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> | |
29 | #include <linux/pagevec.h> | |
30 | #include <linux/backing-dev.h> | |
31 | #include <linux/rmap.h> | |
32 | #include <linux/topology.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/cpuset.h> | |
3e7d3449 | 35 | #include <linux/compaction.h> |
1da177e4 LT |
36 | #include <linux/notifier.h> |
37 | #include <linux/rwsem.h> | |
248a0301 | 38 | #include <linux/delay.h> |
3218ae14 | 39 | #include <linux/kthread.h> |
7dfb7103 | 40 | #include <linux/freezer.h> |
66e1707b | 41 | #include <linux/memcontrol.h> |
873b4771 | 42 | #include <linux/delayacct.h> |
af936a16 | 43 | #include <linux/sysctl.h> |
929bea7c | 44 | #include <linux/oom.h> |
268bb0ce | 45 | #include <linux/prefetch.h> |
1da177e4 LT |
46 | |
47 | #include <asm/tlbflush.h> | |
48 | #include <asm/div64.h> | |
49 | ||
50 | #include <linux/swapops.h> | |
51 | ||
0f8053a5 NP |
52 | #include "internal.h" |
53 | ||
33906bc5 MG |
54 | #define CREATE_TRACE_POINTS |
55 | #include <trace/events/vmscan.h> | |
56 | ||
ee64fc93 | 57 | /* |
f3a310bc MG |
58 | * reclaim_mode determines how the inactive list is shrunk |
59 | * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages | |
60 | * RECLAIM_MODE_ASYNC: Do not block | |
61 | * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback | |
62 | * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference | |
ee64fc93 MG |
63 | * page from the LRU and reclaim all pages within a |
64 | * naturally aligned range | |
f3a310bc | 65 | * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of |
3e7d3449 | 66 | * order-0 pages and then compact the zone |
ee64fc93 | 67 | */ |
f3a310bc MG |
68 | typedef unsigned __bitwise__ reclaim_mode_t; |
69 | #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u) | |
70 | #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u) | |
71 | #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u) | |
72 | #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u) | |
73 | #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u) | |
7d3579e8 | 74 | |
1da177e4 | 75 | struct scan_control { |
1da177e4 LT |
76 | /* Incremented by the number of inactive pages that were scanned */ |
77 | unsigned long nr_scanned; | |
78 | ||
a79311c1 RR |
79 | /* Number of pages freed so far during a call to shrink_zones() */ |
80 | unsigned long nr_reclaimed; | |
81 | ||
22fba335 KM |
82 | /* How many pages shrink_list() should reclaim */ |
83 | unsigned long nr_to_reclaim; | |
84 | ||
7b51755c KM |
85 | unsigned long hibernation_mode; |
86 | ||
1da177e4 | 87 | /* This context's GFP mask */ |
6daa0e28 | 88 | gfp_t gfp_mask; |
1da177e4 LT |
89 | |
90 | int may_writepage; | |
91 | ||
a6dc60f8 JW |
92 | /* Can mapped pages be reclaimed? */ |
93 | int may_unmap; | |
f1fd1067 | 94 | |
2e2e4259 KM |
95 | /* Can pages be swapped as part of reclaim? */ |
96 | int may_swap; | |
97 | ||
5ad333eb | 98 | int order; |
66e1707b | 99 | |
5f53e762 | 100 | /* |
415b54e3 NK |
101 | * Intend to reclaim enough continuous memory rather than reclaim |
102 | * enough amount of memory. i.e, mode for high order allocation. | |
5f53e762 | 103 | */ |
f3a310bc | 104 | reclaim_mode_t reclaim_mode; |
5f53e762 | 105 | |
f16015fb JW |
106 | /* |
107 | * The memory cgroup that hit its limit and as a result is the | |
108 | * primary target of this reclaim invocation. | |
109 | */ | |
110 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 111 | |
327c0e96 KH |
112 | /* |
113 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
114 | * are scanned. | |
115 | */ | |
116 | nodemask_t *nodemask; | |
1da177e4 LT |
117 | }; |
118 | ||
f16015fb JW |
119 | struct mem_cgroup_zone { |
120 | struct mem_cgroup *mem_cgroup; | |
121 | struct zone *zone; | |
122 | }; | |
123 | ||
1da177e4 LT |
124 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
125 | ||
126 | #ifdef ARCH_HAS_PREFETCH | |
127 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
128 | do { \ | |
129 | if ((_page)->lru.prev != _base) { \ | |
130 | struct page *prev; \ | |
131 | \ | |
132 | prev = lru_to_page(&(_page->lru)); \ | |
133 | prefetch(&prev->_field); \ | |
134 | } \ | |
135 | } while (0) | |
136 | #else | |
137 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
138 | #endif | |
139 | ||
140 | #ifdef ARCH_HAS_PREFETCHW | |
141 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
142 | do { \ | |
143 | if ((_page)->lru.prev != _base) { \ | |
144 | struct page *prev; \ | |
145 | \ | |
146 | prev = lru_to_page(&(_page->lru)); \ | |
147 | prefetchw(&prev->_field); \ | |
148 | } \ | |
149 | } while (0) | |
150 | #else | |
151 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
152 | #endif | |
153 | ||
154 | /* | |
155 | * From 0 .. 100. Higher means more swappy. | |
156 | */ | |
157 | int vm_swappiness = 60; | |
bd1e22b8 | 158 | long vm_total_pages; /* The total number of pages which the VM controls */ |
1da177e4 LT |
159 | |
160 | static LIST_HEAD(shrinker_list); | |
161 | static DECLARE_RWSEM(shrinker_rwsem); | |
162 | ||
00f0b825 | 163 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
89b5fae5 JW |
164 | static bool global_reclaim(struct scan_control *sc) |
165 | { | |
f16015fb | 166 | return !sc->target_mem_cgroup; |
89b5fae5 JW |
167 | } |
168 | ||
f16015fb | 169 | static bool scanning_global_lru(struct mem_cgroup_zone *mz) |
89b5fae5 | 170 | { |
f16015fb | 171 | return !mz->mem_cgroup; |
89b5fae5 | 172 | } |
91a45470 | 173 | #else |
89b5fae5 JW |
174 | static bool global_reclaim(struct scan_control *sc) |
175 | { | |
176 | return true; | |
177 | } | |
178 | ||
f16015fb | 179 | static bool scanning_global_lru(struct mem_cgroup_zone *mz) |
89b5fae5 JW |
180 | { |
181 | return true; | |
182 | } | |
91a45470 KH |
183 | #endif |
184 | ||
f16015fb | 185 | static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz) |
6e901571 | 186 | { |
f16015fb JW |
187 | if (!scanning_global_lru(mz)) |
188 | return mem_cgroup_get_reclaim_stat(mz->mem_cgroup, mz->zone); | |
3e2f41f1 | 189 | |
f16015fb | 190 | return &mz->zone->reclaim_stat; |
6e901571 KM |
191 | } |
192 | ||
f16015fb JW |
193 | static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz, |
194 | enum lru_list lru) | |
c9f299d9 | 195 | { |
f16015fb JW |
196 | if (!scanning_global_lru(mz)) |
197 | return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup, | |
198 | zone_to_nid(mz->zone), | |
199 | zone_idx(mz->zone), | |
200 | BIT(lru)); | |
a3d8e054 | 201 | |
f16015fb | 202 | return zone_page_state(mz->zone, NR_LRU_BASE + lru); |
c9f299d9 KM |
203 | } |
204 | ||
205 | ||
1da177e4 LT |
206 | /* |
207 | * Add a shrinker callback to be called from the vm | |
208 | */ | |
8e1f936b | 209 | void register_shrinker(struct shrinker *shrinker) |
1da177e4 | 210 | { |
83aeeada | 211 | atomic_long_set(&shrinker->nr_in_batch, 0); |
8e1f936b RR |
212 | down_write(&shrinker_rwsem); |
213 | list_add_tail(&shrinker->list, &shrinker_list); | |
214 | up_write(&shrinker_rwsem); | |
1da177e4 | 215 | } |
8e1f936b | 216 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
217 | |
218 | /* | |
219 | * Remove one | |
220 | */ | |
8e1f936b | 221 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
222 | { |
223 | down_write(&shrinker_rwsem); | |
224 | list_del(&shrinker->list); | |
225 | up_write(&shrinker_rwsem); | |
1da177e4 | 226 | } |
8e1f936b | 227 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 | 228 | |
1495f230 YH |
229 | static inline int do_shrinker_shrink(struct shrinker *shrinker, |
230 | struct shrink_control *sc, | |
231 | unsigned long nr_to_scan) | |
232 | { | |
233 | sc->nr_to_scan = nr_to_scan; | |
234 | return (*shrinker->shrink)(shrinker, sc); | |
235 | } | |
236 | ||
1da177e4 LT |
237 | #define SHRINK_BATCH 128 |
238 | /* | |
239 | * Call the shrink functions to age shrinkable caches | |
240 | * | |
241 | * Here we assume it costs one seek to replace a lru page and that it also | |
242 | * takes a seek to recreate a cache object. With this in mind we age equal | |
243 | * percentages of the lru and ageable caches. This should balance the seeks | |
244 | * generated by these structures. | |
245 | * | |
183ff22b | 246 | * If the vm encountered mapped pages on the LRU it increase the pressure on |
1da177e4 LT |
247 | * slab to avoid swapping. |
248 | * | |
249 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
250 | * | |
251 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
252 | * are eligible for the caller's allocation attempt. It is used for balancing | |
253 | * slab reclaim versus page reclaim. | |
b15e0905 | 254 | * |
255 | * Returns the number of slab objects which we shrunk. | |
1da177e4 | 256 | */ |
a09ed5e0 | 257 | unsigned long shrink_slab(struct shrink_control *shrink, |
1495f230 | 258 | unsigned long nr_pages_scanned, |
a09ed5e0 | 259 | unsigned long lru_pages) |
1da177e4 LT |
260 | { |
261 | struct shrinker *shrinker; | |
69e05944 | 262 | unsigned long ret = 0; |
1da177e4 | 263 | |
1495f230 YH |
264 | if (nr_pages_scanned == 0) |
265 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 266 | |
f06590bd MK |
267 | if (!down_read_trylock(&shrinker_rwsem)) { |
268 | /* Assume we'll be able to shrink next time */ | |
269 | ret = 1; | |
270 | goto out; | |
271 | } | |
1da177e4 LT |
272 | |
273 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
274 | unsigned long long delta; | |
635697c6 KK |
275 | long total_scan; |
276 | long max_pass; | |
09576073 | 277 | int shrink_ret = 0; |
acf92b48 DC |
278 | long nr; |
279 | long new_nr; | |
e9299f50 DC |
280 | long batch_size = shrinker->batch ? shrinker->batch |
281 | : SHRINK_BATCH; | |
1da177e4 | 282 | |
635697c6 KK |
283 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); |
284 | if (max_pass <= 0) | |
285 | continue; | |
286 | ||
acf92b48 DC |
287 | /* |
288 | * copy the current shrinker scan count into a local variable | |
289 | * and zero it so that other concurrent shrinker invocations | |
290 | * don't also do this scanning work. | |
291 | */ | |
83aeeada | 292 | nr = atomic_long_xchg(&shrinker->nr_in_batch, 0); |
acf92b48 DC |
293 | |
294 | total_scan = nr; | |
1495f230 | 295 | delta = (4 * nr_pages_scanned) / shrinker->seeks; |
ea164d73 | 296 | delta *= max_pass; |
1da177e4 | 297 | do_div(delta, lru_pages + 1); |
acf92b48 DC |
298 | total_scan += delta; |
299 | if (total_scan < 0) { | |
88c3bd70 DR |
300 | printk(KERN_ERR "shrink_slab: %pF negative objects to " |
301 | "delete nr=%ld\n", | |
acf92b48 DC |
302 | shrinker->shrink, total_scan); |
303 | total_scan = max_pass; | |
ea164d73 AA |
304 | } |
305 | ||
3567b59a DC |
306 | /* |
307 | * We need to avoid excessive windup on filesystem shrinkers | |
308 | * due to large numbers of GFP_NOFS allocations causing the | |
309 | * shrinkers to return -1 all the time. This results in a large | |
310 | * nr being built up so when a shrink that can do some work | |
311 | * comes along it empties the entire cache due to nr >>> | |
312 | * max_pass. This is bad for sustaining a working set in | |
313 | * memory. | |
314 | * | |
315 | * Hence only allow the shrinker to scan the entire cache when | |
316 | * a large delta change is calculated directly. | |
317 | */ | |
318 | if (delta < max_pass / 4) | |
319 | total_scan = min(total_scan, max_pass / 2); | |
320 | ||
ea164d73 AA |
321 | /* |
322 | * Avoid risking looping forever due to too large nr value: | |
323 | * never try to free more than twice the estimate number of | |
324 | * freeable entries. | |
325 | */ | |
acf92b48 DC |
326 | if (total_scan > max_pass * 2) |
327 | total_scan = max_pass * 2; | |
1da177e4 | 328 | |
acf92b48 | 329 | trace_mm_shrink_slab_start(shrinker, shrink, nr, |
09576073 DC |
330 | nr_pages_scanned, lru_pages, |
331 | max_pass, delta, total_scan); | |
332 | ||
e9299f50 | 333 | while (total_scan >= batch_size) { |
b15e0905 | 334 | int nr_before; |
1da177e4 | 335 | |
1495f230 YH |
336 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); |
337 | shrink_ret = do_shrinker_shrink(shrinker, shrink, | |
e9299f50 | 338 | batch_size); |
1da177e4 LT |
339 | if (shrink_ret == -1) |
340 | break; | |
b15e0905 | 341 | if (shrink_ret < nr_before) |
342 | ret += nr_before - shrink_ret; | |
e9299f50 DC |
343 | count_vm_events(SLABS_SCANNED, batch_size); |
344 | total_scan -= batch_size; | |
1da177e4 LT |
345 | |
346 | cond_resched(); | |
347 | } | |
348 | ||
acf92b48 DC |
349 | /* |
350 | * move the unused scan count back into the shrinker in a | |
351 | * manner that handles concurrent updates. If we exhausted the | |
352 | * scan, there is no need to do an update. | |
353 | */ | |
83aeeada KK |
354 | if (total_scan > 0) |
355 | new_nr = atomic_long_add_return(total_scan, | |
356 | &shrinker->nr_in_batch); | |
357 | else | |
358 | new_nr = atomic_long_read(&shrinker->nr_in_batch); | |
acf92b48 DC |
359 | |
360 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); | |
1da177e4 LT |
361 | } |
362 | up_read(&shrinker_rwsem); | |
f06590bd MK |
363 | out: |
364 | cond_resched(); | |
b15e0905 | 365 | return ret; |
1da177e4 LT |
366 | } |
367 | ||
f3a310bc | 368 | static void set_reclaim_mode(int priority, struct scan_control *sc, |
7d3579e8 KM |
369 | bool sync) |
370 | { | |
f3a310bc | 371 | reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC; |
7d3579e8 KM |
372 | |
373 | /* | |
3e7d3449 MG |
374 | * Initially assume we are entering either lumpy reclaim or |
375 | * reclaim/compaction.Depending on the order, we will either set the | |
376 | * sync mode or just reclaim order-0 pages later. | |
7d3579e8 | 377 | */ |
3e7d3449 | 378 | if (COMPACTION_BUILD) |
f3a310bc | 379 | sc->reclaim_mode = RECLAIM_MODE_COMPACTION; |
3e7d3449 | 380 | else |
f3a310bc | 381 | sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM; |
7d3579e8 KM |
382 | |
383 | /* | |
3e7d3449 MG |
384 | * Avoid using lumpy reclaim or reclaim/compaction if possible by |
385 | * restricting when its set to either costly allocations or when | |
386 | * under memory pressure | |
7d3579e8 KM |
387 | */ |
388 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) | |
f3a310bc | 389 | sc->reclaim_mode |= syncmode; |
7d3579e8 | 390 | else if (sc->order && priority < DEF_PRIORITY - 2) |
f3a310bc | 391 | sc->reclaim_mode |= syncmode; |
7d3579e8 | 392 | else |
f3a310bc | 393 | sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; |
7d3579e8 KM |
394 | } |
395 | ||
f3a310bc | 396 | static void reset_reclaim_mode(struct scan_control *sc) |
7d3579e8 | 397 | { |
f3a310bc | 398 | sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; |
7d3579e8 KM |
399 | } |
400 | ||
1da177e4 LT |
401 | static inline int is_page_cache_freeable(struct page *page) |
402 | { | |
ceddc3a5 JW |
403 | /* |
404 | * A freeable page cache page is referenced only by the caller | |
405 | * that isolated the page, the page cache radix tree and | |
406 | * optional buffer heads at page->private. | |
407 | */ | |
edcf4748 | 408 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
409 | } |
410 | ||
7d3579e8 KM |
411 | static int may_write_to_queue(struct backing_dev_info *bdi, |
412 | struct scan_control *sc) | |
1da177e4 | 413 | { |
930d9152 | 414 | if (current->flags & PF_SWAPWRITE) |
1da177e4 LT |
415 | return 1; |
416 | if (!bdi_write_congested(bdi)) | |
417 | return 1; | |
418 | if (bdi == current->backing_dev_info) | |
419 | return 1; | |
7d3579e8 KM |
420 | |
421 | /* lumpy reclaim for hugepage often need a lot of write */ | |
422 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) | |
423 | return 1; | |
1da177e4 LT |
424 | return 0; |
425 | } | |
426 | ||
427 | /* | |
428 | * We detected a synchronous write error writing a page out. Probably | |
429 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
430 | * fsync(), msync() or close(). | |
431 | * | |
432 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
433 | * prevents it from being freed up. But we have a ref on the page and once | |
434 | * that page is locked, the mapping is pinned. | |
435 | * | |
436 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
437 | * __GFP_FS. | |
438 | */ | |
439 | static void handle_write_error(struct address_space *mapping, | |
440 | struct page *page, int error) | |
441 | { | |
7eaceacc | 442 | lock_page(page); |
3e9f45bd GC |
443 | if (page_mapping(page) == mapping) |
444 | mapping_set_error(mapping, error); | |
1da177e4 LT |
445 | unlock_page(page); |
446 | } | |
447 | ||
04e62a29 CL |
448 | /* possible outcome of pageout() */ |
449 | typedef enum { | |
450 | /* failed to write page out, page is locked */ | |
451 | PAGE_KEEP, | |
452 | /* move page to the active list, page is locked */ | |
453 | PAGE_ACTIVATE, | |
454 | /* page has been sent to the disk successfully, page is unlocked */ | |
455 | PAGE_SUCCESS, | |
456 | /* page is clean and locked */ | |
457 | PAGE_CLEAN, | |
458 | } pageout_t; | |
459 | ||
1da177e4 | 460 | /* |
1742f19f AM |
461 | * pageout is called by shrink_page_list() for each dirty page. |
462 | * Calls ->writepage(). | |
1da177e4 | 463 | */ |
c661b078 | 464 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 465 | struct scan_control *sc) |
1da177e4 LT |
466 | { |
467 | /* | |
468 | * If the page is dirty, only perform writeback if that write | |
469 | * will be non-blocking. To prevent this allocation from being | |
470 | * stalled by pagecache activity. But note that there may be | |
471 | * stalls if we need to run get_block(). We could test | |
472 | * PagePrivate for that. | |
473 | * | |
6aceb53b | 474 | * If this process is currently in __generic_file_aio_write() against |
1da177e4 LT |
475 | * this page's queue, we can perform writeback even if that |
476 | * will block. | |
477 | * | |
478 | * If the page is swapcache, write it back even if that would | |
479 | * block, for some throttling. This happens by accident, because | |
480 | * swap_backing_dev_info is bust: it doesn't reflect the | |
481 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
482 | */ |
483 | if (!is_page_cache_freeable(page)) | |
484 | return PAGE_KEEP; | |
485 | if (!mapping) { | |
486 | /* | |
487 | * Some data journaling orphaned pages can have | |
488 | * page->mapping == NULL while being dirty with clean buffers. | |
489 | */ | |
266cf658 | 490 | if (page_has_private(page)) { |
1da177e4 LT |
491 | if (try_to_free_buffers(page)) { |
492 | ClearPageDirty(page); | |
d40cee24 | 493 | printk("%s: orphaned page\n", __func__); |
