Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/vmscan.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
5 | * | |
6 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
7 | * kswapd added: 7.1.96 sct | |
8 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
9 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
10 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
11 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
12 | */ | |
13 | ||
14 | #include <linux/mm.h> | |
15 | #include <linux/module.h> | |
5a0e3ad6 | 16 | #include <linux/gfp.h> |
1da177e4 LT |
17 | #include <linux/kernel_stat.h> |
18 | #include <linux/swap.h> | |
19 | #include <linux/pagemap.h> | |
20 | #include <linux/init.h> | |
21 | #include <linux/highmem.h> | |
e129b5c2 | 22 | #include <linux/vmstat.h> |
1da177e4 LT |
23 | #include <linux/file.h> |
24 | #include <linux/writeback.h> | |
25 | #include <linux/blkdev.h> | |
26 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
27 | buffer_heads_over_limit */ | |
28 | #include <linux/mm_inline.h> | |
1da177e4 LT |
29 | #include <linux/backing-dev.h> |
30 | #include <linux/rmap.h> | |
31 | #include <linux/topology.h> | |
32 | #include <linux/cpu.h> | |
33 | #include <linux/cpuset.h> | |
3e7d3449 | 34 | #include <linux/compaction.h> |
1da177e4 LT |
35 | #include <linux/notifier.h> |
36 | #include <linux/rwsem.h> | |
248a0301 | 37 | #include <linux/delay.h> |
3218ae14 | 38 | #include <linux/kthread.h> |
7dfb7103 | 39 | #include <linux/freezer.h> |
66e1707b | 40 | #include <linux/memcontrol.h> |
873b4771 | 41 | #include <linux/delayacct.h> |
af936a16 | 42 | #include <linux/sysctl.h> |
929bea7c | 43 | #include <linux/oom.h> |
268bb0ce | 44 | #include <linux/prefetch.h> |
1da177e4 LT |
45 | |
46 | #include <asm/tlbflush.h> | |
47 | #include <asm/div64.h> | |
48 | ||
49 | #include <linux/swapops.h> | |
50 | ||
0f8053a5 NP |
51 | #include "internal.h" |
52 | ||
33906bc5 MG |
53 | #define CREATE_TRACE_POINTS |
54 | #include <trace/events/vmscan.h> | |
55 | ||
ee64fc93 | 56 | /* |
f3a310bc MG |
57 | * reclaim_mode determines how the inactive list is shrunk |
58 | * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages | |
59 | * RECLAIM_MODE_ASYNC: Do not block | |
60 | * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback | |
f3a310bc | 61 | * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of |
3e7d3449 | 62 | * order-0 pages and then compact the zone |
ee64fc93 | 63 | */ |
f3a310bc MG |
64 | typedef unsigned __bitwise__ reclaim_mode_t; |
65 | #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u) | |
66 | #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u) | |
67 | #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u) | |
f3a310bc | 68 | #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u) |
7d3579e8 | 69 | |
1da177e4 | 70 | struct scan_control { |
1da177e4 LT |
71 | /* Incremented by the number of inactive pages that were scanned */ |
72 | unsigned long nr_scanned; | |
73 | ||
a79311c1 RR |
74 | /* Number of pages freed so far during a call to shrink_zones() */ |
75 | unsigned long nr_reclaimed; | |
76 | ||
22fba335 KM |
77 | /* How many pages shrink_list() should reclaim */ |
78 | unsigned long nr_to_reclaim; | |
79 | ||
7b51755c KM |
80 | unsigned long hibernation_mode; |
81 | ||
1da177e4 | 82 | /* This context's GFP mask */ |
6daa0e28 | 83 | gfp_t gfp_mask; |
1da177e4 LT |
84 | |
85 | int may_writepage; | |
86 | ||
a6dc60f8 JW |
87 | /* Can mapped pages be reclaimed? */ |
88 | int may_unmap; | |
f1fd1067 | 89 | |
2e2e4259 KM |
90 | /* Can pages be swapped as part of reclaim? */ |
91 | int may_swap; | |
92 | ||
5ad333eb | 93 | int order; |
66e1707b | 94 | |
5f53e762 | 95 | /* |
415b54e3 NK |
96 | * Intend to reclaim enough continuous memory rather than reclaim |
97 | * enough amount of memory. i.e, mode for high order allocation. | |
5f53e762 | 98 | */ |
f3a310bc | 99 | reclaim_mode_t reclaim_mode; |
5f53e762 | 100 | |
f16015fb JW |
101 | /* |
102 | * The memory cgroup that hit its limit and as a result is the | |
103 | * primary target of this reclaim invocation. | |
104 | */ | |
105 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 106 | |
327c0e96 KH |
107 | /* |
108 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
109 | * are scanned. | |
110 | */ | |
111 | nodemask_t *nodemask; | |
1da177e4 LT |
112 | }; |
113 | ||
f16015fb JW |
114 | struct mem_cgroup_zone { |
115 | struct mem_cgroup *mem_cgroup; | |
116 | struct zone *zone; | |
117 | }; | |
118 | ||
1da177e4 LT |
119 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
120 | ||
121 | #ifdef ARCH_HAS_PREFETCH | |
122 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
123 | do { \ | |
124 | if ((_page)->lru.prev != _base) { \ | |
125 | struct page *prev; \ | |
126 | \ | |
127 | prev = lru_to_page(&(_page->lru)); \ | |
128 | prefetch(&prev->_field); \ | |
129 | } \ | |
130 | } while (0) | |
131 | #else | |
132 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
133 | #endif | |
134 | ||
135 | #ifdef ARCH_HAS_PREFETCHW | |
136 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
137 | do { \ | |
138 | if ((_page)->lru.prev != _base) { \ | |
139 | struct page *prev; \ | |
140 | \ | |
141 | prev = lru_to_page(&(_page->lru)); \ | |
142 | prefetchw(&prev->_field); \ | |
143 | } \ | |
144 | } while (0) | |
145 | #else | |
146 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
147 | #endif | |
148 | ||
149 | /* | |
150 | * From 0 .. 100. Higher means more swappy. | |
151 | */ | |
152 | int vm_swappiness = 60; | |
bd1e22b8 | 153 | long vm_total_pages; /* The total number of pages which the VM controls */ |
1da177e4 LT |
154 | |
155 | static LIST_HEAD(shrinker_list); | |
156 | static DECLARE_RWSEM(shrinker_rwsem); | |
157 | ||
00f0b825 | 158 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
89b5fae5 JW |
159 | static bool global_reclaim(struct scan_control *sc) |
160 | { | |
f16015fb | 161 | return !sc->target_mem_cgroup; |
89b5fae5 JW |
162 | } |
163 | ||
f16015fb | 164 | static bool scanning_global_lru(struct mem_cgroup_zone *mz) |
89b5fae5 | 165 | { |
f16015fb | 166 | return !mz->mem_cgroup; |
89b5fae5 | 167 | } |
91a45470 | 168 | #else |
89b5fae5 JW |
169 | static bool global_reclaim(struct scan_control *sc) |
170 | { | |
171 | return true; | |
172 | } | |
173 | ||
f16015fb | 174 | static bool scanning_global_lru(struct mem_cgroup_zone *mz) |
89b5fae5 JW |
175 | { |
176 | return true; | |
177 | } | |
91a45470 KH |
178 | #endif |
179 | ||
f16015fb | 180 | static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz) |
6e901571 | 181 | { |
f16015fb JW |
182 | if (!scanning_global_lru(mz)) |
183 | return mem_cgroup_get_reclaim_stat(mz->mem_cgroup, mz->zone); | |
3e2f41f1 | 184 | |
f16015fb | 185 | return &mz->zone->reclaim_stat; |
