
99504748 | 3199 | classzone_idx = new_classzone_idx; |
1da177e4 | 3200 | } else { |
d2ebd0f6 AS |
3201 | kswapd_try_to_sleep(pgdat, balanced_order, |
3202 | balanced_classzone_idx); | |
1da177e4 | 3203 | order = pgdat->kswapd_max_order; |
99504748 | 3204 | classzone_idx = pgdat->classzone_idx; |
f0dfcde0 AS |
3205 | new_order = order; |
3206 | new_classzone_idx = classzone_idx; | |
4d40502e | 3207 | pgdat->kswapd_max_order = 0; |
215ddd66 | 3208 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
1da177e4 | 3209 | } |
1da177e4 | 3210 | |
8fe23e05 DR |
3211 | ret = try_to_freeze(); |
3212 | if (kthread_should_stop()) | |
3213 | break; | |
3214 | ||
3215 | /* | |
3216 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3217 | * after returning from the refrigerator | |
3218 | */ | |
33906bc5 MG |
3219 | if (!ret) { |
3220 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
d2ebd0f6 AS |
3221 | balanced_classzone_idx = classzone_idx; |
3222 | balanced_order = balance_pgdat(pgdat, order, | |
3223 | &balanced_classzone_idx); | |
33906bc5 | 3224 | } |
1da177e4 | 3225 | } |
b0a8cc58 TY |
3226 | |
3227 | current->reclaim_state = NULL; | |
1da177e4 LT |
3228 | return 0; |
3229 | } | |
3230 | ||
3231 | /* | |
3232 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3233 | */ | |
99504748 | 3234 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3235 | { |
3236 | pg_data_t *pgdat; | |
3237 | ||
f3fe6512 | 3238 | if (!populated_zone(zone)) |
1da177e4 LT |
3239 | return; |
3240 | ||
88f5acf8 | 3241 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1da177e4 | 3242 | return; |
88f5acf8 | 3243 | pgdat = zone->zone_pgdat; |
99504748 | 3244 | if (pgdat->kswapd_max_order < order) { |
1da177e4 | 3245 | pgdat->kswapd_max_order = order; |
99504748 MG |
3246 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
3247 | } | |
8d0986e2 | 3248 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3249 | return; |
892f795d | 3250 | if (zone_balanced(zone, order, 0, 0)) |
88f5acf8 MG |
3251 | return; |
3252 | ||
3253 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3254 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3255 | } |
3256 | ||
adea02a1 WF |
3257 | /* |
3258 | * The reclaimable count would be mostly accurate. | |
3259 | * The less reclaimable pages may be | |
3260 | * - mlocked pages, which will be moved to unevictable list when encountered | |
3261 | * - mapped pages, which may require several travels to be reclaimed | |
3262 | * - dirty pages, which is not "instantly" reclaimable | |
3263 | */ | |
3264 | unsigned long global_reclaimable_pages(void) | |
4f98a2fe | 3265 | { |
adea02a1 WF |
3266 | int nr; |
3267 | ||
3268 | nr = global_page_state(NR_ACTIVE_FILE) + | |
3269 | global_page_state(NR_INACTIVE_FILE); | |
3270 | ||
ec8acf20 | 3271 | if (get_nr_swap_pages() > 0) |
adea02a1 WF |
3272 | nr += global_page_state(NR_ACTIVE_ANON) + |
3273 | global_page_state(NR_INACTIVE_ANON); | |
3274 | ||
3275 | return nr; | |
3276 | } | |
3277 | ||
c6f37f12 | 3278 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3279 | /* |
7b51755c | 3280 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3281 | * freed pages. |
3282 | * | |
3283 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3284 | * LRU order by reclaiming preferentially | |
3285 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3286 | */ |
7b51755c | 3287 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3288 | { |
d6277db4 | 3289 | struct reclaim_state reclaim_state; |
d6277db4 | 3290 | struct scan_control sc = { |
7b51755c KM |
3291 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
3292 | .may_swap = 1, | |
3293 | .may_unmap = 1, | |
d6277db4 | 3294 | .may_writepage = 1, |
7b51755c KM |
3295 | .nr_to_reclaim = nr_to_reclaim, |
3296 | .hibernation_mode = 1, | |
7b51755c | 3297 | .order = 0, |
9e3b2f8c | 3298 | .priority = DEF_PRIORITY, |
1da177e4 | 3299 | }; |
a09ed5e0 YH |
3300 | struct shrink_control shrink = { |
3301 | .gfp_mask = sc.gfp_mask, | |
3302 | }; | |
3303 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
7b51755c KM |
3304 | struct task_struct *p = current; |
3305 | unsigned long nr_reclaimed; | |
1da177e4 | 3306 | |
7b51755c KM |
3307 | p->flags |= PF_MEMALLOC; |