1da177e4 LT |
494 | return PAGE_CLEAN; |
495 | } | |
496 | } | |
497 | return PAGE_KEEP; | |
498 | } | |
499 | if (mapping->a_ops->writepage == NULL) | |
500 | return PAGE_ACTIVATE; | |
0e093d99 | 501 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) |
1da177e4 LT |
502 | return PAGE_KEEP; |
503 | ||
504 | if (clear_page_dirty_for_io(page)) { | |
505 | int res; | |
506 | struct writeback_control wbc = { | |
507 | .sync_mode = WB_SYNC_NONE, | |
508 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
509 | .range_start = 0, |
510 | .range_end = LLONG_MAX, | |
1da177e4 LT |
511 | .for_reclaim = 1, |
512 | }; | |
513 | ||
514 | SetPageReclaim(page); | |
515 | res = mapping->a_ops->writepage(page, &wbc); | |
516 | if (res < 0) | |
517 | handle_write_error(mapping, page, res); | |
994fc28c | 518 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
519 | ClearPageReclaim(page); |
520 | return PAGE_ACTIVATE; | |
521 | } | |
c661b078 | 522 | |
1da177e4 LT |
523 | if (!PageWriteback(page)) { |
524 | /* synchronous write or broken a_ops? */ | |
525 | ClearPageReclaim(page); | |
526 | } | |
755f0225 | 527 | trace_mm_vmscan_writepage(page, |
f3a310bc | 528 | trace_reclaim_flags(page, sc->reclaim_mode)); |
e129b5c2 | 529 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
530 | return PAGE_SUCCESS; |
531 | } | |
532 | ||
533 | return PAGE_CLEAN; | |
534 | } | |
535 | ||
a649fd92 | 536 | /* |
e286781d NP |
537 | * Same as remove_mapping, but if the page is removed from the mapping, it |
538 | * gets returned with a refcount of 0. | |
a649fd92 | 539 | */ |
e286781d | 540 | static int __remove_mapping(struct address_space *mapping, struct page *page) |
49d2e9cc | 541 | { |
28e4d965 NP |
542 | BUG_ON(!PageLocked(page)); |
543 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 544 | |
19fd6231 | 545 | spin_lock_irq(&mapping->tree_lock); |
49d2e9cc | 546 | /* |
0fd0e6b0 NP |
547 | * The non racy check for a busy page. |
548 | * | |
549 | * Must be careful with the order of the tests. When someone has | |
550 | * a ref to the page, it may be possible that they dirty it then | |
551 | * drop the reference. So if PageDirty is tested before page_count | |
552 | * here, then the following race may occur: | |
553 | * | |
554 | * get_user_pages(&page); | |
555 | * [user mapping goes away] | |
556 | * write_to(page); | |
557 | * !PageDirty(page) [good] | |
558 | * SetPageDirty(page); | |
559 | * put_page(page); | |
560 | * !page_count(page) [good, discard it] | |
561 | * | |
562 | * [oops, our write_to data is lost] | |
563 | * | |
564 | * Reversing the order of the tests ensures such a situation cannot | |
565 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
566 | * load is not satisfied before that of page->_count. | |
567 | * | |
568 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
569 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 570 | */ |
e286781d | 571 | if (!page_freeze_refs(page, 2)) |
49d2e9cc | 572 | goto cannot_free; |
e286781d NP |
573 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
574 | if (unlikely(PageDirty(page))) { | |
575 | page_unfreeze_refs(page, 2); | |
49d2e9cc | 576 | goto cannot_free; |
e286781d | 577 | } |
49d2e9cc CL |
578 | |
579 | if (PageSwapCache(page)) { | |
580 | swp_entry_t swap = { .val = page_private(page) }; | |
581 | __delete_from_swap_cache(page); | |
19fd6231 | 582 | spin_unlock_irq(&mapping->tree_lock); |
cb4b86ba | 583 | swapcache_free(swap, page); |
e286781d | 584 | } else { |
6072d13c LT |
585 | void (*freepage)(struct page *); |
586 | ||
587 | freepage = mapping->a_ops->freepage; | |
588 | ||
e64a782f | 589 | __delete_from_page_cache(page); |
19fd6231 | 590 | spin_unlock_irq(&mapping->tree_lock); |
e767e056 | 591 | mem_cgroup_uncharge_cache_page(page); |
6072d13c LT |
592 | |
593 | if (freepage != NULL) | |
594 | freepage(page); | |
49d2e9cc CL |
595 | } |
596 | ||
49d2e9cc CL |
597 | return 1; |
598 | ||
599 | cannot_free: | |
19fd6231 | 600 | spin_unlock_irq(&mapping->tree_lock); |
49d2e9cc CL |
601 | return 0; |
602 | } | |
603 | ||
e286781d NP |
604 | /* |
605 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
606 | * someone else has a ref on the page, abort and return 0. If it was | |
607 | * successfully detached, return 1. Assumes the caller has a single ref on | |
608 | * this page. | |
609 | */ | |
610 | int remove_mapping(struct address_space *mapping, struct page *page) | |
611 | { | |
612 | if (__remove_mapping(mapping, page)) { | |
613 | /* | |
614 | * Unfreezing the refcount with 1 rather than 2 effectively | |
615 | * drops the pagecache ref for us without requiring another | |
616 | * atomic operation. | |
617 | */ | |
618 | page_unfreeze_refs(page, 1); | |
619 | return 1; | |
620 | } | |
621 | return 0; | |
622 | } | |
623 | ||
894bc310 LS |
624 | /** |
625 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
626 | * @page: page to be put back to appropriate lru list | |
627 | * | |
628 | * Add previously isolated @page to appropriate LRU list. | |
629 | * Page may still be unevictable for other reasons. | |
630 | * | |
631 | * lru_lock must not be held, interrupts must be enabled. | |
632 | */ | |
894bc310 LS |
633 | void putback_lru_page(struct page *page) |
634 | { | |
635 | int lru; | |
636 | int active = !!TestClearPageActive(page); | |
bbfd28ee | 637 | int was_unevictable = PageUnevictable(page); |
894bc310 LS |
638 | |
639 | VM_BUG_ON(PageLRU(page)); | |
640 | ||
641 | redo: | |
642 | ClearPageUnevictable(page); | |
643 | ||
644 | if (page_evictable(page, NULL)) { | |
645 | /* | |
646 | * For evictable pages, we can use the cache. | |
647 | * In event of a race, worst case is we end up with an | |
648 | * unevictable page on [in]active list. | |
649 | * We know how to handle that. | |
650 | */ | |
401a8e1c | 651 | lru = active + page_lru_base_type(page); |
894bc310 LS |
652 | lru_cache_add_lru(page, lru); |
653 | } else { | |
654 | /* | |
655 | * Put unevictable pages directly on zone's unevictable | |
656 | * list. | |
657 | */ | |
658 | lru = LRU_UNEVICTABLE; | |
659 | add_page_to_unevictable_list(page); | |
6a7b9548 | 660 | /* |
21ee9f39 MK |
661 | * When racing with an mlock or AS_UNEVICTABLE clearing |
662 | * (page is unlocked) make sure that if the other thread | |
663 | * does not observe our setting of PG_lru and fails | |
664 | * isolation/check_move_unevictable_page, | |
665 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move | |
6a7b9548 JW |
666 | * the page back to the evictable list. |
667 | * | |
21ee9f39 | 668 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
669 | */ |
670 | smp_mb(); | |
894bc310 | 671 | } |
894bc310 LS |
672 | |
673 | /* | |
674 | * page's status can change while we move it among lru. If an evictable | |
675 | * page is on unevictable list, it never be freed. To avoid that, | |
676 | * check after we added it to the list, again. | |
677 | */ | |
678 | if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { | |
679 | if (!isolate_lru_page(page)) { | |
680 | put_page(page); | |
681 | goto redo; | |
682 | } | |
683 | /* This means someone else dropped this page from LRU | |
684 | * So, it will be freed or putback to LRU again. There is | |
685 | * nothing to do here. | |
686 | */ | |
687 | } | |
688 | ||
bbfd28ee LS |
689 | if (was_unevictable && lru != LRU_UNEVICTABLE) |
690 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
691 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) | |
692 | count_vm_event(UNEVICTABLE_PGCULLED); | |
693 | ||
894bc310 LS |
694 | put_page(page); /* drop ref from isolate */ |
695 | } | |
696 | ||
dfc8d636 JW |
697 | enum page_references { |
698 | PAGEREF_RECLAIM, | |
699 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 700 | PAGEREF_KEEP, |
dfc8d636 JW |
701 | PAGEREF_ACTIVATE, |
702 | }; | |
703 | ||
704 | static enum page_references page_check_references(struct page *page, | |
f16015fb | 705 | struct mem_cgroup_zone *mz, |
dfc8d636 JW |
706 | struct scan_control *sc) |
707 | { | |
64574746 | 708 | int referenced_ptes, referenced_page; |
dfc8d636 | 709 | unsigned long vm_flags; |
dfc8d636 | 710 | |
f16015fb | 711 | referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags); |
64574746 | 712 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 JW |
713 | |
714 | /* Lumpy reclaim - ignore references */ | |
f3a310bc | 715 | if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM) |
dfc8d636 JW |
716 | return PAGEREF_RECLAIM; |
717 | ||
718 | /* | |
719 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
720 | * move the page to the unevictable list. | |
721 | */ | |
722 | if (vm_flags & VM_LOCKED) | |
723 | return PAGEREF_RECLAIM; | |
724 | ||
64574746 JW |
725 | if (referenced_ptes) { |
726 | if (PageAnon(page)) | |
727 | return PAGEREF_ACTIVATE; | |
728 | /* | |
729 | * All mapped pages start out with page table | |
730 | * references from the instantiating fault, so we need | |
731 | * to look twice if a mapped file page is used more | |
732 | * than once. | |
733 | * | |
734 | * Mark it and spare it for another trip around the | |
735 | * inactive list. Another page table reference will | |
736 | * lead to its activation. | |
737 | * | |
738 | * Note: the mark is set for activated pages as well | |
739 | * so that recently deactivated but used pages are | |
740 | * quickly recovered. | |
741 | */ | |
742 | SetPageReferenced(page); | |
743 | ||
34dbc67a | 744 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
745 | return PAGEREF_ACTIVATE; |
746 | ||
c909e993 KK |
747 | /* |
748 | * Activate file-backed executable pages after first usage. | |
749 | */ | |
750 | if (vm_flags & VM_EXEC) | |
751 | return PAGEREF_ACTIVATE; | |
752 | ||
64574746 JW |
753 | return PAGEREF_KEEP; |
754 | } | |
dfc8d636 JW |
755 | |
756 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 757 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
758 | return PAGEREF_RECLAIM_CLEAN; |
759 | ||
760 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
761 | } |
762 | ||
1da177e4 | 763 | /* |
1742f19f | 764 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 765 | */ |
1742f19f | 766 | static unsigned long shrink_page_list(struct list_head *page_list, |
f16015fb | 767 | struct mem_cgroup_zone *mz, |
f84f6e2b | 768 | struct scan_control *sc, |
92df3a72 MG |
769 | int priority, |
770 | unsigned long *ret_nr_dirty, | |
771 | unsigned long *ret_nr_writeback) | |
1da177e4 LT |
772 | { |
773 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 774 | LIST_HEAD(free_pages); |
1da177e4 | 775 | int pgactivate = 0; |
0e093d99 MG |
776 | unsigned long nr_dirty = 0; |
777 | unsigned long nr_congested = 0; | |
05ff5137 | 778 | unsigned long nr_reclaimed = 0; |
92df3a72 | 779 | unsigned long nr_writeback = 0; |
1da177e4 LT |
780 | |
781 | cond_resched(); | |
782 | ||
1da177e4 | 783 | while (!list_empty(page_list)) { |
dfc8d636 | 784 | enum page_references references; |
1da177e4 LT |
785 | struct address_space *mapping; |
786 | struct page *page; | |
787 | int may_enter_fs; | |
1da177e4 LT |
788 | |
789 | cond_resched(); | |
790 | ||
791 | page = lru_to_page(page_list); | |
792 | list_del(&page->lru); | |
793 | ||
529ae9aa | 794 | if (!trylock_page(page)) |
1da177e4 LT |
795 | goto keep; |
796 | ||
725d704e | 797 | VM_BUG_ON(PageActive(page)); |
f16015fb | 798 | VM_BUG_ON(page_zone(page) != mz->zone); |
1da177e4 LT |
799 | |
800 | sc->nr_scanned++; | |
80e43426 | 801 | |
b291f000 NP |
802 | if (unlikely(!page_evictable(page, NULL))) |
803 | goto cull_mlocked; | |
894bc310 | 804 | |
a6dc60f8 | 805 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
806 | goto keep_locked; |
807 | ||
1da177e4 LT |
808 | /* Double the slab pressure for mapped and swapcache pages */ |
809 | if (page_mapped(page) || PageSwapCache(page)) | |
810 | sc->nr_scanned++; | |
811 | ||
c661b078 AW |
812 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
813 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
814 | ||
815 | if (PageWriteback(page)) { | |
92df3a72 | 816 | nr_writeback++; |
c661b078 | 817 | /* |
a18bba06 MG |
818 | * Synchronous reclaim cannot queue pages for |
819 | * writeback due to the possibility of stack overflow | |
820 | * but if it encounters a page under writeback, wait | |
821 | * for the IO to complete. | |
c661b078 | 822 | */ |
f3a310bc | 823 | if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) && |
7d3579e8 | 824 | may_enter_fs) |
c661b078 | 825 | wait_on_page_writeback(page); |
7d3579e8 KM |
826 | else { |
827 | unlock_page(page); | |
828 | goto keep_lumpy; | |
829 | } | |
c661b078 | 830 | } |
1da177e4 | 831 | |
f16015fb | 832 | references = page_check_references(page, mz, sc); |
dfc8d636 JW |
833 | switch (references) { |
834 | case PAGEREF_ACTIVATE: | |
1da177e4 | 835 | goto activate_locked; |
64574746 JW |
836 | case PAGEREF_KEEP: |
837 | goto keep_locked; | |
dfc8d636 JW |
838 | case PAGEREF_RECLAIM: |
839 | case PAGEREF_RECLAIM_CLEAN: | |
840 | ; /* try to reclaim the page below */ | |
841 | } | |
1da177e4 | 842 | |
1da177e4 LT |
843 | /* |
844 | * Anonymous process memory has backing store? | |
845 | * Try to allocate it some swap space here. | |
846 | */ | |
b291f000 | 847 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
848 | if (!(sc->gfp_mask & __GFP_IO)) |
849 | goto keep_locked; | |
ac47b003 | 850 | if (!add_to_swap(page)) |
1da177e4 | 851 | goto activate_locked; |
63eb6b93 | 852 | may_enter_fs = 1; |
b291f000 | 853 | } |
1da177e4 LT |
854 | |
855 | mapping = page_mapping(page); | |
1da177e4 LT |
856 | |
857 | /* | |
858 | * The page is mapped into the page tables of one or more | |
859 | * processes. Try to unmap it here. | |
860 | */ | |
861 | if (page_mapped(page) && mapping) { | |
14fa31b8 | 862 | switch (try_to_unmap(page, TTU_UNMAP)) { |
1da177e4 LT |
863 | case SWAP_FAIL: |
864 | goto activate_locked; | |
865 | case SWAP_AGAIN: | |
866 | goto keep_locked; | |
b291f000 NP |
867 | case SWAP_MLOCK: |
868 | goto cull_mlocked; | |
1da177e4 LT |
869 | case SWAP_SUCCESS: |
870 | ; /* try to free the page below */ | |
871 | } | |
872 | } | |
873 | ||
874 | if (PageDirty(page)) { | |
0e093d99 MG |
875 | nr_dirty++; |
876 | ||
ee72886d MG |
877 | /* |
878 | * Only kswapd can writeback filesystem pages to | |
f84f6e2b MG |
879 | * avoid risk of stack overflow but do not writeback |
880 | * unless under significant pressure. | |
ee72886d | 881 | */ |
f84f6e2b MG |
882 | if (page_is_file_cache(page) && |
883 | (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) { | |
49ea7eb6 MG |
884 | /* |
885 | * Immediately reclaim when written back. | |
886 | * Similar in principal to deactivate_page() | |
887 | * except we already have the page isolated | |
888 | * and know it's dirty | |
889 | */ | |
890 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
891 | SetPageReclaim(page); | |
892 | ||
ee72886d MG |
893 | goto keep_locked; |
894 | } | |
895 | ||
dfc8d636 | 896 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 897 | goto keep_locked; |
4dd4b920 | 898 | if (!may_enter_fs) |
1da177e4 | 899 | goto keep_locked; |
52a8363e | 900 | if (!sc->may_writepage) |
1da177e4 LT |
901 | goto keep_locked; |
902 | ||
903 | /* Page is dirty, try to write it out here */ | |
7d3579e8 | 904 | switch (pageout(page, mapping, sc)) { |
1da177e4 | 905 | case PAGE_KEEP: |
0e093d99 | 906 | nr_congested++; |
1da177e4 LT |
907 | goto keep_locked; |
908 | case PAGE_ACTIVATE: | |
909 | goto activate_locked; | |
910 | case PAGE_SUCCESS: | |
7d3579e8 KM |
911 | if (PageWriteback(page)) |
912 | goto keep_lumpy; | |
913 | if (PageDirty(page)) | |
1da177e4 | 914 | goto keep; |
7d3579e8 | 915 | |
1da177e4 LT |
916 | /* |
917 | * A synchronous write - probably a ramdisk. Go | |
918 | * ahead and try to reclaim the page. | |
919 | */ | |
529ae9aa | 920 | if (!trylock_page(page)) |
1da177e4 LT |
921 | goto keep; |
922 | if (PageDirty(page) || PageWriteback(page)) | |
923 | goto keep_locked; | |
924 | mapping = page_mapping(page); | |
925 | case PAGE_CLEAN: | |
926 | ; /* try to free the page below */ | |
927 | } | |
928 | } | |
929 | ||
930 | /* | |
931 | * If the page has buffers, try to free the buffer mappings | |
932 | * associated with this page. If we succeed we try to free | |
933 | * the page as well. | |
934 | * | |
935 | * We do this even if the page is PageDirty(). | |
936 | * try_to_release_page() does not perform I/O, but it is | |
937 | * possible for a page to have PageDirty set, but it is actually | |
938 | * clean (all its buffers are clean). This happens if the | |
939 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 940 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
941 | * try_to_release_page() will discover that cleanness and will |
942 | * drop the buffers and mark the page clean - it can be freed. | |
943 | * | |
944 | * Rarely, pages can have buffers and no ->mapping. These are | |
945 | * the pages which were not successfully invalidated in | |
946 | * truncate_complete_page(). We try to drop those buffers here | |
947 | * and if that worked, and the page is no longer mapped into | |
948 | * process address space (page_count == 1) it can be freed. | |
949 | * Otherwise, leave the page on the LRU so it is swappable. | |
950 | */ | |
266cf658 | 951 | if (page_has_private(page)) { |
1da177e4 LT |
952 | if (!try_to_release_page(page, sc->gfp_mask)) |
953 | goto activate_locked; | |
e286781d NP |
954 | if (!mapping && page_count(page) == 1) { |
955 | unlock_page(page); | |
956 | if (put_page_testzero(page)) | |
957 | goto free_it; | |
958 | else { | |
959 | /* | |
960 | * rare race with speculative reference. | |
961 | * the speculative reference will free | |
962 | * this page shortly, so we may | |
963 | * increment nr_reclaimed here (and | |
964 | * leave it off the LRU). | |
965 | */ | |
966 | nr_reclaimed++; | |
967 | continue; | |
968 | } | |
969 | } | |
1da177e4 LT |
970 | } |
971 | ||
e286781d | 972 | if (!mapping || !__remove_mapping(mapping, page)) |
49d2e9cc | 973 | goto keep_locked; |
1da177e4 | 974 | |
a978d6f5 NP |
975 | /* |
976 | * At this point, we have no other references and there is | |
977 | * no way to pick any more up (removed from LRU, removed | |
978 | * from pagecache). Can use non-atomic bitops now (and | |
979 | * we obviously don't have to worry about waking up a process | |
980 | * waiting on the page lock, because there are no references. | |