6e901571 KM |
186 | } |
187 | ||
f16015fb JW |
188 | static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz, |
189 | enum lru_list lru) | |
c9f299d9 | 190 | { |
f16015fb JW |
191 | if (!scanning_global_lru(mz)) |
192 | return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup, | |
193 | zone_to_nid(mz->zone), | |
194 | zone_idx(mz->zone), | |
195 | BIT(lru)); | |
a3d8e054 | 196 | |
f16015fb | 197 | return zone_page_state(mz->zone, NR_LRU_BASE + lru); |
c9f299d9 KM |
198 | } |
199 | ||
200 | ||
1da177e4 LT |
201 | /* |
202 | * Add a shrinker callback to be called from the vm | |
203 | */ | |
8e1f936b | 204 | void register_shrinker(struct shrinker *shrinker) |
1da177e4 | 205 | { |
83aeeada | 206 | atomic_long_set(&shrinker->nr_in_batch, 0); |
8e1f936b RR |
207 | down_write(&shrinker_rwsem); |
208 | list_add_tail(&shrinker->list, &shrinker_list); | |
209 | up_write(&shrinker_rwsem); | |
1da177e4 | 210 | } |
8e1f936b | 211 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
212 | |
213 | /* | |
214 | * Remove one | |
215 | */ | |
8e1f936b | 216 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
217 | { |
218 | down_write(&shrinker_rwsem); | |
219 | list_del(&shrinker->list); | |
220 | up_write(&shrinker_rwsem); | |
1da177e4 | 221 | } |
8e1f936b | 222 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 | 223 | |
1495f230 YH |
224 | static inline int do_shrinker_shrink(struct shrinker *shrinker, |
225 | struct shrink_control *sc, | |
226 | unsigned long nr_to_scan) | |
227 | { | |
228 | sc->nr_to_scan = nr_to_scan; | |
229 | return (*shrinker->shrink)(shrinker, sc); | |
230 | } | |
231 | ||
1da177e4 LT |
232 | #define SHRINK_BATCH 128 |
233 | /* | |
234 | * Call the shrink functions to age shrinkable caches | |
235 | * | |
236 | * Here we assume it costs one seek to replace a lru page and that it also | |
237 | * takes a seek to recreate a cache object. With this in mind we age equal | |
238 | * percentages of the lru and ageable caches. This should balance the seeks | |
239 | * generated by these structures. | |
240 | * | |
183ff22b | 241 | * If the vm encountered mapped pages on the LRU it increase the pressure on |
1da177e4 LT |
242 | * slab to avoid swapping. |
243 | * | |
244 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
245 | * | |
246 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
247 | * are eligible for the caller's allocation attempt. It is used for balancing | |
248 | * slab reclaim versus page reclaim. | |
b15e0905 | 249 | * |
250 | * Returns the number of slab objects which we shrunk. | |
1da177e4 | 251 | */ |
a09ed5e0 | 252 | unsigned long shrink_slab(struct shrink_control *shrink, |
1495f230 | 253 | unsigned long nr_pages_scanned, |
a09ed5e0 | 254 | unsigned long lru_pages) |
1da177e4 LT |
255 | { |
256 | struct shrinker *shrinker; | |
69e05944 | 257 | unsigned long ret = 0; |
1da177e4 | 258 | |
1495f230 YH |
259 | if (nr_pages_scanned == 0) |
260 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 261 | |
f06590bd MK |
262 | if (!down_read_trylock(&shrinker_rwsem)) { |
263 | /* Assume we'll be able to shrink next time */ | |
264 | ret = 1; | |
265 | goto out; | |
266 | } | |
1da177e4 LT |
267 | |
268 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
269 | unsigned long long delta; | |
635697c6 KK |
270 | long total_scan; |
271 | long max_pass; | |
09576073 | 272 | int shrink_ret = 0; |
acf92b48 DC |
273 | long nr; |
274 | long new_nr; | |
e9299f50 DC |
275 | long batch_size = shrinker->batch ? shrinker->batch |
276 | : SHRINK_BATCH; | |
1da177e4 | 277 | |
635697c6 KK |
278 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); |
279 | if (max_pass <= 0) | |
280 | continue; | |
281 | ||
acf92b48 DC |
282 | /* |
283 | * copy the current shrinker scan count into a local variable | |
284 | * and zero it so that other concurrent shrinker invocations | |
285 | * don't also do this scanning work. | |
286 | */ | |
83aeeada | 287 | nr = atomic_long_xchg(&shrinker->nr_in_batch, 0); |
acf92b48 DC |
288 | |
289 | total_scan = nr; | |
1495f230 | 290 | delta = (4 * nr_pages_scanned) / shrinker->seeks; |
ea164d73 | 291 | delta *= max_pass; |
1da177e4 | 292 | do_div(delta, lru_pages + 1); |
acf92b48 DC |
293 | total_scan += delta; |
294 | if (total_scan < 0) { | |
88c3bd70 DR |
295 | printk(KERN_ERR "shrink_slab: %pF negative objects to " |
296 | "delete nr=%ld\n", | |
acf92b48 DC |
297 | shrinker->shrink, total_scan); |
298 | total_scan = max_pass; | |
ea164d73 AA |
299 | } |
300 | ||
3567b59a DC |
301 | /* |
302 | * We need to avoid excessive windup on filesystem shrinkers | |
303 | * due to large numbers of GFP_NOFS allocations causing the | |
304 | * shrinkers to return -1 all the time. This results in a large | |
305 | * nr being built up so when a shrink that can do some work | |
306 | * comes along it empties the entire cache due to nr >>> | |
307 | * max_pass. This is bad for sustaining a working set in | |
308 | * memory. | |
309 | * | |
310 | * Hence only allow the shrinker to scan the entire cache when | |
311 | * a large delta change is calculated directly. | |
312 | */ | |
313 | if (delta < max_pass / 4) | |
314 | total_scan = min(total_scan, max_pass / 2); | |
315 | ||
ea164d73 AA |
316 | /* |
317 | * Avoid risking looping forever due to too large nr value: | |
318 | * never try to free more than twice the estimate number of | |
319 | * freeable entries. | |
320 | */ | |
acf92b48 DC |
321 | if (total_scan > max_pass * 2) |
322 | total_scan = max_pass * 2; | |
1da177e4 | 323 | |
acf92b48 | 324 | trace_mm_shrink_slab_start(shrinker, shrink, nr, |
09576073 DC |
325 | nr_pages_scanned, lru_pages, |
326 | max_pass, delta, total_scan); | |
327 | ||
e9299f50 | 328 | while (total_scan >= batch_size) { |
b15e0905 | 329 | int nr_before; |
1da177e4 | 330 | |
1495f230 YH |
331 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); |
332 | shrink_ret = do_shrinker_shrink(shrinker, shrink, | |
e9299f50 | 333 | batch_size); |
1da177e4 LT |
334 | if (shrink_ret == -1) |
335 | break; | |
b15e0905 | 336 | if (shrink_ret < nr_before) |
337 | ret += nr_before - shrink_ret; | |
e9299f50 DC |
338 | count_vm_events(SLABS_SCANNED, batch_size); |
339 | total_scan -= batch_size; | |
1da177e4 LT |
340 | |
341 | cond_resched(); | |
342 | } | |
343 | ||
acf92b48 DC |
344 | /* |
345 | * move the unused scan count back into the shrinker in a | |
346 | * manner that handles concurrent updates. If we exhausted the | |
347 | * scan, there is no need to do an update. | |
348 | */ | |
83aeeada KK |
349 | if (total_scan > 0) |
350 | new_nr = atomic_long_add_return(total_scan, | |
351 | &shrinker->nr_in_batch); | |
352 | else | |
353 | new_nr = atomic_long_read(&shrinker->nr_in_batch); | |
acf92b48 DC |
354 | |
