Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/vmscan.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
5 | * | |
6 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
7 | * kswapd added: 7.1.96 sct | |
8 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
9 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
10 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
11 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
12 | */ | |
13 | ||
14 | #include <linux/mm.h> | |
15 | #include <linux/module.h> | |
5a0e3ad6 | 16 | #include <linux/gfp.h> |
1da177e4 LT |
17 | #include <linux/kernel_stat.h> |
18 | #include <linux/swap.h> | |
19 | #include <linux/pagemap.h> | |
20 | #include <linux/init.h> | |
21 | #include <linux/highmem.h> | |
e129b5c2 | 22 | #include <linux/vmstat.h> |
1da177e4 LT |
23 | #include <linux/file.h> |
24 | #include <linux/writeback.h> | |
25 | #include <linux/blkdev.h> | |
26 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
27 | buffer_heads_over_limit */ | |
28 | #include <linux/mm_inline.h> | |
1da177e4 LT |
29 | #include <linux/backing-dev.h> |
30 | #include <linux/rmap.h> | |
31 | #include <linux/topology.h> | |
32 | #include <linux/cpu.h> | |
33 | #include <linux/cpuset.h> | |
3e7d3449 | 34 | #include <linux/compaction.h> |
1da177e4 LT |
35 | #include <linux/notifier.h> |
36 | #include <linux/rwsem.h> | |
248a0301 | 37 | #include <linux/delay.h> |
3218ae14 | 38 | #include <linux/kthread.h> |
7dfb7103 | 39 | #include <linux/freezer.h> |
66e1707b | 40 | #include <linux/memcontrol.h> |
873b4771 | 41 | #include <linux/delayacct.h> |
af936a16 | 42 | #include <linux/sysctl.h> |
929bea7c | 43 | #include <linux/oom.h> |
268bb0ce | 44 | #include <linux/prefetch.h> |
1da177e4 LT |
45 | |
46 | #include <asm/tlbflush.h> | |
47 | #include <asm/div64.h> | |
48 | ||
49 | #include <linux/swapops.h> | |
50 | ||
0f8053a5 NP |
51 | #include "internal.h" |
52 | ||
33906bc5 MG |
53 | #define CREATE_TRACE_POINTS |
54 | #include <trace/events/vmscan.h> | |
55 | ||
1da177e4 | 56 | struct scan_control { |
1da177e4 LT |
57 | /* Incremented by the number of inactive pages that were scanned */ |
58 | unsigned long nr_scanned; | |
59 | ||
a79311c1 RR |
60 | /* Number of pages freed so far during a call to shrink_zones() */ |
61 | unsigned long nr_reclaimed; | |
62 | ||
22fba335 KM |
63 | /* How many pages shrink_list() should reclaim */ |
64 | unsigned long nr_to_reclaim; | |
65 | ||
7b51755c KM |
66 | unsigned long hibernation_mode; |
67 | ||
1da177e4 | 68 | /* This context's GFP mask */ |
6daa0e28 | 69 | gfp_t gfp_mask; |
1da177e4 LT |
70 | |
71 | int may_writepage; | |
72 | ||
a6dc60f8 JW |
73 | /* Can mapped pages be reclaimed? */ |
74 | int may_unmap; | |
f1fd1067 | 75 | |
2e2e4259 KM |
76 | /* Can pages be swapped as part of reclaim? */ |
77 | int may_swap; | |
78 | ||
5ad333eb | 79 | int order; |
66e1707b | 80 | |
9e3b2f8c KK |
81 | /* Scan (total_size >> priority) pages at once */ |
82 | int priority; | |
83 | ||
f16015fb JW |
84 | /* |
85 | * The memory cgroup that hit its limit and as a result is the | |
86 | * primary target of this reclaim invocation. | |
87 | */ | |
88 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 89 | |
327c0e96 KH |
90 | /* |
91 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
92 | * are scanned. | |
93 | */ | |
94 | nodemask_t *nodemask; | |
1da177e4 LT |
95 | }; |
96 | ||
1da177e4 LT |
97 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
98 | ||
99 | #ifdef ARCH_HAS_PREFETCH | |
100 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
101 | do { \ | |
102 | if ((_page)->lru.prev != _base) { \ | |
103 | struct page *prev; \ | |
104 | \ | |
105 | prev = lru_to_page(&(_page->lru)); \ | |
106 | prefetch(&prev->_field); \ | |
107 | } \ | |
108 | } while (0) | |
109 | #else | |
110 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
111 | #endif | |
112 | ||
113 | #ifdef ARCH_HAS_PREFETCHW | |
114 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
115 | do { \ | |
116 | if ((_page)->lru.prev != _base) { \ | |
117 | struct page *prev; \ | |
118 | \ | |
119 | prev = lru_to_page(&(_page->lru)); \ | |
120 | prefetchw(&prev->_field); \ | |
121 | } \ | |
122 | } while (0) | |
123 | #else | |
124 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
125 | #endif | |
126 | ||
127 | /* | |
128 | * From 0 .. 100. Higher means more swappy. | |
129 | */ | |
130 | int vm_swappiness = 60; | |
bd1e22b8 | 131 | long vm_total_pages; /* The total number of pages which the VM controls */ |
1da177e4 LT |
132 | |
133 | static LIST_HEAD(shrinker_list); | |
134 | static DECLARE_RWSEM(shrinker_rwsem); | |
135 | ||
c255a458 | 136 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
137 | static bool global_reclaim(struct scan_control *sc) |
138 | { | |
f16015fb | 139 | return !sc->target_mem_cgroup; |
89b5fae5 | 140 | } |
91a45470 | 141 | #else |
89b5fae5 JW |
142 | static bool global_reclaim(struct scan_control *sc) |
143 | { | |
144 | return true; | |
145 | } | |
91a45470 KH |
146 | #endif |
147 | ||
4d7dcca2 | 148 | static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) |
c9f299d9 | 149 | { |
c3c787e8 | 150 | if (!mem_cgroup_disabled()) |
4d7dcca2 | 151 | return mem_cgroup_get_lru_size(lruvec, lru); |
a3d8e054 | 152 | |
074291fe | 153 | return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); |
c9f299d9 KM |
154 | } |
155 | ||
1da177e4 LT |
156 | /* |
157 | * Add a shrinker callback to be called from the vm | |
158 | */ | |
8e1f936b | 159 | void register_shrinker(struct shrinker *shrinker) |
1da177e4 | 160 | { |
83aeeada | 161 | atomic_long_set(&shrinker->nr_in_batch, 0); |
8e1f936b RR |
162 | down_write(&shrinker_rwsem); |
163 | list_add_tail(&shrinker->list, &shrinker_list); | |
164 | up_write(&shrinker_rwsem); | |
1da177e4 | 165 | } |
8e1f936b | 166 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
167 | |
168 | /* | |
169 | * Remove one | |
170 | */ | |
8e1f936b | 171 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
172 | { |
173 | down_write(&shrinker_rwsem); | |
174 | list_del(&shrinker->list); | |
175 | up_write(&shrinker_rwsem); | |
1da177e4 | 176 | } |
8e1f936b | 177 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 | 178 | |
1495f230 YH |
179 | static inline int do_shrinker_shrink(struct shrinker *shrinker, |
180 | struct shrink_control *sc, | |
181 | unsigned long nr_to_scan) | |
182 | { | |
183 | sc->nr_to_scan = nr_to_scan; | |
184 | return (*shrinker->shrink)(shrinker, sc); | |
185 | } | |
186 | ||
1da177e4 LT |
187 | #define SHRINK_BATCH 128 |
188 | /* | |
189 | * Call the shrink functions to age shrinkable caches | |
190 | * | |
191 | * Here we assume it costs one seek to replace a lru page and that it also | |
192 | * takes a seek to recreate a cache object. With this in mind we age equal | |
193 | * percentages of the lru and ageable caches. This should balance the seeks | |
194 | * generated by these structures. | |
195 | * | |
183ff22b | 196 | * If the vm encountered mapped pages on the LRU it increase the pressure on |
1da177e4 LT |
197 | * slab to avoid swapping. |
198 | * | |
199 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
200 | * | |
201 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
202 | * are eligible for the caller's allocation attempt. It is used for balancing | |
203 | * slab reclaim versus page reclaim. | |
b15e0905 | 204 | * |
205 | * Returns the number of slab objects which we shrunk. | |
1da177e4 | 206 | */ |
a09ed5e0 | 207 | unsigned long shrink_slab(struct shrink_control *shrink, |
1495f230 | 208 | unsigned long nr_pages_scanned, |
a09ed5e0 | 209 | unsigned long lru_pages) |
1da177e4 LT |
210 | { |
211 | struct shrinker *shrinker; | |
69e05944 | 212 | unsigned long ret = 0; |
1da177e4 | 213 | |
1495f230 YH |
214 | if (nr_pages_scanned == 0) |
215 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 216 | |
f06590bd MK |
217 | if (!down_read_trylock(&shrinker_rwsem)) { |
218 | /* Assume we'll be able to shrink next time */ | |
219 | ret = 1; | |
220 | goto out; | |
221 | } | |
1da177e4 LT |
222 | |
223 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
224 | unsigned long long delta; | |
635697c6 KK |
225 | long total_scan; |
226 | long max_pass; | |
09576073 | 227 | int shrink_ret = 0; |
acf92b48 DC |
228 | long nr; |
229 | long new_nr; | |
e9299f50 DC |
230 | long batch_size = shrinker->batch ? shrinker->batch |
231 | : SHRINK_BATCH; | |
1da177e4 | 232 | |
635697c6 KK |
233 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); |
234 | if (max_pass <= 0) | |
235 | continue; | |
236 | ||
acf92b48 DC |
237 | /* |
238 | * copy the current shrinker scan count into a local variable | |
239 | * and zero it so that other concurrent shrinker invocations | |
240 | * don't also do this scanning work. | |
241 | */ | |
83aeeada | 242 | nr = atomic_long_xchg(&shrinker->nr_in_batch, 0); |
acf92b48 DC |
243 | |
244 | total_scan = nr; | |
1495f230 | 245 | delta = (4 * nr_pages_scanned) / shrinker->seeks; |
ea164d73 | 246 | delta *= max_pass; |
1da177e4 | 247 | do_div(delta, lru_pages + 1); |
acf92b48 DC |
248 | total_scan += delta; |
249 | if (total_scan < 0) { | |
88c3bd70 DR |
250 | printk(KERN_ERR "shrink_slab: %pF negative objects to " |
251 | "delete nr=%ld\n", | |
acf92b48 DC |
252 | shrinker->shrink, total_scan); |
253 | total_scan = max_pass; | |
ea164d73 AA |
254 | } |
255 | ||
3567b59a DC |
256 | /* |
257 | * We need to avoid excessive windup on filesystem shrinkers | |
258 | * due to large numbers of GFP_NOFS allocations causing the | |
259 | * shrinkers to return -1 all the time. This results in a large | |
260 | * nr being built up so when a shrink that can do some work | |
261 | * comes along it empties the entire cache due to nr >>> | |
262 | * max_pass. This is bad for sustaining a working set in | |
263 | * memory. | |
264 | * | |
265 | * Hence only allow the shrinker to scan the entire cache when | |
266 | * a large delta change is calculated directly. | |
267 | */ | |
268 | if (delta < max_pass / 4) | |
269 | total_scan = min(total_scan, max_pass / 2); | |
270 | ||
ea164d73 AA |
271 | /* |
272 | * Avoid risking looping forever due to too large nr value: | |
273 | * never try to free more than twice the estimate number of | |
274 | * freeable entries. | |
275 | */ | |
acf92b48 DC |
276 | if (total_scan > max_pass * 2) |
277 | total_scan = max_pass * 2; | |
1da177e4 | 278 | |
acf92b48 | 279 | trace_mm_shrink_slab_start(shrinker, shrink, nr, |
09576073 DC |
280 | nr_pages_scanned, lru_pages, |
281 | max_pass, delta, total_scan); | |
282 | ||
e9299f50 | 283 | while (total_scan >= batch_size) { |
b15e0905 | 284 | int nr_before; |
1da177e4 | 285 | |
1495f230 YH |
286 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); |
287 | shrink_ret = do_shrinker_shrink(shrinker, shrink, | |
e9299f50 | 288 | batch_size); |
1da177e4 LT |
289 | if (shrink_ret == -1) |
290 | break; | |
b15e0905 | 291 | if (shrink_ret < nr_before) |
292 | ret += nr_before - shrink_ret; | |
e9299f50 DC |
293 | count_vm_events(SLABS_SCANNED, batch_size); |
294 | total_scan -= batch_size; | |
1da177e4 LT |
295 | |
296 | cond_resched(); | |
297 | } | |
298 | ||
acf92b48 DC |
299 | /* |
300 | * move the unused scan count back into the shrinker in a | |
301 | * manner that handles concurrent updates. If we exhausted the | |
302 | * scan, there is no need to do an update. | |
303 | */ | |
83aeeada KK |
304 | if (total_scan > 0) |
305 | new_nr = atomic_long_add_return(total_scan, | |
306 | &shrinker->nr_in_batch); | |
307 | else | |
308 | new_nr = atomic_long_read(&shrinker->nr_in_batch); | |
acf92b48 DC |
309 | |
310 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); | |
1da177e4 LT |
311 | } |
312 | up_read(&shrinker_rwsem); | |
f06590bd MK |
313 | out: |
314 | cond_resched(); | |
b15e0905 | 315 | return ret; |
1da177e4 LT |
316 | } |
317 | ||
1da177e4 LT |
318 | static inline int is_page_cache_freeable(struct page *page) |
319 | { | |
ceddc3a5 JW |
320 | /* |
321 | * A freeable page cache page is referenced only by the caller | |
322 | * that isolated the page, the page cache radix tree and | |
323 | * optional buffer heads at page->private. | |
324 | */ | |
edcf4748 | 325 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
326 | } |
327 | ||
7d3579e8 KM |
328 | static int may_write_to_queue(struct backing_dev_info *bdi, |
329 | struct scan_control *sc) | |
1da177e4 | 330 | { |
930d9152 | 331 | if (current->flags & PF_SWAPWRITE) |
1da177e4 LT |
332 | return 1; |
333 | if (!bdi_write_congested(bdi)) | |
334 | return 1; | |
335 | if (bdi == current->backing_dev_info) | |
336 | return 1; | |
337 | return 0; | |
338 | } | |
339 | ||
340 | /* | |
341 | * We detected a synchronous write error writing a page out. Probably | |
342 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
343 | * fsync(), msync() or close(). | |
344 | * | |
345 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
346 | * prevents it from being freed up. But we have a ref on the page and once | |
347 | * that page is locked, the mapping is pinned. | |
348 | * | |
349 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
350 | * __GFP_FS. | |
351 | */ | |
352 | static void handle_write_error(struct address_space *mapping, | |
353 | struct page *page, int error) | |
354 | { | |
7eaceacc | 355 | lock_page(page); |
3e9f45bd GC |
356 | if (page_mapping(page) == mapping) |
357 | mapping_set_error(mapping, error); | |
1da177e4 LT |
358 | unlock_page(page); |
359 | } | |
360 | ||
04e62a29 CL |
361 | /* possible outcome of pageout() */ |
362 | typedef enum { | |
363 | /* failed to write page out, page is locked */ | |
364 | PAGE_KEEP, | |
365 | /* move page to the active list, page is locked */ | |
366 | PAGE_ACTIVATE, | |
367 | /* page has been sent to the disk successfully, page is unlocked */ | |
368 | PAGE_SUCCESS, | |
369 | /* page is clean and locked */ | |
370 | PAGE_CLEAN, | |
371 | } pageout_t; | |
372 | ||
1da177e4 | 373 | /* |
1742f19f AM |
374 | * pageout is called by shrink_page_list() for each dirty page. |
375 | * Calls ->writepage(). | |
1da177e4 | 376 | */ |
c661b078 | 377 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 378 | struct scan_control *sc) |
1da177e4 LT |
379 | { |
380 | /* | |
381 | * If the page is dirty, only perform writeback if that write | |
382 | * will be non-blocking. To prevent this allocation from being | |
383 | * stalled by pagecache activity. But note that there may be | |
384 | * stalls if we need to run get_block(). We could test | |
385 | * PagePrivate for that. | |
386 | * | |
6aceb53b | 387 | * If this process is currently in __generic_file_aio_write() against |
1da177e4 LT |
388 | * this page's queue, we can perform writeback even if that |
389 | * will block. | |
390 | * | |
391 | * If the page is swapcache, write it back even if that would | |
392 | * block, for some throttling. This happens by accident, because | |
393 | * swap_backing_dev_info is bust: it doesn't reflect the | |
394 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
395 | */ |
396 | if (!is_page_cache_freeable(page)) | |
397 | return PAGE_KEEP; | |
398 | if (!mapping) { | |
399 | /* | |
400 | * Some data journaling orphaned pages can have | |
401 | * page->mapping == NULL while being dirty with clean buffers. | |
402 | */ | |
266cf658 | 403 | if (page_has_private(page)) { |
1da177e4 LT |
404 | if (try_to_free_buffers(page)) { |
405 | ClearPageDirty(page); | |
d40cee24 | 406 | printk("%s: orphaned page\n", __func__); |
1da177e4 LT |
407 | return PAGE_CLEAN; |
408 | } | |
409 | } | |
410 | return PAGE_KEEP; | |
411 | } | |
412 | if (mapping->a_ops->writepage == NULL) | |
413 | return PAGE_ACTIVATE; | |
0e093d99 | 414 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) |
1da177e4 LT |
415 | return PAGE_KEEP; |
416 | ||
417 | if (clear_page_dirty_for_io(page)) { | |
418 | int res; | |
419 | struct writeback_control wbc = { | |
420 | .sync_mode = WB_SYNC_NONE, | |
421 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
422 | .range_start = 0, |
423 | .range_end = LLONG_MAX, | |
1da177e4 LT |
424 | .for_reclaim = 1, |
425 | }; | |
426 | ||
427 | SetPageReclaim(page); | |
428 | res = mapping->a_ops->writepage(page, &wbc); | |
429 | if (res < 0) | |
430 | handle_write_error(mapping, page, res); | |
994fc28c | 431 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
432 | ClearPageReclaim(page); |
433 | return PAGE_ACTIVATE; | |
434 | } | |
c661b078 | 435 | |
1da177e4 LT |
436 | if (!PageWriteback(page)) { |
437 | /* synchronous write or broken a_ops? */ | |
438 | ClearPageReclaim(page); | |
439 | } | |
23b9da55 | 440 | trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); |
e129b5c2 | 441 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
442 | return PAGE_SUCCESS; |
443 | } | |
444 | ||
445 | return PAGE_CLEAN; | |
446 | } | |
447 | ||
a649fd92 | 448 | /* |
e286781d NP |
449 | * Same as remove_mapping, but if the page is removed from the mapping, it |
450 | * gets returned with a refcount of 0. | |
a649fd92 | 451 | */ |
e286781d | 452 | static int __remove_mapping(struct address_space *mapping, struct page *page) |
49d2e9cc | 453 | { |
28e4d965 NP |
454 | BUG_ON(!PageLocked(page)); |
455 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 456 | |
19fd6231 | 457 | spin_lock_irq(&mapping->tree_lock); |
49d2e9cc | 458 | /* |
0fd0e6b0 NP |
459 | * The non racy check for a busy page. |
460 | * | |
461 | * Must be careful with the order of the tests. When someone has | |
462 | * a ref to the page, it may be possible that they dirty it then | |
463 | * drop the reference. So if PageDirty is tested before page_count | |
464 | * here, then the following race may occur: | |
465 | * | |
466 | * get_user_pages(&page); | |
467 | * [user mapping goes away] | |
468 | * write_to(page); | |
469 | * !PageDirty(page) [good] | |
470 | * SetPageDirty(page); | |
471 | * put_page(page); | |
472 | * !page_count(page) [good, discard it] | |
473 | * | |
474 | * [oops, our write_to data is lost] | |