981 | */ | |
982 | __clear_page_locked(page); | |
e286781d | 983 | free_it: |
05ff5137 | 984 | nr_reclaimed++; |
abe4c3b5 MG |
985 | |
986 | /* | |
987 | * Is there need to periodically free_page_list? It would | |
988 | * appear not as the counts should be low | |
989 | */ | |
990 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
991 | continue; |
992 | ||
b291f000 | 993 | cull_mlocked: |
63d6c5ad HD |
994 | if (PageSwapCache(page)) |
995 | try_to_free_swap(page); | |
b291f000 NP |
996 | unlock_page(page); |
997 | putback_lru_page(page); | |
f3a310bc | 998 | reset_reclaim_mode(sc); |
b291f000 NP |
999 | continue; |
1000 | ||
1da177e4 | 1001 | activate_locked: |
68a22394 RR |
1002 | /* Not a candidate for swapping, so reclaim swap space. */ |
1003 | if (PageSwapCache(page) && vm_swap_full()) | |
a2c43eed | 1004 | try_to_free_swap(page); |
894bc310 | 1005 | VM_BUG_ON(PageActive(page)); |
1da177e4 LT |
1006 | SetPageActive(page); |
1007 | pgactivate++; | |
1008 | keep_locked: | |
1009 | unlock_page(page); | |
1010 | keep: | |
f3a310bc | 1011 | reset_reclaim_mode(sc); |
7d3579e8 | 1012 | keep_lumpy: |
1da177e4 | 1013 | list_add(&page->lru, &ret_pages); |
b291f000 | 1014 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); |
1da177e4 | 1015 | } |
abe4c3b5 | 1016 | |
0e093d99 MG |
1017 | /* |
1018 | * Tag a zone as congested if all the dirty pages encountered were | |
1019 | * backed by a congested BDI. In this case, reclaimers should just | |
1020 | * back off and wait for congestion to clear because further reclaim | |
1021 | * will encounter the same problem | |
1022 | */ | |
89b5fae5 | 1023 | if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc)) |
f16015fb | 1024 | zone_set_flag(mz->zone, ZONE_CONGESTED); |
0e093d99 | 1025 | |
cc59850e | 1026 | free_hot_cold_page_list(&free_pages, 1); |
abe4c3b5 | 1027 | |
1da177e4 | 1028 | list_splice(&ret_pages, page_list); |
f8891e5e | 1029 | count_vm_events(PGACTIVATE, pgactivate); |
92df3a72 MG |
1030 | *ret_nr_dirty += nr_dirty; |
1031 | *ret_nr_writeback += nr_writeback; | |
05ff5137 | 1032 | return nr_reclaimed; |
1da177e4 LT |
1033 | } |
1034 | ||
5ad333eb AW |
1035 | /* |
1036 | * Attempt to remove the specified page from its LRU. Only take this page | |
1037 | * if it is of the appropriate PageActive status. Pages which are being | |
1038 | * freed elsewhere are also ignored. | |
1039 | * | |
1040 | * page: page to consider | |
1041 | * mode: one of the LRU isolation modes defined above | |
1042 | * | |
1043 | * returns 0 on success, -ve errno on failure. | |
1044 | */ | |
4356f21d | 1045 | int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file) |
5ad333eb | 1046 | { |
4356f21d | 1047 | bool all_lru_mode; |
5ad333eb AW |
1048 | int ret = -EINVAL; |
1049 | ||
1050 | /* Only take pages on the LRU. */ | |
1051 | if (!PageLRU(page)) | |
1052 | return ret; | |
1053 | ||
4356f21d MK |
1054 | all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) == |
1055 | (ISOLATE_ACTIVE|ISOLATE_INACTIVE); | |
1056 | ||
5ad333eb AW |
1057 | /* |
1058 | * When checking the active state, we need to be sure we are | |
1059 | * dealing with comparible boolean values. Take the logical not | |
1060 | * of each. | |
1061 | */ | |
4356f21d | 1062 | if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE)) |
5ad333eb AW |
1063 | return ret; |
1064 | ||
4356f21d | 1065 | if (!all_lru_mode && !!page_is_file_cache(page) != file) |
4f98a2fe RR |
1066 | return ret; |
1067 | ||
894bc310 LS |
1068 | /* |
1069 | * When this function is being called for lumpy reclaim, we | |
1070 | * initially look into all LRU pages, active, inactive and | |
1071 | * unevictable; only give shrink_page_list evictable pages. | |
1072 | */ | |
1073 | if (PageUnevictable(page)) | |
1074 | return ret; | |
1075 | ||
5ad333eb | 1076 | ret = -EBUSY; |
08e552c6 | 1077 | |
c8244935 MG |
1078 | /* |
1079 | * To minimise LRU disruption, the caller can indicate that it only | |
1080 | * wants to isolate pages it will be able to operate on without | |
1081 | * blocking - clean pages for the most part. | |
1082 | * | |
1083 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1084 | * is used by reclaim when it is cannot write to backing storage | |
1085 | * | |
1086 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1087 | * that it is possible to migrate without blocking | |
1088 | */ | |
1089 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1090 | /* All the caller can do on PageWriteback is block */ | |
1091 | if (PageWriteback(page)) | |
1092 | return ret; | |
1093 | ||
1094 | if (PageDirty(page)) { | |
1095 | struct address_space *mapping; | |
1096 | ||
1097 | /* ISOLATE_CLEAN means only clean pages */ | |
1098 | if (mode & ISOLATE_CLEAN) | |
1099 | return ret; | |
1100 | ||
1101 | /* | |
1102 | * Only pages without mappings or that have a | |
1103 | * ->migratepage callback are possible to migrate | |
1104 | * without blocking | |
1105 | */ | |
1106 | mapping = page_mapping(page); | |
1107 | if (mapping && !mapping->a_ops->migratepage) | |
1108 | return ret; | |
1109 | } | |
1110 | } | |
39deaf85 | 1111 | |
f80c0673 MK |
1112 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1113 | return ret; | |
1114 | ||
5ad333eb AW |
1115 | if (likely(get_page_unless_zero(page))) { |
1116 | /* | |
1117 | * Be careful not to clear PageLRU until after we're | |
1118 | * sure the page is not being freed elsewhere -- the | |
1119 | * page release code relies on it. | |
1120 | */ | |
1121 | ClearPageLRU(page); | |
1122 | ret = 0; | |
1123 | } | |
1124 | ||
1125 | return ret; | |
1126 | } | |
1127 | ||
1da177e4 LT |
1128 | /* |
1129 | * zone->lru_lock is heavily contended. Some of the functions that | |
1130 | * shrink the lists perform better by taking out a batch of pages | |
1131 | * and working on them outside the LRU lock. | |
1132 | * | |
1133 | * For pagecache intensive workloads, this function is the hottest | |
1134 | * spot in the kernel (apart from copy_*_user functions). | |
1135 | * | |
1136 | * Appropriate locks must be held before calling this function. | |
1137 | * | |
1138 | * @nr_to_scan: The number of pages to look through on the list. | |
1139 | * @src: The LRU list to pull pages off. | |
1140 | * @dst: The temp list to put pages on to. | |
1141 | * @scanned: The number of pages that were scanned. | |
5ad333eb AW |
1142 | * @order: The caller's attempted allocation order |
1143 | * @mode: One of the LRU isolation modes | |
4f98a2fe | 1144 | * @file: True [1] if isolating file [!anon] pages |
1da177e4 LT |
1145 | * |
1146 | * returns how many pages were moved onto *@dst. | |
1147 | */ | |
69e05944 AM |
1148 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
1149 | struct list_head *src, struct list_head *dst, | |
4356f21d MK |
1150 | unsigned long *scanned, int order, isolate_mode_t mode, |
1151 | int file) | |
1da177e4 | 1152 | { |
69e05944 | 1153 | unsigned long nr_taken = 0; |
a8a94d15 MG |
1154 | unsigned long nr_lumpy_taken = 0; |
1155 | unsigned long nr_lumpy_dirty = 0; | |
1156 | unsigned long nr_lumpy_failed = 0; | |
c9b02d97 | 1157 | unsigned long scan; |
1da177e4 | 1158 | |
c9b02d97 | 1159 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
5ad333eb AW |
1160 | struct page *page; |
1161 | unsigned long pfn; | |
1162 | unsigned long end_pfn; | |
1163 | unsigned long page_pfn; | |
1164 | int zone_id; | |
1165 | ||
1da177e4 LT |
1166 | page = lru_to_page(src); |
1167 | prefetchw_prev_lru_page(page, src, flags); | |
1168 | ||
725d704e | 1169 | VM_BUG_ON(!PageLRU(page)); |
8d438f96 | 1170 | |
4f98a2fe | 1171 | switch (__isolate_lru_page(page, mode, file)) { |
5ad333eb | 1172 | case 0: |
925b7673 | 1173 | mem_cgroup_lru_del(page); |
5ad333eb | 1174 | list_move(&page->lru, dst); |
2c888cfb | 1175 | nr_taken += hpage_nr_pages(page); |
5ad333eb AW |
1176 | break; |
1177 | ||
1178 | case -EBUSY: | |
1179 | /* else it is being freed elsewhere */ | |
1180 | list_move(&page->lru, src); | |
1181 | continue; | |
46453a6e | 1182 | |
5ad333eb AW |
1183 | default: |
1184 | BUG(); | |
1185 | } | |
1186 | ||
1187 | if (!order) | |
1188 | continue; | |
1189 | ||
1190 | /* | |
1191 | * Attempt to take all pages in the order aligned region | |
1192 | * surrounding the tag page. Only take those pages of | |
1193 | * the same active state as that tag page. We may safely | |
1194 | * round the target page pfn down to the requested order | |
25985edc | 1195 | * as the mem_map is guaranteed valid out to MAX_ORDER, |
5ad333eb AW |
1196 | * where that page is in a different zone we will detect |
1197 | * it from its zone id and abort this block scan. | |
1198 | */ | |
1199 | zone_id = page_zone_id(page); | |
1200 | page_pfn = page_to_pfn(page); | |
1201 | pfn = page_pfn & ~((1 << order) - 1); | |
1202 | end_pfn = pfn + (1 << order); | |
1203 | for (; pfn < end_pfn; pfn++) { | |
1204 | struct page *cursor_page; | |
1205 | ||
1206 | /* The target page is in the block, ignore it. */ | |
1207 | if (unlikely(pfn == page_pfn)) | |
1208 | continue; | |
1209 | ||
1210 | /* Avoid holes within the zone. */ | |
1211 | if (unlikely(!pfn_valid_within(pfn))) | |
1212 | break; | |
1213 | ||
1214 | cursor_page = pfn_to_page(pfn); | |
4f98a2fe | 1215 | |
5ad333eb AW |
1216 | /* Check that we have not crossed a zone boundary. */ |
1217 | if (unlikely(page_zone_id(cursor_page) != zone_id)) | |
08fc468f | 1218 | break; |
de2e7567 MK |
1219 | |
1220 | /* | |
1221 | * If we don't have enough swap space, reclaiming of | |
1222 | * anon page which don't already have a swap slot is | |
1223 | * pointless. | |
1224 | */ | |
043bcbe5 | 1225 | if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) && |
08fc468f KM |
1226 | !PageSwapCache(cursor_page)) |
1227 | break; | |
de2e7567 | 1228 | |
ee993b13 | 1229 | if (__isolate_lru_page(cursor_page, mode, file) == 0) { |
50134731 AA |
1230 | unsigned int isolated_pages; |
1231 | ||
925b7673 | 1232 | mem_cgroup_lru_del(cursor_page); |
5ad333eb | 1233 | list_move(&cursor_page->lru, dst); |
50134731 AA |
1234 | isolated_pages = hpage_nr_pages(cursor_page); |
1235 | nr_taken += isolated_pages; | |
1236 | nr_lumpy_taken += isolated_pages; | |
a8a94d15 | 1237 | if (PageDirty(cursor_page)) |
50134731 | 1238 | nr_lumpy_dirty += isolated_pages; |
5ad333eb | 1239 | scan++; |
50134731 | 1240 | pfn += isolated_pages - 1; |
a8a94d15 | 1241 | } else { |
d179e84b AA |
1242 | /* |
1243 | * Check if the page is freed already. | |
1244 | * | |
1245 | * We can't use page_count() as that | |
1246 | * requires compound_head and we don't | |
1247 | * have a pin on the page here. If a | |
1248 | * page is tail, we may or may not | |
1249 | * have isolated the head, so assume | |
1250 | * it's not free, it'd be tricky to | |
1251 | * track the head status without a | |
1252 | * page pin. | |
1253 | */ | |
1254 | if (!PageTail(cursor_page) && | |
1255 | !atomic_read(&cursor_page->_count)) | |
08fc468f KM |
1256 | continue; |
1257 | break; | |
5ad333eb AW |
1258 | } |
1259 | } | |
08fc468f KM |
1260 | |
1261 | /* If we break out of the loop above, lumpy reclaim failed */ | |
1262 | if (pfn < end_pfn) | |
1263 | nr_lumpy_failed++; | |
1da177e4 LT |
1264 | } |
1265 | ||
1266 | *scanned = scan; | |
a8a94d15 MG |
1267 | |
1268 | trace_mm_vmscan_lru_isolate(order, | |
1269 | nr_to_scan, scan, | |
1270 | nr_taken, | |
1271 | nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed, | |
ea4d349f | 1272 | mode, file); |
1da177e4 LT |
1273 | return nr_taken; |
1274 | } | |
1275 | ||
925b7673 JW |
1276 | static unsigned long isolate_pages(unsigned long nr, struct mem_cgroup_zone *mz, |
1277 | struct list_head *dst, | |
1278 | unsigned long *scanned, int order, | |
1279 | isolate_mode_t mode, int active, int file) | |
66e1707b | 1280 | { |
925b7673 | 1281 | struct lruvec *lruvec; |
4f98a2fe | 1282 | int lru = LRU_BASE; |
925b7673 JW |
1283 | |
1284 | lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup); | |
66e1707b | 1285 | if (active) |
4f98a2fe RR |
1286 | lru += LRU_ACTIVE; |
1287 | if (file) | |
1288 | lru += LRU_FILE; | |
925b7673 | 1289 | return isolate_lru_pages(nr, &lruvec->lists[lru], dst, |
6290df54 | 1290 | scanned, order, mode, file); |
66e1707b BS |
1291 | } |
1292 | ||
5ad333eb AW |
1293 | /* |
1294 | * clear_active_flags() is a helper for shrink_active_list(), clearing | |
1295 | * any active bits from the pages in the list. | |
1296 | */ | |
4f98a2fe RR |
1297 | static unsigned long clear_active_flags(struct list_head *page_list, |
1298 | unsigned int *count) | |
5ad333eb AW |
1299 | { |
1300 | int nr_active = 0; | |
4f98a2fe | 1301 | int lru; |
5ad333eb AW |
1302 | struct page *page; |
1303 | ||
4f98a2fe | 1304 | list_for_each_entry(page, page_list, lru) { |
2c888cfb | 1305 | int numpages = hpage_nr_pages(page); |
401a8e1c | 1306 | lru = page_lru_base_type(page); |
5ad333eb | 1307 | if (PageActive(page)) { |
4f98a2fe | 1308 | lru += LRU_ACTIVE; |
5ad333eb | 1309 | ClearPageActive(page); |
2c888cfb | 1310 | nr_active += numpages; |
5ad333eb | 1311 | } |
1489fa14 | 1312 | if (count) |
2c888cfb | 1313 | count[lru] += numpages; |
4f98a2fe | 1314 | } |
5ad333eb AW |
1315 | |
1316 | return nr_active; | |
1317 | } | |
1318 | ||
62695a84 NP |
1319 | /** |
1320 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1321 | * @page: page to isolate from its LRU list | |
1322 | * | |
1323 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1324 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1325 | * | |
1326 | * Returns 0 if the page was removed from an LRU list. | |
1327 | * Returns -EBUSY if the page was not on an LRU list. | |
1328 | * | |
1329 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1330 | * the active list, it will have PageActive set. If it was found on |
1331 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1332 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1333 | * |
1334 | * The vmstat statistic corresponding to the list on which the page was | |
1335 | * found will be decremented. | |
1336 | * | |
1337 | * Restrictions: | |
1338 | * (1) Must be called with an elevated refcount on the page. This is a | |
1339 | * fundamentnal difference from isolate_lru_pages (which is called | |
1340 | * without a stable reference). | |
1341 | * (2) the lru_lock must not be held. | |
1342 | * (3) interrupts must be enabled. | |
1343 | */ | |
1344 | int isolate_lru_page(struct page *page) | |
1345 | { | |
1346 | int ret = -EBUSY; | |
1347 | ||
0c917313 KK |
1348 | VM_BUG_ON(!page_count(page)); |
1349 | ||
62695a84 NP |
1350 | if (PageLRU(page)) { |
1351 | struct zone *zone = page_zone(page); | |
1352 | ||
1353 | spin_lock_irq(&zone->lru_lock); | |
0c917313 | 1354 | if (PageLRU(page)) { |
894bc310 | 1355 | int lru = page_lru(page); |
62695a84 | 1356 | ret = 0; |
0c917313 | 1357 | get_page(page); |
62695a84 | 1358 | ClearPageLRU(page); |
4f98a2fe | 1359 | |
4f98a2fe | 1360 | del_page_from_lru_list(zone, page, lru); |
62695a84 NP |
1361 | } |
1362 | spin_unlock_irq(&zone->lru_lock); | |
1363 | } | |
1364 | return ret; | |
1365 | } | |
1366 | ||
35cd7815 RR |
1367 | /* |
1368 | * Are there way too many processes in the direct reclaim path already? | |
1369 | */ | |
1370 | static int too_many_isolated(struct zone *zone, int file, | |
1371 | struct scan_control *sc) | |
1372 | { | |
1373 | unsigned long inactive, isolated; | |
1374 | ||
1375 | if (current_is_kswapd()) | |
1376 | return 0; | |
1377 | ||
89b5fae5 | 1378 | if (!global_reclaim(sc)) |
35cd7815 RR |
1379 | return 0; |
1380 | ||
1381 | if (file) { | |
1382 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1383 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1384 | } else { | |
1385 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1386 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1387 | } | |
1388 | ||
1389 | return isolated > inactive; | |
1390 | } | |
1391 | ||
66635629 MG |
1392 | /* |
1393 | * TODO: Try merging with migrations version of putback_lru_pages | |
1394 | */ | |
1395 | static noinline_for_stack void | |
f16015fb JW |
1396 | putback_lru_pages(struct mem_cgroup_zone *mz, struct scan_control *sc, |
1397 | unsigned long nr_anon, unsigned long nr_file, | |
1398 | struct list_head *page_list) | |
66635629 MG |
1399 | { |
1400 | struct page *page; | |
2bcf8879 | 1401 | LIST_HEAD(pages_to_free); |
f16015fb JW |
1402 | struct zone *zone = mz->zone; |
1403 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); | |
66635629 | 1404 | |
66635629 MG |
1405 | /* |
1406 | * Put back any unfreeable pages. | |
1407 | */ | |
1408 | spin_lock(&zone->lru_lock); | |
1409 | while (!list_empty(page_list)) { | |
1410 | int lru; | |
1411 | page = lru_to_page(page_list); | |
1412 | VM_BUG_ON(PageLRU(page)); | |
1413 | list_del(&page->lru); | |
1414 | if (unlikely(!page_evictable(page, NULL))) { | |
1415 | spin_unlock_irq(&zone->lru_lock); | |
1416 | putback_lru_page(page); | |
1417 | spin_lock_irq(&zone->lru_lock); | |
1418 | continue; | |
1419 | } | |
7a608572 | 1420 | SetPageLRU(page); |
66635629 | 1421 | lru = page_lru(page); |
7a608572 | 1422 | add_page_to_lru_list(zone, page, lru); |
66635629 MG |
1423 | if (is_active_lru(lru)) { |
1424 | int file = is_file_lru(lru); | |
9992af10 RR |
1425 | int numpages = hpage_nr_pages(page); |
1426 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1427 | } |
2bcf8879 HD |
1428 | if (put_page_testzero(page)) { |
1429 | __ClearPageLRU(page); | |
1430 | __ClearPageActive(page); | |
1431 | del_page_from_lru_list(zone, page, lru); | |
1432 | ||
1433 | if (unlikely(PageCompound(page))) { | |
1434 | spin_unlock_irq(&zone->lru_lock); | |
1435 | (*get_compound_page_dtor(page))(page); | |
1436 | spin_lock_irq(&zone->lru_lock); | |
1437 | } else | |
1438 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1439 | } |
1440 | } | |
1441 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon); | |
1442 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file); | |
1443 | ||
1444 | spin_unlock_irq(&zone->lru_lock); | |
2bcf8879 | 1445 | free_hot_cold_page_list(&pages_to_free, 1); |
66635629 MG |
1446 | } |
1447 | ||
f16015fb JW |
1448 | static noinline_for_stack void |
1449 | update_isolated_counts(struct mem_cgroup_zone *mz, | |
1450 | struct scan_control *sc, | |
1451 | unsigned long *nr_anon, | |
1452 | unsigned long *nr_file, | |
1453 | struct list_head *isolated_list) | |
1489fa14 MG |
1454 | { |