355 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); | |
1da177e4 LT |
356 | } |
357 | up_read(&shrinker_rwsem); | |
f06590bd MK |
358 | out: |
359 | cond_resched(); | |
b15e0905 | 360 | return ret; |
1da177e4 LT |
361 | } |
362 | ||
f3a310bc | 363 | static void set_reclaim_mode(int priority, struct scan_control *sc, |
7d3579e8 KM |
364 | bool sync) |
365 | { | |
c53919ad | 366 | /* Sync reclaim used only for compaction */ |
f3a310bc | 367 | reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC; |
7d3579e8 KM |
368 | |
369 | /* | |
c53919ad | 370 | * Restrict reclaim/compaction to costly allocations or when |
3e7d3449 | 371 | * under memory pressure |
7d3579e8 | 372 | */ |
c53919ad MG |
373 | if (COMPACTION_BUILD && sc->order && |
374 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || | |
375 | priority < DEF_PRIORITY - 2)) | |
376 | sc->reclaim_mode = RECLAIM_MODE_COMPACTION | syncmode; | |
7d3579e8 | 377 | else |
f3a310bc | 378 | sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; |
7d3579e8 KM |
379 | } |
380 | ||
f3a310bc | 381 | static void reset_reclaim_mode(struct scan_control *sc) |
7d3579e8 | 382 | { |
f3a310bc | 383 | sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; |
7d3579e8 KM |
384 | } |
385 | ||
1da177e4 LT |
386 | static inline int is_page_cache_freeable(struct page *page) |
387 | { | |
ceddc3a5 JW |
388 | /* |
389 | * A freeable page cache page is referenced only by the caller | |
390 | * that isolated the page, the page cache radix tree and | |
391 | * optional buffer heads at page->private. | |
392 | */ | |
edcf4748 | 393 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
394 | } |
395 | ||
7d3579e8 KM |
396 | static int may_write_to_queue(struct backing_dev_info *bdi, |
397 | struct scan_control *sc) | |
1da177e4 | 398 | { |
930d9152 | 399 | if (current->flags & PF_SWAPWRITE) |
1da177e4 LT |
400 | return 1; |
401 | if (!bdi_write_congested(bdi)) | |
402 | return 1; | |
403 | if (bdi == current->backing_dev_info) | |
404 | return 1; | |
405 | return 0; | |
406 | } | |
407 | ||
408 | /* | |
409 | * We detected a synchronous write error writing a page out. Probably | |
410 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
411 | * fsync(), msync() or close(). | |
412 | * | |
413 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
414 | * prevents it from being freed up. But we have a ref on the page and once | |
415 | * that page is locked, the mapping is pinned. | |
416 | * | |
417 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
418 | * __GFP_FS. | |
419 | */ | |
420 | static void handle_write_error(struct address_space *mapping, | |
421 | struct page *page, int error) | |
422 | { | |
7eaceacc | 423 | lock_page(page); |
3e9f45bd GC |
424 | if (page_mapping(page) == mapping) |
425 | mapping_set_error(mapping, error); | |
1da177e4 LT |
426 | unlock_page(page); |
427 | } | |
428 | ||
04e62a29 CL |
429 | /* possible outcome of pageout() */ |
430 | typedef enum { | |
431 | /* failed to write page out, page is locked */ | |
432 | PAGE_KEEP, | |
433 | /* move page to the active list, page is locked */ | |
434 | PAGE_ACTIVATE, | |
435 | /* page has been sent to the disk successfully, page is unlocked */ | |
436 | PAGE_SUCCESS, | |
437 | /* page is clean and locked */ | |
438 | PAGE_CLEAN, | |
439 | } pageout_t; | |
440 | ||
1da177e4 | 441 | /* |
1742f19f AM |
442 | * pageout is called by shrink_page_list() for each dirty page. |
443 | * Calls ->writepage(). | |
1da177e4 | 444 | */ |
c661b078 | 445 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 446 | struct scan_control *sc) |
1da177e4 LT |
447 | { |
448 | /* | |
449 | * If the page is dirty, only perform writeback if that write | |
450 | * will be non-blocking. To prevent this allocation from being | |
451 | * stalled by pagecache activity. But note that there may be | |
452 | * stalls if we need to run get_block(). We could test | |
453 | * PagePrivate for that. | |
454 | * | |
6aceb53b | 455 | * If this process is currently in __generic_file_aio_write() against |
1da177e4 LT |
456 | * this page's queue, we can perform writeback even if that |
457 | * will block. | |
458 | * | |
459 | * If the page is swapcache, write it back even if that would | |
460 | * block, for some throttling. This happens by accident, because | |
461 | * swap_backing_dev_info is bust: it doesn't reflect the | |
462 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
463 | */ |
464 | if (!is_page_cache_freeable(page)) | |
465 | return PAGE_KEEP; | |
466 | if (!mapping) { | |
467 | /* | |
468 | * Some data journaling orphaned pages can have | |
469 | * page->mapping == NULL while being dirty with clean buffers. | |
470 | */ | |
266cf658 | 471 | if (page_has_private(page)) { |
1da177e4 LT |
472 | if (try_to_free_buffers(page)) { |
473 | ClearPageDirty(page); | |
d40cee24 | 474 | printk("%s: orphaned page\n", __func__); |
1da177e4 LT |
475 | return PAGE_CLEAN; |
476 | } | |
477 | } | |
478 | return PAGE_KEEP; | |
479 | } | |
480 | if (mapping->a_ops->writepage == NULL) | |
481 | return PAGE_ACTIVATE; | |
0e093d99 | 482 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) |
1da177e4 LT |
483 | return PAGE_KEEP; |
484 | ||
485 | if (clear_page_dirty_for_io(page)) { | |
486 | int res; | |
487 | struct writeback_control wbc = { | |
488 | .sync_mode = WB_SYNC_NONE, | |
489 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
490 | .range_start = 0, |
491 | .range_end = LLONG_MAX, | |
1da177e4 LT |
492 | .for_reclaim = 1, |
493 | }; | |
494 | ||
495 | SetPageReclaim(page); | |
496 | res = mapping->a_ops->writepage(page, &wbc); | |
497 | if (res < 0) | |
498 | handle_write_error(mapping, page, res); | |
994fc28c | 499 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
500 | ClearPageReclaim(page); |
501 | return PAGE_ACTIVATE; | |
502 | } | |
c661b078 | 503 | |
1da177e4 LT |
504 | if (!PageWriteback(page)) { |
505 | /* synchronous write or broken a_ops? */ | |
506 | ClearPageReclaim(page); | |
507 | } | |
755f0225 | 508 | trace_mm_vmscan_writepage(page, |
f3a310bc | 509 | trace_reclaim_flags(page, sc->reclaim_mode)); |
e129b5c2 | 510 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
511 | return PAGE_SUCCESS; |
512 | } | |
513 | ||
514 | return PAGE_CLEAN; | |
515 | } | |
516 | ||
a649fd92 | 517 | /* |
e286781d NP |
518 | * Same as remove_mapping, but if the page is removed from the mapping, it |
519 | * gets returned with a refcount of 0. | |
a649fd92 | 520 | */ |
e286781d | 521 | static int __remove_mapping(struct address_space *mapping, struct page *page) |
49d2e9cc | 522 | { |
28e4d965 NP |
523 | BUG_ON(!PageLocked(page)); |
524 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 525 | |
19fd6231 | 526 | spin_lock_irq(&mapping->tree_lock); |
49d2e9cc | 527 | /* |