3308 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3309 | reclaim_state.reclaimed_slab = 0; | |
3310 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3311 | |
a09ed5e0 | 3312 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
d979677c | 3313 | |
7b51755c KM |
3314 | p->reclaim_state = NULL; |
3315 | lockdep_clear_current_reclaim_state(); | |
3316 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3317 | |
7b51755c | 3318 | return nr_reclaimed; |
1da177e4 | 3319 | } |
c6f37f12 | 3320 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3321 | |
1da177e4 LT |
3322 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3323 | not required for correctness. So if the last cpu in a node goes | |
3324 | away, we get changed to run anywhere: as the first one comes back, | |
3325 | restore their cpu bindings. */ | |
fcb35a9b GKH |
3326 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
3327 | void *hcpu) | |
1da177e4 | 3328 | { |
58c0a4a7 | 3329 | int nid; |
1da177e4 | 3330 | |
8bb78442 | 3331 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
48fb2e24 | 3332 | for_each_node_state(nid, N_MEMORY) { |
c5f59f08 | 3333 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3334 | const struct cpumask *mask; |
3335 | ||
3336 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3337 | |
3e597945 | 3338 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3339 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3340 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3341 | } |
3342 | } | |
3343 | return NOTIFY_OK; | |
3344 | } | |
1da177e4 | 3345 | |
3218ae14 YG |
3346 | /* |
3347 | * This kswapd start function will be called by init and node-hot-add. | |
3348 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3349 | */ | |
3350 | int kswapd_run(int nid) | |
3351 | { | |
3352 | pg_data_t *pgdat = NODE_DATA(nid); | |
3353 | int ret = 0; | |
3354 | ||
3355 | if (pgdat->kswapd) | |
3356 | return 0; | |
3357 | ||
3358 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3359 | if (IS_ERR(pgdat->kswapd)) { | |
3360 | /* failure at boot is fatal */ | |
3361 | BUG_ON(system_state == SYSTEM_BOOTING); | |
d5dc0ad9 GS |
3362 | pr_err("Failed to start kswapd on node %d\n", nid); |
3363 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3364 | pgdat->kswapd = NULL; |
3218ae14 YG |
3365 | } |
3366 | return ret; | |
3367 | } | |
3368 | ||
8fe23e05 | 3369 | /* |
d8adde17 JL |
3370 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
3371 | * hold lock_memory_hotplug(). | |
8fe23e05 DR |
3372 | */ |
3373 | void kswapd_stop(int nid) | |
3374 | { | |
3375 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3376 | ||
d8adde17 | 3377 | if (kswapd) { |
8fe23e05 | 3378 | kthread_stop(kswapd); |
d8adde17 JL |
3379 | NODE_DATA(nid)->kswapd = NULL; |
3380 | } | |
8fe23e05 DR |
3381 | } |
3382 | ||
1da177e4 LT |
3383 | static int __init kswapd_init(void) |
3384 | { | |
3218ae14 | 3385 | int nid; |
69e05944 | 3386 | |
1da177e4 | 3387 | swap_setup(); |
48fb2e24 | 3388 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3389 | kswapd_run(nid); |
1da177e4 LT |
3390 | hotcpu_notifier(cpu_callback, 0); |
3391 | return 0; | |
3392 | } | |
3393 | ||
3394 | module_init(kswapd_init) | |
9eeff239 CL |
3395 | |
3396 | #ifdef CONFIG_NUMA | |
3397 | /* | |
3398 | * Zone reclaim mode | |
3399 | * | |
3400 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3401 | * the watermarks. | |
9eeff239 CL |
3402 | */ |
3403 | int zone_reclaim_mode __read_mostly; | |
3404 | ||
1b2ffb78 | 3405 | #define RECLAIM_OFF 0 |
7d03431c | 3406 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 CL |
3407 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
3408 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3409 | ||
a92f7126 CL |
3410 | /* |
3411 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3412 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3413 | * a zone. | |
3414 | */ | |
3415 | #define ZONE_RECLAIM_PRIORITY 4 | |
3416 | ||
9614634f CL |
3417 | /* |
3418 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3419 | * occur. | |
3420 | */ | |
3421 | int sysctl_min_unmapped_ratio = 1; | |
3422 | ||
0ff38490 CL |
3423 | /* |
3424 | * If the number of slab pages in a zone grows beyond this percentage then | |
3425 | * slab reclaim needs to occur. | |