1455 | unsigned long nr_active; | |
f16015fb | 1456 | struct zone *zone = mz->zone; |
1489fa14 | 1457 | unsigned int count[NR_LRU_LISTS] = { 0, }; |
f16015fb | 1458 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
1489fa14 MG |
1459 | |
1460 | nr_active = clear_active_flags(isolated_list, count); | |
1461 | __count_vm_events(PGDEACTIVATE, nr_active); | |
1462 | ||
1463 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, | |
1464 | -count[LRU_ACTIVE_FILE]); | |
1465 | __mod_zone_page_state(zone, NR_INACTIVE_FILE, | |
1466 | -count[LRU_INACTIVE_FILE]); | |
1467 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, | |
1468 | -count[LRU_ACTIVE_ANON]); | |
1469 | __mod_zone_page_state(zone, NR_INACTIVE_ANON, | |
1470 | -count[LRU_INACTIVE_ANON]); | |
1471 | ||
1472 | *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON]; | |
1473 | *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE]; | |
1474 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon); | |
1475 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file); | |
1476 | ||
1477 | reclaim_stat->recent_scanned[0] += *nr_anon; | |
1478 | reclaim_stat->recent_scanned[1] += *nr_file; | |
1479 | } | |
1480 | ||
e31f3698 | 1481 | /* |
a18bba06 | 1482 | * Returns true if a direct reclaim should wait on pages under writeback. |
e31f3698 WF |
1483 | * |
1484 | * If we are direct reclaiming for contiguous pages and we do not reclaim | |
1485 | * everything in the list, try again and wait for writeback IO to complete. | |
1486 | * This will stall high-order allocations noticeably. Only do that when really | |
1487 | * need to free the pages under high memory pressure. | |
1488 | */ | |
1489 | static inline bool should_reclaim_stall(unsigned long nr_taken, | |
1490 | unsigned long nr_freed, | |
1491 | int priority, | |
1492 | struct scan_control *sc) | |
1493 | { | |
1494 | int lumpy_stall_priority; | |
1495 | ||
1496 | /* kswapd should not stall on sync IO */ | |
1497 | if (current_is_kswapd()) | |
1498 | return false; | |
1499 | ||
1500 | /* Only stall on lumpy reclaim */ | |
f3a310bc | 1501 | if (sc->reclaim_mode & RECLAIM_MODE_SINGLE) |
e31f3698 WF |
1502 | return false; |
1503 | ||
81d66c70 | 1504 | /* If we have reclaimed everything on the isolated list, no stall */ |
e31f3698 WF |
1505 | if (nr_freed == nr_taken) |
1506 | return false; | |
1507 | ||
1508 | /* | |
1509 | * For high-order allocations, there are two stall thresholds. | |
1510 | * High-cost allocations stall immediately where as lower | |
1511 | * order allocations such as stacks require the scanning | |
1512 | * priority to be much higher before stalling. | |
1513 | */ | |
1514 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) | |
1515 | lumpy_stall_priority = DEF_PRIORITY; | |
1516 | else | |
1517 | lumpy_stall_priority = DEF_PRIORITY / 3; | |
1518 | ||
1519 | return priority <= lumpy_stall_priority; | |
1520 | } | |
1521 | ||
1da177e4 | 1522 | /* |
1742f19f AM |
1523 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1524 | * of reclaimed pages | |
1da177e4 | 1525 | */ |
66635629 | 1526 | static noinline_for_stack unsigned long |
f16015fb JW |
1527 | shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz, |
1528 | struct scan_control *sc, int priority, int file) | |
1da177e4 LT |
1529 | { |
1530 | LIST_HEAD(page_list); | |
e247dbce | 1531 | unsigned long nr_scanned; |
05ff5137 | 1532 | unsigned long nr_reclaimed = 0; |
e247dbce | 1533 | unsigned long nr_taken; |
e247dbce KM |
1534 | unsigned long nr_anon; |
1535 | unsigned long nr_file; | |
92df3a72 MG |
1536 | unsigned long nr_dirty = 0; |
1537 | unsigned long nr_writeback = 0; | |
4356f21d | 1538 | isolate_mode_t reclaim_mode = ISOLATE_INACTIVE; |
f16015fb | 1539 | struct zone *zone = mz->zone; |
78dc583d | 1540 | |
35cd7815 | 1541 | while (unlikely(too_many_isolated(zone, file, sc))) { |
58355c78 | 1542 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1543 | |
1544 | /* We are about to die and free our memory. Return now. */ | |
1545 | if (fatal_signal_pending(current)) | |
1546 | return SWAP_CLUSTER_MAX; | |
1547 | } | |
1548 | ||
f3a310bc | 1549 | set_reclaim_mode(priority, sc, false); |
4356f21d MK |
1550 | if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM) |
1551 | reclaim_mode |= ISOLATE_ACTIVE; | |
1552 | ||
1da177e4 | 1553 | lru_add_drain(); |
f80c0673 MK |
1554 | |
1555 | if (!sc->may_unmap) | |
1556 | reclaim_mode |= ISOLATE_UNMAPPED; | |
1557 | if (!sc->may_writepage) | |
1558 | reclaim_mode |= ISOLATE_CLEAN; | |
1559 | ||
1da177e4 | 1560 | spin_lock_irq(&zone->lru_lock); |
b35ea17b | 1561 | |
925b7673 JW |
1562 | nr_taken = isolate_pages(nr_to_scan, mz, &page_list, |
1563 | &nr_scanned, sc->order, | |
1564 | reclaim_mode, 0, file); | |
89b5fae5 | 1565 | if (global_reclaim(sc)) { |
e247dbce KM |
1566 | zone->pages_scanned += nr_scanned; |
1567 | if (current_is_kswapd()) | |
1568 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, | |
1569 | nr_scanned); | |
1570 | else | |
1571 | __count_zone_vm_events(PGSCAN_DIRECT, zone, | |
1572 | nr_scanned); | |
e247dbce | 1573 | } |
b35ea17b | 1574 | |
66635629 MG |
1575 | if (nr_taken == 0) { |
1576 | spin_unlock_irq(&zone->lru_lock); | |
1577 | return 0; | |
1578 | } | |
5ad333eb | 1579 | |
f16015fb | 1580 | update_isolated_counts(mz, sc, &nr_anon, &nr_file, &page_list); |
1da177e4 | 1581 | |
e247dbce | 1582 | spin_unlock_irq(&zone->lru_lock); |
c661b078 | 1583 | |
f16015fb | 1584 | nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority, |
92df3a72 | 1585 | &nr_dirty, &nr_writeback); |
c661b078 | 1586 | |
e31f3698 WF |
1587 | /* Check if we should syncronously wait for writeback */ |
1588 | if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) { | |
f3a310bc | 1589 | set_reclaim_mode(priority, sc, true); |
f16015fb | 1590 | nr_reclaimed += shrink_page_list(&page_list, mz, sc, |
92df3a72 | 1591 | priority, &nr_dirty, &nr_writeback); |
e247dbce | 1592 | } |
b35ea17b | 1593 | |
e247dbce KM |
1594 | local_irq_disable(); |
1595 | if (current_is_kswapd()) | |
1596 | __count_vm_events(KSWAPD_STEAL, nr_reclaimed); | |
1597 | __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed); | |
a74609fa | 1598 | |
f16015fb | 1599 | putback_lru_pages(mz, sc, nr_anon, nr_file, &page_list); |
e11da5b4 | 1600 | |
92df3a72 MG |
1601 | /* |
1602 | * If reclaim is isolating dirty pages under writeback, it implies | |
1603 | * that the long-lived page allocation rate is exceeding the page | |
1604 | * laundering rate. Either the global limits are not being effective | |
1605 | * at throttling processes due to the page distribution throughout | |
1606 | * zones or there is heavy usage of a slow backing device. The | |
1607 | * only option is to throttle from reclaim context which is not ideal | |
1608 | * as there is no guarantee the dirtying process is throttled in the | |
1609 | * same way balance_dirty_pages() manages. | |
1610 | * | |
1611 | * This scales the number of dirty pages that must be under writeback | |
1612 | * before throttling depending on priority. It is a simple backoff | |
1613 | * function that has the most effect in the range DEF_PRIORITY to | |
1614 | * DEF_PRIORITY-2 which is the priority reclaim is considered to be | |
1615 | * in trouble and reclaim is considered to be in trouble. | |
1616 | * | |
1617 | * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle | |
1618 | * DEF_PRIORITY-1 50% must be PageWriteback | |
1619 | * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble | |
1620 | * ... | |
1621 | * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any | |
1622 | * isolated page is PageWriteback | |
1623 | */ | |
1624 | if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority))) | |
1625 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); | |
1626 | ||
e11da5b4 MG |
1627 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, |
1628 | zone_idx(zone), | |
1629 | nr_scanned, nr_reclaimed, | |
1630 | priority, | |
f3a310bc | 1631 | trace_shrink_flags(file, sc->reclaim_mode)); |
05ff5137 | 1632 | return nr_reclaimed; |
1da177e4 LT |
1633 | } |
1634 | ||
1635 | /* | |
1636 | * This moves pages from the active list to the inactive list. | |
1637 | * | |
1638 | * We move them the other way if the page is referenced by one or more | |
1639 | * processes, from rmap. | |
1640 | * | |
1641 | * If the pages are mostly unmapped, the processing is fast and it is | |
1642 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1643 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1644 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1645 | * this, so instead we remove the pages from the LRU while processing them. | |
1646 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1647 | * nobody will play with that bit on a non-LRU page. | |
1648 | * | |
1649 | * The downside is that we have to touch page->_count against each page. | |
1650 | * But we had to alter page->flags anyway. | |
1651 | */ | |
1cfb419b | 1652 | |
3eb4140f WF |
1653 | static void move_active_pages_to_lru(struct zone *zone, |
1654 | struct list_head *list, | |
2bcf8879 | 1655 | struct list_head *pages_to_free, |
3eb4140f WF |
1656 | enum lru_list lru) |
1657 | { | |
1658 | unsigned long pgmoved = 0; | |
3eb4140f WF |
1659 | struct page *page; |
1660 | ||
2bcf8879 HD |
1661 | if (buffer_heads_over_limit) { |
1662 | spin_unlock_irq(&zone->lru_lock); | |
1663 | list_for_each_entry(page, list, lru) { | |
1664 | if (page_has_private(page) && trylock_page(page)) { | |
1665 | if (page_has_private(page)) | |
1666 | try_to_release_page(page, 0); | |
1667 | unlock_page(page); | |
1668 | } | |
1669 | } | |
1670 | spin_lock_irq(&zone->lru_lock); | |
1671 | } | |
3eb4140f WF |
1672 | |
1673 | while (!list_empty(list)) { | |
925b7673 JW |
1674 | struct lruvec *lruvec; |
1675 | ||
3eb4140f | 1676 | page = lru_to_page(list); |
3eb4140f WF |
1677 | |
1678 | VM_BUG_ON(PageLRU(page)); | |
1679 | SetPageLRU(page); | |
1680 | ||
925b7673 JW |
1681 | lruvec = mem_cgroup_lru_add_list(zone, page, lru); |
1682 | list_move(&page->lru, &lruvec->lists[lru]); | |
2c888cfb | 1683 | pgmoved += hpage_nr_pages(page); |
3eb4140f | 1684 | |
2bcf8879 HD |
1685 | if (put_page_testzero(page)) { |
1686 | __ClearPageLRU(page); | |
1687 | __ClearPageActive(page); | |
1688 | del_page_from_lru_list(zone, page, lru); | |
1689 | ||
1690 | if (unlikely(PageCompound(page))) { | |
1691 | spin_unlock_irq(&zone->lru_lock); | |
1692 | (*get_compound_page_dtor(page))(page); | |
1693 | spin_lock_irq(&zone->lru_lock); | |
1694 | } else | |
1695 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1696 | } |
1697 | } | |
1698 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1699 | if (!is_active_lru(lru)) | |
1700 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1701 | } | |
1cfb419b | 1702 | |
f16015fb JW |
1703 | static void shrink_active_list(unsigned long nr_pages, |
1704 | struct mem_cgroup_zone *mz, | |
1705 | struct scan_control *sc, | |
1706 | int priority, int file) | |
1da177e4 | 1707 | { |
44c241f1 | 1708 | unsigned long nr_taken; |
69e05944 | 1709 | unsigned long pgscanned; |
6fe6b7e3 | 1710 | unsigned long vm_flags; |
1da177e4 | 1711 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1712 | LIST_HEAD(l_active); |
b69408e8 | 1713 | LIST_HEAD(l_inactive); |
1da177e4 | 1714 | struct page *page; |
f16015fb | 1715 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
44c241f1 | 1716 | unsigned long nr_rotated = 0; |
f80c0673 | 1717 | isolate_mode_t reclaim_mode = ISOLATE_ACTIVE; |
f16015fb | 1718 | struct zone *zone = mz->zone; |
1da177e4 LT |
1719 | |
1720 | lru_add_drain(); | |
f80c0673 MK |
1721 | |
1722 | if (!sc->may_unmap) | |
1723 | reclaim_mode |= ISOLATE_UNMAPPED; | |
1724 | if (!sc->may_writepage) | |
1725 | reclaim_mode |= ISOLATE_CLEAN; | |
1726 | ||
1da177e4 | 1727 | spin_lock_irq(&zone->lru_lock); |
925b7673 JW |
1728 | |
1729 | nr_taken = isolate_pages(nr_pages, mz, &l_hold, | |
1730 | &pgscanned, sc->order, | |
1731 | reclaim_mode, 1, file); | |
8b25c6d2 | 1732 | |
89b5fae5 JW |
1733 | if (global_reclaim(sc)) |
1734 | zone->pages_scanned += pgscanned; | |
1735 | ||
b7c46d15 | 1736 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1737 | |
3eb4140f | 1738 | __count_zone_vm_events(PGREFILL, zone, pgscanned); |
4f98a2fe | 1739 | if (file) |
44c241f1 | 1740 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken); |
4f98a2fe | 1741 | else |
44c241f1 | 1742 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken); |
a731286d | 1743 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
1da177e4 LT |
1744 | spin_unlock_irq(&zone->lru_lock); |
1745 | ||
1da177e4 LT |
1746 | while (!list_empty(&l_hold)) { |
1747 | cond_resched(); | |
1748 | page = lru_to_page(&l_hold); | |
1749 | list_del(&page->lru); | |
7e9cd484 | 1750 | |
894bc310 LS |
1751 | if (unlikely(!page_evictable(page, NULL))) { |
1752 | putback_lru_page(page); | |
1753 | continue; | |
1754 | } | |
1755 | ||
f16015fb | 1756 | if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) { |
9992af10 | 1757 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1758 | /* |
1759 | * Identify referenced, file-backed active pages and | |
1760 | * give them one more trip around the active list. So | |
1761 | * that executable code get better chances to stay in | |
1762 | * memory under moderate memory pressure. Anon pages | |
1763 | * are not likely to be evicted by use-once streaming | |
1764 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1765 | * so we ignore them here. | |
1766 | */ | |
41e20983 | 1767 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1768 | list_add(&page->lru, &l_active); |
1769 | continue; | |
1770 | } | |
1771 | } | |
7e9cd484 | 1772 | |
5205e56e | 1773 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1774 | list_add(&page->lru, &l_inactive); |
1775 | } | |
1776 | ||
b555749a | 1777 | /* |
8cab4754 | 1778 | * Move pages back to the lru list. |
b555749a | 1779 | */ |
2a1dc509 | 1780 | spin_lock_irq(&zone->lru_lock); |
556adecb | 1781 | /* |
8cab4754 WF |
1782 | * Count referenced pages from currently used mappings as rotated, |
1783 | * even though only some of them are actually re-activated. This | |
1784 | * helps balance scan pressure between file and anonymous pages in | |
1785 | * get_scan_ratio. | |
7e9cd484 | 1786 | */ |
b7c46d15 | 1787 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1788 | |
2bcf8879 | 1789 | move_active_pages_to_lru(zone, &l_active, &l_hold, |
3eb4140f | 1790 | LRU_ACTIVE + file * LRU_FILE); |
2bcf8879 | 1791 | move_active_pages_to_lru(zone, &l_inactive, &l_hold, |
3eb4140f | 1792 | LRU_BASE + file * LRU_FILE); |
a731286d | 1793 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
f8891e5e | 1794 | spin_unlock_irq(&zone->lru_lock); |
2bcf8879 HD |
1795 | |
1796 | free_hot_cold_page_list(&l_hold, 1); | |
1da177e4 LT |
1797 | } |
1798 | ||
74e3f3c3 | 1799 | #ifdef CONFIG_SWAP |
14797e23 | 1800 | static int inactive_anon_is_low_global(struct zone *zone) |
f89eb90e KM |
1801 | { |
1802 | unsigned long active, inactive; | |
1803 | ||
1804 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1805 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1806 | ||
1807 | if (inactive * zone->inactive_ratio < active) | |
1808 | return 1; | |
1809 | ||
1810 | return 0; | |
1811 | } | |
1812 | ||
14797e23 KM |
1813 | /** |
1814 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
1815 | * @zone: zone to check | |
1816 | * @sc: scan control of this context | |
1817 | * | |
1818 | * Returns true if the zone does not have enough inactive anon pages, | |
1819 | * meaning some active anon pages need to be deactivated. | |
1820 | */ | |
f16015fb | 1821 | static int inactive_anon_is_low(struct mem_cgroup_zone *mz) |
14797e23 | 1822 | { |
74e3f3c3 MK |
1823 | /* |
1824 | * If we don't have swap space, anonymous page deactivation | |
1825 | * is pointless. | |
1826 | */ | |
1827 | if (!total_swap_pages) | |
1828 | return 0; | |
1829 | ||
f16015fb JW |
1830 | if (!scanning_global_lru(mz)) |
1831 | return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup, | |
1832 | mz->zone); | |
1833 | ||
1834 | return inactive_anon_is_low_global(mz->zone); | |
14797e23 | 1835 | } |
74e3f3c3 | 1836 | #else |
f16015fb | 1837 | static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz) |
74e3f3c3 MK |
1838 | { |
1839 | return 0; | |
1840 | } | |
1841 | #endif | |
14797e23 | 1842 | |
56e49d21 RR |
1843 | static int inactive_file_is_low_global(struct zone *zone) |
1844 | { | |
1845 | unsigned long active, inactive; | |
1846 | ||
1847 | active = zone_page_state(zone, NR_ACTIVE_FILE); | |
1848 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1849 | ||
1850 | return (active > inactive); | |
1851 | } | |
1852 | ||
1853 | /** | |
1854 | * inactive_file_is_low - check if file pages need to be deactivated | |
f16015fb | 1855 | * @mz: memory cgroup and zone to check |
56e49d21 RR |
1856 | * |
1857 | * When the system is doing streaming IO, memory pressure here | |
1858 | * ensures that active file pages get deactivated, until more | |
1859 | * than half of the file pages are on the inactive list. | |
1860 | * | |
1861 | * Once we get to that situation, protect the system's working | |
1862 | * set from being evicted by disabling active file page aging. | |
1863 | * | |
1864 | * This uses a different ratio than the anonymous pages, because | |
1865 | * the page cache uses a use-once replacement algorithm. | |
1866 | */ | |
f16015fb | 1867 | static int inactive_file_is_low(struct mem_cgroup_zone *mz) |
56e49d21 | 1868 | { |
f16015fb JW |
1869 | if (!scanning_global_lru(mz)) |
1870 | return mem_cgroup_inactive_file_is_low(mz->mem_cgroup, | |
1871 | mz->zone); | |
56e49d21 | 1872 | |
f16015fb | 1873 | return inactive_file_is_low_global(mz->zone); |
56e49d21 RR |
1874 | } |
1875 | ||
f16015fb | 1876 | static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file) |
b39415b2 RR |
1877 | { |
1878 | if (file) | |
f16015fb | 1879 | return inactive_file_is_low(mz); |
b39415b2 | 1880 | else |
f16015fb | 1881 | return inactive_anon_is_low(mz); |
b39415b2 RR |
1882 | } |
1883 | ||
4f98a2fe | 1884 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
f16015fb JW |
1885 | struct mem_cgroup_zone *mz, |
1886 | struct scan_control *sc, int priority) | |
b69408e8 | 1887 | { |
4f98a2fe RR |
1888 | int file = is_file_lru(lru); |
1889 | ||
b39415b2 | 1890 | if (is_active_lru(lru)) { |