0fd0e6b0 NP |
528 | * The non racy check for a busy page. |
529 | * | |
530 | * Must be careful with the order of the tests. When someone has | |
531 | * a ref to the page, it may be possible that they dirty it then | |
532 | * drop the reference. So if PageDirty is tested before page_count | |
533 | * here, then the following race may occur: | |
534 | * | |
535 | * get_user_pages(&page); | |
536 | * [user mapping goes away] | |
537 | * write_to(page); | |
538 | * !PageDirty(page) [good] | |
539 | * SetPageDirty(page); | |
540 | * put_page(page); | |
541 | * !page_count(page) [good, discard it] | |
542 | * | |
543 | * [oops, our write_to data is lost] | |
544 | * | |
545 | * Reversing the order of the tests ensures such a situation cannot | |
546 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
547 | * load is not satisfied before that of page->_count. | |
548 | * | |
549 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
550 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 551 | */ |
e286781d | 552 | if (!page_freeze_refs(page, 2)) |
49d2e9cc | 553 | goto cannot_free; |
e286781d NP |
554 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
555 | if (unlikely(PageDirty(page))) { | |
556 | page_unfreeze_refs(page, 2); | |
49d2e9cc | 557 | goto cannot_free; |
e286781d | 558 | } |
49d2e9cc CL |
559 | |
560 | if (PageSwapCache(page)) { | |
561 | swp_entry_t swap = { .val = page_private(page) }; | |
562 | __delete_from_swap_cache(page); | |
19fd6231 | 563 | spin_unlock_irq(&mapping->tree_lock); |
cb4b86ba | 564 | swapcache_free(swap, page); |
e286781d | 565 | } else { |
6072d13c LT |
566 | void (*freepage)(struct page *); |
567 | ||
568 | freepage = mapping->a_ops->freepage; | |
569 | ||
e64a782f | 570 | __delete_from_page_cache(page); |
19fd6231 | 571 | spin_unlock_irq(&mapping->tree_lock); |
e767e056 | 572 | mem_cgroup_uncharge_cache_page(page); |
6072d13c LT |
573 | |
574 | if (freepage != NULL) | |
575 | freepage(page); | |
49d2e9cc CL |
576 | } |
577 | ||
49d2e9cc CL |
578 | return 1; |
579 | ||
580 | cannot_free: | |
19fd6231 | 581 | spin_unlock_irq(&mapping->tree_lock); |
49d2e9cc CL |
582 | return 0; |
583 | } | |
584 | ||
e286781d NP |
585 | /* |
586 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
587 | * someone else has a ref on the page, abort and return 0. If it was | |
588 | * successfully detached, return 1. Assumes the caller has a single ref on | |
589 | * this page. | |
590 | */ | |
591 | int remove_mapping(struct address_space *mapping, struct page *page) | |
592 | { | |
593 | if (__remove_mapping(mapping, page)) { | |
594 | /* | |
595 | * Unfreezing the refcount with 1 rather than 2 effectively | |
596 | * drops the pagecache ref for us without requiring another | |
597 | * atomic operation. | |
598 | */ | |
599 | page_unfreeze_refs(page, 1); | |
600 | return 1; | |
601 | } | |
602 | return 0; | |
603 | } | |
604 | ||
894bc310 LS |
605 | /** |
606 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
607 | * @page: page to be put back to appropriate lru list | |
608 | * | |
609 | * Add previously isolated @page to appropriate LRU list. | |
610 | * Page may still be unevictable for other reasons. | |
611 | * | |
612 | * lru_lock must not be held, interrupts must be enabled. | |
613 | */ | |
894bc310 LS |
614 | void putback_lru_page(struct page *page) |
615 | { | |
616 | int lru; | |
617 | int active = !!TestClearPageActive(page); | |
bbfd28ee | 618 | int was_unevictable = PageUnevictable(page); |
894bc310 LS |
619 | |
620 | VM_BUG_ON(PageLRU(page)); | |
621 | ||
622 | redo: | |
623 | ClearPageUnevictable(page); | |
624 | ||
625 | if (page_evictable(page, NULL)) { | |
626 | /* | |
627 | * For evictable pages, we can use the cache. | |
628 | * In event of a race, worst case is we end up with an | |
629 | * unevictable page on [in]active list. | |
630 | * We know how to handle that. | |
631 | */ | |
401a8e1c | 632 | lru = active + page_lru_base_type(page); |
894bc310 LS |
633 | lru_cache_add_lru(page, lru); |
634 | } else { | |
635 | /* | |
636 | * Put unevictable pages directly on zone's unevictable | |
637 | * list. | |
638 | */ | |
639 | lru = LRU_UNEVICTABLE; | |
640 | add_page_to_unevictable_list(page); | |
6a7b9548 | 641 | /* |
21ee9f39 MK |
642 | * When racing with an mlock or AS_UNEVICTABLE clearing |
643 | * (page is unlocked) make sure that if the other thread | |
644 | * does not observe our setting of PG_lru and fails | |
24513264 | 645 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 646 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
647 | * the page back to the evictable list. |
648 | * | |
21ee9f39 | 649 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
650 | */ |
651 | smp_mb(); | |
894bc310 | 652 | } |
894bc310 LS |
653 | |
654 | /* | |
655 | * page's status can change while we move it among lru. If an evictable | |
656 | * page is on unevictable list, it never be freed. To avoid that, | |
657 | * check after we added it to the list, again. | |
658 | */ | |
659 | if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { | |
660 | if (!isolate_lru_page(page)) { | |
661 | put_page(page); | |
662 | goto redo; | |
663 | } | |
664 | /* This means someone else dropped this page from LRU | |
665 | * So, it will be freed or putback to LRU again. There is | |
666 | * nothing to do here. | |
667 | */ | |
668 | } | |
669 | ||
bbfd28ee LS |
670 | if (was_unevictable && lru != LRU_UNEVICTABLE) |
671 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
672 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) | |
673 | count_vm_event(UNEVICTABLE_PGCULLED); | |
674 | ||
894bc310 LS |
675 | put_page(page); /* drop ref from isolate */ |
676 | } | |
677 | ||
dfc8d636 JW |
678 | enum page_references { |
679 | PAGEREF_RECLAIM, | |
680 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 681 | PAGEREF_KEEP, |
dfc8d636 JW |
682 | PAGEREF_ACTIVATE, |
683 | }; | |
684 | ||
685 | static enum page_references page_check_references(struct page *page, | |
f16015fb | 686 | struct mem_cgroup_zone *mz, |
dfc8d636 JW |
687 | struct scan_control *sc) |
688 | { | |
64574746 | 689 | int referenced_ptes, referenced_page; |
dfc8d636 | 690 | unsigned long vm_flags; |
dfc8d636 | 691 | |
f16015fb | 692 | referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags); |
64574746 | 693 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 694 | |
dfc8d636 JW |
695 | /* |
696 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
697 | * move the page to the unevictable list. | |
698 | */ | |
699 | if (vm_flags & VM_LOCKED) | |
700 | return PAGEREF_RECLAIM; | |
701 | ||
64574746 JW |
702 | if (referenced_ptes) { |
703 | if (PageAnon(page)) | |
704 | return PAGEREF_ACTIVATE; | |
705 | /* | |
706 | * All mapped pages start out with page table | |