475 | * | |
476 | * Reversing the order of the tests ensures such a situation cannot | |
477 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
478 | * load is not satisfied before that of page->_count. | |
479 | * | |
480 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
481 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 482 | */ |
e286781d | 483 | if (!page_freeze_refs(page, 2)) |
49d2e9cc | 484 | goto cannot_free; |
e286781d NP |
485 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
486 | if (unlikely(PageDirty(page))) { | |
487 | page_unfreeze_refs(page, 2); | |
49d2e9cc | 488 | goto cannot_free; |
e286781d | 489 | } |
49d2e9cc CL |
490 | |
491 | if (PageSwapCache(page)) { | |
492 | swp_entry_t swap = { .val = page_private(page) }; | |
493 | __delete_from_swap_cache(page); | |
19fd6231 | 494 | spin_unlock_irq(&mapping->tree_lock); |
cb4b86ba | 495 | swapcache_free(swap, page); |
e286781d | 496 | } else { |
6072d13c LT |
497 | void (*freepage)(struct page *); |
498 | ||
499 | freepage = mapping->a_ops->freepage; | |
500 | ||
e64a782f | 501 | __delete_from_page_cache(page); |
19fd6231 | 502 | spin_unlock_irq(&mapping->tree_lock); |
e767e056 | 503 | mem_cgroup_uncharge_cache_page(page); |
6072d13c LT |
504 | |
505 | if (freepage != NULL) | |
506 | freepage(page); | |
49d2e9cc CL |
507 | } |
508 | ||
49d2e9cc CL |
509 | return 1; |
510 | ||
511 | cannot_free: | |
19fd6231 | 512 | spin_unlock_irq(&mapping->tree_lock); |
49d2e9cc CL |
513 | return 0; |
514 | } | |
515 | ||
e286781d NP |
516 | /* |
517 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
518 | * someone else has a ref on the page, abort and return 0. If it was | |
519 | * successfully detached, return 1. Assumes the caller has a single ref on | |
520 | * this page. | |
521 | */ | |
522 | int remove_mapping(struct address_space *mapping, struct page *page) | |
523 | { | |
524 | if (__remove_mapping(mapping, page)) { | |
525 | /* | |
526 | * Unfreezing the refcount with 1 rather than 2 effectively | |
527 | * drops the pagecache ref for us without requiring another | |
528 | * atomic operation. | |
529 | */ | |
530 | page_unfreeze_refs(page, 1); | |
531 | return 1; | |
532 | } | |
533 | return 0; | |
534 | } | |
535 | ||
894bc310 LS |
536 | /** |
537 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
538 | * @page: page to be put back to appropriate lru list | |
539 | * | |
540 | * Add previously isolated @page to appropriate LRU list. | |
541 | * Page may still be unevictable for other reasons. | |
542 | * | |
543 | * lru_lock must not be held, interrupts must be enabled. | |
544 | */ | |
894bc310 LS |
545 | void putback_lru_page(struct page *page) |
546 | { | |
547 | int lru; | |
548 | int active = !!TestClearPageActive(page); | |
bbfd28ee | 549 | int was_unevictable = PageUnevictable(page); |
894bc310 LS |
550 | |
551 | VM_BUG_ON(PageLRU(page)); | |
552 | ||
553 | redo: | |
554 | ClearPageUnevictable(page); | |
555 | ||
39b5f29a | 556 | if (page_evictable(page)) { |
894bc310 LS |
557 | /* |
558 | * For evictable pages, we can use the cache. | |
559 | * In event of a race, worst case is we end up with an | |
560 | * unevictable page on [in]active list. | |
561 | * We know how to handle that. | |
562 | */ | |
401a8e1c | 563 | lru = active + page_lru_base_type(page); |
894bc310 LS |
564 | lru_cache_add_lru(page, lru); |
565 | } else { | |
566 | /* | |
567 | * Put unevictable pages directly on zone's unevictable | |
568 | * list. | |
569 | */ | |
570 | lru = LRU_UNEVICTABLE; | |
571 | add_page_to_unevictable_list(page); | |
6a7b9548 | 572 | /* |
21ee9f39 MK |
573 | * When racing with an mlock or AS_UNEVICTABLE clearing |
574 | * (page is unlocked) make sure that if the other thread | |
575 | * does not observe our setting of PG_lru and fails | |
24513264 | 576 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 577 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
578 | * the page back to the evictable list. |
579 | * | |
21ee9f39 | 580 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
581 | */ |
582 | smp_mb(); | |
894bc310 | 583 | } |
894bc310 LS |
584 | |
585 | /* | |
586 | * page's status can change while we move it among lru. If an evictable | |
587 | * page is on unevictable list, it never be freed. To avoid that, | |
588 | * check after we added it to the list, again. | |
589 | */ | |
39b5f29a | 590 | if (lru == LRU_UNEVICTABLE && page_evictable(page)) { |
894bc310 LS |
591 | if (!isolate_lru_page(page)) { |
592 | put_page(page); | |
593 | goto redo; | |
594 | } | |
595 | /* This means someone else dropped this page from LRU | |
596 | * So, it will be freed or putback to LRU again. There is | |
597 | * nothing to do here. | |
598 | */ | |
599 | } | |
600 | ||
bbfd28ee LS |
601 | if (was_unevictable && lru != LRU_UNEVICTABLE) |
602 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
603 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) | |
604 | count_vm_event(UNEVICTABLE_PGCULLED); | |
605 | ||
894bc310 LS |
606 | put_page(page); /* drop ref from isolate */ |
607 | } | |
608 | ||
dfc8d636 JW |
609 | enum page_references { |
610 | PAGEREF_RECLAIM, | |
611 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 612 | PAGEREF_KEEP, |
dfc8d636 JW |
613 | PAGEREF_ACTIVATE, |
614 | }; | |
615 | ||
616 | static enum page_references page_check_references(struct page *page, | |
617 | struct scan_control *sc) | |
618 | { | |
64574746 | 619 | int referenced_ptes, referenced_page; |
dfc8d636 | 620 | unsigned long vm_flags; |
dfc8d636 | 621 | |
c3ac9a8a JW |
622 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
623 | &vm_flags); | |
64574746 | 624 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 625 | |
dfc8d636 JW |
626 | /* |
627 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
628 | * move the page to the unevictable list. | |
629 | */ | |
630 | if (vm_flags & VM_LOCKED) | |
631 | return PAGEREF_RECLAIM; | |
632 | ||
64574746 | 633 | if (referenced_ptes) { |
e4898273 | 634 | if (PageSwapBacked(page)) |
64574746 JW |
635 | return PAGEREF_ACTIVATE; |
636 | /* | |
637 | * All mapped pages start out with page table | |
638 | * references from the instantiating fault, so we need | |
639 | * to look twice if a mapped file page is used more | |
640 | * than once. | |
641 | * | |
642 | * Mark it and spare it for another trip around the | |
643 | * inactive list. Another page table reference will | |
644 | * lead to its activation. | |
645 | * | |
646 | * Note: the mark is set for activated pages as well | |
647 | * so that recently deactivated but used pages are | |
648 | * quickly recovered. | |
649 | */ | |
650 | SetPageReferenced(page); | |
651 | ||
34dbc67a | 652 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
653 | return PAGEREF_ACTIVATE; |
654 | ||
c909e993 KK |
655 | /* |
656 | * Activate file-backed executable pages after first usage. | |
657 | */ | |
658 | if (vm_flags & VM_EXEC) | |
659 | return PAGEREF_ACTIVATE; | |
660 | ||
64574746 JW |
661 | return PAGEREF_KEEP; |
662 | } | |
dfc8d636 JW |
663 | |
664 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 665 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
666 | return PAGEREF_RECLAIM_CLEAN; |
667 | ||
668 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
669 | } |
670 | ||
1da177e4 | 671 | /* |
1742f19f | 672 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 673 | */ |
1742f19f | 674 | static unsigned long shrink_page_list(struct list_head *page_list, |
6a18adb3 | 675 | struct zone *zone, |
f84f6e2b | 676 | struct scan_control *sc, |
02c6de8d | 677 | enum ttu_flags ttu_flags, |
92df3a72 | 678 | unsigned long *ret_nr_dirty, |
02c6de8d MK |
679 | unsigned long *ret_nr_writeback, |
680 | bool force_reclaim) | |
1da177e4 LT |
681 | { |
682 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 683 | LIST_HEAD(free_pages); |
1da177e4 | 684 | int pgactivate = 0; |
0e093d99 MG |
685 | unsigned long nr_dirty = 0; |
686 | unsigned long nr_congested = 0; | |
05ff5137 | 687 | unsigned long nr_reclaimed = 0; |
92df3a72 | 688 | unsigned long nr_writeback = 0; |
1da177e4 LT |
689 | |
690 | cond_resched(); | |
691 | ||
69980e31 | 692 | mem_cgroup_uncharge_start(); |
1da177e4 LT |
693 | while (!list_empty(page_list)) { |
694 | struct address_space *mapping; | |
695 | struct page *page; | |
696 | int may_enter_fs; | |
02c6de8d | 697 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
1da177e4 LT |
698 | |
699 | cond_resched(); | |
700 | ||
701 | page = lru_to_page(page_list); | |
702 | list_del(&page->lru); | |
703 | ||
529ae9aa | 704 | if (!trylock_page(page)) |
1da177e4 LT |
705 | goto keep; |
706 | ||
725d704e | 707 | VM_BUG_ON(PageActive(page)); |
6a18adb3 | 708 | VM_BUG_ON(page_zone(page) != zone); |
1da177e4 LT |
709 | |
710 | sc->nr_scanned++; | |
80e43426 | 711 | |
39b5f29a | 712 | if (unlikely(!page_evictable(page))) |
b291f000 | 713 | goto cull_mlocked; |
894bc310 | 714 | |
a6dc60f8 | 715 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
716 | goto keep_locked; |
717 | ||
1da177e4 LT |
718 | /* Double the slab pressure for mapped and swapcache pages */ |
719 | if (page_mapped(page) || PageSwapCache(page)) | |
720 | sc->nr_scanned++; | |
721 | ||
c661b078 AW |
722 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
723 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
724 | ||
725 | if (PageWriteback(page)) { | |
e62e384e MH |
726 | /* |
727 | * memcg doesn't have any dirty pages throttling so we | |
728 | * could easily OOM just because too many pages are in | |
c3b94f44 | 729 | * writeback and there is nothing else to reclaim. |
e62e384e | 730 | * |
c3b94f44 | 731 | * Check __GFP_IO, certainly because a loop driver |
e62e384e MH |
732 | * thread might enter reclaim, and deadlock if it waits |
733 | * on a page for which it is needed to do the write | |
734 | * (loop masks off __GFP_IO|__GFP_FS for this reason); | |
735 | * but more thought would probably show more reasons. | |
c3b94f44 HD |
736 | * |
737 | * Don't require __GFP_FS, since we're not going into | |
738 | * the FS, just waiting on its writeback completion. | |
739 | * Worryingly, ext4 gfs2 and xfs allocate pages with | |
740 | * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so | |
741 | * testing may_enter_fs here is liable to OOM on them. | |
e62e384e | 742 | */ |
c3b94f44 HD |
743 | if (global_reclaim(sc) || |
744 | !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) { | |
745 | /* | |
746 | * This is slightly racy - end_page_writeback() | |
747 | * might have just cleared PageReclaim, then | |
748 | * setting PageReclaim here end up interpreted | |
749 | * as PageReadahead - but that does not matter | |
750 | * enough to care. What we do want is for this | |
751 | * page to have PageReclaim set next time memcg | |
752 | * reclaim reaches the tests above, so it will | |
753 | * then wait_on_page_writeback() to avoid OOM; | |
754 | * and it's also appropriate in global reclaim. | |
755 | */ | |
756 | SetPageReclaim(page); | |
e62e384e | 757 | nr_writeback++; |
c3b94f44 | 758 | goto keep_locked; |
e62e384e | 759 | } |
c3b94f44 | 760 | wait_on_page_writeback(page); |
c661b078 | 761 | } |
1da177e4 | 762 | |
02c6de8d MK |
763 | if (!force_reclaim) |
764 | references = page_check_references(page, sc); | |
765 | ||
dfc8d636 JW |
766 | switch (references) { |
767 | case PAGEREF_ACTIVATE: | |
1da177e4 | 768 | goto activate_locked; |
64574746 JW |
769 | case PAGEREF_KEEP: |
770 | goto keep_locked; | |
dfc8d636 JW |
771 | case PAGEREF_RECLAIM: |
772 | case PAGEREF_RECLAIM_CLEAN: | |
773 | ; /* try to reclaim the page below */ | |
774 | } | |
1da177e4 | 775 | |
1da177e4 LT |
776 | /* |
777 | * Anonymous process memory has backing store? | |
778 | * Try to allocate it some swap space here. | |
779 | */ | |
b291f000 | 780 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
781 | if (!(sc->gfp_mask & __GFP_IO)) |
782 | goto keep_locked; | |
ac47b003 | 783 | if (!add_to_swap(page)) |
1da177e4 | 784 | goto activate_locked; |
63eb6b93 | 785 | may_enter_fs = 1; |
b291f000 | 786 | } |
1da177e4 LT |
787 | |
788 | mapping = page_mapping(page); | |
1da177e4 LT |
789 | |
790 | /* | |
791 | * The page is mapped into the page tables of one or more | |
792 | * processes. Try to unmap it here. | |
793 | */ | |
794 | if (page_mapped(page) && mapping) { | |
02c6de8d | 795 | switch (try_to_unmap(page, ttu_flags)) { |
1da177e4 LT |
796 | case SWAP_FAIL: |
797 | goto activate_locked; | |
798 | case SWAP_AGAIN: | |
799 | goto keep_locked; | |
b291f000 NP |
800 | case SWAP_MLOCK: |
801 | goto cull_mlocked; | |
1da177e4 LT |
802 | case SWAP_SUCCESS: |
803 | ; /* try to free the page below */ | |
804 | } | |
805 | } | |
806 | ||
807 | if (PageDirty(page)) { | |
0e093d99 MG |
808 | nr_dirty++; |
809 | ||
ee72886d MG |
810 | /* |
811 | * Only kswapd can writeback filesystem pages to | |
f84f6e2b MG |
812 | * avoid risk of stack overflow but do not writeback |
813 | * unless under significant pressure. | |
ee72886d | 814 | */ |
f84f6e2b | 815 | if (page_is_file_cache(page) && |
9e3b2f8c KK |
816 | (!current_is_kswapd() || |
817 | sc->priority >= DEF_PRIORITY - 2)) { | |
49ea7eb6 MG |
818 | /* |
819 | * Immediately reclaim when written back. | |
820 | * Similar in principal to deactivate_page() | |
821 | * except we already have the page isolated | |
822 | * and know it's dirty | |
823 | */ | |
824 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
825 | SetPageReclaim(page); | |
826 | ||
ee72886d MG |
827 | goto keep_locked; |
828 | } | |
829 | ||
dfc8d636 | 830 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 831 | goto keep_locked; |
4dd4b920 | 832 | if (!may_enter_fs) |
1da177e4 | 833 | goto keep_locked; |
52a8363e | 834 | if (!sc->may_writepage) |
1da177e4 LT |
835 | goto keep_locked; |
836 | ||
837 | /* Page is dirty, try to write it out here */ | |
7d3579e8 | 838 | switch (pageout(page, mapping, sc)) { |
1da177e4 | 839 | case PAGE_KEEP: |
0e093d99 | 840 | nr_congested++; |
1da177e4 LT |
841 | goto keep_locked; |
842 | case PAGE_ACTIVATE: | |
843 | goto activate_locked; | |
844 | case PAGE_SUCCESS: | |
7d3579e8 | 845 | if (PageWriteback(page)) |
41ac1999 | 846 | goto keep; |
7d3579e8 | 847 | if (PageDirty(page)) |
1da177e4 | 848 | goto keep; |
7d3579e8 | 849 | |
1da177e4 LT |
850 | /* |
851 | * A synchronous write - probably a ramdisk. Go | |
852 | * ahead and try to reclaim the page. | |
853 | */ | |
529ae9aa | 854 | if (!trylock_page(page)) |
1da177e4 LT |
855 | goto keep; |
856 | if (PageDirty(page) || PageWriteback(page)) | |
857 | goto keep_locked; | |
858 | mapping = page_mapping(page); | |
859 | case PAGE_CLEAN: | |
860 | ; /* try to free the page below */ | |
861 | } | |
862 | } | |
863 | ||
864 | /* | |
865 | * If the page has buffers, try to free the buffer mappings | |
866 | * associated with this page. If we succeed we try to free | |
867 | * the page as well. | |
868 | * | |
869 | * We do this even if the page is PageDirty(). | |
870 | * try_to_release_page() does not perform I/O, but it is | |
871 | * possible for a page to have PageDirty set, but it is actually | |
872 | * clean (all its buffers are clean). This happens if the | |
873 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 874 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
875 | * try_to_release_page() will discover that cleanness and will |
876 | * drop the buffers and mark the page clean - it can be freed. | |
877 | * | |
878 | * Rarely, pages can have buffers and no ->mapping. These are | |
879 | * the pages which were not successfully invalidated in | |
880 | * truncate_complete_page(). We try to drop those buffers here | |
881 | * and if that worked, and the page is no longer mapped into | |
882 | * process address space (page_count == 1) it can be freed. | |
883 | * Otherwise, leave the page on the LRU so it is swappable. | |
884 | */ | |
266cf658 | 885 | if (page_has_private(page)) { |
1da177e4 LT |
886 | if (!try_to_release_page(page, sc->gfp_mask)) |
887 | goto activate_locked; | |
e286781d NP |
888 | if (!mapping && page_count(page) == 1) { |
889 | unlock_page(page); | |
890 | if (put_page_testzero(page)) | |
891 | goto free_it; | |
892 | else { | |
893 | /* | |
894 | * rare race with speculative reference. | |
895 | * the speculative reference will free | |
896 | * this page shortly, so we may | |
897 | * increment nr_reclaimed here (and | |
898 | * leave it off the LRU). | |
899 | */ | |
900 | nr_reclaimed++; | |
901 | continue; | |
902 | } | |
903 | } | |
1da177e4 LT |
904 | } |
905 | ||
e286781d | 906 | if (!mapping || !__remove_mapping(mapping, page)) |
49d2e9cc | 907 | goto keep_locked; |
1da177e4 | 908 | |
a978d6f5 NP |
909 | /* |
910 | * At this point, we have no other references and there is | |
911 | * no way to pick any more up (removed from LRU, removed | |
912 | * from pagecache). Can use non-atomic bitops now (and | |
913 | * we obviously don't have to worry about waking up a process | |
914 | * waiting on the page lock, because there are no references. | |
915 | */ | |
916 | __clear_page_locked(page); | |
e286781d | 917 | free_it: |
05ff5137 | 918 | nr_reclaimed++; |
abe4c3b5 MG |
919 | |
920 | /* | |
921 | * Is there need to periodically free_page_list? It would | |
922 | * appear not as the counts should be low | |
923 | */ | |
924 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
925 | continue; |
926 | ||
b291f000 | 927 | cull_mlocked: |
63d6c5ad HD |
928 | if (PageSwapCache(page)) |
929 | try_to_free_swap(page); | |
b291f000 NP |
930 | unlock_page(page); |
931 | putback_lru_page(page); | |
932 | continue; | |
933 | ||
1da177e4 | 934 | activate_locked: |
68a22394 RR |
935 | /* Not a candidate for swapping, so reclaim swap space. */ |
936 | if (PageSwapCache(page) && vm_swap_full()) | |
a2c43eed | 937 | try_to_free_swap(page); |
894bc310 | 938 | VM_BUG_ON(PageActive(page)); |
1da177e4 LT |
939 | SetPageActive(page); |
940 | pgactivate++; | |
941 | keep_locked: | |
942 | unlock_page(page); | |
943 | keep: | |