3426 | */ | |
3427 | int sysctl_min_slab_ratio = 5; | |
3428 | ||
90afa5de MG |
3429 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
3430 | { | |
3431 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3432 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3433 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3434 | ||
3435 | /* | |
3436 | * It's possible for there to be more file mapped pages than | |
3437 | * accounted for by the pages on the file LRU lists because | |
3438 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3439 | */ | |
3440 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3441 | } | |
3442 | ||
3443 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3444 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3445 | { | |
3446 | long nr_pagecache_reclaimable; | |
3447 | long delta = 0; | |
3448 | ||
3449 | /* | |
3450 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3451 | * potentially reclaimable. Otherwise, we have to worry about | |
3452 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3453 | * a better estimate | |
3454 | */ | |
3455 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3456 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3457 | else | |
3458 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3459 | ||
3460 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3461 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3462 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3463 | ||
3464 | /* Watch for any possible underflows due to delta */ | |
3465 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3466 | delta = nr_pagecache_reclaimable; | |
3467 | ||
3468 | return nr_pagecache_reclaimable - delta; | |
3469 | } | |
3470 | ||
9eeff239 CL |
3471 | /* |
3472 | * Try to free up some pages from this zone through reclaim. | |
3473 | */ | |
179e9639 | 3474 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3475 | { |
7fb2d46d | 3476 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3477 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3478 | struct task_struct *p = current; |
3479 | struct reclaim_state reclaim_state; | |
179e9639 AM |
3480 | struct scan_control sc = { |
3481 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
a6dc60f8 | 3482 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), |
2e2e4259 | 3483 | .may_swap = 1, |
62b726c1 | 3484 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
21caf2fc | 3485 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
bd2f6199 | 3486 | .order = order, |
9e3b2f8c | 3487 | .priority = ZONE_RECLAIM_PRIORITY, |
179e9639 | 3488 | }; |
a09ed5e0 YH |
3489 | struct shrink_control shrink = { |
3490 | .gfp_mask = sc.gfp_mask, | |
3491 | }; | |
15748048 | 3492 | unsigned long nr_slab_pages0, nr_slab_pages1; |
9eeff239 | 3493 | |
9eeff239 | 3494 | cond_resched(); |
d4f7796e CL |
3495 | /* |
3496 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3497 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3498 | * and RECLAIM_SWAP. | |
3499 | */ | |
3500 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3501 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3502 | reclaim_state.reclaimed_slab = 0; |
3503 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3504 | |
90afa5de | 3505 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
3506 | /* |
3507 | * Free memory by calling shrink zone with increasing | |
3508 | * priorities until we have enough memory freed. | |
3509 | */ | |
0ff38490 | 3510 | do { |
9e3b2f8c KK |
3511 | shrink_zone(zone, &sc); |
3512 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); | |
0ff38490 | 3513 | } |
c84db23c | 3514 | |
15748048 KM |
3515 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3516 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
2a16e3f4 | 3517 | /* |
7fb2d46d | 3518 | * shrink_slab() does not currently allow us to determine how |
0ff38490 CL |
3519 | * many pages were freed in this zone. So we take the current |
3520 | * number of slab pages and shake the slab until it is reduced | |
3521 | * by the same nr_pages that we used for reclaiming unmapped | |
3522 | * pages. | |
2a16e3f4 | 3523 | * |
0ff38490 CL |
3524 | * Note that shrink_slab will free memory on all zones and may |