f16015fb JW |
1891 | if (inactive_list_is_low(mz, file)) |
1892 | shrink_active_list(nr_to_scan, mz, sc, priority, file); | |
556adecb RR |
1893 | return 0; |
1894 | } | |
1895 | ||
f16015fb | 1896 | return shrink_inactive_list(nr_to_scan, mz, sc, priority, file); |
4f98a2fe RR |
1897 | } |
1898 | ||
f16015fb JW |
1899 | static int vmscan_swappiness(struct mem_cgroup_zone *mz, |
1900 | struct scan_control *sc) | |
1f4c025b | 1901 | { |
89b5fae5 | 1902 | if (global_reclaim(sc)) |
1f4c025b | 1903 | return vm_swappiness; |
f16015fb | 1904 | return mem_cgroup_swappiness(mz->mem_cgroup); |
1f4c025b KH |
1905 | } |
1906 | ||
4f98a2fe RR |
1907 | /* |
1908 | * Determine how aggressively the anon and file LRU lists should be | |
1909 | * scanned. The relative value of each set of LRU lists is determined | |
1910 | * by looking at the fraction of the pages scanned we did rotate back | |
1911 | * onto the active list instead of evict. | |
1912 | * | |
76a33fc3 | 1913 | * nr[0] = anon pages to scan; nr[1] = file pages to scan |
4f98a2fe | 1914 | */ |
f16015fb JW |
1915 | static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc, |
1916 | unsigned long *nr, int priority) | |
4f98a2fe RR |
1917 | { |
1918 | unsigned long anon, file, free; | |
1919 | unsigned long anon_prio, file_prio; | |
1920 | unsigned long ap, fp; | |
f16015fb | 1921 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
76a33fc3 SL |
1922 | u64 fraction[2], denominator; |
1923 | enum lru_list l; | |
1924 | int noswap = 0; | |
a4d3e9e7 | 1925 | bool force_scan = false; |
246e87a9 | 1926 | |
f11c0ca5 JW |
1927 | /* |
1928 | * If the zone or memcg is small, nr[l] can be 0. This | |
1929 | * results in no scanning on this priority and a potential | |
1930 | * priority drop. Global direct reclaim can go to the next | |
1931 | * zone and tends to have no problems. Global kswapd is for | |
1932 | * zone balancing and it needs to scan a minimum amount. When | |
1933 | * reclaiming for a memcg, a priority drop can cause high | |
1934 | * latencies, so it's better to scan a minimum amount there as | |
1935 | * well. | |
1936 | */ | |
b95a2f2d | 1937 | if (current_is_kswapd() && mz->zone->all_unreclaimable) |
a4d3e9e7 | 1938 | force_scan = true; |
89b5fae5 | 1939 | if (!global_reclaim(sc)) |
a4d3e9e7 | 1940 | force_scan = true; |
76a33fc3 SL |
1941 | |
1942 | /* If we have no swap space, do not bother scanning anon pages. */ | |
1943 | if (!sc->may_swap || (nr_swap_pages <= 0)) { | |
1944 | noswap = 1; | |
1945 | fraction[0] = 0; | |
1946 | fraction[1] = 1; | |
1947 | denominator = 1; | |
1948 | goto out; | |
1949 | } | |
4f98a2fe | 1950 | |
f16015fb JW |
1951 | anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) + |
1952 | zone_nr_lru_pages(mz, LRU_INACTIVE_ANON); | |
1953 | file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) + | |
1954 | zone_nr_lru_pages(mz, LRU_INACTIVE_FILE); | |
a4d3e9e7 | 1955 | |
89b5fae5 | 1956 | if (global_reclaim(sc)) { |
f16015fb | 1957 | free = zone_page_state(mz->zone, NR_FREE_PAGES); |
eeee9a8c KM |
1958 | /* If we have very few page cache pages, |
1959 | force-scan anon pages. */ | |
f16015fb | 1960 | if (unlikely(file + free <= high_wmark_pages(mz->zone))) { |
76a33fc3 SL |
1961 | fraction[0] = 1; |
1962 | fraction[1] = 0; | |
1963 | denominator = 1; | |
1964 | goto out; | |
eeee9a8c | 1965 | } |
4f98a2fe RR |
1966 | } |
1967 | ||
58c37f6e KM |
1968 | /* |
1969 | * With swappiness at 100, anonymous and file have the same priority. | |
1970 | * This scanning priority is essentially the inverse of IO cost. | |
1971 | */ | |
f16015fb JW |
1972 | anon_prio = vmscan_swappiness(mz, sc); |
1973 | file_prio = 200 - vmscan_swappiness(mz, sc); | |
58c37f6e | 1974 | |
4f98a2fe RR |
1975 | /* |
1976 | * OK, so we have swap space and a fair amount of page cache | |
1977 | * pages. We use the recently rotated / recently scanned | |
1978 | * ratios to determine how valuable each cache is. | |
1979 | * | |
1980 | * Because workloads change over time (and to avoid overflow) | |
1981 | * we keep these statistics as a floating average, which ends | |
1982 | * up weighing recent references more than old ones. | |
1983 | * | |
1984 | * anon in [0], file in [1] | |
1985 | */ | |
f16015fb | 1986 | spin_lock_irq(&mz->zone->lru_lock); |
6e901571 | 1987 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
1988 | reclaim_stat->recent_scanned[0] /= 2; |
1989 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
1990 | } |
1991 | ||
6e901571 | 1992 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
1993 | reclaim_stat->recent_scanned[1] /= 2; |
1994 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
1995 | } |
1996 | ||
4f98a2fe | 1997 | /* |
00d8089c RR |
1998 | * The amount of pressure on anon vs file pages is inversely |
1999 | * proportional to the fraction of recently scanned pages on | |
2000 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2001 | */ |
6e901571 KM |
2002 | ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1); |
2003 | ap /= reclaim_stat->recent_rotated[0] + 1; | |
4f98a2fe | 2004 | |
6e901571 KM |
2005 | fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1); |
2006 | fp /= reclaim_stat->recent_rotated[1] + 1; | |
f16015fb | 2007 | spin_unlock_irq(&mz->zone->lru_lock); |
4f98a2fe | 2008 | |
76a33fc3 SL |
2009 | fraction[0] = ap; |
2010 | fraction[1] = fp; | |
2011 | denominator = ap + fp + 1; | |
2012 | out: | |
2013 | for_each_evictable_lru(l) { | |
2014 | int file = is_file_lru(l); | |
2015 | unsigned long scan; | |
6e08a369 | 2016 | |
f16015fb | 2017 | scan = zone_nr_lru_pages(mz, l); |
76a33fc3 SL |
2018 | if (priority || noswap) { |
2019 | scan >>= priority; | |
f11c0ca5 JW |
2020 | if (!scan && force_scan) |
2021 | scan = SWAP_CLUSTER_MAX; | |
76a33fc3 SL |
2022 | scan = div64_u64(scan * fraction[file], denominator); |
2023 | } | |
246e87a9 | 2024 | nr[l] = scan; |
76a33fc3 | 2025 | } |
6e08a369 | 2026 | } |
4f98a2fe | 2027 | |
3e7d3449 MG |
2028 | /* |
2029 | * Reclaim/compaction depends on a number of pages being freed. To avoid | |
2030 | * disruption to the system, a small number of order-0 pages continue to be | |
2031 | * rotated and reclaimed in the normal fashion. However, by the time we get | |
2032 | * back to the allocator and call try_to_compact_zone(), we ensure that | |
2033 | * there are enough free pages for it to be likely successful | |
2034 | */ | |
f16015fb | 2035 | static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz, |
3e7d3449 MG |
2036 | unsigned long nr_reclaimed, |
2037 | unsigned long nr_scanned, | |
2038 | struct scan_control *sc) | |
2039 | { | |
2040 | unsigned long pages_for_compaction; | |
2041 | unsigned long inactive_lru_pages; | |
2042 | ||
2043 | /* If not in reclaim/compaction mode, stop */ | |
f3a310bc | 2044 | if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION)) |
3e7d3449 MG |
2045 | return false; |
2046 | ||
2876592f MG |
2047 | /* Consider stopping depending on scan and reclaim activity */ |
2048 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2049 | /* | |
2050 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2051 | * full LRU list has been scanned and we are still failing | |
2052 | * to reclaim pages. This full LRU scan is potentially | |
2053 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2054 | */ | |
2055 | if (!nr_reclaimed && !nr_scanned) | |
2056 | return false; | |
2057 | } else { | |
2058 | /* | |
2059 | * For non-__GFP_REPEAT allocations which can presumably | |
2060 | * fail without consequence, stop if we failed to reclaim | |
2061 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2062 | * pages that were scanned. This will return to the | |
2063 | * caller faster at the risk reclaim/compaction and | |
2064 | * the resulting allocation attempt fails | |
2065 | */ | |
2066 | if (!nr_reclaimed) | |
2067 | return false; | |
2068 | } | |
3e7d3449 MG |
2069 | |
2070 | /* | |
2071 | * If we have not reclaimed enough pages for compaction and the | |
2072 | * inactive lists are large enough, continue reclaiming | |
2073 | */ | |
2074 | pages_for_compaction = (2UL << sc->order); | |
f16015fb | 2075 | inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE); |
86cfd3a4 | 2076 | if (nr_swap_pages > 0) |
f16015fb | 2077 | inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON); |
3e7d3449 MG |
2078 | if (sc->nr_reclaimed < pages_for_compaction && |
2079 | inactive_lru_pages > pages_for_compaction) | |
2080 | return true; | |
2081 | ||
2082 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
f16015fb | 2083 | switch (compaction_suitable(mz->zone, sc->order)) { |
3e7d3449 MG |
2084 | case COMPACT_PARTIAL: |
2085 | case COMPACT_CONTINUE: | |
2086 | return false; | |
2087 | default: | |
2088 | return true; | |
2089 | } | |
2090 | } | |
2091 | ||
1da177e4 LT |
2092 | /* |
2093 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
2094 | */ | |
f16015fb JW |
2095 | static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz, |
2096 | struct scan_control *sc) | |
1da177e4 | 2097 | { |
b69408e8 | 2098 | unsigned long nr[NR_LRU_LISTS]; |
8695949a | 2099 | unsigned long nr_to_scan; |
b69408e8 | 2100 | enum lru_list l; |
f0fdc5e8 | 2101 | unsigned long nr_reclaimed, nr_scanned; |
22fba335 | 2102 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; |
3da367c3 | 2103 | struct blk_plug plug; |
e0f79b8f | 2104 | |
3e7d3449 MG |
2105 | restart: |
2106 | nr_reclaimed = 0; | |
f0fdc5e8 | 2107 | nr_scanned = sc->nr_scanned; |
f16015fb | 2108 | get_scan_count(mz, sc, nr, priority); |
1da177e4 | 2109 | |
3da367c3 | 2110 | blk_start_plug(&plug); |
556adecb RR |
2111 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || |
2112 | nr[LRU_INACTIVE_FILE]) { | |
894bc310 | 2113 | for_each_evictable_lru(l) { |
b69408e8 | 2114 | if (nr[l]) { |
ece74b2e KM |
2115 | nr_to_scan = min_t(unsigned long, |
2116 | nr[l], SWAP_CLUSTER_MAX); | |
b69408e8 | 2117 | nr[l] -= nr_to_scan; |
1da177e4 | 2118 | |
01dbe5c9 | 2119 | nr_reclaimed += shrink_list(l, nr_to_scan, |
f16015fb | 2120 | mz, sc, priority); |
b69408e8 | 2121 | } |
1da177e4 | 2122 | } |
a79311c1 RR |
2123 | /* |
2124 | * On large memory systems, scan >> priority can become | |
2125 | * really large. This is fine for the starting priority; | |
2126 | * we want to put equal scanning pressure on each zone. | |
2127 | * However, if the VM has a harder time of freeing pages, | |
2128 | * with multiple processes reclaiming pages, the total | |
2129 | * freeing target can get unreasonably large. | |
2130 | */ | |
338fde90 | 2131 | if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY) |
a79311c1 | 2132 | break; |
1da177e4 | 2133 | } |
3da367c3 | 2134 | blk_finish_plug(&plug); |
3e7d3449 | 2135 | sc->nr_reclaimed += nr_reclaimed; |
01dbe5c9 | 2136 | |
556adecb RR |
2137 | /* |
2138 | * Even if we did not try to evict anon pages at all, we want to | |
2139 | * rebalance the anon lru active/inactive ratio. | |
2140 | */ | |
f16015fb JW |
2141 | if (inactive_anon_is_low(mz)) |
2142 | shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0); | |
556adecb | 2143 | |
3e7d3449 | 2144 | /* reclaim/compaction might need reclaim to continue */ |
f16015fb | 2145 | if (should_continue_reclaim(mz, nr_reclaimed, |
3e7d3449 MG |
2146 | sc->nr_scanned - nr_scanned, sc)) |
2147 | goto restart; | |
2148 | ||
232ea4d6 | 2149 | throttle_vm_writeout(sc->gfp_mask); |
1da177e4 LT |
2150 | } |
2151 | ||
f16015fb JW |
2152 | static void shrink_zone(int priority, struct zone *zone, |
2153 | struct scan_control *sc) | |
2154 | { | |
5660048c JW |
2155 | struct mem_cgroup *root = sc->target_mem_cgroup; |
2156 | struct mem_cgroup_reclaim_cookie reclaim = { | |
f16015fb | 2157 | .zone = zone, |
5660048c | 2158 | .priority = priority, |
f16015fb | 2159 | }; |
5660048c JW |
2160 | struct mem_cgroup *memcg; |
2161 | ||
5660048c JW |
2162 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2163 | do { | |
2164 | struct mem_cgroup_zone mz = { | |
2165 | .mem_cgroup = memcg, | |
2166 | .zone = zone, | |
2167 | }; | |
f16015fb | 2168 | |
5660048c JW |
2169 | shrink_mem_cgroup_zone(priority, &mz, sc); |
2170 | /* | |
2171 | * Limit reclaim has historically picked one memcg and | |
2172 | * scanned it with decreasing priority levels until | |
2173 | * nr_to_reclaim had been reclaimed. This priority | |
2174 | * cycle is thus over after a single memcg. | |
b95a2f2d JW |
2175 | * |
2176 | * Direct reclaim and kswapd, on the other hand, have | |
2177 | * to scan all memory cgroups to fulfill the overall | |
2178 | * scan target for the zone. | |
5660048c JW |
2179 | */ |
2180 | if (!global_reclaim(sc)) { | |
2181 | mem_cgroup_iter_break(root, memcg); | |
2182 | break; | |
2183 | } | |
2184 | memcg = mem_cgroup_iter(root, memcg, &reclaim); | |
2185 | } while (memcg); | |
f16015fb JW |
2186 | } |
2187 | ||
fe4b1b24 MG |
2188 | /* Returns true if compaction should go ahead for a high-order request */ |
2189 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
2190 | { | |
2191 | unsigned long balance_gap, watermark; | |
2192 | bool watermark_ok; | |
2193 | ||
2194 | /* Do not consider compaction for orders reclaim is meant to satisfy */ | |
2195 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
2196 | return false; | |
2197 | ||
2198 | /* | |
2199 | * Compaction takes time to run and there are potentially other | |
2200 | * callers using the pages just freed. Continue reclaiming until | |
2201 | * there is a buffer of free pages available to give compaction | |
2202 | * a reasonable chance of completing and allocating the page | |
2203 | */ | |
2204 | balance_gap = min(low_wmark_pages(zone), | |
2205 | (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2206 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
2207 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order); | |
2208 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); | |
2209 | ||
2210 | /* | |
2211 | * If compaction is deferred, reclaim up to a point where | |
2212 | * compaction will have a chance of success when re-enabled | |
2213 | */ | |
2214 | if (compaction_deferred(zone)) | |
2215 | return watermark_ok; | |
2216 | ||
2217 | /* If compaction is not ready to start, keep reclaiming */ | |
2218 | if (!compaction_suitable(zone, sc->order)) | |
2219 | return false; | |
2220 | ||
2221 | return watermark_ok; | |
2222 | } | |
2223 | ||
1da177e4 LT |
2224 | /* |
2225 | * This is the direct reclaim path, for page-allocating processes. We only | |
2226 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2227 | * request. | |
2228 | * | |
41858966 MG |
2229 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
2230 | * Because: | |
1da177e4 LT |
2231 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
2232 | * allocation or | |
41858966 MG |
2233 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
2234 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2235 | * zone defense algorithm. | |
1da177e4 | 2236 | * |
1da177e4 LT |
2237 | * If a zone is deemed to be full of pinned pages then just give it a light |
2238 | * scan then give up on it. | |
e0c23279 MG |
2239 | * |
2240 | * This function returns true if a zone is being reclaimed for a costly | |
fe4b1b24 | 2241 | * high-order allocation and compaction is ready to begin. This indicates to |
0cee34fd MG |
2242 | * the caller that it should consider retrying the allocation instead of |
2243 | * further reclaim. | |
1da177e4 | 2244 | */ |
e0c23279 | 2245 | static bool shrink_zones(int priority, struct zonelist *zonelist, |
05ff5137 | 2246 | struct scan_control *sc) |
1da177e4 | 2247 | { |
dd1a239f | 2248 | struct zoneref *z; |
54a6eb5c | 2249 | struct zone *zone; |
d149e3b2 YH |
2250 | unsigned long nr_soft_reclaimed; |
2251 | unsigned long nr_soft_scanned; | |
0cee34fd | 2252 | bool aborted_reclaim = false; |
1cfb419b | 2253 | |
d4debc66 MG |
2254 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
2255 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
f3fe6512 | 2256 | if (!populated_zone(zone)) |
1da177e4 | 2257 | continue; |
1cfb419b KH |
2258 | /* |
2259 | * Take care memory controller reclaiming has small influence | |
2260 | * to global LRU. | |
2261 | */ | |
89b5fae5 | 2262 | if (global_reclaim(sc)) { |
1cfb419b KH |
2263 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2264 | continue; | |
93e4a89a | 2265 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
1cfb419b | 2266 | continue; /* Let kswapd poll it */ |
e0887c19 RR |
2267 | if (COMPACTION_BUILD) { |
2268 | /* | |
e0c23279 MG |
2269 | * If we already have plenty of memory free for |
2270 | * compaction in this zone, don't free any more. | |
2271 | * Even though compaction is invoked for any | |
2272 | * non-zero order, only frequent costly order | |
2273 | * reclamation is disruptive enough to become a | |
2274 | * noticable problem, like transparent huge page | |
2275 | * allocations. | |
e0887c19 | 2276 | */ |
fe4b1b24 | 2277 | if (compaction_ready(zone, sc)) { |
0cee34fd | 2278 | aborted_reclaim = true; |
e0887c19 | 2279 | continue; |
e0c23279 | 2280 | } |
e0887c19 | 2281 | } |
ac34a1a3 KH |
2282 | /* |
2283 | * This steals pages from memory cgroups over softlimit | |
2284 | * and returns the number of reclaimed pages and | |
2285 | * scanned pages. This works for global memory pressure | |
2286 | * and balancing, not for a memcg's limit. | |
2287 | */ | |
2288 | nr_soft_scanned = 0; | |
2289 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2290 | sc->order, sc->gfp_mask, | |
2291 | &nr_soft_scanned); | |
2292 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2293 | sc->nr_scanned += nr_soft_scanned; | |
2294 | /* need some check for avoid more shrink_zone() */ | |
1cfb419b | 2295 | } |
408d8544 | 2296 | |
a79311c1 | 2297 | shrink_zone(priority, zone, sc); |
1da177e4 | 2298 | } |
e0c23279 | 2299 | |
0cee34fd | 2300 | return aborted_reclaim; |
d1908362 MK |
2301 | } |
2302 | ||
2303 | static bool zone_reclaimable(struct zone *zone) | |
2304 | { | |
2305 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
2306 | } | |
2307 | ||
929bea7c | 2308 | /* All zones in zonelist are unreclaimable? */ |
d1908362 MK |
2309 | static bool all_unreclaimable(struct zonelist *zonelist, |
2310 | struct scan_control *sc) | |
2311 | { | |
2312 | struct zoneref *z; | |
2313 | struct zone *zone; | |
d1908362 MK |
2314 | |