944 | list_add(&page->lru, &ret_pages); | |
b291f000 | 945 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); |
1da177e4 | 946 | } |
abe4c3b5 | 947 | |
0e093d99 MG |
948 | /* |
949 | * Tag a zone as congested if all the dirty pages encountered were | |
950 | * backed by a congested BDI. In this case, reclaimers should just | |
951 | * back off and wait for congestion to clear because further reclaim | |
952 | * will encounter the same problem | |
953 | */ | |
89b5fae5 | 954 | if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc)) |
6a18adb3 | 955 | zone_set_flag(zone, ZONE_CONGESTED); |
0e093d99 | 956 | |
cc59850e | 957 | free_hot_cold_page_list(&free_pages, 1); |
abe4c3b5 | 958 | |
1da177e4 | 959 | list_splice(&ret_pages, page_list); |
f8891e5e | 960 | count_vm_events(PGACTIVATE, pgactivate); |
69980e31 | 961 | mem_cgroup_uncharge_end(); |
92df3a72 MG |
962 | *ret_nr_dirty += nr_dirty; |
963 | *ret_nr_writeback += nr_writeback; | |
05ff5137 | 964 | return nr_reclaimed; |
1da177e4 LT |
965 | } |
966 | ||
02c6de8d MK |
967 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
968 | struct list_head *page_list) | |
969 | { | |
970 | struct scan_control sc = { | |
971 | .gfp_mask = GFP_KERNEL, | |
972 | .priority = DEF_PRIORITY, | |
973 | .may_unmap = 1, | |
974 | }; | |
975 | unsigned long ret, dummy1, dummy2; | |
976 | struct page *page, *next; | |
977 | LIST_HEAD(clean_pages); | |
978 | ||
979 | list_for_each_entry_safe(page, next, page_list, lru) { | |
980 | if (page_is_file_cache(page) && !PageDirty(page)) { | |
981 | ClearPageActive(page); | |
982 | list_move(&page->lru, &clean_pages); | |
983 | } | |
984 | } | |
985 | ||
986 | ret = shrink_page_list(&clean_pages, zone, &sc, | |
987 | TTU_UNMAP|TTU_IGNORE_ACCESS, | |
988 | &dummy1, &dummy2, true); | |
989 | list_splice(&clean_pages, page_list); | |
990 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); | |
991 | return ret; | |
992 | } | |
993 | ||
5ad333eb AW |
994 | /* |
995 | * Attempt to remove the specified page from its LRU. Only take this page | |
996 | * if it is of the appropriate PageActive status. Pages which are being | |
997 | * freed elsewhere are also ignored. | |
998 | * | |
999 | * page: page to consider | |
1000 | * mode: one of the LRU isolation modes defined above | |
1001 | * | |
1002 | * returns 0 on success, -ve errno on failure. | |
1003 | */ | |
f3fd4a61 | 1004 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1005 | { |
1006 | int ret = -EINVAL; | |
1007 | ||
1008 | /* Only take pages on the LRU. */ | |
1009 | if (!PageLRU(page)) | |
1010 | return ret; | |
1011 | ||
e46a2879 MK |
1012 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1013 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1014 | return ret; |
1015 | ||
5ad333eb | 1016 | ret = -EBUSY; |
08e552c6 | 1017 | |
c8244935 MG |
1018 | /* |
1019 | * To minimise LRU disruption, the caller can indicate that it only | |
1020 | * wants to isolate pages it will be able to operate on without | |
1021 | * blocking - clean pages for the most part. | |
1022 | * | |
1023 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1024 | * is used by reclaim when it is cannot write to backing storage | |
1025 | * | |
1026 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1027 | * that it is possible to migrate without blocking | |
1028 | */ | |
1029 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1030 | /* All the caller can do on PageWriteback is block */ | |
1031 | if (PageWriteback(page)) | |
1032 | return ret; | |
1033 | ||
1034 | if (PageDirty(page)) { | |
1035 | struct address_space *mapping; | |
1036 | ||
1037 | /* ISOLATE_CLEAN means only clean pages */ | |
1038 | if (mode & ISOLATE_CLEAN) | |
1039 | return ret; | |
1040 | ||
1041 | /* | |
1042 | * Only pages without mappings or that have a | |
1043 | * ->migratepage callback are possible to migrate | |
1044 | * without blocking | |
1045 | */ | |
1046 | mapping = page_mapping(page); | |
1047 | if (mapping && !mapping->a_ops->migratepage) | |
1048 | return ret; | |
1049 | } | |
1050 | } | |
39deaf85 | 1051 | |
f80c0673 MK |
1052 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1053 | return ret; | |
1054 | ||
5ad333eb AW |
1055 | if (likely(get_page_unless_zero(page))) { |
1056 | /* | |
1057 | * Be careful not to clear PageLRU until after we're | |
1058 | * sure the page is not being freed elsewhere -- the | |
1059 | * page release code relies on it. | |
1060 | */ | |
1061 | ClearPageLRU(page); | |
1062 | ret = 0; | |
1063 | } | |
1064 | ||
1065 | return ret; | |
1066 | } | |
1067 | ||
1da177e4 LT |
1068 | /* |
1069 | * zone->lru_lock is heavily contended. Some of the functions that | |
1070 | * shrink the lists perform better by taking out a batch of pages | |
1071 | * and working on them outside the LRU lock. | |
1072 | * | |
1073 | * For pagecache intensive workloads, this function is the hottest | |
1074 | * spot in the kernel (apart from copy_*_user functions). | |
1075 | * | |
1076 | * Appropriate locks must be held before calling this function. | |
1077 | * | |
1078 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1079 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1080 | * @dst: The temp list to put pages on to. |
f626012d | 1081 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1082 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1083 | * @mode: One of the LRU isolation modes |
3cb99451 | 1084 | * @lru: LRU list id for isolating |
1da177e4 LT |
1085 | * |
1086 | * returns how many pages were moved onto *@dst. | |
1087 | */ | |
69e05944 | 1088 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1089 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1090 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1091 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1092 | { |
75b00af7 | 1093 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1094 | unsigned long nr_taken = 0; |
c9b02d97 | 1095 | unsigned long scan; |
1da177e4 | 1096 | |
c9b02d97 | 1097 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
5ad333eb | 1098 | struct page *page; |
fa9add64 | 1099 | int nr_pages; |
5ad333eb | 1100 | |
1da177e4 LT |
1101 | page = lru_to_page(src); |
1102 | prefetchw_prev_lru_page(page, src, flags); | |
1103 | ||
725d704e | 1104 | VM_BUG_ON(!PageLRU(page)); |
8d438f96 | 1105 | |
f3fd4a61 | 1106 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1107 | case 0: |
fa9add64 HD |
1108 | nr_pages = hpage_nr_pages(page); |
1109 | mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); | |
5ad333eb | 1110 | list_move(&page->lru, dst); |
fa9add64 | 1111 | nr_taken += nr_pages; |
5ad333eb AW |
1112 | break; |
1113 | ||
1114 | case -EBUSY: | |
1115 | /* else it is being freed elsewhere */ | |
1116 | list_move(&page->lru, src); | |
1117 | continue; | |
46453a6e | 1118 | |
5ad333eb AW |
1119 | default: |
1120 | BUG(); | |
1121 | } | |
1da177e4 LT |
1122 | } |
1123 | ||
f626012d | 1124 | *nr_scanned = scan; |
75b00af7 HD |
1125 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, |
1126 | nr_taken, mode, is_file_lru(lru)); | |
1da177e4 LT |
1127 | return nr_taken; |
1128 | } | |
1129 | ||
62695a84 NP |
1130 | /** |
1131 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1132 | * @page: page to isolate from its LRU list | |
1133 | * | |
1134 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1135 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1136 | * | |
1137 | * Returns 0 if the page was removed from an LRU list. | |
1138 | * Returns -EBUSY if the page was not on an LRU list. | |
1139 | * | |
1140 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1141 | * the active list, it will have PageActive set. If it was found on |
1142 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1143 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1144 | * |
1145 | * The vmstat statistic corresponding to the list on which the page was | |
1146 | * found will be decremented. | |
1147 | * | |
1148 | * Restrictions: | |
1149 | * (1) Must be called with an elevated refcount on the page. This is a | |
1150 | * fundamentnal difference from isolate_lru_pages (which is called | |
1151 | * without a stable reference). | |
1152 | * (2) the lru_lock must not be held. | |
1153 | * (3) interrupts must be enabled. | |
1154 | */ | |
1155 | int isolate_lru_page(struct page *page) | |
1156 | { | |
1157 | int ret = -EBUSY; | |
1158 | ||
0c917313 KK |
1159 | VM_BUG_ON(!page_count(page)); |
1160 | ||
62695a84 NP |
1161 | if (PageLRU(page)) { |
1162 | struct zone *zone = page_zone(page); | |
fa9add64 | 1163 | struct lruvec *lruvec; |
62695a84 NP |
1164 | |
1165 | spin_lock_irq(&zone->lru_lock); | |
fa9add64 | 1166 | lruvec = mem_cgroup_page_lruvec(page, zone); |
0c917313 | 1167 | if (PageLRU(page)) { |
894bc310 | 1168 | int lru = page_lru(page); |
0c917313 | 1169 | get_page(page); |
62695a84 | 1170 | ClearPageLRU(page); |
fa9add64 HD |
1171 | del_page_from_lru_list(page, lruvec, lru); |
1172 | ret = 0; | |
62695a84 NP |
1173 | } |
1174 | spin_unlock_irq(&zone->lru_lock); | |
1175 | } | |
1176 | return ret; | |
1177 | } | |
1178 | ||
35cd7815 | 1179 | /* |
d37dd5dc FW |
1180 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1181 | * then get resheduled. When there are massive number of tasks doing page | |
1182 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1183 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1184 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 RR |
1185 | */ |
1186 | static int too_many_isolated(struct zone *zone, int file, | |
1187 | struct scan_control *sc) | |
1188 | { | |
1189 | unsigned long inactive, isolated; | |
1190 | ||
1191 | if (current_is_kswapd()) | |
1192 | return 0; | |
1193 | ||
89b5fae5 | 1194 | if (!global_reclaim(sc)) |
35cd7815 RR |
1195 | return 0; |
1196 | ||
1197 | if (file) { | |
1198 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1199 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1200 | } else { | |
1201 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1202 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1203 | } | |
1204 | ||
3cf23841 FW |
1205 | /* |
1206 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1207 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1208 | * deadlock. | |
1209 | */ | |
1210 | if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS) | |
1211 | inactive >>= 3; | |
1212 | ||
35cd7815 RR |
1213 | return isolated > inactive; |
1214 | } | |
1215 | ||
66635629 | 1216 | static noinline_for_stack void |
75b00af7 | 1217 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1218 | { |
27ac81d8 KK |
1219 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1220 | struct zone *zone = lruvec_zone(lruvec); | |
3f79768f | 1221 | LIST_HEAD(pages_to_free); |
66635629 | 1222 | |
66635629 MG |
1223 | /* |
1224 | * Put back any unfreeable pages. | |
1225 | */ | |
66635629 | 1226 | while (!list_empty(page_list)) { |
3f79768f | 1227 | struct page *page = lru_to_page(page_list); |
66635629 | 1228 | int lru; |
3f79768f | 1229 | |
66635629 MG |
1230 | VM_BUG_ON(PageLRU(page)); |
1231 | list_del(&page->lru); | |
39b5f29a | 1232 | if (unlikely(!page_evictable(page))) { |
66635629 MG |
1233 | spin_unlock_irq(&zone->lru_lock); |
1234 | putback_lru_page(page); | |
1235 | spin_lock_irq(&zone->lru_lock); | |
1236 | continue; | |
1237 | } | |
fa9add64 HD |
1238 | |
1239 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1240 | ||
7a608572 | 1241 | SetPageLRU(page); |
66635629 | 1242 | lru = page_lru(page); |
fa9add64 HD |
1243 | add_page_to_lru_list(page, lruvec, lru); |
1244 | ||
66635629 MG |
1245 | if (is_active_lru(lru)) { |
1246 | int file = is_file_lru(lru); | |
9992af10 RR |
1247 | int numpages = hpage_nr_pages(page); |
1248 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1249 | } |
2bcf8879 HD |
1250 | if (put_page_testzero(page)) { |
1251 | __ClearPageLRU(page); | |
1252 | __ClearPageActive(page); | |
fa9add64 | 1253 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1254 | |
1255 | if (unlikely(PageCompound(page))) { | |
1256 | spin_unlock_irq(&zone->lru_lock); | |
1257 | (*get_compound_page_dtor(page))(page); | |
1258 | spin_lock_irq(&zone->lru_lock); | |
1259 | } else | |
1260 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1261 | } |
1262 | } | |
66635629 | 1263 | |
3f79768f HD |
1264 | /* |
1265 | * To save our caller's stack, now use input list for pages to free. | |
1266 | */ | |
1267 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1268 | } |
1269 | ||
1da177e4 | 1270 | /* |
1742f19f AM |
1271 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1272 | * of reclaimed pages | |
1da177e4 | 1273 | */ |
66635629 | 1274 | static noinline_for_stack unsigned long |
1a93be0e | 1275 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1276 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1277 | { |
1278 | LIST_HEAD(page_list); | |
e247dbce | 1279 | unsigned long nr_scanned; |
05ff5137 | 1280 | unsigned long nr_reclaimed = 0; |
e247dbce | 1281 | unsigned long nr_taken; |
92df3a72 MG |
1282 | unsigned long nr_dirty = 0; |
1283 | unsigned long nr_writeback = 0; | |
f3fd4a61 | 1284 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1285 | int file = is_file_lru(lru); |
1a93be0e KK |
1286 | struct zone *zone = lruvec_zone(lruvec); |
1287 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
78dc583d | 1288 | |
35cd7815 | 1289 | while (unlikely(too_many_isolated(zone, file, sc))) { |
58355c78 | 1290 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1291 | |
1292 | /* We are about to die and free our memory. Return now. */ | |
1293 | if (fatal_signal_pending(current)) | |
1294 | return SWAP_CLUSTER_MAX; | |
1295 | } | |
1296 | ||
1da177e4 | 1297 | lru_add_drain(); |
f80c0673 MK |
1298 | |
1299 | if (!sc->may_unmap) | |
61317289 | 1300 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1301 | if (!sc->may_writepage) |
61317289 | 1302 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1303 | |
1da177e4 | 1304 | spin_lock_irq(&zone->lru_lock); |
b35ea17b | 1305 | |
5dc35979 KK |
1306 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1307 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc KK |
1308 | |
1309 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1310 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1311 | ||
89b5fae5 | 1312 | if (global_reclaim(sc)) { |
e247dbce KM |
1313 | zone->pages_scanned += nr_scanned; |
1314 | if (current_is_kswapd()) | |
75b00af7 | 1315 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); |
e247dbce | 1316 | else |
75b00af7 | 1317 | __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); |
e247dbce | 1318 | } |
d563c050 | 1319 | spin_unlock_irq(&zone->lru_lock); |
b35ea17b | 1320 | |
d563c050 | 1321 | if (nr_taken == 0) |
66635629 | 1322 | return 0; |
5ad333eb | 1323 | |
02c6de8d MK |
1324 | nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, |
1325 | &nr_dirty, &nr_writeback, false); | |
c661b078 | 1326 | |
3f79768f HD |
1327 | spin_lock_irq(&zone->lru_lock); |
1328 | ||
95d918fc | 1329 | reclaim_stat->recent_scanned[file] += nr_taken; |
d563c050 | 1330 | |
904249aa YH |
1331 | if (global_reclaim(sc)) { |
1332 | if (current_is_kswapd()) | |
1333 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1334 | nr_reclaimed); | |
1335 | else | |
1336 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1337 | nr_reclaimed); | |
1338 | } | |
a74609fa | 1339 | |
27ac81d8 | 1340 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1341 | |
95d918fc | 1342 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f HD |
1343 | |
1344 | spin_unlock_irq(&zone->lru_lock); | |
1345 | ||
1346 | free_hot_cold_page_list(&page_list, 1); | |
e11da5b4 | 1347 | |
92df3a72 MG |
1348 | /* |
1349 | * If reclaim is isolating dirty pages under writeback, it implies | |
1350 | * that the long-lived page allocation rate is exceeding the page | |
1351 | * laundering rate. Either the global limits are not being effective | |
1352 | * at throttling processes due to the page distribution throughout | |
1353 | * zones or there is heavy usage of a slow backing device. The | |
1354 | * only option is to throttle from reclaim context which is not ideal | |
1355 | * as there is no guarantee the dirtying process is throttled in the | |
1356 | * same way balance_dirty_pages() manages. | |
1357 | * | |
1358 | * This scales the number of dirty pages that must be under writeback | |
1359 | * before throttling depending on priority. It is a simple backoff | |
1360 | * function that has the most effect in the range DEF_PRIORITY to | |
1361 | * DEF_PRIORITY-2 which is the priority reclaim is considered to be | |
1362 | * in trouble and reclaim is considered to be in trouble. | |
1363 | * | |
1364 | * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle | |
1365 | * DEF_PRIORITY-1 50% must be PageWriteback | |
1366 | * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble | |
1367 | * ... | |
1368 | * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any | |
1369 | * isolated page is PageWriteback | |
1370 | */ | |
9e3b2f8c KK |
1371 | if (nr_writeback && nr_writeback >= |
1372 | (nr_taken >> (DEF_PRIORITY - sc->priority))) | |
92df3a72 MG |
1373 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); |
1374 | ||
e11da5b4 MG |
1375 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, |
1376 | zone_idx(zone), | |
1377 | nr_scanned, nr_reclaimed, | |
9e3b2f8c | 1378 | sc->priority, |
23b9da55 | 1379 | trace_shrink_flags(file)); |
05ff5137 | 1380 | return nr_reclaimed; |
1da177e4 LT |
1381 | } |
1382 | ||
1383 | /* | |
1384 | * This moves pages from the active list to the inactive list. | |
1385 | * | |
1386 | * We move them the other way if the page is referenced by one or more | |
1387 | * processes, from rmap. | |
1388 | * | |
1389 | * If the pages are mostly unmapped, the processing is fast and it is | |
1390 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1391 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1392 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1393 | * this, so instead we remove the pages from the LRU while processing them. | |