707 | * references from the instantiating fault, so we need | |
708 | * to look twice if a mapped file page is used more | |
709 | * than once. | |
710 | * | |
711 | * Mark it and spare it for another trip around the | |
712 | * inactive list. Another page table reference will | |
713 | * lead to its activation. | |
714 | * | |
715 | * Note: the mark is set for activated pages as well | |
716 | * so that recently deactivated but used pages are | |
717 | * quickly recovered. | |
718 | */ | |
719 | SetPageReferenced(page); | |
720 | ||
34dbc67a | 721 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
722 | return PAGEREF_ACTIVATE; |
723 | ||
c909e993 KK |
724 | /* |
725 | * Activate file-backed executable pages after first usage. | |
726 | */ | |
727 | if (vm_flags & VM_EXEC) | |
728 | return PAGEREF_ACTIVATE; | |
729 | ||
64574746 JW |
730 | return PAGEREF_KEEP; |
731 | } | |
dfc8d636 JW |
732 | |
733 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 734 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
735 | return PAGEREF_RECLAIM_CLEAN; |
736 | ||
737 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
738 | } |
739 | ||
1da177e4 | 740 | /* |
1742f19f | 741 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 742 | */ |
1742f19f | 743 | static unsigned long shrink_page_list(struct list_head *page_list, |
f16015fb | 744 | struct mem_cgroup_zone *mz, |
f84f6e2b | 745 | struct scan_control *sc, |
92df3a72 MG |
746 | int priority, |
747 | unsigned long *ret_nr_dirty, | |
748 | unsigned long *ret_nr_writeback) | |
1da177e4 LT |
749 | { |
750 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 751 | LIST_HEAD(free_pages); |
1da177e4 | 752 | int pgactivate = 0; |
0e093d99 MG |
753 | unsigned long nr_dirty = 0; |
754 | unsigned long nr_congested = 0; | |
05ff5137 | 755 | unsigned long nr_reclaimed = 0; |
92df3a72 | 756 | unsigned long nr_writeback = 0; |
1da177e4 LT |
757 | |
758 | cond_resched(); | |
759 | ||
1da177e4 | 760 | while (!list_empty(page_list)) { |
dfc8d636 | 761 | enum page_references references; |
1da177e4 LT |
762 | struct address_space *mapping; |
763 | struct page *page; | |
764 | int may_enter_fs; | |
1da177e4 LT |
765 | |
766 | cond_resched(); | |
767 | ||
768 | page = lru_to_page(page_list); | |
769 | list_del(&page->lru); | |
770 | ||
529ae9aa | 771 | if (!trylock_page(page)) |
1da177e4 LT |
772 | goto keep; |
773 | ||
725d704e | 774 | VM_BUG_ON(PageActive(page)); |
f16015fb | 775 | VM_BUG_ON(page_zone(page) != mz->zone); |
1da177e4 LT |
776 | |
777 | sc->nr_scanned++; | |
80e43426 | 778 | |
b291f000 NP |
779 | if (unlikely(!page_evictable(page, NULL))) |
780 | goto cull_mlocked; | |
894bc310 | 781 | |
a6dc60f8 | 782 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
783 | goto keep_locked; |
784 | ||
1da177e4 LT |
785 | /* Double the slab pressure for mapped and swapcache pages */ |
786 | if (page_mapped(page) || PageSwapCache(page)) | |
787 | sc->nr_scanned++; | |
788 | ||
c661b078 AW |
789 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
790 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
791 | ||
792 | if (PageWriteback(page)) { | |
92df3a72 | 793 | nr_writeback++; |
c661b078 | 794 | /* |
a18bba06 MG |
795 | * Synchronous reclaim cannot queue pages for |
796 | * writeback due to the possibility of stack overflow | |
797 | * but if it encounters a page under writeback, wait | |
798 | * for the IO to complete. | |
c661b078 | 799 | */ |
f3a310bc | 800 | if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) && |
7d3579e8 | 801 | may_enter_fs) |
c661b078 | 802 | wait_on_page_writeback(page); |
7d3579e8 KM |
803 | else { |
804 | unlock_page(page); | |
c53919ad | 805 | goto keep_reclaim_mode; |
7d3579e8 | 806 | } |
c661b078 | 807 | } |
1da177e4 | 808 | |
f16015fb | 809 | references = page_check_references(page, mz, sc); |
dfc8d636 JW |
810 | switch (references) { |
811 | case PAGEREF_ACTIVATE: | |
1da177e4 | 812 | goto activate_locked; |
64574746 JW |
813 | case PAGEREF_KEEP: |
814 | goto keep_locked; | |
dfc8d636 JW |
815 | case PAGEREF_RECLAIM: |
816 | case PAGEREF_RECLAIM_CLEAN: | |
817 | ; /* try to reclaim the page below */ | |
818 | } | |
1da177e4 | 819 | |
1da177e4 LT |
820 | /* |
821 | * Anonymous process memory has backing store? | |
822 | * Try to allocate it some swap space here. | |
823 | */ | |
b291f000 | 824 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
825 | if (!(sc->gfp_mask & __GFP_IO)) |
826 | goto keep_locked; | |
ac47b003 | 827 | if (!add_to_swap(page)) |
1da177e4 | 828 | goto activate_locked; |
63eb6b93 | 829 | may_enter_fs = 1; |
b291f000 | 830 | } |
1da177e4 LT |
831 | |
832 | mapping = page_mapping(page); | |
1da177e4 LT |
833 | |
834 | /* | |
835 | * The page is mapped into the page tables of one or more | |
836 | * processes. Try to unmap it here. | |
837 | */ | |
838 | if (page_mapped(page) && mapping) { | |
14fa31b8 | 839 | switch (try_to_unmap(page, TTU_UNMAP)) { |
1da177e4 LT |
840 | case SWAP_FAIL: |
841 | goto activate_locked; | |
842 | case SWAP_AGAIN: | |
843 | goto keep_locked; | |
b291f000 NP |
844 | case SWAP_MLOCK: |
845 | goto cull_mlocked; | |
1da177e4 LT |
846 | case SWAP_SUCCESS: |
847 | ; /* try to free the page below */ | |
848 | } | |
849 | } | |
850 | ||
851 | if (PageDirty(page)) { | |
0e093d99 MG |
852 | nr_dirty++; |
853 | ||
ee72886d MG |
854 | /* |
855 | * Only kswapd can writeback filesystem pages to | |
f84f6e2b MG |
856 | * avoid risk of stack overflow but do not writeback |
857 | * unless under significant pressure. | |
ee72886d | 858 | */ |
f84f6e2b MG |
859 | if (page_is_file_cache(page) && |
860 | (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) { | |
49ea7eb6 MG |
861 | /* |
862 | * Immediately reclaim when written back. | |
863 | * Similar in principal to deactivate_page() | |
864 | * except we already have the page isolated | |
865 | * and know it's dirty | |
866 | */ | |
867 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
868 | SetPageReclaim(page); | |
869 | ||
ee72886d MG |
870 | goto keep_locked; |
871 | } | |
872 | ||
dfc8d636 | 873 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 874 | goto keep_locked; |
4dd4b920 | 875 | if (!may_enter_fs) |
1da177e4 | 876 | goto keep_locked; |
52a8363e | 877 | if (!sc->may_writepage) |
1da177e4 LT |
878 | goto keep_locked; |
879 | ||
880 | /* Page is dirty, try to write it out here */ | |
7d3579e8 | 881 | switch (pageout(page, mapping, sc)) { |
1da177e4 | 882 | case PAGE_KEEP: |
0e093d99 | 883 | nr_congested++; |
1da177e4 LT |
884 | goto keep_locked; |
885 | case PAGE_ACTIVATE: | |
886 | goto activate_locked; | |
887 | case PAGE_SUCCESS: | |
7d3579e8 | 888 | if (PageWriteback(page)) |
c53919ad | 889 | goto keep_reclaim_mode; |
7d3579e8 | 890 | if (PageDirty(page)) |