3525 | * take a long time. | |
2a16e3f4 | 3526 | */ |
4dc4b3d9 KM |
3527 | for (;;) { |
3528 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3529 | ||
3530 | /* No reclaimable slab or very low memory pressure */ | |
1495f230 | 3531 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) |
4dc4b3d9 KM |
3532 | break; |
3533 | ||
3534 | /* Freed enough memory */ | |
3535 | nr_slab_pages1 = zone_page_state(zone, | |
3536 | NR_SLAB_RECLAIMABLE); | |
3537 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3538 | break; | |
3539 | } | |
83e33a47 CL |
3540 | |
3541 | /* | |
3542 | * Update nr_reclaimed by the number of slab pages we | |
3543 | * reclaimed from this zone. | |
3544 | */ | |
15748048 KM |
3545 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3546 | if (nr_slab_pages1 < nr_slab_pages0) | |
3547 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
2a16e3f4 CL |
3548 | } |
3549 | ||
9eeff239 | 3550 | p->reclaim_state = NULL; |
d4f7796e | 3551 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3552 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3553 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3554 | } |
179e9639 AM |
3555 | |
3556 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3557 | { | |
179e9639 | 3558 | int node_id; |
d773ed6b | 3559 | int ret; |
179e9639 AM |
3560 | |
3561 | /* | |
0ff38490 CL |
3562 | * Zone reclaim reclaims unmapped file backed pages and |
3563 | * slab pages if we are over the defined limits. | |
34aa1330 | 3564 | * |
9614634f CL |
3565 | * A small portion of unmapped file backed pages is needed for |
3566 | * file I/O otherwise pages read by file I/O will be immediately | |
3567 | * thrown out if the zone is overallocated. So we do not reclaim | |
3568 | * if less than a specified percentage of the zone is used by | |
3569 | * unmapped file backed pages. | |
179e9639 | 3570 | */ |
90afa5de MG |
3571 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
3572 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 3573 | return ZONE_RECLAIM_FULL; |
179e9639 | 3574 | |
6e543d57 | 3575 | if (!zone_reclaimable(zone)) |
fa5e084e | 3576 | return ZONE_RECLAIM_FULL; |
d773ed6b | 3577 | |
179e9639 | 3578 | /* |
d773ed6b | 3579 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3580 | */ |
d773ed6b | 3581 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 3582 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
3583 | |
3584 | /* | |
3585 | * Only run zone reclaim on the local zone or on zones that do not | |
3586 | * have associated processors. This will favor the local processor | |
3587 | * over remote processors and spread off node memory allocations | |
3588 | * as wide as possible. | |
3589 | */ | |
89fa3024 | 3590 | node_id = zone_to_nid(zone); |
37c0708d | 3591 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 3592 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b DR |
3593 | |
3594 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
fa5e084e MG |
3595 | return ZONE_RECLAIM_NOSCAN; |
3596 | ||
d773ed6b DR |
3597 | ret = __zone_reclaim(zone, gfp_mask, order); |
3598 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3599 | ||
24cf7251 MG |
3600 | if (!ret) |
3601 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3602 | ||
d773ed6b | 3603 | return ret; |
179e9639 | 3604 | } |
9eeff239 | 3605 | #endif |
894bc310 | 3606 | |
894bc310 LS |
3607 | /* |
3608 | * page_evictable - test whether a page is evictable | |
3609 | * @page: the page to test | |
894bc310 LS |
3610 | * |
3611 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3612 | * lists vs unevictable list. |
894bc310 LS |
3613 | * |
3614 | * Reasons page might not be evictable: | |
ba9ddf49 | 3615 | * (1) page's mapping marked unevictable |
b291f000 | 3616 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3617 | * |
894bc310 | 3618 | */ |
39b5f29a | 3619 | int page_evictable(struct page *page) |
894bc310 | 3620 | { |
39b5f29a | 3621 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
894bc310 | 3622 | } |
89e004ea | 3623 | |
85046579 | 3624 | #ifdef CONFIG_SHMEM |
89e004ea | 3625 | /** |
24513264 HD |
3626 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3627 | * @pages: array of pages to check | |
3628 | * @nr_pages: number of pages to check | |