1394 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1395 | * nobody will play with that bit on a non-LRU page. | |
1396 | * | |
1397 | * The downside is that we have to touch page->_count against each page. | |
1398 | * But we had to alter page->flags anyway. | |
1399 | */ | |
1cfb419b | 1400 | |
fa9add64 | 1401 | static void move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1402 | struct list_head *list, |
2bcf8879 | 1403 | struct list_head *pages_to_free, |
3eb4140f WF |
1404 | enum lru_list lru) |
1405 | { | |
fa9add64 | 1406 | struct zone *zone = lruvec_zone(lruvec); |
3eb4140f | 1407 | unsigned long pgmoved = 0; |
3eb4140f | 1408 | struct page *page; |
fa9add64 | 1409 | int nr_pages; |
3eb4140f | 1410 | |
3eb4140f WF |
1411 | while (!list_empty(list)) { |
1412 | page = lru_to_page(list); | |
fa9add64 | 1413 | lruvec = mem_cgroup_page_lruvec(page, zone); |
3eb4140f WF |
1414 | |
1415 | VM_BUG_ON(PageLRU(page)); | |
1416 | SetPageLRU(page); | |
1417 | ||
fa9add64 HD |
1418 | nr_pages = hpage_nr_pages(page); |
1419 | mem_cgroup_update_lru_size(lruvec, lru, nr_pages); | |
925b7673 | 1420 | list_move(&page->lru, &lruvec->lists[lru]); |
fa9add64 | 1421 | pgmoved += nr_pages; |
3eb4140f | 1422 | |
2bcf8879 HD |
1423 | if (put_page_testzero(page)) { |
1424 | __ClearPageLRU(page); | |
1425 | __ClearPageActive(page); | |
fa9add64 | 1426 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1427 | |
1428 | if (unlikely(PageCompound(page))) { | |
1429 | spin_unlock_irq(&zone->lru_lock); | |
1430 | (*get_compound_page_dtor(page))(page); | |
1431 | spin_lock_irq(&zone->lru_lock); | |
1432 | } else | |
1433 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1434 | } |
1435 | } | |
1436 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1437 | if (!is_active_lru(lru)) | |
1438 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1439 | } | |
1cfb419b | 1440 | |
f626012d | 1441 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1442 | struct lruvec *lruvec, |
f16015fb | 1443 | struct scan_control *sc, |
9e3b2f8c | 1444 | enum lru_list lru) |
1da177e4 | 1445 | { |
44c241f1 | 1446 | unsigned long nr_taken; |
f626012d | 1447 | unsigned long nr_scanned; |
6fe6b7e3 | 1448 | unsigned long vm_flags; |
1da177e4 | 1449 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1450 | LIST_HEAD(l_active); |
b69408e8 | 1451 | LIST_HEAD(l_inactive); |
1da177e4 | 1452 | struct page *page; |
1a93be0e | 1453 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
44c241f1 | 1454 | unsigned long nr_rotated = 0; |
f3fd4a61 | 1455 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1456 | int file = is_file_lru(lru); |
1a93be0e | 1457 | struct zone *zone = lruvec_zone(lruvec); |
1da177e4 LT |
1458 | |
1459 | lru_add_drain(); | |
f80c0673 MK |
1460 | |
1461 | if (!sc->may_unmap) | |
61317289 | 1462 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1463 | if (!sc->may_writepage) |
61317289 | 1464 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1465 | |
1da177e4 | 1466 | spin_lock_irq(&zone->lru_lock); |
925b7673 | 1467 | |
5dc35979 KK |
1468 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1469 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1470 | if (global_reclaim(sc)) |
f626012d | 1471 | zone->pages_scanned += nr_scanned; |
89b5fae5 | 1472 | |
b7c46d15 | 1473 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1474 | |
f626012d | 1475 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); |
3cb99451 | 1476 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); |
a731286d | 1477 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
1da177e4 LT |
1478 | spin_unlock_irq(&zone->lru_lock); |
1479 | ||
1da177e4 LT |
1480 | while (!list_empty(&l_hold)) { |
1481 | cond_resched(); | |
1482 | page = lru_to_page(&l_hold); | |
1483 | list_del(&page->lru); | |
7e9cd484 | 1484 | |
39b5f29a | 1485 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
1486 | putback_lru_page(page); |
1487 | continue; | |
1488 | } | |
1489 | ||
cc715d99 MG |
1490 | if (unlikely(buffer_heads_over_limit)) { |
1491 | if (page_has_private(page) && trylock_page(page)) { | |
1492 | if (page_has_private(page)) | |
1493 | try_to_release_page(page, 0); | |
1494 | unlock_page(page); | |
1495 | } | |
1496 | } | |
1497 | ||
c3ac9a8a JW |
1498 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1499 | &vm_flags)) { | |
9992af10 | 1500 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1501 | /* |
1502 | * Identify referenced, file-backed active pages and | |
1503 | * give them one more trip around the active list. So | |
1504 | * that executable code get better chances to stay in | |
1505 | * memory under moderate memory pressure. Anon pages | |
1506 | * are not likely to be evicted by use-once streaming | |
1507 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1508 | * so we ignore them here. | |
1509 | */ | |
41e20983 | 1510 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1511 | list_add(&page->lru, &l_active); |
1512 | continue; | |
1513 | } | |
1514 | } | |
7e9cd484 | 1515 | |
5205e56e | 1516 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1517 | list_add(&page->lru, &l_inactive); |
1518 | } | |
1519 | ||
b555749a | 1520 | /* |
8cab4754 | 1521 | * Move pages back to the lru list. |
b555749a | 1522 | */ |
2a1dc509 | 1523 | spin_lock_irq(&zone->lru_lock); |
556adecb | 1524 | /* |
8cab4754 WF |
1525 | * Count referenced pages from currently used mappings as rotated, |
1526 | * even though only some of them are actually re-activated. This | |
1527 | * helps balance scan pressure between file and anonymous pages in | |
1528 | * get_scan_ratio. | |
7e9cd484 | 1529 | */ |
b7c46d15 | 1530 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1531 | |
fa9add64 HD |
1532 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
1533 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
a731286d | 1534 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
f8891e5e | 1535 | spin_unlock_irq(&zone->lru_lock); |
2bcf8879 HD |
1536 | |
1537 | free_hot_cold_page_list(&l_hold, 1); | |
1da177e4 LT |
1538 | } |
1539 | ||
74e3f3c3 | 1540 | #ifdef CONFIG_SWAP |
14797e23 | 1541 | static int inactive_anon_is_low_global(struct zone *zone) |
f89eb90e KM |
1542 | { |
1543 | unsigned long active, inactive; | |
1544 | ||
1545 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1546 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1547 | ||
1548 | if (inactive * zone->inactive_ratio < active) | |
1549 | return 1; | |
1550 | ||
1551 | return 0; | |
1552 | } | |
1553 | ||
14797e23 KM |
1554 | /** |
1555 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
c56d5c7d | 1556 | * @lruvec: LRU vector to check |
14797e23 KM |
1557 | * |
1558 | * Returns true if the zone does not have enough inactive anon pages, | |
1559 | * meaning some active anon pages need to be deactivated. | |
1560 | */ | |
c56d5c7d | 1561 | static int inactive_anon_is_low(struct lruvec *lruvec) |
14797e23 | 1562 | { |
74e3f3c3 MK |
1563 | /* |
1564 | * If we don't have swap space, anonymous page deactivation | |
1565 | * is pointless. | |
1566 | */ | |
1567 | if (!total_swap_pages) | |
1568 | return 0; | |
1569 | ||
c3c787e8 | 1570 | if (!mem_cgroup_disabled()) |
c56d5c7d | 1571 | return mem_cgroup_inactive_anon_is_low(lruvec); |
f16015fb | 1572 | |
c56d5c7d | 1573 | return inactive_anon_is_low_global(lruvec_zone(lruvec)); |
14797e23 | 1574 | } |
74e3f3c3 | 1575 | #else |
c56d5c7d | 1576 | static inline int inactive_anon_is_low(struct lruvec *lruvec) |
74e3f3c3 MK |
1577 | { |
1578 | return 0; | |
1579 | } | |
1580 | #endif | |
14797e23 | 1581 | |
56e49d21 RR |
1582 | static int inactive_file_is_low_global(struct zone *zone) |
1583 | { | |
1584 | unsigned long active, inactive; | |
1585 | ||
1586 | active = zone_page_state(zone, NR_ACTIVE_FILE); | |
1587 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1588 | ||
1589 | return (active > inactive); | |
1590 | } | |
1591 | ||
1592 | /** | |
1593 | * inactive_file_is_low - check if file pages need to be deactivated | |
c56d5c7d | 1594 | * @lruvec: LRU vector to check |
56e49d21 RR |
1595 | * |
1596 | * When the system is doing streaming IO, memory pressure here | |
1597 | * ensures that active file pages get deactivated, until more | |
1598 | * than half of the file pages are on the inactive list. | |
1599 | * | |
1600 | * Once we get to that situation, protect the system's working | |
1601 | * set from being evicted by disabling active file page aging. | |
1602 | * | |
1603 | * This uses a different ratio than the anonymous pages, because | |
1604 | * the page cache uses a use-once replacement algorithm. | |
1605 | */ | |
c56d5c7d | 1606 | static int inactive_file_is_low(struct lruvec *lruvec) |
56e49d21 | 1607 | { |
c3c787e8 | 1608 | if (!mem_cgroup_disabled()) |
c56d5c7d | 1609 | return mem_cgroup_inactive_file_is_low(lruvec); |
56e49d21 | 1610 | |
c56d5c7d | 1611 | return inactive_file_is_low_global(lruvec_zone(lruvec)); |
56e49d21 RR |
1612 | } |
1613 | ||
75b00af7 | 1614 | static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) |
b39415b2 | 1615 | { |
75b00af7 | 1616 | if (is_file_lru(lru)) |
c56d5c7d | 1617 | return inactive_file_is_low(lruvec); |
b39415b2 | 1618 | else |
c56d5c7d | 1619 | return inactive_anon_is_low(lruvec); |
b39415b2 RR |
1620 | } |
1621 | ||
4f98a2fe | 1622 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
1a93be0e | 1623 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 1624 | { |
b39415b2 | 1625 | if (is_active_lru(lru)) { |
75b00af7 | 1626 | if (inactive_list_is_low(lruvec, lru)) |
1a93be0e | 1627 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
1628 | return 0; |
1629 | } | |
1630 | ||
1a93be0e | 1631 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
1632 | } |
1633 | ||
3d58ab5c | 1634 | static int vmscan_swappiness(struct scan_control *sc) |
1f4c025b | 1635 | { |
89b5fae5 | 1636 | if (global_reclaim(sc)) |
1f4c025b | 1637 | return vm_swappiness; |
3d58ab5c | 1638 | return mem_cgroup_swappiness(sc->target_mem_cgroup); |
1f4c025b KH |
1639 | } |
1640 | ||
9a265114 JW |
1641 | enum scan_balance { |
1642 | SCAN_EQUAL, | |
1643 | SCAN_FRACT, | |
1644 | SCAN_ANON, | |
1645 | SCAN_FILE, | |
1646 | }; | |
1647 | ||
4f98a2fe RR |
1648 | /* |
1649 | * Determine how aggressively the anon and file LRU lists should be | |
1650 | * scanned. The relative value of each set of LRU lists is determined | |
1651 | * by looking at the fraction of the pages scanned we did rotate back | |
1652 | * onto the active list instead of evict. | |
1653 | * | |
be7bd59d WL |
1654 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
1655 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 1656 | */ |
90126375 | 1657 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
9e3b2f8c | 1658 | unsigned long *nr) |
4f98a2fe | 1659 | { |
9a265114 JW |
1660 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1661 | u64 fraction[2]; | |
1662 | u64 denominator = 0; /* gcc */ | |
1663 | struct zone *zone = lruvec_zone(lruvec); | |
4f98a2fe | 1664 | unsigned long anon_prio, file_prio; |
9a265114 JW |
1665 | enum scan_balance scan_balance; |
1666 | unsigned long anon, file, free; | |
1667 | bool force_scan = false; | |
4f98a2fe | 1668 | unsigned long ap, fp; |
4111304d | 1669 | enum lru_list lru; |
246e87a9 | 1670 | |
f11c0ca5 JW |
1671 | /* |
1672 | * If the zone or memcg is small, nr[l] can be 0. This | |
1673 | * results in no scanning on this priority and a potential | |
1674 | * priority drop. Global direct reclaim can go to the next | |
1675 | * zone and tends to have no problems. Global kswapd is for | |
1676 | * zone balancing and it needs to scan a minimum amount. When | |
1677 | * reclaiming for a memcg, a priority drop can cause high | |
1678 | * latencies, so it's better to scan a minimum amount there as | |
1679 | * well. | |
1680 | */ | |
90126375 | 1681 | if (current_is_kswapd() && zone->all_unreclaimable) |
a4d3e9e7 | 1682 | force_scan = true; |
89b5fae5 | 1683 | if (!global_reclaim(sc)) |
a4d3e9e7 | 1684 | force_scan = true; |
76a33fc3 SL |
1685 | |
1686 | /* If we have no swap space, do not bother scanning anon pages. */ | |
1687 | if (!sc->may_swap || (nr_swap_pages <= 0)) { | |
9a265114 | 1688 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
1689 | goto out; |
1690 | } | |
4f98a2fe | 1691 | |
10316b31 JW |
1692 | /* |
1693 | * Global reclaim will swap to prevent OOM even with no | |
1694 | * swappiness, but memcg users want to use this knob to | |
1695 | * disable swapping for individual groups completely when | |
1696 | * using the memory controller's swap limit feature would be | |
1697 | * too expensive. | |
1698 | */ | |
1699 | if (!global_reclaim(sc) && !vmscan_swappiness(sc)) { | |
9a265114 | 1700 | scan_balance = SCAN_FILE; |
10316b31 JW |
1701 | goto out; |
1702 | } | |
1703 | ||
1704 | /* | |
1705 | * Do not apply any pressure balancing cleverness when the | |
1706 | * system is close to OOM, scan both anon and file equally | |
1707 | * (unless the swappiness setting disagrees with swapping). | |
1708 | */ | |
1709 | if (!sc->priority && vmscan_swappiness(sc)) { | |
9a265114 | 1710 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
1711 | goto out; |
1712 | } | |
1713 | ||
4d7dcca2 HD |
1714 | anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + |
1715 | get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
1716 | file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
1717 | get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
a4d3e9e7 | 1718 | |
11d16c25 JW |
1719 | /* |
1720 | * If it's foreseeable that reclaiming the file cache won't be | |
1721 | * enough to get the zone back into a desirable shape, we have | |
1722 | * to swap. Better start now and leave the - probably heavily | |
1723 | * thrashing - remaining file pages alone. | |
1724 | */ | |
89b5fae5 | 1725 | if (global_reclaim(sc)) { |
11d16c25 | 1726 | free = zone_page_state(zone, NR_FREE_PAGES); |
90126375 | 1727 | if (unlikely(file + free <= high_wmark_pages(zone))) { |
9a265114 | 1728 | scan_balance = SCAN_ANON; |
76a33fc3 | 1729 | goto out; |
eeee9a8c | 1730 | } |
4f98a2fe RR |
1731 | } |
1732 | ||
7c5bd705 JW |
1733 | /* |
1734 | * There is enough inactive page cache, do not reclaim | |
1735 | * anything from the anonymous working set right now. | |
1736 | */ | |
1737 | if (!inactive_file_is_low(lruvec)) { | |
9a265114 | 1738 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
1739 | goto out; |
1740 | } | |
1741 | ||
9a265114 JW |
1742 | scan_balance = SCAN_FRACT; |
1743 | ||
58c37f6e KM |
1744 | /* |
1745 | * With swappiness at 100, anonymous and file have the same priority. | |
1746 | * This scanning priority is essentially the inverse of IO cost. | |
1747 | */ | |
3d58ab5c | 1748 | anon_prio = vmscan_swappiness(sc); |
75b00af7 | 1749 | file_prio = 200 - anon_prio; |
58c37f6e | 1750 | |
4f98a2fe RR |
1751 | /* |
1752 | * OK, so we have swap space and a fair amount of page cache | |
1753 | * pages. We use the recently rotated / recently scanned | |
1754 | * ratios to determine how valuable each cache is. | |
1755 | * | |
1756 | * Because workloads change over time (and to avoid overflow) | |
1757 | * we keep these statistics as a floating average, which ends | |
1758 | * up weighing recent references more than old ones. | |
1759 | * | |
1760 | * anon in [0], file in [1] | |
1761 | */ | |
90126375 | 1762 | spin_lock_irq(&zone->lru_lock); |
6e901571 | 1763 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
1764 | reclaim_stat->recent_scanned[0] /= 2; |
1765 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
1766 | } |
1767 | ||
6e901571 | 1768 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
1769 | reclaim_stat->recent_scanned[1] /= 2; |
1770 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
1771 | } |
1772 | ||
4f98a2fe | 1773 | /* |
00d8089c RR |
1774 | * The amount of pressure on anon vs file pages is inversely |
1775 | * proportional to the fraction of recently scanned pages on | |
1776 | * each list that were recently referenced and in active use. | |
4f98a2fe | 1777 | */ |
fe35004f | 1778 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 1779 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 1780 | |
fe35004f | 1781 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 1782 | fp /= reclaim_stat->recent_rotated[1] + 1; |
90126375 | 1783 | spin_unlock_irq(&zone->lru_lock); |
4f98a2fe | 1784 | |
76a33fc3 SL |
1785 | fraction[0] = ap; |
1786 | fraction[1] = fp; | |
1787 | denominator = ap + fp + 1; | |
1788 | out: | |
4111304d HD |
1789 | for_each_evictable_lru(lru) { |
1790 | int file = is_file_lru(lru); | |
d778df51 | 1791 | unsigned long size; |
76a33fc3 | 1792 | unsigned long scan; |
6e08a369 | 1793 | |
d778df51 | 1794 | size = get_lru_size(lruvec, lru); |
10316b31 | 1795 | scan = size >> sc->priority; |
9a265114 | 1796 | |
10316b31 JW |
1797 | if (!scan && force_scan) |
1798 | scan = min(size, SWAP_CLUSTER_MAX); | |
9a265114 JW |
1799 | |
1800 | switch (scan_balance) { | |
1801 | case SCAN_EQUAL: | |
1802 | /* Scan lists relative to size */ | |
1803 | break; | |
1804 | case SCAN_FRACT: | |
1805 | /* | |
1806 | * Scan types proportional to swappiness and | |
1807 | * their relative recent reclaim efficiency. | |
1808 | */ | |
1809 | scan = div64_u64(scan * fraction[file], denominator); | |
1810 | break; | |
1811 | case SCAN_FILE: | |
1812 | case SCAN_ANON: | |
1813 | /* Scan one type exclusively */ | |
1814 | if ((scan_balance == SCAN_FILE) != file) | |
1815 | scan = 0; | |
1816 | break; | |
1817 | default: | |
1818 | /* Look ma, no brain */ | |
1819 | BUG(); | |
1820 | } | |
4111304d | 1821 | nr[lru] = scan; |
76a33fc3 | 1822 | } |
6e08a369 | 1823 | } |
4f98a2fe | 1824 | |
9b4f98cd JW |
1825 | /* |
1826 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1827 | */ | |
1828 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) | |