1da177e4 | 891 | goto keep; |
7d3579e8 | 892 | |
1da177e4 LT |
893 | /* |
894 | * A synchronous write - probably a ramdisk. Go | |
895 | * ahead and try to reclaim the page. | |
896 | */ | |
529ae9aa | 897 | if (!trylock_page(page)) |
1da177e4 LT |
898 | goto keep; |
899 | if (PageDirty(page) || PageWriteback(page)) | |
900 | goto keep_locked; | |
901 | mapping = page_mapping(page); | |
902 | case PAGE_CLEAN: | |
903 | ; /* try to free the page below */ | |
904 | } | |
905 | } | |
906 | ||
907 | /* | |
908 | * If the page has buffers, try to free the buffer mappings | |
909 | * associated with this page. If we succeed we try to free | |
910 | * the page as well. | |
911 | * | |
912 | * We do this even if the page is PageDirty(). | |
913 | * try_to_release_page() does not perform I/O, but it is | |
914 | * possible for a page to have PageDirty set, but it is actually | |
915 | * clean (all its buffers are clean). This happens if the | |
916 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 917 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
918 | * try_to_release_page() will discover that cleanness and will |
919 | * drop the buffers and mark the page clean - it can be freed. | |
920 | * | |
921 | * Rarely, pages can have buffers and no ->mapping. These are | |
922 | * the pages which were not successfully invalidated in | |
923 | * truncate_complete_page(). We try to drop those buffers here | |
924 | * and if that worked, and the page is no longer mapped into | |
925 | * process address space (page_count == 1) it can be freed. | |
926 | * Otherwise, leave the page on the LRU so it is swappable. | |
927 | */ | |
266cf658 | 928 | if (page_has_private(page)) { |
1da177e4 LT |
929 | if (!try_to_release_page(page, sc->gfp_mask)) |
930 | goto activate_locked; | |
e286781d NP |
931 | if (!mapping && page_count(page) == 1) { |
932 | unlock_page(page); | |
933 | if (put_page_testzero(page)) | |
934 | goto free_it; | |
935 | else { | |
936 | /* | |
937 | * rare race with speculative reference. | |
938 | * the speculative reference will free | |
939 | * this page shortly, so we may | |
940 | * increment nr_reclaimed here (and | |
941 | * leave it off the LRU). | |
942 | */ | |
943 | nr_reclaimed++; | |
944 | continue; | |
945 | } | |
946 | } | |
1da177e4 LT |
947 | } |
948 | ||
e286781d | 949 | if (!mapping || !__remove_mapping(mapping, page)) |
49d2e9cc | 950 | goto keep_locked; |
1da177e4 | 951 | |
a978d6f5 NP |
952 | /* |
953 | * At this point, we have no other references and there is | |
954 | * no way to pick any more up (removed from LRU, removed | |
955 | * from pagecache). Can use non-atomic bitops now (and | |
956 | * we obviously don't have to worry about waking up a process | |
957 | * waiting on the page lock, because there are no references. | |
958 | */ | |
959 | __clear_page_locked(page); | |
e286781d | 960 | free_it: |
05ff5137 | 961 | nr_reclaimed++; |
abe4c3b5 MG |
962 | |
963 | /* | |
964 | * Is there need to periodically free_page_list? It would | |
965 | * appear not as the counts should be low | |
966 | */ | |
967 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
968 | continue; |
969 | ||
b291f000 | 970 | cull_mlocked: |
63d6c5ad HD |
971 | if (PageSwapCache(page)) |
972 | try_to_free_swap(page); | |
b291f000 NP |
973 | unlock_page(page); |
974 | putback_lru_page(page); | |
f3a310bc | 975 | reset_reclaim_mode(sc); |
b291f000 NP |
976 | continue; |
977 | ||
1da177e4 | 978 | activate_locked: |
68a22394 RR |
979 | /* Not a candidate for swapping, so reclaim swap space. */ |
980 | if (PageSwapCache(page) && vm_swap_full()) | |
a2c43eed | 981 | try_to_free_swap(page); |
894bc310 | 982 | VM_BUG_ON(PageActive(page)); |
1da177e4 LT |
983 | SetPageActive(page); |
984 | pgactivate++; | |
985 | keep_locked: | |
986 | unlock_page(page); | |
987 | keep: | |
f3a310bc | 988 | reset_reclaim_mode(sc); |
c53919ad | 989 | keep_reclaim_mode: |
1da177e4 | 990 | list_add(&page->lru, &ret_pages); |
b291f000 | 991 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); |
1da177e4 | 992 | } |
abe4c3b5 | 993 | |
0e093d99 MG |
994 | /* |
995 | * Tag a zone as congested if all the dirty pages encountered were | |
996 | * backed by a congested BDI. In this case, reclaimers should just | |
997 | * back off and wait for congestion to clear because further reclaim | |
998 | * will encounter the same problem | |
999 | */ | |
89b5fae5 | 1000 | if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc)) |
f16015fb | 1001 | zone_set_flag(mz->zone, ZONE_CONGESTED); |
0e093d99 | 1002 | |
cc59850e | 1003 | free_hot_cold_page_list(&free_pages, 1); |
abe4c3b5 | 1004 | |
1da177e4 | 1005 | list_splice(&ret_pages, page_list); |
f8891e5e | 1006 | count_vm_events(PGACTIVATE, pgactivate); |
92df3a72 MG |
1007 | *ret_nr_dirty += nr_dirty; |
1008 | *ret_nr_writeback += nr_writeback; | |
05ff5137 | 1009 | return nr_reclaimed; |
1da177e4 LT |
1010 | } |
1011 | ||
5ad333eb AW |
1012 | /* |
1013 | * Attempt to remove the specified page from its LRU. Only take this page | |
1014 | * if it is of the appropriate PageActive status. Pages which are being | |
1015 | * freed elsewhere are also ignored. | |
1016 | * | |
1017 | * page: page to consider | |
1018 | * mode: one of the LRU isolation modes defined above | |
1019 | * | |
1020 | * returns 0 on success, -ve errno on failure. | |
1021 | */ | |
4356f21d | 1022 | int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file) |
5ad333eb | 1023 | { |
4356f21d | 1024 | bool all_lru_mode; |
5ad333eb AW |
1025 | int ret = -EINVAL; |
1026 | ||
1027 | /* Only take pages on the LRU. */ | |
1028 | if (!PageLRU(page)) | |
1029 | return ret; | |
1030 | ||
4356f21d MK |
1031 | all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) == |
1032 | (ISOLATE_ACTIVE|ISOLATE_INACTIVE); | |
1033 | ||
5ad333eb AW |
1034 | /* |
1035 | * When checking the active state, we need to be sure we are | |
1036 | * dealing with comparible boolean values. Take the logical not | |
1037 | * of each. | |
1038 | */ | |
4356f21d | 1039 | if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE)) |
5ad333eb AW |
1040 | return ret; |
1041 | ||
4356f21d | 1042 | if (!all_lru_mode && !!page_is_file_cache(page) != file) |
4f98a2fe RR |
1043 | return ret; |
1044 | ||
c53919ad | 1045 | /* Do not give back unevictable pages for compaction */ |
894bc310 LS |
1046 | if (PageUnevictable(page)) |
1047 | return ret; | |
1048 | ||
5ad333eb | 1049 | ret = -EBUSY; |
08e552c6 | 1050 | |
c8244935 MG |
1051 | /* |
1052 | * To minimise LRU disruption, the caller can indicate that it only | |
1053 | * wants to isolate pages it will be able to operate on without | |
1054 | * blocking - clean pages for the most part. | |
1055 | * | |