2315 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2316 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2317 | if (!populated_zone(zone)) | |
2318 | continue; | |
2319 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2320 | continue; | |
929bea7c KM |
2321 | if (!zone->all_unreclaimable) |
2322 | return false; | |
d1908362 MK |
2323 | } |
2324 | ||
929bea7c | 2325 | return true; |
1da177e4 | 2326 | } |
4f98a2fe | 2327 | |
1da177e4 LT |
2328 | /* |
2329 | * This is the main entry point to direct page reclaim. | |
2330 | * | |
2331 | * If a full scan of the inactive list fails to free enough memory then we | |
2332 | * are "out of memory" and something needs to be killed. | |
2333 | * | |
2334 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2335 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2336 | * caller can't do much about. We kick the writeback threads and take explicit |
2337 | * naps in the hope that some of these pages can be written. But if the | |
2338 | * allocating task holds filesystem locks which prevent writeout this might not | |
2339 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2340 | * |
2341 | * returns: 0, if no pages reclaimed | |
2342 | * else, the number of pages reclaimed | |
1da177e4 | 2343 | */ |
dac1d27b | 2344 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
a09ed5e0 YH |
2345 | struct scan_control *sc, |
2346 | struct shrink_control *shrink) | |
1da177e4 LT |
2347 | { |
2348 | int priority; | |
69e05944 | 2349 | unsigned long total_scanned = 0; |
1da177e4 | 2350 | struct reclaim_state *reclaim_state = current->reclaim_state; |
dd1a239f | 2351 | struct zoneref *z; |
54a6eb5c | 2352 | struct zone *zone; |
22fba335 | 2353 | unsigned long writeback_threshold; |
0cee34fd | 2354 | bool aborted_reclaim; |
1da177e4 | 2355 | |
c0ff7453 | 2356 | get_mems_allowed(); |
873b4771 KK |
2357 | delayacct_freepages_start(); |
2358 | ||
89b5fae5 | 2359 | if (global_reclaim(sc)) |
1cfb419b | 2360 | count_vm_event(ALLOCSTALL); |
1da177e4 LT |
2361 | |
2362 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { | |
66e1707b | 2363 | sc->nr_scanned = 0; |
f7b7fd8f | 2364 | if (!priority) |
f16015fb | 2365 | disable_swap_token(sc->target_mem_cgroup); |
0cee34fd | 2366 | aborted_reclaim = shrink_zones(priority, zonelist, sc); |
e0c23279 | 2367 | |
66e1707b BS |
2368 | /* |
2369 | * Don't shrink slabs when reclaiming memory from | |
2370 | * over limit cgroups | |
2371 | */ | |
89b5fae5 | 2372 | if (global_reclaim(sc)) { |
c6a8a8c5 | 2373 | unsigned long lru_pages = 0; |
d4debc66 MG |
2374 | for_each_zone_zonelist(zone, z, zonelist, |
2375 | gfp_zone(sc->gfp_mask)) { | |
c6a8a8c5 KM |
2376 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2377 | continue; | |
2378 | ||
2379 | lru_pages += zone_reclaimable_pages(zone); | |
2380 | } | |
2381 | ||
1495f230 | 2382 | shrink_slab(shrink, sc->nr_scanned, lru_pages); |
91a45470 | 2383 | if (reclaim_state) { |
a79311c1 | 2384 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
91a45470 KH |
2385 | reclaim_state->reclaimed_slab = 0; |
2386 | } | |
1da177e4 | 2387 | } |
66e1707b | 2388 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2389 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
1da177e4 | 2390 | goto out; |
1da177e4 LT |
2391 | |
2392 | /* | |
2393 | * Try to write back as many pages as we just scanned. This | |
2394 | * tends to cause slow streaming writers to write data to the | |
2395 | * disk smoothly, at the dirtying rate, which is nice. But | |
2396 | * that's undesirable in laptop mode, where we *want* lumpy | |
2397 | * writeout. So in laptop mode, write out the whole world. | |
2398 | */ | |
22fba335 KM |
2399 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2400 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2401 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2402 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2403 | sc->may_writepage = 1; |
1da177e4 LT |
2404 | } |
2405 | ||
2406 | /* Take a nap, wait for some writeback to complete */ | |
7b51755c | 2407 | if (!sc->hibernation_mode && sc->nr_scanned && |
0e093d99 MG |
2408 | priority < DEF_PRIORITY - 2) { |
2409 | struct zone *preferred_zone; | |
2410 | ||
2411 | first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), | |
f33261d7 DR |
2412 | &cpuset_current_mems_allowed, |
2413 | &preferred_zone); | |
0e093d99 MG |
2414 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); |
2415 | } | |
1da177e4 | 2416 | } |
bb21c7ce | 2417 | |
1da177e4 | 2418 | out: |
873b4771 | 2419 | delayacct_freepages_end(); |
c0ff7453 | 2420 | put_mems_allowed(); |
873b4771 | 2421 | |
bb21c7ce KM |
2422 | if (sc->nr_reclaimed) |
2423 | return sc->nr_reclaimed; | |
2424 | ||
929bea7c KM |
2425 | /* |
2426 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2427 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2428 | * check. | |
2429 | */ | |
2430 | if (oom_killer_disabled) | |
2431 | return 0; | |
2432 | ||
0cee34fd MG |
2433 | /* Aborted reclaim to try compaction? don't OOM, then */ |
2434 | if (aborted_reclaim) | |
7335084d MG |
2435 | return 1; |
2436 | ||
bb21c7ce | 2437 | /* top priority shrink_zones still had more to do? don't OOM, then */ |
89b5fae5 | 2438 | if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc)) |
bb21c7ce KM |
2439 | return 1; |
2440 | ||
2441 | return 0; | |
1da177e4 LT |
2442 | } |
2443 | ||
dac1d27b | 2444 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2445 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2446 | { |
33906bc5 | 2447 | unsigned long nr_reclaimed; |
66e1707b BS |
2448 | struct scan_control sc = { |
2449 | .gfp_mask = gfp_mask, | |
2450 | .may_writepage = !laptop_mode, | |
22fba335 | 2451 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
a6dc60f8 | 2452 | .may_unmap = 1, |
2e2e4259 | 2453 | .may_swap = 1, |
66e1707b | 2454 | .order = order, |
f16015fb | 2455 | .target_mem_cgroup = NULL, |
327c0e96 | 2456 | .nodemask = nodemask, |
66e1707b | 2457 | }; |
a09ed5e0 YH |
2458 | struct shrink_control shrink = { |
2459 | .gfp_mask = sc.gfp_mask, | |
2460 | }; | |
66e1707b | 2461 | |
33906bc5 MG |
2462 | trace_mm_vmscan_direct_reclaim_begin(order, |
2463 | sc.may_writepage, | |
2464 | gfp_mask); | |
2465 | ||
a09ed5e0 | 2466 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
33906bc5 MG |
2467 | |
2468 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2469 | ||
2470 | return nr_reclaimed; | |
66e1707b BS |
2471 | } |
2472 | ||
00f0b825 | 2473 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
66e1707b | 2474 | |
72835c86 | 2475 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, |
4e416953 | 2476 | gfp_t gfp_mask, bool noswap, |
0ae5e89c YH |
2477 | struct zone *zone, |
2478 | unsigned long *nr_scanned) | |
4e416953 BS |
2479 | { |
2480 | struct scan_control sc = { | |
0ae5e89c | 2481 | .nr_scanned = 0, |
b8f5c566 | 2482 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
4e416953 BS |
2483 | .may_writepage = !laptop_mode, |
2484 | .may_unmap = 1, | |
2485 | .may_swap = !noswap, | |
4e416953 | 2486 | .order = 0, |
72835c86 | 2487 | .target_mem_cgroup = memcg, |
4e416953 | 2488 | }; |
5660048c | 2489 | struct mem_cgroup_zone mz = { |
72835c86 | 2490 | .mem_cgroup = memcg, |
5660048c JW |
2491 | .zone = zone, |
2492 | }; | |
0ae5e89c | 2493 | |
4e416953 BS |
2494 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2495 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e KM |
2496 | |
2497 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(0, | |
2498 | sc.may_writepage, | |
2499 | sc.gfp_mask); | |
2500 | ||
4e416953 BS |
2501 | /* |
2502 | * NOTE: Although we can get the priority field, using it | |
2503 | * here is not a good idea, since it limits the pages we can scan. | |
2504 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2505 | * will pick up pages from other mem cgroup's as well. We hack | |
2506 | * the priority and make it zero. | |
2507 | */ | |
5660048c | 2508 | shrink_mem_cgroup_zone(0, &mz, &sc); |
bdce6d9e KM |
2509 | |
2510 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2511 | ||
0ae5e89c | 2512 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2513 | return sc.nr_reclaimed; |
2514 | } | |
2515 | ||
72835c86 | 2516 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
a7885eb8 | 2517 | gfp_t gfp_mask, |
185efc0f | 2518 | bool noswap) |
66e1707b | 2519 | { |
4e416953 | 2520 | struct zonelist *zonelist; |
bdce6d9e | 2521 | unsigned long nr_reclaimed; |
889976db | 2522 | int nid; |
66e1707b | 2523 | struct scan_control sc = { |
66e1707b | 2524 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2525 | .may_unmap = 1, |
2e2e4259 | 2526 | .may_swap = !noswap, |
22fba335 | 2527 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
66e1707b | 2528 | .order = 0, |
72835c86 | 2529 | .target_mem_cgroup = memcg, |
327c0e96 | 2530 | .nodemask = NULL, /* we don't care the placement */ |
a09ed5e0 YH |
2531 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2532 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2533 | }; | |
2534 | struct shrink_control shrink = { | |
2535 | .gfp_mask = sc.gfp_mask, | |
66e1707b | 2536 | }; |
66e1707b | 2537 | |
889976db YH |
2538 | /* |
2539 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2540 | * take care of from where we get pages. So the node where we start the | |
2541 | * scan does not need to be the current node. | |
2542 | */ | |
72835c86 | 2543 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2544 | |
2545 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2546 | |
2547 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2548 | sc.may_writepage, | |
2549 | sc.gfp_mask); | |
2550 | ||
a09ed5e0 | 2551 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
bdce6d9e KM |
2552 | |
2553 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2554 | ||
2555 | return nr_reclaimed; | |
66e1707b BS |
2556 | } |
2557 | #endif | |
2558 | ||
f16015fb JW |
2559 | static void age_active_anon(struct zone *zone, struct scan_control *sc, |
2560 | int priority) | |
2561 | { | |
b95a2f2d | 2562 | struct mem_cgroup *memcg; |
f16015fb | 2563 | |
b95a2f2d JW |
2564 | if (!total_swap_pages) |
2565 | return; | |
2566 | ||
2567 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2568 | do { | |
2569 | struct mem_cgroup_zone mz = { | |
2570 | .mem_cgroup = memcg, | |
2571 | .zone = zone, | |
2572 | }; | |
2573 | ||
2574 | if (inactive_anon_is_low(&mz)) | |
2575 | shrink_active_list(SWAP_CLUSTER_MAX, &mz, | |
2576 | sc, priority, 0); | |
2577 | ||
2578 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2579 | } while (memcg); | |
f16015fb JW |
2580 | } |
2581 | ||
1741c877 MG |
2582 | /* |
2583 | * pgdat_balanced is used when checking if a node is balanced for high-order | |
2584 | * allocations. Only zones that meet watermarks and are in a zone allowed | |
2585 | * by the callers classzone_idx are added to balanced_pages. The total of | |
2586 | * balanced pages must be at least 25% of the zones allowed by classzone_idx | |
2587 | * for the node to be considered balanced. Forcing all zones to be balanced | |
2588 | * for high orders can cause excessive reclaim when there are imbalanced zones. | |
2589 | * The choice of 25% is due to | |
2590 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2591 | * reasonable sized machine | |
2592 | * o On all other machines, the top zone must be at least a reasonable | |
25985edc | 2593 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
1741c877 MG |
2594 | * would need to be at least 256M for it to be balance a whole node. |
2595 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2596 | * to balance a node on its own. These seemed like reasonable ratios. | |
2597 | */ | |
2598 | static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages, | |
2599 | int classzone_idx) | |
2600 | { | |
2601 | unsigned long present_pages = 0; | |
2602 | int i; | |
2603 | ||
2604 | for (i = 0; i <= classzone_idx; i++) | |
2605 | present_pages += pgdat->node_zones[i].present_pages; | |
2606 | ||
4746efde SL |
2607 | /* A special case here: if zone has no page, we think it's balanced */ |
2608 | return balanced_pages >= (present_pages >> 2); | |
1741c877 MG |
2609 | } |
2610 | ||
f50de2d3 | 2611 | /* is kswapd sleeping prematurely? */ |
dc83edd9 MG |
2612 | static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining, |
2613 | int classzone_idx) | |
f50de2d3 | 2614 | { |
bb3ab596 | 2615 | int i; |
1741c877 MG |
2616 | unsigned long balanced = 0; |
2617 | bool all_zones_ok = true; | |
f50de2d3 MG |
2618 | |
2619 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ | |
2620 | if (remaining) | |
dc83edd9 | 2621 | return true; |
f50de2d3 | 2622 | |
0abdee2b | 2623 | /* Check the watermark levels */ |
08951e54 | 2624 | for (i = 0; i <= classzone_idx; i++) { |
bb3ab596 KM |
2625 | struct zone *zone = pgdat->node_zones + i; |
2626 | ||
2627 | if (!populated_zone(zone)) | |
2628 | continue; | |
2629 | ||
355b09c4 MG |
2630 | /* |
2631 | * balance_pgdat() skips over all_unreclaimable after | |
2632 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2633 | * they must be considered balanced here as well if kswapd | |
2634 | * is to sleep | |
2635 | */ | |
2636 | if (zone->all_unreclaimable) { | |
2637 | balanced += zone->present_pages; | |
de3fab39 | 2638 | continue; |
355b09c4 | 2639 | } |
de3fab39 | 2640 | |
88f5acf8 | 2641 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone), |
da175d06 | 2642 | i, 0)) |
1741c877 MG |
2643 | all_zones_ok = false; |
2644 | else | |
2645 | balanced += zone->present_pages; | |
bb3ab596 | 2646 | } |
f50de2d3 | 2647 | |
1741c877 MG |
2648 | /* |
2649 | * For high-order requests, the balanced zones must contain at least | |
2650 | * 25% of the nodes pages for kswapd to sleep. For order-0, all zones | |
2651 | * must be balanced | |
2652 | */ | |
2653 | if (order) | |
afc7e326 | 2654 | return !pgdat_balanced(pgdat, balanced, classzone_idx); |
1741c877 MG |
2655 | else |
2656 | return !all_zones_ok; | |
f50de2d3 MG |
2657 | } |
2658 | ||
1da177e4 LT |
2659 | /* |
2660 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 2661 | * they are all at high_wmark_pages(zone). |
1da177e4 | 2662 | * |
0abdee2b | 2663 | * Returns the final order kswapd was reclaiming at |
1da177e4 LT |
2664 | * |
2665 | * There is special handling here for zones which are full of pinned pages. | |
2666 | * This can happen if the pages are all mlocked, or if they are all used by | |
2667 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
2668 | * What we do is to detect the case where all pages in the zone have been | |
2669 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
2670 | * dead and from now on, only perform a short scan. Basically we're polling | |
2671 | * the zone for when the problem goes away. | |
2672 | * | |
2673 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
2674 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
2675 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
2676 | * lower zones regardless of the number of free pages in the lower zones. This | |
2677 | * interoperates with the page allocator fallback scheme to ensure that aging | |
2678 | * of pages is balanced across the zones. | |
1da177e4 | 2679 | */ |
99504748 | 2680 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, |
dc83edd9 | 2681 | int *classzone_idx) |
1da177e4 | 2682 | { |
1da177e4 | 2683 | int all_zones_ok; |
1741c877 | 2684 | unsigned long balanced; |
1da177e4 LT |
2685 | int priority; |
2686 | int i; | |
99504748 | 2687 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
69e05944 | 2688 | unsigned long total_scanned; |
1da177e4 | 2689 | struct reclaim_state *reclaim_state = current->reclaim_state; |
0ae5e89c YH |
2690 | unsigned long nr_soft_reclaimed; |
2691 | unsigned long nr_soft_scanned; | |
179e9639 AM |
2692 | struct scan_control sc = { |
2693 | .gfp_mask = GFP_KERNEL, | |
a6dc60f8 | 2694 | .may_unmap = 1, |
2e2e4259 | 2695 | .may_swap = 1, |
22fba335 KM |
2696 | /* |
2697 | * kswapd doesn't want to be bailed out while reclaim. because | |
2698 | * we want to put equal scanning pressure on each zone. | |
2699 | */ | |
2700 | .nr_to_reclaim = ULONG_MAX, | |
5ad333eb | 2701 | .order = order, |
f16015fb | 2702 | .target_mem_cgroup = NULL, |
179e9639 | 2703 | }; |
a09ed5e0 YH |
2704 | struct shrink_control shrink = { |
2705 | .gfp_mask = sc.gfp_mask, | |
2706 | }; | |
1da177e4 LT |
2707 | loop_again: |
2708 | total_scanned = 0; | |
a79311c1 | 2709 | sc.nr_reclaimed = 0; |
c0bbbc73 | 2710 | sc.may_writepage = !laptop_mode; |
f8891e5e | 2711 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 2712 | |
1da177e4 | 2713 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { |
1da177e4 | 2714 | unsigned long lru_pages = 0; |
bb3ab596 | 2715 | int has_under_min_watermark_zone = 0; |
1da177e4 | 2716 | |
f7b7fd8f RR |
2717 | /* The swap token gets in the way of swapout... */ |
2718 | if (!priority) | |
a433658c | 2719 | disable_swap_token(NULL); |
f7b7fd8f | 2720 | |
1da177e4 | 2721 | all_zones_ok = 1; |
1741c877 | 2722 | balanced = 0; |
1da177e4 | 2723 | |
d6277db4 RW |
2724 | /* |
2725 | * Scan in the highmem->dma direction for the highest | |
2726 | * zone which needs scanning | |
2727 | */ | |
2728 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
2729 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 2730 | |
d6277db4 RW |
2731 | if (!populated_zone(zone)) |
2732 | continue; | |
1da177e4 | 2733 | |
93e4a89a | 2734 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
d6277db4 | 2735 | continue; |
1da177e4 | 2736 | |
556adecb RR |
2737 | /* |
2738 | * Do some background aging of the anon list, to give | |
2739 | * pages a chance to be referenced before reclaiming. | |
2740 | */ | |
f16015fb | 2741 | age_active_anon(zone, &sc, priority); |
556adecb | 2742 | |
88f5acf8 | 2743 | if (!zone_watermark_ok_safe(zone, order, |
41858966 | 2744 | high_wmark_pages(zone), 0, 0)) { |
d6277db4 | 2745 | end_zone = i; |
e1dbeda6 | 2746 | break; |
439423f6 SL |
2747 | } else { |
2748 | /* If balanced, clear the congested flag */ | |
2749 | zone_clear_flag(zone, ZONE_CONGESTED); | |
1da177e4 | 2750 | } |
1da177e4 | 2751 | } |
e1dbeda6 AM |
2752 | if (i < 0) |
2753 | goto out; | |
2754 | ||
1da177e4 LT |
2755 | for (i = 0; i <= end_zone; i++) { |
2756 | struct zone *zone = pgdat->node_zones + i; | |
2757 | ||