1829 | { | |
1830 | unsigned long nr[NR_LRU_LISTS]; | |
1831 | unsigned long nr_to_scan; | |
1832 | enum lru_list lru; | |
1833 | unsigned long nr_reclaimed = 0; | |
1834 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
1835 | struct blk_plug plug; | |
1836 | ||
1837 | get_scan_count(lruvec, sc, nr); | |
1838 | ||
1839 | blk_start_plug(&plug); | |
1840 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
1841 | nr[LRU_INACTIVE_FILE]) { | |
1842 | for_each_evictable_lru(lru) { | |
1843 | if (nr[lru]) { | |
1844 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
1845 | nr[lru] -= nr_to_scan; | |
1846 | ||
1847 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
1848 | lruvec, sc); | |
1849 | } | |
1850 | } | |
1851 | /* | |
1852 | * On large memory systems, scan >> priority can become | |
1853 | * really large. This is fine for the starting priority; | |
1854 | * we want to put equal scanning pressure on each zone. | |
1855 | * However, if the VM has a harder time of freeing pages, | |
1856 | * with multiple processes reclaiming pages, the total | |
1857 | * freeing target can get unreasonably large. | |
1858 | */ | |
1859 | if (nr_reclaimed >= nr_to_reclaim && | |
1860 | sc->priority < DEF_PRIORITY) | |
1861 | break; | |
1862 | } | |
1863 | blk_finish_plug(&plug); | |
1864 | sc->nr_reclaimed += nr_reclaimed; | |
1865 | ||
1866 | /* | |
1867 | * Even if we did not try to evict anon pages at all, we want to | |
1868 | * rebalance the anon lru active/inactive ratio. | |
1869 | */ | |
1870 | if (inactive_anon_is_low(lruvec)) | |
1871 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
1872 | sc, LRU_ACTIVE_ANON); | |
1873 | ||
1874 | throttle_vm_writeout(sc->gfp_mask); | |
1875 | } | |
1876 | ||
23b9da55 | 1877 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 1878 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 1879 | { |
d84da3f9 | 1880 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 1881 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 1882 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
1883 | return true; |
1884 | ||
1885 | return false; | |
1886 | } | |
1887 | ||
3e7d3449 | 1888 | /* |
23b9da55 MG |
1889 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
1890 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
1891 | * true if more pages should be reclaimed such that when the page allocator | |
1892 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
1893 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 1894 | */ |
9b4f98cd | 1895 | static inline bool should_continue_reclaim(struct zone *zone, |
3e7d3449 MG |
1896 | unsigned long nr_reclaimed, |
1897 | unsigned long nr_scanned, | |
1898 | struct scan_control *sc) | |
1899 | { | |
1900 | unsigned long pages_for_compaction; | |
1901 | unsigned long inactive_lru_pages; | |
1902 | ||
1903 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 1904 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
1905 | return false; |
1906 | ||
2876592f MG |
1907 | /* Consider stopping depending on scan and reclaim activity */ |
1908 | if (sc->gfp_mask & __GFP_REPEAT) { | |
1909 | /* | |
1910 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
1911 | * full LRU list has been scanned and we are still failing | |
1912 | * to reclaim pages. This full LRU scan is potentially | |
1913 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
1914 | */ | |
1915 | if (!nr_reclaimed && !nr_scanned) | |
1916 | return false; | |
1917 | } else { | |
1918 | /* | |
1919 | * For non-__GFP_REPEAT allocations which can presumably | |
1920 | * fail without consequence, stop if we failed to reclaim | |
1921 | * any pages from the last SWAP_CLUSTER_MAX number of | |
1922 | * pages that were scanned. This will return to the | |
1923 | * caller faster at the risk reclaim/compaction and | |
1924 | * the resulting allocation attempt fails | |
1925 | */ | |
1926 | if (!nr_reclaimed) | |
1927 | return false; | |
1928 | } | |
3e7d3449 MG |
1929 | |
1930 | /* | |
1931 | * If we have not reclaimed enough pages for compaction and the | |
1932 | * inactive lists are large enough, continue reclaiming | |
1933 | */ | |
1934 | pages_for_compaction = (2UL << sc->order); | |
9b4f98cd | 1935 | inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); |
86cfd3a4 | 1936 | if (nr_swap_pages > 0) |
9b4f98cd | 1937 | inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); |
3e7d3449 MG |
1938 | if (sc->nr_reclaimed < pages_for_compaction && |
1939 | inactive_lru_pages > pages_for_compaction) | |
1940 | return true; | |
1941 | ||
1942 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
9b4f98cd | 1943 | switch (compaction_suitable(zone, sc->order)) { |
3e7d3449 MG |
1944 | case COMPACT_PARTIAL: |
1945 | case COMPACT_CONTINUE: | |
1946 | return false; | |
1947 | default: | |
1948 | return true; | |
1949 | } | |
1950 | } | |
1951 | ||
9b4f98cd | 1952 | static void shrink_zone(struct zone *zone, struct scan_control *sc) |
1da177e4 | 1953 | { |
f0fdc5e8 | 1954 | unsigned long nr_reclaimed, nr_scanned; |
1da177e4 | 1955 | |
9b4f98cd JW |
1956 | do { |
1957 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
1958 | struct mem_cgroup_reclaim_cookie reclaim = { | |
1959 | .zone = zone, | |
1960 | .priority = sc->priority, | |
1961 | }; | |
1962 | struct mem_cgroup *memcg; | |
3e7d3449 | 1963 | |
9b4f98cd JW |
1964 | nr_reclaimed = sc->nr_reclaimed; |
1965 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 1966 | |
9b4f98cd JW |
1967 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
1968 | do { | |
1969 | struct lruvec *lruvec; | |
5660048c | 1970 | |
9b4f98cd | 1971 | lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
f9be23d6 | 1972 | |
9b4f98cd | 1973 | shrink_lruvec(lruvec, sc); |
f16015fb | 1974 | |
9b4f98cd JW |
1975 | /* |
1976 | * Limit reclaim has historically picked one | |
1977 | * memcg and scanned it with decreasing | |
1978 | * priority levels until nr_to_reclaim had | |
1979 | * been reclaimed. This priority cycle is | |
1980 | * thus over after a single memcg. | |
1981 | * | |
1982 | * Direct reclaim and kswapd, on the other | |
1983 | * hand, have to scan all memory cgroups to | |
1984 | * fulfill the overall scan target for the | |
1985 | * zone. | |
1986 | */ | |
1987 | if (!global_reclaim(sc)) { | |
1988 | mem_cgroup_iter_break(root, memcg); | |
1989 | break; | |
1990 | } | |
1991 | memcg = mem_cgroup_iter(root, memcg, &reclaim); | |
1992 | } while (memcg); | |
1993 | } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, | |
1994 | sc->nr_scanned - nr_scanned, sc)); | |
f16015fb JW |
1995 | } |
1996 | ||
fe4b1b24 MG |
1997 | /* Returns true if compaction should go ahead for a high-order request */ |
1998 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
1999 | { | |
2000 | unsigned long balance_gap, watermark; | |
2001 | bool watermark_ok; | |
2002 | ||
2003 | /* Do not consider compaction for orders reclaim is meant to satisfy */ | |
2004 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
2005 | return false; | |
2006 | ||
2007 | /* | |
2008 | * Compaction takes time to run and there are potentially other | |
2009 | * callers using the pages just freed. Continue reclaiming until | |
2010 | * there is a buffer of free pages available to give compaction | |
2011 | * a reasonable chance of completing and allocating the page | |
2012 | */ | |
2013 | balance_gap = min(low_wmark_pages(zone), | |
2014 | (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2015 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
2016 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order); | |
2017 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); | |
2018 | ||
2019 | /* | |
2020 | * If compaction is deferred, reclaim up to a point where | |
2021 | * compaction will have a chance of success when re-enabled | |
2022 | */ | |
aff62249 | 2023 | if (compaction_deferred(zone, sc->order)) |
fe4b1b24 MG |
2024 | return watermark_ok; |
2025 | ||
2026 | /* If compaction is not ready to start, keep reclaiming */ | |
2027 | if (!compaction_suitable(zone, sc->order)) | |
2028 | return false; | |
2029 | ||
2030 | return watermark_ok; | |
2031 | } | |
2032 | ||
1da177e4 LT |
2033 | /* |
2034 | * This is the direct reclaim path, for page-allocating processes. We only | |
2035 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2036 | * request. | |
2037 | * | |
41858966 MG |
2038 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
2039 | * Because: | |
1da177e4 LT |
2040 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
2041 | * allocation or | |
41858966 MG |
2042 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
2043 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2044 | * zone defense algorithm. | |
1da177e4 | 2045 | * |
1da177e4 LT |
2046 | * If a zone is deemed to be full of pinned pages then just give it a light |
2047 | * scan then give up on it. | |
e0c23279 MG |
2048 | * |
2049 | * This function returns true if a zone is being reclaimed for a costly | |
fe4b1b24 | 2050 | * high-order allocation and compaction is ready to begin. This indicates to |
0cee34fd MG |
2051 | * the caller that it should consider retrying the allocation instead of |
2052 | * further reclaim. | |
1da177e4 | 2053 | */ |
9e3b2f8c | 2054 | static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2055 | { |
dd1a239f | 2056 | struct zoneref *z; |
54a6eb5c | 2057 | struct zone *zone; |
d149e3b2 YH |
2058 | unsigned long nr_soft_reclaimed; |
2059 | unsigned long nr_soft_scanned; | |
0cee34fd | 2060 | bool aborted_reclaim = false; |
1cfb419b | 2061 | |
cc715d99 MG |
2062 | /* |
2063 | * If the number of buffer_heads in the machine exceeds the maximum | |
2064 | * allowed level, force direct reclaim to scan the highmem zone as | |
2065 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2066 | */ | |
2067 | if (buffer_heads_over_limit) | |
2068 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2069 | ||
d4debc66 MG |
2070 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
2071 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
f3fe6512 | 2072 | if (!populated_zone(zone)) |
1da177e4 | 2073 | continue; |
1cfb419b KH |
2074 | /* |
2075 | * Take care memory controller reclaiming has small influence | |
2076 | * to global LRU. | |
2077 | */ | |
89b5fae5 | 2078 | if (global_reclaim(sc)) { |
1cfb419b KH |
2079 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2080 | continue; | |
9e3b2f8c KK |
2081 | if (zone->all_unreclaimable && |
2082 | sc->priority != DEF_PRIORITY) | |
1cfb419b | 2083 | continue; /* Let kswapd poll it */ |
d84da3f9 | 2084 | if (IS_ENABLED(CONFIG_COMPACTION)) { |
e0887c19 | 2085 | /* |
e0c23279 MG |
2086 | * If we already have plenty of memory free for |
2087 | * compaction in this zone, don't free any more. | |
2088 | * Even though compaction is invoked for any | |
2089 | * non-zero order, only frequent costly order | |
2090 | * reclamation is disruptive enough to become a | |
c7cfa37b CA |
2091 | * noticeable problem, like transparent huge |
2092 | * page allocations. | |
e0887c19 | 2093 | */ |
fe4b1b24 | 2094 | if (compaction_ready(zone, sc)) { |
0cee34fd | 2095 | aborted_reclaim = true; |
e0887c19 | 2096 | continue; |
e0c23279 | 2097 | } |
e0887c19 | 2098 | } |
ac34a1a3 KH |
2099 | /* |
2100 | * This steals pages from memory cgroups over softlimit | |
2101 | * and returns the number of reclaimed pages and | |
2102 | * scanned pages. This works for global memory pressure | |
2103 | * and balancing, not for a memcg's limit. | |
2104 | */ | |
2105 | nr_soft_scanned = 0; | |
2106 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2107 | sc->order, sc->gfp_mask, | |
2108 | &nr_soft_scanned); | |
2109 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2110 | sc->nr_scanned += nr_soft_scanned; | |
2111 | /* need some check for avoid more shrink_zone() */ | |
1cfb419b | 2112 | } |
408d8544 | 2113 | |
9e3b2f8c | 2114 | shrink_zone(zone, sc); |
1da177e4 | 2115 | } |
e0c23279 | 2116 | |
0cee34fd | 2117 | return aborted_reclaim; |
d1908362 MK |
2118 | } |
2119 | ||
2120 | static bool zone_reclaimable(struct zone *zone) | |
2121 | { | |
2122 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
2123 | } | |
2124 | ||
929bea7c | 2125 | /* All zones in zonelist are unreclaimable? */ |
d1908362 MK |
2126 | static bool all_unreclaimable(struct zonelist *zonelist, |
2127 | struct scan_control *sc) | |
2128 | { | |
2129 | struct zoneref *z; | |
2130 | struct zone *zone; | |
d1908362 MK |
2131 | |
2132 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2133 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2134 | if (!populated_zone(zone)) | |
2135 | continue; | |
2136 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2137 | continue; | |
929bea7c KM |
2138 | if (!zone->all_unreclaimable) |
2139 | return false; | |
d1908362 MK |
2140 | } |
2141 | ||
929bea7c | 2142 | return true; |
1da177e4 | 2143 | } |
4f98a2fe | 2144 | |
1da177e4 LT |
2145 | /* |
2146 | * This is the main entry point to direct page reclaim. | |
2147 | * | |
2148 | * If a full scan of the inactive list fails to free enough memory then we | |
2149 | * are "out of memory" and something needs to be killed. | |
2150 | * | |
2151 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2152 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2153 | * caller can't do much about. We kick the writeback threads and take explicit |
2154 | * naps in the hope that some of these pages can be written. But if the | |
2155 | * allocating task holds filesystem locks which prevent writeout this might not | |
2156 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2157 | * |
2158 | * returns: 0, if no pages reclaimed | |
2159 | * else, the number of pages reclaimed | |
1da177e4 | 2160 | */ |
dac1d27b | 2161 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
a09ed5e0 YH |
2162 | struct scan_control *sc, |
2163 | struct shrink_control *shrink) | |
1da177e4 | 2164 | { |
69e05944 | 2165 | unsigned long total_scanned = 0; |
1da177e4 | 2166 | struct reclaim_state *reclaim_state = current->reclaim_state; |
dd1a239f | 2167 | struct zoneref *z; |
54a6eb5c | 2168 | struct zone *zone; |
22fba335 | 2169 | unsigned long writeback_threshold; |
0cee34fd | 2170 | bool aborted_reclaim; |
1da177e4 | 2171 | |
873b4771 KK |
2172 | delayacct_freepages_start(); |
2173 | ||
89b5fae5 | 2174 | if (global_reclaim(sc)) |
1cfb419b | 2175 | count_vm_event(ALLOCSTALL); |
1da177e4 | 2176 | |
9e3b2f8c | 2177 | do { |
66e1707b | 2178 | sc->nr_scanned = 0; |
9e3b2f8c | 2179 | aborted_reclaim = shrink_zones(zonelist, sc); |
e0c23279 | 2180 | |
66e1707b BS |
2181 | /* |
2182 | * Don't shrink slabs when reclaiming memory from | |
2183 | * over limit cgroups | |
2184 | */ | |
89b5fae5 | 2185 | if (global_reclaim(sc)) { |
c6a8a8c5 | 2186 | unsigned long lru_pages = 0; |
d4debc66 MG |
2187 | for_each_zone_zonelist(zone, z, zonelist, |
2188 | gfp_zone(sc->gfp_mask)) { | |
c6a8a8c5 KM |
2189 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2190 | continue; | |
2191 | ||
2192 | lru_pages += zone_reclaimable_pages(zone); | |
2193 | } | |
2194 | ||
1495f230 | 2195 | shrink_slab(shrink, sc->nr_scanned, lru_pages); |
91a45470 | 2196 | if (reclaim_state) { |
a79311c1 | 2197 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
91a45470 KH |
2198 | reclaim_state->reclaimed_slab = 0; |
2199 | } | |
1da177e4 | 2200 | } |
66e1707b | 2201 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2202 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
1da177e4 | 2203 | goto out; |
1da177e4 LT |
2204 | |
2205 | /* | |
2206 | * Try to write back as many pages as we just scanned. This | |
2207 | * tends to cause slow streaming writers to write data to the | |
2208 | * disk smoothly, at the dirtying rate, which is nice. But | |
2209 | * that's undesirable in laptop mode, where we *want* lumpy | |
2210 | * writeout. So in laptop mode, write out the whole world. | |
2211 | */ | |
22fba335 KM |
2212 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2213 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2214 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2215 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2216 | sc->may_writepage = 1; |
1da177e4 LT |
2217 | } |
2218 | ||
2219 | /* Take a nap, wait for some writeback to complete */ | |
7b51755c | 2220 | if (!sc->hibernation_mode && sc->nr_scanned && |
9e3b2f8c | 2221 | sc->priority < DEF_PRIORITY - 2) { |
0e093d99 MG |
2222 | struct zone *preferred_zone; |
2223 | ||
2224 | first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), | |
f33261d7 DR |
2225 | &cpuset_current_mems_allowed, |
2226 | &preferred_zone); | |
0e093d99 MG |
2227 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); |
2228 | } | |
9e3b2f8c | 2229 | } while (--sc->priority >= 0); |
bb21c7ce | 2230 | |
1da177e4 | 2231 | out: |
873b4771 KK |
2232 | delayacct_freepages_end(); |
2233 | ||
bb21c7ce KM |
2234 | if (sc->nr_reclaimed) |
2235 | return sc->nr_reclaimed; | |
2236 | ||
929bea7c KM |
2237 | /* |
2238 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2239 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2240 | * check. | |
2241 | */ | |
2242 | if (oom_killer_disabled) | |
2243 | return 0; | |
2244 | ||
0cee34fd MG |
2245 | /* Aborted reclaim to try compaction? don't OOM, then */ |
2246 | if (aborted_reclaim) | |
7335084d MG |
2247 | return 1; |
2248 | ||
bb21c7ce | 2249 | /* top priority shrink_zones still had more to do? don't OOM, then */ |
89b5fae5 | 2250 | if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc)) |
bb21c7ce KM |
2251 | return 1; |
2252 | ||
2253 | return 0; | |
1da177e4 LT |
2254 | } |
2255 | ||
5515061d MG |
2256 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) |
2257 | { | |
2258 | struct zone *zone; | |
2259 | unsigned long pfmemalloc_reserve = 0; | |
2260 | unsigned long free_pages = 0; | |
2261 | int i; | |
2262 | bool wmark_ok; | |
2263 | ||
2264 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2265 | zone = &pgdat->node_zones[i]; | |