1056 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1057 | * is used by reclaim when it is cannot write to backing storage | |
1058 | * | |
1059 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1060 | * that it is possible to migrate without blocking | |
1061 | */ | |
1062 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1063 | /* All the caller can do on PageWriteback is block */ | |
1064 | if (PageWriteback(page)) | |
1065 | return ret; | |
1066 | ||
1067 | if (PageDirty(page)) { | |
1068 | struct address_space *mapping; | |
1069 | ||
1070 | /* ISOLATE_CLEAN means only clean pages */ | |
1071 | if (mode & ISOLATE_CLEAN) | |
1072 | return ret; | |
1073 | ||
1074 | /* | |
1075 | * Only pages without mappings or that have a | |
1076 | * ->migratepage callback are possible to migrate | |
1077 | * without blocking | |
1078 | */ | |
1079 | mapping = page_mapping(page); | |
1080 | if (mapping && !mapping->a_ops->migratepage) | |
1081 | return ret; | |
1082 | } | |
1083 | } | |
39deaf85 | 1084 | |
f80c0673 MK |
1085 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1086 | return ret; | |
1087 | ||
5ad333eb AW |
1088 | if (likely(get_page_unless_zero(page))) { |
1089 | /* | |
1090 | * Be careful not to clear PageLRU until after we're | |
1091 | * sure the page is not being freed elsewhere -- the | |
1092 | * page release code relies on it. | |
1093 | */ | |
1094 | ClearPageLRU(page); | |
1095 | ret = 0; | |
1096 | } | |
1097 | ||
1098 | return ret; | |
1099 | } | |
1100 | ||
1da177e4 LT |
1101 | /* |
1102 | * zone->lru_lock is heavily contended. Some of the functions that | |
1103 | * shrink the lists perform better by taking out a batch of pages | |
1104 | * and working on them outside the LRU lock. | |
1105 | * | |
1106 | * For pagecache intensive workloads, this function is the hottest | |
1107 | * spot in the kernel (apart from copy_*_user functions). | |
1108 | * | |
1109 | * Appropriate locks must be held before calling this function. | |
1110 | * | |
1111 | * @nr_to_scan: The number of pages to look through on the list. | |
f626012d | 1112 | * @mz: The mem_cgroup_zone to pull pages from. |
1da177e4 | 1113 | * @dst: The temp list to put pages on to. |
f626012d | 1114 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1115 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1116 | * @mode: One of the LRU isolation modes |
f626012d | 1117 | * @active: True [1] if isolating active pages |
4f98a2fe | 1118 | * @file: True [1] if isolating file [!anon] pages |
1da177e4 LT |
1119 | * |
1120 | * returns how many pages were moved onto *@dst. | |
1121 | */ | |
69e05944 | 1122 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
f626012d | 1123 | struct mem_cgroup_zone *mz, struct list_head *dst, |
fe2c2a10 RR |
1124 | unsigned long *nr_scanned, struct scan_control *sc, |
1125 | isolate_mode_t mode, int active, int file) | |
1da177e4 | 1126 | { |
f626012d HD |
1127 | struct lruvec *lruvec; |
1128 | struct list_head *src; | |
69e05944 | 1129 | unsigned long nr_taken = 0; |
c9b02d97 | 1130 | unsigned long scan; |
f626012d HD |
1131 | int lru = LRU_BASE; |
1132 | ||
1133 | lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup); | |
1134 | if (active) | |
1135 | lru += LRU_ACTIVE; | |
1136 | if (file) | |
1137 | lru += LRU_FILE; | |
1138 | src = &lruvec->lists[lru]; | |
1da177e4 | 1139 | |
c9b02d97 | 1140 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
5ad333eb | 1141 | struct page *page; |
5ad333eb | 1142 | |
1da177e4 LT |
1143 | page = lru_to_page(src); |
1144 | prefetchw_prev_lru_page(page, src, flags); | |
1145 | ||
725d704e | 1146 | VM_BUG_ON(!PageLRU(page)); |
8d438f96 | 1147 | |
4f98a2fe | 1148 | switch (__isolate_lru_page(page, mode, file)) { |
5ad333eb | 1149 | case 0: |
925b7673 | 1150 | mem_cgroup_lru_del(page); |
5ad333eb | 1151 | list_move(&page->lru, dst); |
2c888cfb | 1152 | nr_taken += hpage_nr_pages(page); |
5ad333eb AW |
1153 | break; |
1154 | ||
1155 | case -EBUSY: | |
1156 | /* else it is being freed elsewhere */ | |
1157 | list_move(&page->lru, src); | |
1158 | continue; | |
46453a6e | 1159 | |
5ad333eb AW |
1160 | default: |
1161 | BUG(); | |
1162 | } | |
1da177e4 LT |
1163 | } |
1164 | ||
f626012d | 1165 | *nr_scanned = scan; |
a8a94d15 | 1166 | |
fe2c2a10 | 1167 | trace_mm_vmscan_lru_isolate(sc->order, |
a8a94d15 MG |
1168 | nr_to_scan, scan, |
1169 | nr_taken, | |
ea4d349f | 1170 | mode, file); |
1da177e4 LT |
1171 | return nr_taken; |
1172 | } | |
1173 | ||
62695a84 NP |
1174 | /** |
1175 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1176 | * @page: page to isolate from its LRU list | |
1177 | * | |
1178 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1179 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1180 | * | |
1181 | * Returns 0 if the page was removed from an LRU list. | |
1182 | * Returns -EBUSY if the page was not on an LRU list. | |
1183 | * | |
1184 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1185 | * the active list, it will have PageActive set. If it was found on |
1186 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1187 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1188 | * |
1189 | * The vmstat statistic corresponding to the list on which the page was | |
1190 | * found will be decremented. | |
1191 | * | |
1192 | * Restrictions: | |
1193 | * (1) Must be called with an elevated refcount on the page. This is a | |
1194 | * fundamentnal difference from isolate_lru_pages (which is called | |
1195 | * without a stable reference). | |
1196 | * (2) the lru_lock must not be held. | |
1197 | * (3) interrupts must be enabled. | |
1198 | */ | |
1199 | int isolate_lru_page(struct page *page) | |
1200 | { | |
1201 | int ret = -EBUSY; | |
1202 | ||
0c917313 KK |
1203 | VM_BUG_ON(!page_count(page)); |
1204 | ||
62695a84 NP |
1205 | if (PageLRU(page)) { |
1206 | struct zone *zone = page_zone(page); | |
1207 | ||
1208 | spin_lock_irq(&zone->lru_lock); | |
0c917313 | 1209 | if (PageLRU(page)) { |
894bc310 | 1210 | int lru = page_lru(page); |
62695a84 | 1211 | ret = 0; |
0c917313 | 1212 | get_page(page); |
62695a84 | 1213 | ClearPageLRU(page); |
4f98a2fe | 1214 | |
4f98a2fe | 1215 | del_page_from_lru_list(zone, page, lru); |
62695a84 NP |
1216 | } |
1217 | spin_unlock_irq(&zone->lru_lock); | |
1218 | } | |
1219 | return ret; | |
1220 | } | |
1221 | ||
35cd7815 RR |
1222 | /* |
1223 | * Are there way too many processes in the direct reclaim path already? | |
1224 | */ | |
1225 | static int too_many_isolated(struct zone *zone, int file, | |
1226 | struct scan_control *sc) | |
1227 | { | |
1228 | unsigned long inactive, isolated; | |
1229 | ||
1230 | if (current_is_kswapd()) | |
1231 | return 0; | |
1232 | ||
89b5fae5 | 1233 | if (!global_reclaim(sc)) |
35cd7815 RR |
1234 | return 0; |
1235 | ||