adea02a1 | 2758 | lru_pages += zone_reclaimable_pages(zone); |
1da177e4 LT |
2759 | } |
2760 | ||
2761 | /* | |
2762 | * Now scan the zone in the dma->highmem direction, stopping | |
2763 | * at the last zone which needs scanning. | |
2764 | * | |
2765 | * We do this because the page allocator works in the opposite | |
2766 | * direction. This prevents the page allocator from allocating | |
2767 | * pages behind kswapd's direction of progress, which would | |
2768 | * cause too much scanning of the lower zones. | |
2769 | */ | |
2770 | for (i = 0; i <= end_zone; i++) { | |
2771 | struct zone *zone = pgdat->node_zones + i; | |
b15e0905 | 2772 | int nr_slab; |
8afdcece | 2773 | unsigned long balance_gap; |
1da177e4 | 2774 | |
f3fe6512 | 2775 | if (!populated_zone(zone)) |
1da177e4 LT |
2776 | continue; |
2777 | ||
93e4a89a | 2778 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
1da177e4 LT |
2779 | continue; |
2780 | ||
1da177e4 | 2781 | sc.nr_scanned = 0; |
4e416953 | 2782 | |
0ae5e89c | 2783 | nr_soft_scanned = 0; |
4e416953 BS |
2784 | /* |
2785 | * Call soft limit reclaim before calling shrink_zone. | |
4e416953 | 2786 | */ |
0ae5e89c YH |
2787 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, |
2788 | order, sc.gfp_mask, | |
2789 | &nr_soft_scanned); | |
2790 | sc.nr_reclaimed += nr_soft_reclaimed; | |
2791 | total_scanned += nr_soft_scanned; | |
00918b6a | 2792 | |
32a4330d | 2793 | /* |
8afdcece MG |
2794 | * We put equal pressure on every zone, unless |
2795 | * one zone has way too many pages free | |
2796 | * already. The "too many pages" is defined | |
2797 | * as the high wmark plus a "gap" where the | |
2798 | * gap is either the low watermark or 1% | |
2799 | * of the zone, whichever is smaller. | |
32a4330d | 2800 | */ |
8afdcece MG |
2801 | balance_gap = min(low_wmark_pages(zone), |
2802 | (zone->present_pages + | |
2803 | KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2804 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
88f5acf8 | 2805 | if (!zone_watermark_ok_safe(zone, order, |
8afdcece | 2806 | high_wmark_pages(zone) + balance_gap, |
d7868dae | 2807 | end_zone, 0)) { |
a79311c1 | 2808 | shrink_zone(priority, zone, &sc); |
5a03b051 | 2809 | |
d7868dae MG |
2810 | reclaim_state->reclaimed_slab = 0; |
2811 | nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages); | |
2812 | sc.nr_reclaimed += reclaim_state->reclaimed_slab; | |
2813 | total_scanned += sc.nr_scanned; | |
2814 | ||
2815 | if (nr_slab == 0 && !zone_reclaimable(zone)) | |
2816 | zone->all_unreclaimable = 1; | |
2817 | } | |
2818 | ||
1da177e4 LT |
2819 | /* |
2820 | * If we've done a decent amount of scanning and | |
2821 | * the reclaim ratio is low, start doing writepage | |
2822 | * even in laptop mode | |
2823 | */ | |
2824 | if (total_scanned > SWAP_CLUSTER_MAX * 2 && | |
a79311c1 | 2825 | total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) |
1da177e4 | 2826 | sc.may_writepage = 1; |
bb3ab596 | 2827 | |
215ddd66 MG |
2828 | if (zone->all_unreclaimable) { |
2829 | if (end_zone && end_zone == i) | |
2830 | end_zone--; | |
d7868dae | 2831 | continue; |
215ddd66 | 2832 | } |
d7868dae | 2833 | |
88f5acf8 | 2834 | if (!zone_watermark_ok_safe(zone, order, |
45973d74 MK |
2835 | high_wmark_pages(zone), end_zone, 0)) { |
2836 | all_zones_ok = 0; | |
2837 | /* | |
2838 | * We are still under min water mark. This | |
2839 | * means that we have a GFP_ATOMIC allocation | |
2840 | * failure risk. Hurry up! | |
2841 | */ | |
88f5acf8 | 2842 | if (!zone_watermark_ok_safe(zone, order, |
45973d74 MK |
2843 | min_wmark_pages(zone), end_zone, 0)) |
2844 | has_under_min_watermark_zone = 1; | |
0e093d99 MG |
2845 | } else { |
2846 | /* | |
2847 | * If a zone reaches its high watermark, | |
2848 | * consider it to be no longer congested. It's | |
2849 | * possible there are dirty pages backed by | |
2850 | * congested BDIs but as pressure is relieved, | |
2851 | * spectulatively avoid congestion waits | |
2852 | */ | |
2853 | zone_clear_flag(zone, ZONE_CONGESTED); | |
dc83edd9 | 2854 | if (i <= *classzone_idx) |
1741c877 | 2855 | balanced += zone->present_pages; |
45973d74 | 2856 | } |
bb3ab596 | 2857 | |
1da177e4 | 2858 | } |
dc83edd9 | 2859 | if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx))) |
1da177e4 LT |
2860 | break; /* kswapd: all done */ |
2861 | /* | |
2862 | * OK, kswapd is getting into trouble. Take a nap, then take | |
2863 | * another pass across the zones. | |
2864 | */ | |
bb3ab596 KM |
2865 | if (total_scanned && (priority < DEF_PRIORITY - 2)) { |
2866 | if (has_under_min_watermark_zone) | |
2867 | count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT); | |
2868 | else | |
2869 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2870 | } | |
1da177e4 LT |
2871 | |
2872 | /* | |
2873 | * We do this so kswapd doesn't build up large priorities for | |
2874 | * example when it is freeing in parallel with allocators. It | |
2875 | * matches the direct reclaim path behaviour in terms of impact | |
2876 | * on zone->*_priority. | |
2877 | */ | |
a79311c1 | 2878 | if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) |
1da177e4 LT |
2879 | break; |
2880 | } | |
2881 | out: | |
99504748 MG |
2882 | |
2883 | /* | |
2884 | * order-0: All zones must meet high watermark for a balanced node | |
1741c877 MG |
2885 | * high-order: Balanced zones must make up at least 25% of the node |
2886 | * for the node to be balanced | |
99504748 | 2887 | */ |
dc83edd9 | 2888 | if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) { |
1da177e4 | 2889 | cond_resched(); |
8357376d RW |
2890 | |
2891 | try_to_freeze(); | |
2892 | ||
73ce02e9 KM |
2893 | /* |
2894 | * Fragmentation may mean that the system cannot be | |
2895 | * rebalanced for high-order allocations in all zones. | |
2896 | * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, | |
2897 | * it means the zones have been fully scanned and are still | |
2898 | * not balanced. For high-order allocations, there is | |
2899 | * little point trying all over again as kswapd may | |
2900 | * infinite loop. | |
2901 | * | |
2902 | * Instead, recheck all watermarks at order-0 as they | |
2903 | * are the most important. If watermarks are ok, kswapd will go | |
2904 | * back to sleep. High-order users can still perform direct | |
2905 | * reclaim if they wish. | |
2906 | */ | |
2907 | if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) | |
2908 | order = sc.order = 0; | |
2909 | ||
1da177e4 LT |
2910 | goto loop_again; |
2911 | } | |
2912 | ||
99504748 MG |
2913 | /* |
2914 | * If kswapd was reclaiming at a higher order, it has the option of | |
2915 | * sleeping without all zones being balanced. Before it does, it must | |
2916 | * ensure that the watermarks for order-0 on *all* zones are met and | |
2917 | * that the congestion flags are cleared. The congestion flag must | |
2918 | * be cleared as kswapd is the only mechanism that clears the flag | |
2919 | * and it is potentially going to sleep here. | |
2920 | */ | |
2921 | if (order) { | |
2922 | for (i = 0; i <= end_zone; i++) { | |
2923 | struct zone *zone = pgdat->node_zones + i; | |
2924 | ||
2925 | if (!populated_zone(zone)) | |
2926 | continue; | |
2927 | ||
2928 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) | |
2929 | continue; | |
2930 | ||
2931 | /* Confirm the zone is balanced for order-0 */ | |
2932 | if (!zone_watermark_ok(zone, 0, | |
2933 | high_wmark_pages(zone), 0, 0)) { | |
2934 | order = sc.order = 0; | |
2935 | goto loop_again; | |
2936 | } | |
2937 | ||
2938 | /* If balanced, clear the congested flag */ | |
2939 | zone_clear_flag(zone, ZONE_CONGESTED); | |
16fb9512 SL |
2940 | if (i <= *classzone_idx) |
2941 | balanced += zone->present_pages; | |
99504748 MG |
2942 | } |
2943 | } | |
2944 | ||
0abdee2b MG |
2945 | /* |
2946 | * Return the order we were reclaiming at so sleeping_prematurely() | |
2947 | * makes a decision on the order we were last reclaiming at. However, | |
2948 | * if another caller entered the allocator slow path while kswapd | |
2949 | * was awake, order will remain at the higher level | |
2950 | */ | |
dc83edd9 | 2951 | *classzone_idx = end_zone; |
0abdee2b | 2952 | return order; |
1da177e4 LT |
2953 | } |
2954 | ||
dc83edd9 | 2955 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f0bc0a60 KM |
2956 | { |
2957 | long remaining = 0; | |
2958 | DEFINE_WAIT(wait); | |
2959 | ||
2960 | if (freezing(current) || kthread_should_stop()) | |
2961 | return; | |
2962 | ||
2963 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2964 | ||
2965 | /* Try to sleep for a short interval */ | |
dc83edd9 | 2966 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
2967 | remaining = schedule_timeout(HZ/10); |
2968 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2969 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2970 | } | |
2971 | ||
2972 | /* | |
2973 | * After a short sleep, check if it was a premature sleep. If not, then | |
2974 | * go fully to sleep until explicitly woken up. | |
2975 | */ | |
dc83edd9 | 2976 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
2977 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
2978 | ||
2979 | /* | |
2980 | * vmstat counters are not perfectly accurate and the estimated | |
2981 | * value for counters such as NR_FREE_PAGES can deviate from the | |
2982 | * true value by nr_online_cpus * threshold. To avoid the zone | |
2983 | * watermarks being breached while under pressure, we reduce the | |
2984 | * per-cpu vmstat threshold while kswapd is awake and restore | |
2985 | * them before going back to sleep. | |
2986 | */ | |
2987 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
2988 | schedule(); | |
2989 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); | |
2990 | } else { | |
2991 | if (remaining) | |
2992 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
2993 | else | |
2994 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
2995 | } | |
2996 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2997 | } | |
2998 | ||
1da177e4 LT |
2999 | /* |
3000 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3001 | * from the init process. |
1da177e4 LT |
3002 | * |
3003 | * This basically trickles out pages so that we have _some_ | |
3004 | * free memory available even if there is no other activity | |
3005 | * that frees anything up. This is needed for things like routing | |
3006 | * etc, where we otherwise might have all activity going on in | |
3007 | * asynchronous contexts that cannot page things out. | |
3008 | * | |
3009 | * If there are applications that are active memory-allocators | |
3010 | * (most normal use), this basically shouldn't matter. | |
3011 | */ | |
3012 | static int kswapd(void *p) | |
3013 | { | |
215ddd66 | 3014 | unsigned long order, new_order; |
d2ebd0f6 | 3015 | unsigned balanced_order; |
215ddd66 | 3016 | int classzone_idx, new_classzone_idx; |
d2ebd0f6 | 3017 | int balanced_classzone_idx; |
1da177e4 LT |
3018 | pg_data_t *pgdat = (pg_data_t*)p; |
3019 | struct task_struct *tsk = current; | |
f0bc0a60 | 3020 | |
1da177e4 LT |
3021 | struct reclaim_state reclaim_state = { |
3022 | .reclaimed_slab = 0, | |
3023 | }; | |
a70f7302 | 3024 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3025 | |
cf40bd16 NP |
3026 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3027 | ||
174596a0 | 3028 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3029 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3030 | current->reclaim_state = &reclaim_state; |
3031 | ||
3032 | /* | |
3033 | * Tell the memory management that we're a "memory allocator", | |
3034 | * and that if we need more memory we should get access to it | |
3035 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3036 | * never get caught in the normal page freeing logic. | |
3037 | * | |
3038 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3039 | * you need a small amount of memory in order to be able to | |
3040 | * page out something else, and this flag essentially protects | |
3041 | * us from recursively trying to free more memory as we're | |
3042 | * trying to free the first piece of memory in the first place). | |
3043 | */ | |
930d9152 | 3044 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3045 | set_freezable(); |
1da177e4 | 3046 | |
215ddd66 | 3047 | order = new_order = 0; |
d2ebd0f6 | 3048 | balanced_order = 0; |
215ddd66 | 3049 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; |
d2ebd0f6 | 3050 | balanced_classzone_idx = classzone_idx; |
1da177e4 | 3051 | for ( ; ; ) { |
8fe23e05 | 3052 | int ret; |
3e1d1d28 | 3053 | |
215ddd66 MG |
3054 | /* |
3055 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
3056 | * new request of a similar or harder type will succeed soon | |
3057 | * so consider going to sleep on the basis we reclaimed at | |
3058 | */ | |
d2ebd0f6 AS |
3059 | if (balanced_classzone_idx >= new_classzone_idx && |
3060 | balanced_order == new_order) { | |
215ddd66 MG |
3061 | new_order = pgdat->kswapd_max_order; |
3062 | new_classzone_idx = pgdat->classzone_idx; | |
3063 | pgdat->kswapd_max_order = 0; | |
3064 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3065 | } | |
3066 | ||
99504748 | 3067 | if (order < new_order || classzone_idx > new_classzone_idx) { |
1da177e4 LT |
3068 | /* |
3069 | * Don't sleep if someone wants a larger 'order' | |
99504748 | 3070 | * allocation or has tigher zone constraints |
1da177e4 LT |
3071 | */ |
3072 | order = new_order; | |
99504748 | 3073 | classzone_idx = new_classzone_idx; |
1da177e4 | 3074 | } else { |
d2ebd0f6 AS |
3075 | kswapd_try_to_sleep(pgdat, balanced_order, |
3076 | balanced_classzone_idx); | |
1da177e4 | 3077 | order = pgdat->kswapd_max_order; |
99504748 | 3078 | classzone_idx = pgdat->classzone_idx; |
f0dfcde0 AS |
3079 | new_order = order; |
3080 | new_classzone_idx = classzone_idx; | |
4d40502e | 3081 | pgdat->kswapd_max_order = 0; |
215ddd66 | 3082 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
1da177e4 | 3083 | } |
1da177e4 | 3084 | |
8fe23e05 DR |
3085 | ret = try_to_freeze(); |
3086 | if (kthread_should_stop()) | |
3087 | break; | |
3088 | ||
3089 | /* | |
3090 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3091 | * after returning from the refrigerator | |
3092 | */ | |
33906bc5 MG |
3093 | if (!ret) { |
3094 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
d2ebd0f6 AS |
3095 | balanced_classzone_idx = classzone_idx; |
3096 | balanced_order = balance_pgdat(pgdat, order, | |
3097 | &balanced_classzone_idx); | |
33906bc5 | 3098 | } |
1da177e4 LT |
3099 | } |
3100 | return 0; | |
3101 | } | |
3102 | ||
3103 | /* | |
3104 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3105 | */ | |
99504748 | 3106 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3107 | { |
3108 | pg_data_t *pgdat; | |
3109 | ||
f3fe6512 | 3110 | if (!populated_zone(zone)) |
1da177e4 LT |
3111 | return; |
3112 | ||
88f5acf8 | 3113 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1da177e4 | 3114 | return; |
88f5acf8 | 3115 | pgdat = zone->zone_pgdat; |
99504748 | 3116 | if (pgdat->kswapd_max_order < order) { |
1da177e4 | 3117 | pgdat->kswapd_max_order = order; |
99504748 MG |
3118 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
3119 | } | |
8d0986e2 | 3120 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3121 | return; |
88f5acf8 MG |
3122 | if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) |
3123 | return; | |
3124 | ||
3125 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3126 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3127 | } |
3128 | ||
adea02a1 WF |
3129 | /* |
3130 | * The reclaimable count would be mostly accurate. | |
3131 | * The less reclaimable pages may be | |
3132 | * - mlocked pages, which will be moved to unevictable list when encountered | |
3133 | * - mapped pages, which may require several travels to be reclaimed | |
3134 | * - dirty pages, which is not "instantly" reclaimable | |
3135 | */ | |
3136 | unsigned long global_reclaimable_pages(void) | |
4f98a2fe | 3137 | { |
adea02a1 WF |
3138 | int nr; |
3139 | ||
3140 | nr = global_page_state(NR_ACTIVE_FILE) + | |
3141 | global_page_state(NR_INACTIVE_FILE); | |
3142 | ||
3143 | if (nr_swap_pages > 0) | |
3144 | nr += global_page_state(NR_ACTIVE_ANON) + | |
3145 | global_page_state(NR_INACTIVE_ANON); | |
3146 | ||
3147 | return nr; | |
3148 | } | |
3149 | ||
3150 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
3151 | { | |
3152 | int nr; | |
3153 | ||
3154 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
3155 | zone_page_state(zone, NR_INACTIVE_FILE); | |
3156 | ||
3157 | if (nr_swap_pages > 0) | |
3158 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
3159 | zone_page_state(zone, NR_INACTIVE_ANON); | |
3160 | ||
3161 | return nr; | |
4f98a2fe RR |
3162 | } |
3163 | ||
c6f37f12 | 3164 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3165 | /* |
7b51755c | 3166 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3167 | * freed pages. |
3168 | * | |
3169 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3170 | * LRU order by reclaiming preferentially | |
3171 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3172 | */ |
7b51755c | 3173 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3174 | { |
d6277db4 | 3175 | struct reclaim_state reclaim_state; |
d6277db4 | 3176 | struct scan_control sc = { |
7b51755c KM |
3177 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
3178 | .may_swap = 1, | |
3179 | .may_unmap = 1, | |
d6277db4 | 3180 | .may_writepage = 1, |
7b51755c KM |
3181 | .nr_to_reclaim = nr_to_reclaim, |
3182 | .hibernation_mode = 1, | |
7b51755c | 3183 | .order = 0, |
1da177e4 | 3184 | }; |
a09ed5e0 YH |
3185 | struct shrink_control shrink = { |
3186 | .gfp_mask = sc.gfp_mask, | |
3187 | }; | |
3188 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
7b51755c KM |
3189 | struct task_struct *p = current; |
3190 | unsigned long nr_reclaimed; | |
1da177e4 | 3191 | |
7b51755c KM |
3192 | p->flags |= PF_MEMALLOC; |
3193 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3194 | reclaim_state.reclaimed_slab = 0; | |
3195 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3196 | |
a09ed5e0 | 3197 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
d979677c | 3198 | |
7b51755c KM |
3199 | p->reclaim_state = NULL; |
3200 | lockdep_clear_current_reclaim_state(); | |