2266 | pfmemalloc_reserve += min_wmark_pages(zone); | |
2267 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2268 | } | |
2269 | ||
2270 | wmark_ok = free_pages > pfmemalloc_reserve / 2; | |
2271 | ||
2272 | /* kswapd must be awake if processes are being throttled */ | |
2273 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2274 | pgdat->classzone_idx = min(pgdat->classzone_idx, | |
2275 | (enum zone_type)ZONE_NORMAL); | |
2276 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2277 | } | |
2278 | ||
2279 | return wmark_ok; | |
2280 | } | |
2281 | ||
2282 | /* | |
2283 | * Throttle direct reclaimers if backing storage is backed by the network | |
2284 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2285 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2286 | * when the low watermark is reached. |
2287 | * | |
2288 | * Returns true if a fatal signal was delivered during throttling. If this | |
2289 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2290 | */ |
50694c28 | 2291 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2292 | nodemask_t *nodemask) |
2293 | { | |
2294 | struct zone *zone; | |
2295 | int high_zoneidx = gfp_zone(gfp_mask); | |
2296 | pg_data_t *pgdat; | |
2297 | ||
2298 | /* | |
2299 | * Kernel threads should not be throttled as they may be indirectly | |
2300 | * responsible for cleaning pages necessary for reclaim to make forward | |
2301 | * progress. kjournald for example may enter direct reclaim while | |
2302 | * committing a transaction where throttling it could forcing other | |
2303 | * processes to block on log_wait_commit(). | |
2304 | */ | |
2305 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2306 | goto out; |
2307 | ||
2308 | /* | |
2309 | * If a fatal signal is pending, this process should not throttle. | |
2310 | * It should return quickly so it can exit and free its memory | |
2311 | */ | |
2312 | if (fatal_signal_pending(current)) | |
2313 | goto out; | |
5515061d MG |
2314 | |
2315 | /* Check if the pfmemalloc reserves are ok */ | |
2316 | first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone); | |
2317 | pgdat = zone->zone_pgdat; | |
2318 | if (pfmemalloc_watermark_ok(pgdat)) | |
50694c28 | 2319 | goto out; |
5515061d | 2320 | |
68243e76 MG |
2321 | /* Account for the throttling */ |
2322 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2323 | ||
5515061d MG |
2324 | /* |
2325 | * If the caller cannot enter the filesystem, it's possible that it | |
2326 | * is due to the caller holding an FS lock or performing a journal | |
2327 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2328 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2329 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2330 | * second before continuing. | |
2331 | */ | |
2332 | if (!(gfp_mask & __GFP_FS)) { | |
2333 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2334 | pfmemalloc_watermark_ok(pgdat), HZ); | |
50694c28 MG |
2335 | |
2336 | goto check_pending; | |
5515061d MG |
2337 | } |
2338 | ||
2339 | /* Throttle until kswapd wakes the process */ | |
2340 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2341 | pfmemalloc_watermark_ok(pgdat)); | |
50694c28 MG |
2342 | |
2343 | check_pending: | |
2344 | if (fatal_signal_pending(current)) | |
2345 | return true; | |
2346 | ||
2347 | out: | |
2348 | return false; | |
5515061d MG |
2349 | } |
2350 | ||
dac1d27b | 2351 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2352 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2353 | { |
33906bc5 | 2354 | unsigned long nr_reclaimed; |
66e1707b BS |
2355 | struct scan_control sc = { |
2356 | .gfp_mask = gfp_mask, | |
2357 | .may_writepage = !laptop_mode, | |
22fba335 | 2358 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
a6dc60f8 | 2359 | .may_unmap = 1, |
2e2e4259 | 2360 | .may_swap = 1, |
66e1707b | 2361 | .order = order, |
9e3b2f8c | 2362 | .priority = DEF_PRIORITY, |
f16015fb | 2363 | .target_mem_cgroup = NULL, |
327c0e96 | 2364 | .nodemask = nodemask, |
66e1707b | 2365 | }; |
a09ed5e0 YH |
2366 | struct shrink_control shrink = { |
2367 | .gfp_mask = sc.gfp_mask, | |
2368 | }; | |
66e1707b | 2369 | |
5515061d | 2370 | /* |
50694c28 MG |
2371 | * Do not enter reclaim if fatal signal was delivered while throttled. |
2372 | * 1 is returned so that the page allocator does not OOM kill at this | |
2373 | * point. | |
5515061d | 2374 | */ |
50694c28 | 2375 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) |
5515061d MG |
2376 | return 1; |
2377 | ||
33906bc5 MG |
2378 | trace_mm_vmscan_direct_reclaim_begin(order, |
2379 | sc.may_writepage, | |
2380 | gfp_mask); | |
2381 | ||
a09ed5e0 | 2382 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
33906bc5 MG |
2383 | |
2384 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2385 | ||
2386 | return nr_reclaimed; | |
66e1707b BS |
2387 | } |
2388 | ||
c255a458 | 2389 | #ifdef CONFIG_MEMCG |
66e1707b | 2390 | |
72835c86 | 2391 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, |
4e416953 | 2392 | gfp_t gfp_mask, bool noswap, |
0ae5e89c YH |
2393 | struct zone *zone, |
2394 | unsigned long *nr_scanned) | |
4e416953 BS |
2395 | { |
2396 | struct scan_control sc = { | |
0ae5e89c | 2397 | .nr_scanned = 0, |
b8f5c566 | 2398 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
4e416953 BS |
2399 | .may_writepage = !laptop_mode, |
2400 | .may_unmap = 1, | |
2401 | .may_swap = !noswap, | |
4e416953 | 2402 | .order = 0, |
9e3b2f8c | 2403 | .priority = 0, |
72835c86 | 2404 | .target_mem_cgroup = memcg, |
4e416953 | 2405 | }; |
f9be23d6 | 2406 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
0ae5e89c | 2407 | |
4e416953 BS |
2408 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2409 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 2410 | |
9e3b2f8c | 2411 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e KM |
2412 | sc.may_writepage, |
2413 | sc.gfp_mask); | |
2414 | ||
4e416953 BS |
2415 | /* |
2416 | * NOTE: Although we can get the priority field, using it | |
2417 | * here is not a good idea, since it limits the pages we can scan. | |
2418 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2419 | * will pick up pages from other mem cgroup's as well. We hack | |
2420 | * the priority and make it zero. | |
2421 | */ | |
f9be23d6 | 2422 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
2423 | |
2424 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2425 | ||
0ae5e89c | 2426 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2427 | return sc.nr_reclaimed; |
2428 | } | |
2429 | ||
72835c86 | 2430 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
a7885eb8 | 2431 | gfp_t gfp_mask, |
185efc0f | 2432 | bool noswap) |
66e1707b | 2433 | { |
4e416953 | 2434 | struct zonelist *zonelist; |
bdce6d9e | 2435 | unsigned long nr_reclaimed; |
889976db | 2436 | int nid; |
66e1707b | 2437 | struct scan_control sc = { |
66e1707b | 2438 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2439 | .may_unmap = 1, |
2e2e4259 | 2440 | .may_swap = !noswap, |
22fba335 | 2441 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
66e1707b | 2442 | .order = 0, |
9e3b2f8c | 2443 | .priority = DEF_PRIORITY, |
72835c86 | 2444 | .target_mem_cgroup = memcg, |
327c0e96 | 2445 | .nodemask = NULL, /* we don't care the placement */ |
a09ed5e0 YH |
2446 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2447 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2448 | }; | |
2449 | struct shrink_control shrink = { | |
2450 | .gfp_mask = sc.gfp_mask, | |
66e1707b | 2451 | }; |
66e1707b | 2452 | |
889976db YH |
2453 | /* |
2454 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2455 | * take care of from where we get pages. So the node where we start the | |
2456 | * scan does not need to be the current node. | |
2457 | */ | |
72835c86 | 2458 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2459 | |
2460 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2461 | |
2462 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2463 | sc.may_writepage, | |
2464 | sc.gfp_mask); | |
2465 | ||
a09ed5e0 | 2466 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
bdce6d9e KM |
2467 | |
2468 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2469 | ||
2470 | return nr_reclaimed; | |
66e1707b BS |
2471 | } |
2472 | #endif | |
2473 | ||
9e3b2f8c | 2474 | static void age_active_anon(struct zone *zone, struct scan_control *sc) |
f16015fb | 2475 | { |
b95a2f2d | 2476 | struct mem_cgroup *memcg; |
f16015fb | 2477 | |
b95a2f2d JW |
2478 | if (!total_swap_pages) |
2479 | return; | |
2480 | ||
2481 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2482 | do { | |
c56d5c7d | 2483 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
b95a2f2d | 2484 | |
c56d5c7d | 2485 | if (inactive_anon_is_low(lruvec)) |
1a93be0e | 2486 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 2487 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
2488 | |
2489 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2490 | } while (memcg); | |
f16015fb JW |
2491 | } |
2492 | ||
60cefed4 JW |
2493 | static bool zone_balanced(struct zone *zone, int order, |
2494 | unsigned long balance_gap, int classzone_idx) | |
2495 | { | |
2496 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + | |
2497 | balance_gap, classzone_idx, 0)) | |
2498 | return false; | |
2499 | ||
d84da3f9 KS |
2500 | if (IS_ENABLED(CONFIG_COMPACTION) && order && |
2501 | !compaction_suitable(zone, order)) | |
60cefed4 JW |
2502 | return false; |
2503 | ||
2504 | return true; | |
2505 | } | |
2506 | ||
1741c877 | 2507 | /* |
4ae0a48b ZC |
2508 | * pgdat_balanced() is used when checking if a node is balanced. |
2509 | * | |
2510 | * For order-0, all zones must be balanced! | |
2511 | * | |
2512 | * For high-order allocations only zones that meet watermarks and are in a | |
2513 | * zone allowed by the callers classzone_idx are added to balanced_pages. The | |
2514 | * total of balanced pages must be at least 25% of the zones allowed by | |
2515 | * classzone_idx for the node to be considered balanced. Forcing all zones to | |
2516 | * be balanced for high orders can cause excessive reclaim when there are | |
2517 | * imbalanced zones. | |
1741c877 MG |
2518 | * The choice of 25% is due to |
2519 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2520 | * reasonable sized machine | |
2521 | * o On all other machines, the top zone must be at least a reasonable | |
25985edc | 2522 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
1741c877 MG |
2523 | * would need to be at least 256M for it to be balance a whole node. |
2524 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2525 | * to balance a node on its own. These seemed like reasonable ratios. | |
2526 | */ | |
4ae0a48b | 2527 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) |
1741c877 MG |
2528 | { |
2529 | unsigned long present_pages = 0; | |
4ae0a48b | 2530 | unsigned long balanced_pages = 0; |
1741c877 MG |
2531 | int i; |
2532 | ||
4ae0a48b ZC |
2533 | /* Check the watermark levels */ |
2534 | for (i = 0; i <= classzone_idx; i++) { | |
2535 | struct zone *zone = pgdat->node_zones + i; | |
1741c877 | 2536 | |
4ae0a48b ZC |
2537 | if (!populated_zone(zone)) |
2538 | continue; | |
2539 | ||
2540 | present_pages += zone->present_pages; | |
2541 | ||
2542 | /* | |
2543 | * A special case here: | |
2544 | * | |
2545 | * balance_pgdat() skips over all_unreclaimable after | |
2546 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2547 | * they must be considered balanced here as well! | |
2548 | */ | |
2549 | if (zone->all_unreclaimable) { | |
2550 | balanced_pages += zone->present_pages; | |
2551 | continue; | |
2552 | } | |
2553 | ||
2554 | if (zone_balanced(zone, order, 0, i)) | |
2555 | balanced_pages += zone->present_pages; | |
2556 | else if (!order) | |
2557 | return false; | |
2558 | } | |
2559 | ||
2560 | if (order) | |
2561 | return balanced_pages >= (present_pages >> 2); | |
2562 | else | |
2563 | return true; | |
1741c877 MG |
2564 | } |
2565 | ||
5515061d MG |
2566 | /* |
2567 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
2568 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
2569 | * | |
2570 | * Returns true if kswapd is ready to sleep | |
2571 | */ | |
2572 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, | |
dc83edd9 | 2573 | int classzone_idx) |
f50de2d3 | 2574 | { |
f50de2d3 MG |
2575 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ |
2576 | if (remaining) | |
5515061d MG |
2577 | return false; |
2578 | ||
2579 | /* | |
2580 | * There is a potential race between when kswapd checks its watermarks | |
2581 | * and a process gets throttled. There is also a potential race if | |
2582 | * processes get throttled, kswapd wakes, a large process exits therby | |
2583 | * balancing the zones that causes kswapd to miss a wakeup. If kswapd | |
2584 | * is going to sleep, no process should be sleeping on pfmemalloc_wait | |
2585 | * so wake them now if necessary. If necessary, processes will wake | |
2586 | * kswapd and get throttled again | |
2587 | */ | |
2588 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) { | |
2589 | wake_up(&pgdat->pfmemalloc_wait); | |
2590 | return false; | |
2591 | } | |
f50de2d3 | 2592 | |
4ae0a48b | 2593 | return pgdat_balanced(pgdat, order, classzone_idx); |
f50de2d3 MG |
2594 | } |
2595 | ||
1da177e4 LT |
2596 | /* |
2597 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 2598 | * they are all at high_wmark_pages(zone). |
1da177e4 | 2599 | * |
0abdee2b | 2600 | * Returns the final order kswapd was reclaiming at |
1da177e4 LT |
2601 | * |
2602 | * There is special handling here for zones which are full of pinned pages. | |
2603 | * This can happen if the pages are all mlocked, or if they are all used by | |
2604 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
2605 | * What we do is to detect the case where all pages in the zone have been | |
2606 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
2607 | * dead and from now on, only perform a short scan. Basically we're polling | |
2608 | * the zone for when the problem goes away. | |
2609 | * | |
2610 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
2611 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
2612 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
2613 | * lower zones regardless of the number of free pages in the lower zones. This | |
2614 | * interoperates with the page allocator fallback scheme to ensure that aging | |
2615 | * of pages is balanced across the zones. | |
1da177e4 | 2616 | */ |
99504748 | 2617 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, |
dc83edd9 | 2618 | int *classzone_idx) |
1da177e4 | 2619 | { |
cda73a10 | 2620 | struct zone *unbalanced_zone; |
1da177e4 | 2621 | int i; |
99504748 | 2622 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
69e05944 | 2623 | unsigned long total_scanned; |
1da177e4 | 2624 | struct reclaim_state *reclaim_state = current->reclaim_state; |
0ae5e89c YH |
2625 | unsigned long nr_soft_reclaimed; |
2626 | unsigned long nr_soft_scanned; | |
179e9639 AM |
2627 | struct scan_control sc = { |
2628 | .gfp_mask = GFP_KERNEL, | |
a6dc60f8 | 2629 | .may_unmap = 1, |
2e2e4259 | 2630 | .may_swap = 1, |
22fba335 KM |
2631 | /* |
2632 | * kswapd doesn't want to be bailed out while reclaim. because | |
2633 | * we want to put equal scanning pressure on each zone. | |
2634 | */ | |
2635 | .nr_to_reclaim = ULONG_MAX, | |
5ad333eb | 2636 | .order = order, |
f16015fb | 2637 | .target_mem_cgroup = NULL, |
179e9639 | 2638 | }; |
a09ed5e0 YH |
2639 | struct shrink_control shrink = { |
2640 | .gfp_mask = sc.gfp_mask, | |
2641 | }; | |
1da177e4 LT |
2642 | loop_again: |
2643 | total_scanned = 0; | |
9e3b2f8c | 2644 | sc.priority = DEF_PRIORITY; |
a79311c1 | 2645 | sc.nr_reclaimed = 0; |
c0bbbc73 | 2646 | sc.may_writepage = !laptop_mode; |
f8891e5e | 2647 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 2648 | |
9e3b2f8c | 2649 | do { |
1da177e4 | 2650 | unsigned long lru_pages = 0; |
bb3ab596 | 2651 | int has_under_min_watermark_zone = 0; |
1da177e4 | 2652 | |
cda73a10 | 2653 | unbalanced_zone = NULL; |
1da177e4 | 2654 | |
d6277db4 RW |
2655 | /* |
2656 | * Scan in the highmem->dma direction for the highest | |
2657 | * zone which needs scanning | |
2658 | */ | |
2659 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
2660 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 2661 | |
d6277db4 RW |
2662 | if (!populated_zone(zone)) |
2663 | continue; | |
1da177e4 | 2664 | |
9e3b2f8c KK |
2665 | if (zone->all_unreclaimable && |
2666 | sc.priority != DEF_PRIORITY) | |
d6277db4 | 2667 | continue; |
1da177e4 | 2668 | |
556adecb RR |
2669 | /* |
2670 | * Do some background aging of the anon list, to give | |
2671 | * pages a chance to be referenced before reclaiming. | |
2672 | */ | |
9e3b2f8c | 2673 | age_active_anon(zone, &sc); |
556adecb | 2674 | |
cc715d99 MG |
2675 | /* |
2676 | * If the number of buffer_heads in the machine | |
2677 | * exceeds the maximum allowed level and this node | |
2678 | * has a highmem zone, force kswapd to reclaim from | |
2679 | * it to relieve lowmem pressure. | |
2680 | */ | |
2681 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
2682 | end_zone = i; | |
2683 | break; | |
2684 | } | |
2685 | ||
60cefed4 | 2686 | if (!zone_balanced(zone, order, 0, 0)) { |
d6277db4 | 2687 | end_zone = i; |
e1dbeda6 | 2688 | break; |
439423f6 SL |
2689 | } else { |
2690 | /* If balanced, clear the congested flag */ | |
2691 | zone_clear_flag(zone, ZONE_CONGESTED); | |
1da177e4 | 2692 | } |
1da177e4 | 2693 | } |
e1dbeda6 AM |
2694 | if (i < 0) |
2695 | goto out; | |
2696 | ||
1da177e4 LT |
2697 | for (i = 0; i <= end_zone; i++) { |
2698 | struct zone *zone = pgdat->node_zones + i; | |
2699 | ||
adea02a1 | 2700 | lru_pages += zone_reclaimable_pages(zone); |
1da177e4 LT |
2701 | } |
2702 | ||
2703 | /* | |
2704 | * Now scan the zone in the dma->highmem direction, stopping | |
2705 | * at the last zone which needs scanning. | |