89e004ea | 3629 | * |
24513264 | 3630 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3631 | * |
3632 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3633 | */ |
24513264 | 3634 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3635 | { |
925b7673 | 3636 | struct lruvec *lruvec; |
24513264 HD |
3637 | struct zone *zone = NULL; |
3638 | int pgscanned = 0; | |
3639 | int pgrescued = 0; | |
3640 | int i; | |
89e004ea | 3641 | |
24513264 HD |
3642 | for (i = 0; i < nr_pages; i++) { |
3643 | struct page *page = pages[i]; | |
3644 | struct zone *pagezone; | |
89e004ea | 3645 | |
24513264 HD |
3646 | pgscanned++; |
3647 | pagezone = page_zone(page); | |
3648 | if (pagezone != zone) { | |
3649 | if (zone) | |
3650 | spin_unlock_irq(&zone->lru_lock); | |
3651 | zone = pagezone; | |
3652 | spin_lock_irq(&zone->lru_lock); | |
3653 | } | |
fa9add64 | 3654 | lruvec = mem_cgroup_page_lruvec(page, zone); |
89e004ea | 3655 | |
24513264 HD |
3656 | if (!PageLRU(page) || !PageUnevictable(page)) |
3657 | continue; | |
89e004ea | 3658 | |
39b5f29a | 3659 | if (page_evictable(page)) { |
24513264 HD |
3660 | enum lru_list lru = page_lru_base_type(page); |
3661 | ||
3662 | VM_BUG_ON(PageActive(page)); | |
3663 | ClearPageUnevictable(page); | |
fa9add64 HD |
3664 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
3665 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 3666 | pgrescued++; |
89e004ea | 3667 | } |
24513264 | 3668 | } |
89e004ea | 3669 | |
24513264 HD |
3670 | if (zone) { |
3671 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3672 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3673 | spin_unlock_irq(&zone->lru_lock); | |
89e004ea | 3674 | } |
89e004ea | 3675 | } |
85046579 | 3676 | #endif /* CONFIG_SHMEM */ |
af936a16 | 3677 | |
264e56d8 | 3678 | static void warn_scan_unevictable_pages(void) |
af936a16 | 3679 | { |
264e56d8 | 3680 | printk_once(KERN_WARNING |
25bd91bd | 3681 | "%s: The scan_unevictable_pages sysctl/node-interface has been " |
264e56d8 | 3682 | "disabled for lack of a legitimate use case. If you have " |
25bd91bd KM |
3683 | "one, please send an email to linux-mm@kvack.org.\n", |
3684 | current->comm); | |
af936a16 LS |
3685 | } |
3686 | ||
3687 | /* | |
3688 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3689 | * all nodes' unevictable lists for evictable pages | |
3690 | */ | |
3691 | unsigned long scan_unevictable_pages; | |
3692 | ||
3693 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
8d65af78 | 3694 | void __user *buffer, |
af936a16 LS |
3695 | size_t *length, loff_t *ppos) |
3696 | { | |
264e56d8 | 3697 | warn_scan_unevictable_pages(); |
8d65af78 | 3698 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
af936a16 LS |
3699 | scan_unevictable_pages = 0; |
3700 | return 0; | |
3701 | } | |
3702 | ||
e4455abb | 3703 | #ifdef CONFIG_NUMA |
af936a16 LS |
3704 | /* |
3705 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3706 | * a specified node's per zone unevictable lists for evictable pages. | |
3707 | */ | |
3708 | ||
10fbcf4c KS |
3709 | static ssize_t read_scan_unevictable_node(struct device *dev, |
3710 | struct device_attribute *attr, | |
af936a16 LS |
3711 | char *buf) |
3712 | { | |
264e56d8 | 3713 | warn_scan_unevictable_pages(); |
af936a16 LS |
3714 | return sprintf(buf, "0\n"); /* always zero; should fit... */ |
3715 | } | |
3716 | ||
10fbcf4c KS |
3717 | static ssize_t write_scan_unevictable_node(struct device *dev, |
3718 | struct device_attribute *attr, | |
af936a16 LS |
3719 | const char *buf, size_t count) |
3720 | { | |
264e56d8 | 3721 | warn_scan_unevictable_pages(); |
af936a16 LS |
3722 | return 1; |
3723 | } | |
3724 | ||
3725 | ||
10fbcf4c | 3726 | static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, |
af936a16 LS |
3727 | read_scan_unevictable_node, |
3728 | write_scan_unevictable_node); | |
3729 | ||
3730 | int scan_unevictable_register_node(struct node *node) | |
3731 | { | |
10fbcf4c | 3732 | return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 LS |
3733 | } |
3734 | ||
3735 | void scan_unevictable_unregister_node(struct node *node) | |
3736 | { | |
10fbcf4c | 3737 | device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 | 3738 | } |
e4455abb | 3739 | #endif |