3201 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3202 | |
7b51755c | 3203 | return nr_reclaimed; |
1da177e4 | 3204 | } |
c6f37f12 | 3205 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3206 | |
1da177e4 LT |
3207 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3208 | not required for correctness. So if the last cpu in a node goes | |
3209 | away, we get changed to run anywhere: as the first one comes back, | |
3210 | restore their cpu bindings. */ | |
9c7b216d | 3211 | static int __devinit cpu_callback(struct notifier_block *nfb, |
69e05944 | 3212 | unsigned long action, void *hcpu) |
1da177e4 | 3213 | { |
58c0a4a7 | 3214 | int nid; |
1da177e4 | 3215 | |
8bb78442 | 3216 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
58c0a4a7 | 3217 | for_each_node_state(nid, N_HIGH_MEMORY) { |
c5f59f08 | 3218 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3219 | const struct cpumask *mask; |
3220 | ||
3221 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3222 | |
3e597945 | 3223 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3224 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3225 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3226 | } |
3227 | } | |
3228 | return NOTIFY_OK; | |
3229 | } | |
1da177e4 | 3230 | |
3218ae14 YG |
3231 | /* |
3232 | * This kswapd start function will be called by init and node-hot-add. | |
3233 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3234 | */ | |
3235 | int kswapd_run(int nid) | |
3236 | { | |
3237 | pg_data_t *pgdat = NODE_DATA(nid); | |
3238 | int ret = 0; | |
3239 | ||
3240 | if (pgdat->kswapd) | |
3241 | return 0; | |
3242 | ||
3243 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3244 | if (IS_ERR(pgdat->kswapd)) { | |
3245 | /* failure at boot is fatal */ | |
3246 | BUG_ON(system_state == SYSTEM_BOOTING); | |
3247 | printk("Failed to start kswapd on node %d\n",nid); | |
3248 | ret = -1; | |
3249 | } | |
3250 | return ret; | |
3251 | } | |
3252 | ||
8fe23e05 DR |
3253 | /* |
3254 | * Called by memory hotplug when all memory in a node is offlined. | |
3255 | */ | |
3256 | void kswapd_stop(int nid) | |
3257 | { | |
3258 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3259 | ||
3260 | if (kswapd) | |
3261 | kthread_stop(kswapd); | |
3262 | } | |
3263 | ||
1da177e4 LT |
3264 | static int __init kswapd_init(void) |
3265 | { | |
3218ae14 | 3266 | int nid; |
69e05944 | 3267 | |
1da177e4 | 3268 | swap_setup(); |
9422ffba | 3269 | for_each_node_state(nid, N_HIGH_MEMORY) |
3218ae14 | 3270 | kswapd_run(nid); |
1da177e4 LT |
3271 | hotcpu_notifier(cpu_callback, 0); |
3272 | return 0; | |
3273 | } | |
3274 | ||
3275 | module_init(kswapd_init) | |
9eeff239 CL |
3276 | |
3277 | #ifdef CONFIG_NUMA | |
3278 | /* | |
3279 | * Zone reclaim mode | |
3280 | * | |
3281 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3282 | * the watermarks. | |
9eeff239 CL |
3283 | */ |
3284 | int zone_reclaim_mode __read_mostly; | |
3285 | ||
1b2ffb78 | 3286 | #define RECLAIM_OFF 0 |
7d03431c | 3287 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 CL |
3288 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
3289 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3290 | ||
a92f7126 CL |
3291 | /* |
3292 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3293 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3294 | * a zone. | |
3295 | */ | |
3296 | #define ZONE_RECLAIM_PRIORITY 4 | |
3297 | ||
9614634f CL |
3298 | /* |
3299 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3300 | * occur. | |
3301 | */ | |
3302 | int sysctl_min_unmapped_ratio = 1; | |
3303 | ||
0ff38490 CL |
3304 | /* |
3305 | * If the number of slab pages in a zone grows beyond this percentage then | |
3306 | * slab reclaim needs to occur. | |
3307 | */ | |
3308 | int sysctl_min_slab_ratio = 5; | |
3309 | ||
90afa5de MG |
3310 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
3311 | { | |
3312 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3313 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3314 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3315 | ||
3316 | /* | |
3317 | * It's possible for there to be more file mapped pages than | |
3318 | * accounted for by the pages on the file LRU lists because | |
3319 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3320 | */ | |
3321 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3322 | } | |
3323 | ||
3324 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3325 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3326 | { | |
3327 | long nr_pagecache_reclaimable; | |
3328 | long delta = 0; | |
3329 | ||
3330 | /* | |
3331 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3332 | * potentially reclaimable. Otherwise, we have to worry about | |
3333 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3334 | * a better estimate | |
3335 | */ | |
3336 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3337 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3338 | else | |
3339 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3340 | ||
3341 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3342 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3343 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3344 | ||
3345 | /* Watch for any possible underflows due to delta */ | |
3346 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3347 | delta = nr_pagecache_reclaimable; | |
3348 | ||
3349 | return nr_pagecache_reclaimable - delta; | |
3350 | } | |
3351 | ||
9eeff239 CL |
3352 | /* |
3353 | * Try to free up some pages from this zone through reclaim. | |
3354 | */ | |
179e9639 | 3355 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3356 | { |
7fb2d46d | 3357 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3358 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3359 | struct task_struct *p = current; |
3360 | struct reclaim_state reclaim_state; | |
8695949a | 3361 | int priority; |
179e9639 AM |
3362 | struct scan_control sc = { |
3363 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
a6dc60f8 | 3364 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), |
2e2e4259 | 3365 | .may_swap = 1, |
22fba335 KM |
3366 | .nr_to_reclaim = max_t(unsigned long, nr_pages, |
3367 | SWAP_CLUSTER_MAX), | |
179e9639 | 3368 | .gfp_mask = gfp_mask, |
bd2f6199 | 3369 | .order = order, |
179e9639 | 3370 | }; |
a09ed5e0 YH |
3371 | struct shrink_control shrink = { |
3372 | .gfp_mask = sc.gfp_mask, | |
3373 | }; | |
15748048 | 3374 | unsigned long nr_slab_pages0, nr_slab_pages1; |
9eeff239 | 3375 | |
9eeff239 | 3376 | cond_resched(); |
d4f7796e CL |
3377 | /* |
3378 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3379 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3380 | * and RECLAIM_SWAP. | |
3381 | */ | |
3382 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3383 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3384 | reclaim_state.reclaimed_slab = 0; |
3385 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3386 | |
90afa5de | 3387 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
3388 | /* |
3389 | * Free memory by calling shrink zone with increasing | |
3390 | * priorities until we have enough memory freed. | |
3391 | */ | |
3392 | priority = ZONE_RECLAIM_PRIORITY; | |
3393 | do { | |
a79311c1 | 3394 | shrink_zone(priority, zone, &sc); |
0ff38490 | 3395 | priority--; |
a79311c1 | 3396 | } while (priority >= 0 && sc.nr_reclaimed < nr_pages); |
0ff38490 | 3397 | } |
c84db23c | 3398 | |
15748048 KM |
3399 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3400 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
2a16e3f4 | 3401 | /* |
7fb2d46d | 3402 | * shrink_slab() does not currently allow us to determine how |
0ff38490 CL |
3403 | * many pages were freed in this zone. So we take the current |
3404 | * number of slab pages and shake the slab until it is reduced | |
3405 | * by the same nr_pages that we used for reclaiming unmapped | |
3406 | * pages. | |
2a16e3f4 | 3407 | * |
0ff38490 CL |
3408 | * Note that shrink_slab will free memory on all zones and may |
3409 | * take a long time. | |
2a16e3f4 | 3410 | */ |
4dc4b3d9 KM |
3411 | for (;;) { |
3412 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3413 | ||
3414 | /* No reclaimable slab or very low memory pressure */ | |
1495f230 | 3415 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) |
4dc4b3d9 KM |
3416 | break; |
3417 | ||
3418 | /* Freed enough memory */ | |
3419 | nr_slab_pages1 = zone_page_state(zone, | |
3420 | NR_SLAB_RECLAIMABLE); | |
3421 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3422 | break; | |
3423 | } | |
83e33a47 CL |
3424 | |
3425 | /* | |
3426 | * Update nr_reclaimed by the number of slab pages we | |
3427 | * reclaimed from this zone. | |
3428 | */ | |
15748048 KM |
3429 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3430 | if (nr_slab_pages1 < nr_slab_pages0) | |
3431 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
2a16e3f4 CL |
3432 | } |
3433 | ||
9eeff239 | 3434 | p->reclaim_state = NULL; |
d4f7796e | 3435 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3436 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3437 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3438 | } |
179e9639 AM |
3439 | |
3440 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3441 | { | |
179e9639 | 3442 | int node_id; |
d773ed6b | 3443 | int ret; |
179e9639 AM |
3444 | |
3445 | /* | |
0ff38490 CL |
3446 | * Zone reclaim reclaims unmapped file backed pages and |
3447 | * slab pages if we are over the defined limits. | |
34aa1330 | 3448 | * |
9614634f CL |
3449 | * A small portion of unmapped file backed pages is needed for |
3450 | * file I/O otherwise pages read by file I/O will be immediately | |
3451 | * thrown out if the zone is overallocated. So we do not reclaim | |
3452 | * if less than a specified percentage of the zone is used by | |
3453 | * unmapped file backed pages. | |
179e9639 | 3454 | */ |
90afa5de MG |
3455 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
3456 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 3457 | return ZONE_RECLAIM_FULL; |
179e9639 | 3458 | |
93e4a89a | 3459 | if (zone->all_unreclaimable) |
fa5e084e | 3460 | return ZONE_RECLAIM_FULL; |
d773ed6b | 3461 | |
179e9639 | 3462 | /* |
d773ed6b | 3463 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3464 | */ |
d773ed6b | 3465 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 3466 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
3467 | |
3468 | /* | |
3469 | * Only run zone reclaim on the local zone or on zones that do not | |
3470 | * have associated processors. This will favor the local processor | |
3471 | * over remote processors and spread off node memory allocations | |
3472 | * as wide as possible. | |
3473 | */ | |
89fa3024 | 3474 | node_id = zone_to_nid(zone); |
37c0708d | 3475 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 3476 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b DR |
3477 | |
3478 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
fa5e084e MG |
3479 | return ZONE_RECLAIM_NOSCAN; |
3480 | ||
d773ed6b DR |
3481 | ret = __zone_reclaim(zone, gfp_mask, order); |
3482 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3483 | ||
24cf7251 MG |
3484 | if (!ret) |
3485 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3486 | ||
d773ed6b | 3487 | return ret; |
179e9639 | 3488 | } |
9eeff239 | 3489 | #endif |
894bc310 | 3490 | |
894bc310 LS |
3491 | /* |
3492 | * page_evictable - test whether a page is evictable | |
3493 | * @page: the page to test | |
3494 | * @vma: the VMA in which the page is or will be mapped, may be NULL | |
3495 | * | |
3496 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
b291f000 NP |
3497 | * lists vs unevictable list. The vma argument is !NULL when called from the |
3498 | * fault path to determine how to instantate a new page. | |
894bc310 LS |
3499 | * |
3500 | * Reasons page might not be evictable: | |
ba9ddf49 | 3501 | * (1) page's mapping marked unevictable |
b291f000 | 3502 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3503 | * |
894bc310 LS |
3504 | */ |
3505 | int page_evictable(struct page *page, struct vm_area_struct *vma) | |
3506 | { | |
3507 | ||
ba9ddf49 LS |
3508 | if (mapping_unevictable(page_mapping(page))) |
3509 | return 0; | |
3510 | ||
b291f000 NP |
3511 | if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page))) |
3512 | return 0; | |
894bc310 LS |
3513 | |
3514 | return 1; | |
3515 | } | |
89e004ea LS |
3516 | |
3517 | /** | |
3518 | * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list | |
3519 | * @page: page to check evictability and move to appropriate lru list | |
3520 | * @zone: zone page is in | |
3521 | * | |
3522 | * Checks a page for evictability and moves the page to the appropriate | |
3523 | * zone lru list. | |
3524 | * | |
3525 | * Restrictions: zone->lru_lock must be held, page must be on LRU and must | |
3526 | * have PageUnevictable set. | |
3527 | */ | |
3528 | static void check_move_unevictable_page(struct page *page, struct zone *zone) | |
3529 | { | |
925b7673 | 3530 | struct lruvec *lruvec; |
89e004ea | 3531 | |
925b7673 | 3532 | VM_BUG_ON(PageActive(page)); |
89e004ea LS |
3533 | retry: |
3534 | ClearPageUnevictable(page); | |
3535 | if (page_evictable(page, NULL)) { | |
401a8e1c | 3536 | enum lru_list l = page_lru_base_type(page); |
af936a16 | 3537 | |
89e004ea | 3538 | __dec_zone_state(zone, NR_UNEVICTABLE); |
925b7673 JW |
3539 | lruvec = mem_cgroup_lru_move_lists(zone, page, |
3540 | LRU_UNEVICTABLE, l); | |
3541 | list_move(&page->lru, &lruvec->lists[l]); | |
89e004ea LS |
3542 | __inc_zone_state(zone, NR_INACTIVE_ANON + l); |
3543 | __count_vm_event(UNEVICTABLE_PGRESCUED); | |
3544 | } else { | |
3545 | /* | |
3546 | * rotate unevictable list | |
3547 | */ | |
3548 | SetPageUnevictable(page); | |
925b7673 JW |
3549 | lruvec = mem_cgroup_lru_move_lists(zone, page, LRU_UNEVICTABLE, |
3550 | LRU_UNEVICTABLE); | |
3551 | list_move(&page->lru, &lruvec->lists[LRU_UNEVICTABLE]); | |
89e004ea LS |
3552 | if (page_evictable(page, NULL)) |
3553 | goto retry; | |
3554 | } | |
3555 | } | |
3556 | ||
3557 | /** | |
3558 | * scan_mapping_unevictable_pages - scan an address space for evictable pages | |
3559 | * @mapping: struct address_space to scan for evictable pages | |
3560 | * | |
3561 | * Scan all pages in mapping. Check unevictable pages for | |
3562 | * evictability and move them to the appropriate zone lru list. | |
3563 | */ | |
3564 | void scan_mapping_unevictable_pages(struct address_space *mapping) | |
3565 | { | |
3566 | pgoff_t next = 0; | |
3567 | pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >> | |
3568 | PAGE_CACHE_SHIFT; | |
3569 | struct zone *zone; | |
3570 | struct pagevec pvec; | |
3571 | ||
3572 | if (mapping->nrpages == 0) | |
3573 | return; | |
3574 | ||
3575 | pagevec_init(&pvec, 0); | |
3576 | while (next < end && | |
3577 | pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { | |
3578 | int i; | |
3579 | int pg_scanned = 0; | |
3580 | ||
3581 | zone = NULL; | |
3582 | ||
3583 | for (i = 0; i < pagevec_count(&pvec); i++) { | |
3584 | struct page *page = pvec.pages[i]; | |
3585 | pgoff_t page_index = page->index; | |
3586 | struct zone *pagezone = page_zone(page); | |
3587 | ||
3588 | pg_scanned++; | |
3589 | if (page_index > next) | |
3590 | next = page_index; | |
3591 | next++; | |
3592 | ||
3593 | if (pagezone != zone) { | |
3594 | if (zone) | |
3595 | spin_unlock_irq(&zone->lru_lock); | |
3596 | zone = pagezone; | |
3597 | spin_lock_irq(&zone->lru_lock); | |
3598 | } | |
3599 | ||
3600 | if (PageLRU(page) && PageUnevictable(page)) | |
3601 | check_move_unevictable_page(page, zone); | |
3602 | } | |
3603 | if (zone) | |
3604 | spin_unlock_irq(&zone->lru_lock); | |
3605 | pagevec_release(&pvec); | |
3606 | ||
3607 | count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned); | |
3608 | } | |
3609 | ||
3610 | } | |
af936a16 | 3611 | |
264e56d8 | 3612 | static void warn_scan_unevictable_pages(void) |
af936a16 | 3613 | { |
264e56d8 | 3614 | printk_once(KERN_WARNING |
25bd91bd | 3615 | "%s: The scan_unevictable_pages sysctl/node-interface has been " |
264e56d8 | 3616 | "disabled for lack of a legitimate use case. If you have " |
25bd91bd KM |
3617 | "one, please send an email to linux-mm@kvack.org.\n", |
3618 | current->comm); | |
af936a16 LS |
3619 | } |
3620 | ||
3621 | /* | |
3622 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3623 | * all nodes' unevictable lists for evictable pages | |
3624 | */ | |
3625 | unsigned long scan_unevictable_pages; | |
3626 | ||
3627 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
8d65af78 | 3628 | void __user *buffer, |
af936a16 LS |
3629 | size_t *length, loff_t *ppos) |
3630 | { | |
264e56d8 | 3631 | warn_scan_unevictable_pages(); |
8d65af78 | 3632 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
af936a16 LS |
3633 | scan_unevictable_pages = 0; |
3634 | return 0; | |
3635 | } | |
3636 | ||
e4455abb | 3637 | #ifdef CONFIG_NUMA |
af936a16 LS |
3638 | /* |
3639 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3640 | * a specified node's per zone unevictable lists for evictable pages. | |
3641 | */ | |
3642 | ||
10fbcf4c KS |
3643 | static ssize_t read_scan_unevictable_node(struct device *dev, |
3644 | struct device_attribute *attr, | |
af936a16 LS |
3645 | char *buf) |
3646 | { | |
264e56d8 | 3647 | warn_scan_unevictable_pages(); |
af936a16 LS |
3648 | return sprintf(buf, "0\n"); /* always zero; should fit... */ |
3649 | } | |
3650 | ||
10fbcf4c KS |
3651 | static ssize_t write_scan_unevictable_node(struct device *dev, |
3652 | struct device_attribute *attr, | |
af936a16 LS |
3653 | const char *buf, size_t count) |
3654 | { | |
264e56d8 | 3655 | warn_scan_unevictable_pages(); |
af936a16 LS |
3656 | return 1; |
3657 | } | |
3658 | ||
3659 | ||
10fbcf4c | 3660 | static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, |
af936a16 LS |
3661 | read_scan_unevictable_node, |
3662 | write_scan_unevictable_node); | |
3663 | ||
3664 | int scan_unevictable_register_node(struct node *node) | |
3665 | { | |
10fbcf4c | 3666 | return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 LS |
3667 | } |
3668 | ||
3669 | void scan_unevictable_unregister_node(struct node *node) | |
3670 | { | |
10fbcf4c | 3671 | device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 | 3672 | } |
e4455abb | 3673 | #endif |