2706 | * | |
2707 | * We do this because the page allocator works in the opposite | |
2708 | * direction. This prevents the page allocator from allocating | |
2709 | * pages behind kswapd's direction of progress, which would | |
2710 | * cause too much scanning of the lower zones. | |
2711 | */ | |
2712 | for (i = 0; i <= end_zone; i++) { | |
2713 | struct zone *zone = pgdat->node_zones + i; | |
fe2c2a10 | 2714 | int nr_slab, testorder; |
8afdcece | 2715 | unsigned long balance_gap; |
1da177e4 | 2716 | |
f3fe6512 | 2717 | if (!populated_zone(zone)) |
1da177e4 LT |
2718 | continue; |
2719 | ||
9e3b2f8c KK |
2720 | if (zone->all_unreclaimable && |
2721 | sc.priority != DEF_PRIORITY) | |
1da177e4 LT |
2722 | continue; |
2723 | ||
1da177e4 | 2724 | sc.nr_scanned = 0; |
4e416953 | 2725 | |
0ae5e89c | 2726 | nr_soft_scanned = 0; |
4e416953 BS |
2727 | /* |
2728 | * Call soft limit reclaim before calling shrink_zone. | |
4e416953 | 2729 | */ |
0ae5e89c YH |
2730 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, |
2731 | order, sc.gfp_mask, | |
2732 | &nr_soft_scanned); | |
2733 | sc.nr_reclaimed += nr_soft_reclaimed; | |
2734 | total_scanned += nr_soft_scanned; | |
00918b6a | 2735 | |
32a4330d | 2736 | /* |
8afdcece MG |
2737 | * We put equal pressure on every zone, unless |
2738 | * one zone has way too many pages free | |
2739 | * already. The "too many pages" is defined | |
2740 | * as the high wmark plus a "gap" where the | |
2741 | * gap is either the low watermark or 1% | |
2742 | * of the zone, whichever is smaller. | |
32a4330d | 2743 | */ |
8afdcece MG |
2744 | balance_gap = min(low_wmark_pages(zone), |
2745 | (zone->present_pages + | |
2746 | KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2747 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
fe2c2a10 RR |
2748 | /* |
2749 | * Kswapd reclaims only single pages with compaction | |
2750 | * enabled. Trying too hard to reclaim until contiguous | |
2751 | * free pages have become available can hurt performance | |
2752 | * by evicting too much useful data from memory. | |
2753 | * Do not reclaim more than needed for compaction. | |
2754 | */ | |
2755 | testorder = order; | |
d84da3f9 | 2756 | if (IS_ENABLED(CONFIG_COMPACTION) && order && |
fe2c2a10 RR |
2757 | compaction_suitable(zone, order) != |
2758 | COMPACT_SKIPPED) | |
2759 | testorder = 0; | |
2760 | ||
cc715d99 | 2761 | if ((buffer_heads_over_limit && is_highmem_idx(i)) || |
60cefed4 JW |
2762 | !zone_balanced(zone, testorder, |
2763 | balance_gap, end_zone)) { | |
9e3b2f8c | 2764 | shrink_zone(zone, &sc); |
5a03b051 | 2765 | |
d7868dae MG |
2766 | reclaim_state->reclaimed_slab = 0; |
2767 | nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages); | |
2768 | sc.nr_reclaimed += reclaim_state->reclaimed_slab; | |
2769 | total_scanned += sc.nr_scanned; | |
2770 | ||
2771 | if (nr_slab == 0 && !zone_reclaimable(zone)) | |
2772 | zone->all_unreclaimable = 1; | |
2773 | } | |
2774 | ||
1da177e4 LT |
2775 | /* |
2776 | * If we've done a decent amount of scanning and | |
2777 | * the reclaim ratio is low, start doing writepage | |
2778 | * even in laptop mode | |
2779 | */ | |
2780 | if (total_scanned > SWAP_CLUSTER_MAX * 2 && | |
a79311c1 | 2781 | total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) |
1da177e4 | 2782 | sc.may_writepage = 1; |
bb3ab596 | 2783 | |
215ddd66 MG |
2784 | if (zone->all_unreclaimable) { |
2785 | if (end_zone && end_zone == i) | |
2786 | end_zone--; | |
d7868dae | 2787 | continue; |
215ddd66 | 2788 | } |
d7868dae | 2789 | |
60cefed4 | 2790 | if (!zone_balanced(zone, testorder, 0, end_zone)) { |
cda73a10 | 2791 | unbalanced_zone = zone; |
45973d74 MK |
2792 | /* |
2793 | * We are still under min water mark. This | |
2794 | * means that we have a GFP_ATOMIC allocation | |
2795 | * failure risk. Hurry up! | |
2796 | */ | |
88f5acf8 | 2797 | if (!zone_watermark_ok_safe(zone, order, |
45973d74 MK |
2798 | min_wmark_pages(zone), end_zone, 0)) |
2799 | has_under_min_watermark_zone = 1; | |
0e093d99 MG |
2800 | } else { |
2801 | /* | |
2802 | * If a zone reaches its high watermark, | |
2803 | * consider it to be no longer congested. It's | |
2804 | * possible there are dirty pages backed by | |
2805 | * congested BDIs but as pressure is relieved, | |
ab8704b8 | 2806 | * speculatively avoid congestion waits |
0e093d99 MG |
2807 | */ |
2808 | zone_clear_flag(zone, ZONE_CONGESTED); | |
45973d74 | 2809 | } |
bb3ab596 | 2810 | |
1da177e4 | 2811 | } |
5515061d MG |
2812 | |
2813 | /* | |
2814 | * If the low watermark is met there is no need for processes | |
2815 | * to be throttled on pfmemalloc_wait as they should not be | |
2816 | * able to safely make forward progress. Wake them | |
2817 | */ | |
2818 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
2819 | pfmemalloc_watermark_ok(pgdat)) | |
2820 | wake_up(&pgdat->pfmemalloc_wait); | |
2821 | ||
4ae0a48b | 2822 | if (pgdat_balanced(pgdat, order, *classzone_idx)) |
1da177e4 LT |
2823 | break; /* kswapd: all done */ |
2824 | /* | |
2825 | * OK, kswapd is getting into trouble. Take a nap, then take | |
2826 | * another pass across the zones. | |
2827 | */ | |
9e3b2f8c | 2828 | if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) { |
bb3ab596 KM |
2829 | if (has_under_min_watermark_zone) |
2830 | count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT); | |
ecccd124 | 2831 | else if (unbalanced_zone) |
cda73a10 | 2832 | wait_iff_congested(unbalanced_zone, BLK_RW_ASYNC, HZ/10); |
bb3ab596 | 2833 | } |
1da177e4 LT |
2834 | |
2835 | /* | |
2836 | * We do this so kswapd doesn't build up large priorities for | |
2837 | * example when it is freeing in parallel with allocators. It | |
2838 | * matches the direct reclaim path behaviour in terms of impact | |
2839 | * on zone->*_priority. | |
2840 | */ | |
a79311c1 | 2841 | if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) |
1da177e4 | 2842 | break; |
9e3b2f8c | 2843 | } while (--sc.priority >= 0); |
1da177e4 | 2844 | out: |
99504748 | 2845 | |
4ae0a48b | 2846 | if (!pgdat_balanced(pgdat, order, *classzone_idx)) { |
1da177e4 | 2847 | cond_resched(); |
8357376d RW |
2848 | |
2849 | try_to_freeze(); | |
2850 | ||
73ce02e9 KM |
2851 | /* |
2852 | * Fragmentation may mean that the system cannot be | |
2853 | * rebalanced for high-order allocations in all zones. | |
2854 | * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, | |
2855 | * it means the zones have been fully scanned and are still | |
2856 | * not balanced. For high-order allocations, there is | |
2857 | * little point trying all over again as kswapd may | |
2858 | * infinite loop. | |
2859 | * | |
2860 | * Instead, recheck all watermarks at order-0 as they | |
2861 | * are the most important. If watermarks are ok, kswapd will go | |
2862 | * back to sleep. High-order users can still perform direct | |
2863 | * reclaim if they wish. | |
2864 | */ | |
2865 | if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) | |
2866 | order = sc.order = 0; | |
2867 | ||
1da177e4 LT |
2868 | goto loop_again; |
2869 | } | |
2870 | ||
99504748 MG |
2871 | /* |
2872 | * If kswapd was reclaiming at a higher order, it has the option of | |
2873 | * sleeping without all zones being balanced. Before it does, it must | |
2874 | * ensure that the watermarks for order-0 on *all* zones are met and | |
2875 | * that the congestion flags are cleared. The congestion flag must | |
2876 | * be cleared as kswapd is the only mechanism that clears the flag | |
2877 | * and it is potentially going to sleep here. | |
2878 | */ | |
2879 | if (order) { | |
7be62de9 RR |
2880 | int zones_need_compaction = 1; |
2881 | ||
99504748 MG |
2882 | for (i = 0; i <= end_zone; i++) { |
2883 | struct zone *zone = pgdat->node_zones + i; | |
2884 | ||
2885 | if (!populated_zone(zone)) | |
2886 | continue; | |
2887 | ||
7be62de9 RR |
2888 | /* Check if the memory needs to be defragmented. */ |
2889 | if (zone_watermark_ok(zone, order, | |
2890 | low_wmark_pages(zone), *classzone_idx, 0)) | |
2891 | zones_need_compaction = 0; | |
99504748 | 2892 | } |
7be62de9 RR |
2893 | |
2894 | if (zones_need_compaction) | |
2895 | compact_pgdat(pgdat, order); | |
99504748 MG |
2896 | } |
2897 | ||
0abdee2b | 2898 | /* |
5515061d | 2899 | * Return the order we were reclaiming at so prepare_kswapd_sleep() |
0abdee2b MG |
2900 | * makes a decision on the order we were last reclaiming at. However, |
2901 | * if another caller entered the allocator slow path while kswapd | |
2902 | * was awake, order will remain at the higher level | |
2903 | */ | |
dc83edd9 | 2904 | *classzone_idx = end_zone; |
0abdee2b | 2905 | return order; |
1da177e4 LT |
2906 | } |
2907 | ||
dc83edd9 | 2908 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f0bc0a60 KM |
2909 | { |
2910 | long remaining = 0; | |
2911 | DEFINE_WAIT(wait); | |
2912 | ||
2913 | if (freezing(current) || kthread_should_stop()) | |
2914 | return; | |
2915 | ||
2916 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2917 | ||
2918 | /* Try to sleep for a short interval */ | |
5515061d | 2919 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
2920 | remaining = schedule_timeout(HZ/10); |
2921 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2922 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2923 | } | |
2924 | ||
2925 | /* | |
2926 | * After a short sleep, check if it was a premature sleep. If not, then | |
2927 | * go fully to sleep until explicitly woken up. | |
2928 | */ | |
5515061d | 2929 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
2930 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
2931 | ||
2932 | /* | |
2933 | * vmstat counters are not perfectly accurate and the estimated | |
2934 | * value for counters such as NR_FREE_PAGES can deviate from the | |
2935 | * true value by nr_online_cpus * threshold. To avoid the zone | |
2936 | * watermarks being breached while under pressure, we reduce the | |
2937 | * per-cpu vmstat threshold while kswapd is awake and restore | |
2938 | * them before going back to sleep. | |
2939 | */ | |
2940 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c | 2941 | |
62997027 MG |
2942 | /* |
2943 | * Compaction records what page blocks it recently failed to | |
2944 | * isolate pages from and skips them in the future scanning. | |
2945 | * When kswapd is going to sleep, it is reasonable to assume | |
2946 | * that pages and compaction may succeed so reset the cache. | |
2947 | */ | |
2948 | reset_isolation_suitable(pgdat); | |
2949 | ||
1c7e7f6c AK |
2950 | if (!kthread_should_stop()) |
2951 | schedule(); | |
2952 | ||
f0bc0a60 KM |
2953 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
2954 | } else { | |
2955 | if (remaining) | |
2956 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
2957 | else | |
2958 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
2959 | } | |
2960 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2961 | } | |
2962 | ||
1da177e4 LT |
2963 | /* |
2964 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 2965 | * from the init process. |
1da177e4 LT |
2966 | * |
2967 | * This basically trickles out pages so that we have _some_ | |
2968 | * free memory available even if there is no other activity | |
2969 | * that frees anything up. This is needed for things like routing | |
2970 | * etc, where we otherwise might have all activity going on in | |
2971 | * asynchronous contexts that cannot page things out. | |
2972 | * | |
2973 | * If there are applications that are active memory-allocators | |
2974 | * (most normal use), this basically shouldn't matter. | |
2975 | */ | |
2976 | static int kswapd(void *p) | |
2977 | { | |
215ddd66 | 2978 | unsigned long order, new_order; |
d2ebd0f6 | 2979 | unsigned balanced_order; |
215ddd66 | 2980 | int classzone_idx, new_classzone_idx; |
d2ebd0f6 | 2981 | int balanced_classzone_idx; |
1da177e4 LT |
2982 | pg_data_t *pgdat = (pg_data_t*)p; |
2983 | struct task_struct *tsk = current; | |
f0bc0a60 | 2984 | |
1da177e4 LT |
2985 | struct reclaim_state reclaim_state = { |
2986 | .reclaimed_slab = 0, | |
2987 | }; | |
a70f7302 | 2988 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 2989 | |
cf40bd16 NP |
2990 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
2991 | ||
174596a0 | 2992 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 2993 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
2994 | current->reclaim_state = &reclaim_state; |
2995 | ||
2996 | /* | |
2997 | * Tell the memory management that we're a "memory allocator", | |
2998 | * and that if we need more memory we should get access to it | |
2999 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3000 | * never get caught in the normal page freeing logic. | |
3001 | * | |
3002 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3003 | * you need a small amount of memory in order to be able to | |
3004 | * page out something else, and this flag essentially protects | |
3005 | * us from recursively trying to free more memory as we're | |
3006 | * trying to free the first piece of memory in the first place). | |
3007 | */ | |
930d9152 | 3008 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3009 | set_freezable(); |
1da177e4 | 3010 | |
215ddd66 | 3011 | order = new_order = 0; |
d2ebd0f6 | 3012 | balanced_order = 0; |
215ddd66 | 3013 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; |
d2ebd0f6 | 3014 | balanced_classzone_idx = classzone_idx; |
1da177e4 | 3015 | for ( ; ; ) { |
6f6313d4 | 3016 | bool ret; |
3e1d1d28 | 3017 | |
215ddd66 MG |
3018 | /* |
3019 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
3020 | * new request of a similar or harder type will succeed soon | |
3021 | * so consider going to sleep on the basis we reclaimed at | |
3022 | */ | |
d2ebd0f6 AS |
3023 | if (balanced_classzone_idx >= new_classzone_idx && |
3024 | balanced_order == new_order) { | |
215ddd66 MG |
3025 | new_order = pgdat->kswapd_max_order; |
3026 | new_classzone_idx = pgdat->classzone_idx; | |
3027 | pgdat->kswapd_max_order = 0; | |
3028 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3029 | } | |
3030 | ||
99504748 | 3031 | if (order < new_order || classzone_idx > new_classzone_idx) { |
1da177e4 LT |
3032 | /* |
3033 | * Don't sleep if someone wants a larger 'order' | |
99504748 | 3034 | * allocation or has tigher zone constraints |
1da177e4 LT |
3035 | */ |
3036 | order = new_order; | |
99504748 | 3037 | classzone_idx = new_classzone_idx; |
1da177e4 | 3038 | } else { |
d2ebd0f6 AS |
3039 | kswapd_try_to_sleep(pgdat, balanced_order, |
3040 | balanced_classzone_idx); | |
1da177e4 | 3041 | order = pgdat->kswapd_max_order; |
99504748 | 3042 | classzone_idx = pgdat->classzone_idx; |
f0dfcde0 AS |
3043 | new_order = order; |
3044 | new_classzone_idx = classzone_idx; | |
4d40502e | 3045 | pgdat->kswapd_max_order = 0; |
215ddd66 | 3046 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
1da177e4 | 3047 | } |
1da177e4 | 3048 | |
8fe23e05 DR |
3049 | ret = try_to_freeze(); |
3050 | if (kthread_should_stop()) | |
3051 | break; | |
3052 | ||
3053 | /* | |
3054 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3055 | * after returning from the refrigerator | |
3056 | */ | |
33906bc5 MG |
3057 | if (!ret) { |
3058 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
d2ebd0f6 AS |
3059 | balanced_classzone_idx = classzone_idx; |
3060 | balanced_order = balance_pgdat(pgdat, order, | |
3061 | &balanced_classzone_idx); | |
33906bc5 | 3062 | } |
1da177e4 | 3063 | } |
b0a8cc58 TY |
3064 | |
3065 | current->reclaim_state = NULL; | |
1da177e4 LT |
3066 | return 0; |
3067 | } | |
3068 | ||
3069 | /* | |
3070 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3071 | */ | |
99504748 | 3072 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3073 | { |
3074 | pg_data_t *pgdat; | |
3075 | ||
f3fe6512 | 3076 | if (!populated_zone(zone)) |
1da177e4 LT |
3077 | return; |
3078 | ||
88f5acf8 | 3079 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1da177e4 | 3080 | return; |
88f5acf8 | 3081 | pgdat = zone->zone_pgdat; |
99504748 | 3082 | if (pgdat->kswapd_max_order < order) { |
1da177e4 | 3083 | pgdat->kswapd_max_order = order; |
99504748 MG |
3084 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
3085 | } | |
8d0986e2 | 3086 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3087 | return; |
88f5acf8 MG |
3088 | if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) |
3089 | return; | |
3090 | ||
3091 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3092 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3093 | } |
3094 | ||
adea02a1 WF |
3095 | /* |
3096 | * The reclaimable count would be mostly accurate. | |
3097 | * The less reclaimable pages may be | |
3098 | * - mlocked pages, which will be moved to unevictable list when encountered | |
3099 | * - mapped pages, which may require several travels to be reclaimed | |
3100 | * - dirty pages, which is not "instantly" reclaimable | |
3101 | */ | |
3102 | unsigned long global_reclaimable_pages(void) | |
4f98a2fe | 3103 | { |
adea02a1 WF |
3104 | int nr; |
3105 | ||
3106 | nr = global_page_state(NR_ACTIVE_FILE) + | |
3107 | global_page_state(NR_INACTIVE_FILE); | |
3108 | ||
3109 | if (nr_swap_pages > 0) | |
3110 | nr += global_page_state(NR_ACTIVE_ANON) + | |
3111 | global_page_state(NR_INACTIVE_ANON); | |
3112 | ||
3113 | return nr; | |
3114 | } | |
3115 | ||
3116 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
3117 | { | |
3118 | int nr; | |
3119 | ||
3120 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
3121 | zone_page_state(zone, NR_INACTIVE_FILE); | |
3122 | ||
3123 | if (nr_swap_pages > 0) | |
3124 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
3125 | zone_page_state(zone, NR_INACTIVE_ANON); | |
3126 | ||
3127 | return nr; | |
4f98a2fe RR |
3128 | } |
3129 | ||
c6f37f12 | 3130 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3131 | /* |
7b51755c | 3132 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3133 | * freed pages. |
3134 | * | |
3135 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3136 | * LRU order by reclaiming preferentially | |
3137 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3138 | */ |
7b51755c | 3139 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3140 | { |
d6277db4 | 3141 | struct reclaim_state reclaim_state; |
d6277db4 | 3142 | struct scan_control sc = { |
7b51755c KM |
3143 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
3144 | .may_swap = 1, | |
3145 | .may_unmap = 1, | |
d6277db4 | 3146 | .may_writepage = 1, |
7b51755c KM |
3147 | .nr_to_reclaim = nr_to_reclaim, |
3148 | .hibernation_mode = 1, | |
7b51755c | 3149 | .order = 0, |
9e3b2f8c | 3150 | .priority = DEF_PRIORITY, |
1da177e4 | 3151 | }; |
a09ed5e0 YH |
3152 | struct shrink_control shrink = { |
3153 | .gfp_mask = sc.gfp_mask, | |
3154 | }; | |
3155 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
7b51755c KM |
3156 | struct task_struct *p = current; |
3157 | unsigned long nr_reclaimed; | |
1da177e4 | 3158 | |
7b51755c KM |
3159 | p->flags |= PF_MEMALLOC; |
3160 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3161 | reclaim_state.reclaimed_slab = 0; | |
3162 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3163 | |
a09ed5e0 | 3164 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
d979677c | 3165 | |
7b51755c KM |
3166 | p->reclaim_state = NULL; |
3167 | lockdep_clear_current_reclaim_state(); | |
3168 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3169 | |
7b51755c | 3170 | return nr_reclaimed; |
1da177e4 | 3171 | } |
c6f37f12 | 3172 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3173 | |
1da177e4 LT |
3174 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3175 | not required for correctness. So if the last cpu in a node goes | |
3176 | away, we get changed to run anywhere: as the first one comes back, | |
3177 | restore their cpu bindings. */ | |
fcb35a9b GKH |
3178 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
3179 | void *hcpu) | |
1da177e4 | 3180 | { |
58c0a4a7 | 3181 | int nid; |
1da177e4 | 3182 | |
8bb78442 | 3183 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
48fb2e24 | 3184 | for_each_node_state(nid, N_MEMORY) { |
c5f59f08 | 3185 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3186 | const struct cpumask *mask; |
3187 | ||
3188 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3189 | |
3e597945 | 3190 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3191 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3192 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3193 | } |
3194 | } | |
3195 | return NOTIFY_OK; | |
3196 | } | |
1da177e4 | 3197 | |
3218ae14 YG |
3198 | /* |
3199 | * This kswapd start function will be called by init and node-hot-add. | |
3200 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3201 | */ | |
3202 | int kswapd_run(int nid) | |
3203 | { | |
3204 | pg_data_t *pgdat = NODE_DATA(nid); | |
3205 | int ret = 0; | |
3206 | ||
3207 | if (pgdat->kswapd) | |
3208 | return 0; | |
3209 | ||
3210 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3211 | if (IS_ERR(pgdat->kswapd)) { | |
3212 | /* failure at boot is fatal */ | |
3213 | BUG_ON(system_state == SYSTEM_BOOTING); | |
18b48d58 | 3214 | pgdat->kswapd = NULL; |
d5dc0ad9 GS |
3215 | pr_err("Failed to start kswapd on node %d\n", nid); |
3216 | ret = PTR_ERR(pgdat->kswapd); | |
3218ae14 YG |
3217 | } |
3218 | return ret; | |
3219 | } | |
3220 | ||
8fe23e05 | 3221 | /* |
d8adde17 JL |
3222 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
3223 | * hold lock_memory_hotplug(). | |
8fe23e05 DR |
3224 | */ |
3225 | void kswapd_stop(int nid) | |
3226 | { | |
3227 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3228 | ||
d8adde17 | 3229 | if (kswapd) { |
8fe23e05 | 3230 | kthread_stop(kswapd); |
d8adde17 JL |
3231 | NODE_DATA(nid)->kswapd = NULL; |
3232 | } | |
8fe23e05 DR |
3233 | } |
3234 | ||
1da177e4 LT |
3235 | static int __init kswapd_init(void) |
3236 | { | |
3218ae14 | 3237 | int nid; |
69e05944 | 3238 | |
1da177e4 | 3239 | swap_setup(); |
48fb2e24 | 3240 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3241 | kswapd_run(nid); |
1da177e4 LT |
3242 | hotcpu_notifier(cpu_callback, 0); |
3243 | return 0; | |
3244 | } | |
3245 | ||
3246 | module_init(kswapd_init) | |
9eeff239 CL |
3247 | |
3248 | #ifdef CONFIG_NUMA | |
3249 | /* | |
3250 | * Zone reclaim mode | |
3251 | * | |
3252 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3253 | * the watermarks. | |
9eeff239 CL |
3254 | */ |
3255 | int zone_reclaim_mode __read_mostly; | |
3256 | ||
1b2ffb78 | 3257 | #define RECLAIM_OFF 0 |
7d03431c | 3258 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 CL |
3259 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
3260 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3261 | ||
a92f7126 CL |
3262 | /* |
3263 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3264 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3265 | * a zone. | |
3266 | */ | |
3267 | #define ZONE_RECLAIM_PRIORITY 4 | |
3268 | ||
9614634f CL |
3269 | /* |
3270 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3271 | * occur. | |
3272 | */ | |
3273 | int sysctl_min_unmapped_ratio = 1; | |
3274 | ||
0ff38490 CL |
3275 | /* |
3276 | * If the number of slab pages in a zone grows beyond this percentage then | |
3277 | * slab reclaim needs to occur. | |
3278 | */ | |
3279 | int sysctl_min_slab_ratio = 5; | |
3280 | ||
90afa5de MG |
3281 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
3282 | { | |
3283 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3284 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3285 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3286 | ||
3287 | /* | |
3288 | * It's possible for there to be more file mapped pages than | |
3289 | * accounted for by the pages on the file LRU lists because | |
3290 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3291 | */ | |
3292 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3293 | } | |
3294 | ||
3295 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3296 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3297 | { | |
3298 | long nr_pagecache_reclaimable; | |
3299 | long delta = 0; | |
3300 | ||
3301 | /* | |
3302 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3303 | * potentially reclaimable. Otherwise, we have to worry about | |
3304 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3305 | * a better estimate | |
3306 | */ | |
3307 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3308 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3309 | else | |
3310 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3311 | ||
3312 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3313 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3314 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3315 | ||
3316 | /* Watch for any possible underflows due to delta */ | |
3317 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3318 | delta = nr_pagecache_reclaimable; | |
3319 | ||
3320 | return nr_pagecache_reclaimable - delta; | |
3321 | } | |
3322 | ||
9eeff239 CL |
3323 | /* |
3324 | * Try to free up some pages from this zone through reclaim. | |
3325 | */ | |
179e9639 | 3326 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3327 | { |
7fb2d46d | 3328 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3329 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3330 | struct task_struct *p = current; |
3331 | struct reclaim_state reclaim_state; | |
179e9639 AM |
3332 | struct scan_control sc = { |
3333 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
a6dc60f8 | 3334 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), |
2e2e4259 | 3335 | .may_swap = 1, |
62b726c1 | 3336 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
179e9639 | 3337 | .gfp_mask = gfp_mask, |
bd2f6199 | 3338 | .order = order, |
9e3b2f8c | 3339 | .priority = ZONE_RECLAIM_PRIORITY, |
179e9639 | 3340 | }; |
a09ed5e0 YH |
3341 | struct shrink_control shrink = { |
3342 | .gfp_mask = sc.gfp_mask, | |
3343 | }; | |
15748048 | 3344 | unsigned long nr_slab_pages0, nr_slab_pages1; |
9eeff239 | 3345 | |
9eeff239 | 3346 | cond_resched(); |
d4f7796e CL |
3347 | /* |
3348 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3349 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3350 | * and RECLAIM_SWAP. | |
3351 | */ | |
3352 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3353 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3354 | reclaim_state.reclaimed_slab = 0; |
3355 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3356 | |
90afa5de | 3357 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
3358 | /* |
3359 | * Free memory by calling shrink zone with increasing | |
3360 | * priorities until we have enough memory freed. | |
3361 | */ | |
0ff38490 | 3362 | do { |
9e3b2f8c KK |
3363 | shrink_zone(zone, &sc); |
3364 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); | |
0ff38490 | 3365 | } |
c84db23c | 3366 | |
15748048 KM |
3367 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3368 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
2a16e3f4 | 3369 | /* |
7fb2d46d | 3370 | * shrink_slab() does not currently allow us to determine how |
0ff38490 CL |
3371 | * many pages were freed in this zone. So we take the current |
3372 | * number of slab pages and shake the slab until it is reduced | |
3373 | * by the same nr_pages that we used for reclaiming unmapped | |
3374 | * pages. | |
2a16e3f4 | 3375 | * |
0ff38490 CL |
3376 | * Note that shrink_slab will free memory on all zones and may |
3377 | * take a long time. | |
2a16e3f4 | 3378 | */ |
4dc4b3d9 KM |
3379 | for (;;) { |
3380 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3381 | ||
3382 | /* No reclaimable slab or very low memory pressure */ | |
1495f230 | 3383 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) |
4dc4b3d9 KM |
3384 | break; |
3385 | ||
3386 | /* Freed enough memory */ | |
3387 | nr_slab_pages1 = zone_page_state(zone, | |
3388 | NR_SLAB_RECLAIMABLE); | |
3389 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3390 | break; | |
3391 | } | |
83e33a47 CL |
3392 | |
3393 | /* | |
3394 | * Update nr_reclaimed by the number of slab pages we | |
3395 | * reclaimed from this zone. | |
3396 | */ | |
15748048 KM |
3397 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3398 | if (nr_slab_pages1 < nr_slab_pages0) | |
3399 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
2a16e3f4 CL |
3400 | } |
3401 | ||
9eeff239 | 3402 | p->reclaim_state = NULL; |
d4f7796e | 3403 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3404 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3405 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3406 | } |
179e9639 AM |
3407 | |
3408 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3409 | { | |
179e9639 | 3410 | int node_id; |
d773ed6b | 3411 | int ret; |
179e9639 AM |
3412 | |
3413 | /* | |
0ff38490 CL |
3414 | * Zone reclaim reclaims unmapped file backed pages and |
3415 | * slab pages if we are over the defined limits. | |
34aa1330 | 3416 | * |
9614634f CL |
3417 | * A small portion of unmapped file backed pages is needed for |
3418 | * file I/O otherwise pages read by file I/O will be immediately | |
3419 | * thrown out if the zone is overallocated. So we do not reclaim | |
3420 | * if less than a specified percentage of the zone is used by | |
3421 | * unmapped file backed pages. | |
179e9639 | 3422 | */ |
90afa5de MG |
3423 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
3424 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 3425 | return ZONE_RECLAIM_FULL; |
179e9639 | 3426 | |
93e4a89a | 3427 | if (zone->all_unreclaimable) |
fa5e084e | 3428 | return ZONE_RECLAIM_FULL; |
d773ed6b | 3429 | |
179e9639 | 3430 | /* |
d773ed6b | 3431 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3432 | */ |
d773ed6b | 3433 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 3434 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
3435 | |
3436 | /* | |
3437 | * Only run zone reclaim on the local zone or on zones that do not | |
3438 | * have associated processors. This will favor the local processor | |
3439 | * over remote processors and spread off node memory allocations | |
3440 | * as wide as possible. | |
3441 | */ | |
89fa3024 | 3442 | node_id = zone_to_nid(zone); |
37c0708d | 3443 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 3444 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b DR |
3445 | |
3446 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
fa5e084e MG |
3447 | return ZONE_RECLAIM_NOSCAN; |
3448 | ||
d773ed6b DR |
3449 | ret = __zone_reclaim(zone, gfp_mask, order); |
3450 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3451 | ||
24cf7251 MG |
3452 | if (!ret) |
3453 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3454 | ||
d773ed6b | 3455 | return ret; |
179e9639 | 3456 | } |
9eeff239 | 3457 | #endif |
894bc310 | 3458 | |
894bc310 LS |
3459 | /* |
3460 | * page_evictable - test whether a page is evictable | |
3461 | * @page: the page to test | |
894bc310 LS |
3462 | * |
3463 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3464 | * lists vs unevictable list. |
894bc310 LS |
3465 | * |
3466 | * Reasons page might not be evictable: | |
ba9ddf49 | 3467 | * (1) page's mapping marked unevictable |
b291f000 | 3468 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3469 | * |
894bc310 | 3470 | */ |
39b5f29a | 3471 | int page_evictable(struct page *page) |
894bc310 | 3472 | { |
39b5f29a | 3473 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
894bc310 | 3474 | } |
89e004ea | 3475 | |
85046579 | 3476 | #ifdef CONFIG_SHMEM |
89e004ea | 3477 | /** |
24513264 HD |
3478 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3479 | * @pages: array of pages to check | |
3480 | * @nr_pages: number of pages to check | |
89e004ea | 3481 | * |
24513264 | 3482 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3483 | * |
3484 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3485 | */ |
24513264 | 3486 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3487 | { |
925b7673 | 3488 | struct lruvec *lruvec; |
24513264 HD |
3489 | struct zone *zone = NULL; |
3490 | int pgscanned = 0; | |
3491 | int pgrescued = 0; | |
3492 | int i; | |
89e004ea | 3493 | |
24513264 HD |
3494 | for (i = 0; i < nr_pages; i++) { |
3495 | struct page *page = pages[i]; | |
3496 | struct zone *pagezone; | |
89e004ea | 3497 | |
24513264 HD |
3498 | pgscanned++; |
3499 | pagezone = page_zone(page); | |
3500 | if (pagezone != zone) { | |
3501 | if (zone) | |
3502 | spin_unlock_irq(&zone->lru_lock); | |
3503 | zone = pagezone; | |
3504 | spin_lock_irq(&zone->lru_lock); | |
3505 | } | |
fa9add64 | 3506 | lruvec = mem_cgroup_page_lruvec(page, zone); |
89e004ea | 3507 | |
24513264 HD |
3508 | if (!PageLRU(page) || !PageUnevictable(page)) |
3509 | continue; | |
89e004ea | 3510 | |
39b5f29a | 3511 | if (page_evictable(page)) { |
24513264 HD |
3512 | enum lru_list lru = page_lru_base_type(page); |
3513 | ||
3514 | VM_BUG_ON(PageActive(page)); | |
3515 | ClearPageUnevictable(page); | |
fa9add64 HD |
3516 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
3517 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 3518 | pgrescued++; |
89e004ea | 3519 | } |
24513264 | 3520 | } |
89e004ea | 3521 | |
24513264 HD |
3522 | if (zone) { |
3523 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3524 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3525 | spin_unlock_irq(&zone->lru_lock); | |
89e004ea | 3526 | } |
89e004ea | 3527 | } |
85046579 | 3528 | #endif /* CONFIG_SHMEM */ |
af936a16 | 3529 | |
264e56d8 | 3530 | static void warn_scan_unevictable_pages(void) |
af936a16 | 3531 | { |
264e56d8 | 3532 | printk_once(KERN_WARNING |
25bd91bd | 3533 | "%s: The scan_unevictable_pages sysctl/node-interface has been " |
264e56d8 | 3534 | "disabled for lack of a legitimate use case. If you have " |
25bd91bd KM |
3535 | "one, please send an email to linux-mm@kvack.org.\n", |
3536 | current->comm); | |
af936a16 LS |
3537 | } |
3538 | ||
3539 | /* | |
3540 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3541 | * all nodes' unevictable lists for evictable pages | |
3542 | */ | |
3543 | unsigned long scan_unevictable_pages; | |
3544 | ||
3545 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
8d65af78 | 3546 | void __user *buffer, |
af936a16 LS |
3547 | size_t *length, loff_t *ppos) |
3548 | { | |
264e56d8 | 3549 | warn_scan_unevictable_pages(); |
8d65af78 | 3550 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
af936a16 LS |
3551 | scan_unevictable_pages = 0; |
3552 | return 0; | |
3553 | } | |
3554 | ||
e4455abb | 3555 | #ifdef CONFIG_NUMA |
af936a16 LS |
3556 | /* |
3557 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3558 | * a specified node's per zone unevictable lists for evictable pages. | |
3559 | */ | |
3560 | ||
10fbcf4c KS |
3561 | static ssize_t read_scan_unevictable_node(struct device *dev, |
3562 | struct device_attribute *attr, | |
af936a16 LS |
3563 | char *buf) |
3564 | { | |
264e56d8 | 3565 | warn_scan_unevictable_pages(); |
af936a16 LS |
3566 | return sprintf(buf, "0\n"); /* always zero; should fit... */ |
3567 | } | |
3568 | ||
10fbcf4c KS |
3569 | static ssize_t write_scan_unevictable_node(struct device *dev, |
3570 | struct device_attribute *attr, | |
af936a16 LS |
3571 | const char *buf, size_t count) |
3572 | { | |
264e56d8 | 3573 | warn_scan_unevictable_pages(); |
af936a16 LS |
3574 | return 1; |
3575 | } | |
3576 | ||
3577 | ||
10fbcf4c | 3578 | static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, |
af936a16 LS |
3579 | read_scan_unevictable_node, |
3580 | write_scan_unevictable_node); | |
3581 | ||
3582 | int scan_unevictable_register_node(struct node *node) | |
3583 | { | |
10fbcf4c | 3584 | return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 LS |
3585 | } |
3586 | ||
3587 | void scan_unevictable_unregister_node(struct node *node) | |
3588 | { | |
10fbcf4c | 3589 | device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 | 3590 | } |
e4455abb | 3591 | #endif |