1236 | if (file) { | |
1237 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1238 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1239 | } else { | |
1240 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1241 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1242 | } | |
1243 | ||
1244 | return isolated > inactive; | |
1245 | } | |
1246 | ||
66635629 | 1247 | static noinline_for_stack void |
3f79768f HD |
1248 | putback_inactive_pages(struct mem_cgroup_zone *mz, |
1249 | struct list_head *page_list) | |
66635629 | 1250 | { |
f16015fb | 1251 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
3f79768f HD |
1252 | struct zone *zone = mz->zone; |
1253 | LIST_HEAD(pages_to_free); | |
66635629 | 1254 | |
66635629 MG |
1255 | /* |
1256 | * Put back any unfreeable pages. | |
1257 | */ | |
66635629 | 1258 | while (!list_empty(page_list)) { |
3f79768f | 1259 | struct page *page = lru_to_page(page_list); |
66635629 | 1260 | int lru; |
3f79768f | 1261 | |
66635629 MG |
1262 | VM_BUG_ON(PageLRU(page)); |
1263 | list_del(&page->lru); | |
1264 | if (unlikely(!page_evictable(page, NULL))) { | |
1265 | spin_unlock_irq(&zone->lru_lock); | |
1266 | putback_lru_page(page); | |
1267 | spin_lock_irq(&zone->lru_lock); | |
1268 | continue; | |
1269 | } | |
7a608572 | 1270 | SetPageLRU(page); |
66635629 | 1271 | lru = page_lru(page); |
7a608572 | 1272 | add_page_to_lru_list(zone, page, lru); |
66635629 MG |
1273 | if (is_active_lru(lru)) { |
1274 | int file = is_file_lru(lru); | |
9992af10 RR |
1275 | int numpages = hpage_nr_pages(page); |
1276 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1277 | } |
2bcf8879 HD |
1278 | if (put_page_testzero(page)) { |
1279 | __ClearPageLRU(page); | |
1280 | __ClearPageActive(page); | |
1281 | del_page_from_lru_list(zone, page, lru); | |
1282 | ||
1283 | if (unlikely(PageCompound(page))) { | |
1284 | spin_unlock_irq(&zone->lru_lock); | |
1285 | (*get_compound_page_dtor(page))(page); | |
1286 | spin_lock_irq(&zone->lru_lock); | |
1287 | } else | |
1288 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1289 | } |
1290 | } | |
66635629 | 1291 | |
3f79768f HD |
1292 | /* |
1293 | * To save our caller's stack, now use input list for pages to free. | |
1294 | */ | |
1295 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1296 | } |
1297 | ||
f16015fb JW |
1298 | static noinline_for_stack void |
1299 | update_isolated_counts(struct mem_cgroup_zone *mz, | |
3f79768f | 1300 | struct list_head *page_list, |
f16015fb | 1301 | unsigned long *nr_anon, |
3f79768f | 1302 | unsigned long *nr_file) |
1489fa14 | 1303 | { |
f16015fb | 1304 | struct zone *zone = mz->zone; |
1489fa14 | 1305 | unsigned int count[NR_LRU_LISTS] = { 0, }; |
3f79768f HD |
1306 | unsigned long nr_active = 0; |
1307 | struct page *page; | |
1308 | int lru; | |
1309 | ||
1310 | /* | |
1311 | * Count pages and clear active flags | |
1312 | */ | |
1313 | list_for_each_entry(page, page_list, lru) { | |
1314 | int numpages = hpage_nr_pages(page); | |
1315 | lru = page_lru_base_type(page); | |
1316 | if (PageActive(page)) { | |
1317 | lru += LRU_ACTIVE; | |
1318 | ClearPageActive(page); | |
1319 | nr_active += numpages; | |
1320 | } | |
1321 | count[lru] += numpages; | |
1322 | } | |
1489fa14 | 1323 | |
d563c050 | 1324 | preempt_disable(); |
1489fa14 MG |
1325 | __count_vm_events(PGDEACTIVATE, nr_active); |
1326 | ||
1327 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, | |
1328 | -count[LRU_ACTIVE_FILE]); | |
1329 | __mod_zone_page_state(zone, NR_INACTIVE_FILE, | |
1330 | -count[LRU_INACTIVE_FILE]); | |
1331 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, | |
1332 | -count[LRU_ACTIVE_ANON]); | |
1333 | __mod_zone_page_state(zone, NR_INACTIVE_ANON, | |
1334 | -count[LRU_INACTIVE_ANON]); | |
1335 | ||
1336 | *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON]; | |
1337 | *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE]; | |
1489fa14 | 1338 | |
d563c050 HD |
1339 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon); |
1340 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file); | |
1341 | preempt_enable(); | |
1489fa14 MG |
1342 | } |
1343 | ||
e31f3698 | 1344 | /* |
a18bba06 | 1345 | * Returns true if a direct reclaim should wait on pages under writeback. |
e31f3698 WF |
1346 | * |
1347 | * If we are direct reclaiming for contiguous pages and we do not reclaim | |
1348 | * everything in the list, try again and wait for writeback IO to complete. | |
1349 | * This will stall high-order allocations noticeably. Only do that when really | |
1350 | * need to free the pages under high memory pressure. | |
1351 | */ | |
1352 | static inline bool should_reclaim_stall(unsigned long nr_taken, | |
1353 | unsigned long nr_freed, | |
1354 | int priority, | |
1355 | struct scan_control *sc) | |
1356 | { | |
c53919ad | 1357 | int stall_priority; |
e31f3698 WF |
1358 | |
1359 | /* kswapd should not stall on sync IO */ | |
1360 | if (current_is_kswapd()) | |
1361 | return false; | |
1362 | ||
c53919ad | 1363 | /* Only stall for memory compaction */ |
f3a310bc | 1364 | if (sc->reclaim_mode & RECLAIM_MODE_SINGLE) |
e31f3698 WF |
1365 | return false; |
1366 | ||
81d66c70 | 1367 | /* If we have reclaimed everything on the isolated list, no stall */ |
e31f3698 WF |
1368 | if (nr_freed == nr_taken) |
1369 | return false; | |
1370 | ||
1371 | /* | |
1372 | * For high-order allocations, there are two stall thresholds. | |
1373 | * High-cost allocations stall immediately where as lower | |
1374 | * order allocations such as stacks require the scanning | |
1375 | * priority to be much higher before stalling. | |
1376 | */ | |
1377 | if (sc->order > PAGE_ALLOC_COSTLY_ORDER) | |
c53919ad | 1378 | stall_priority = DEF_PRIORITY; |
e31f3698 | 1379 | else |
c53919ad | 1380 | stall_priority = DEF_PRIORITY / 3; |
e31f3698 | 1381 | |
c53919ad | 1382 | return priority <= stall_priority; |
e31f3698 WF |
1383 | } |
1384 | ||
1da177e4 | 1385 | /* |
1742f19f AM |
1386 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1387 | * of reclaimed pages | |
1da177e4 | 1388 | */ |
66635629 | 1389 | static noinline_for_stack unsigned long |
f16015fb JW |
1390 | shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz, |
1391 | struct scan_control *sc, int priority, int file) | |
1da177e4 LT |
1392 | { |
1393 | LIST_HEAD(page_list); | |
e247dbce | 1394 | unsigned long nr_scanned; |
05ff5137 | 1395 | unsigned long nr_reclaimed = 0; |
e247dbce | 1396 | unsigned long nr_taken; |
e247dbce KM |
1397 | unsigned long nr_anon; |
1398 | unsigned long nr_file; | |
92df3a72 MG |
1399 | unsigned long nr_dirty = 0; |
1400 | unsigned long nr_writeback = 0; | |
61317289 | 1401 | isolate_mode_t isolate_mode = ISOLATE_INACTIVE; |
f16015fb | 1402 | struct zone *zone = mz->zone; |
d563c050 | 1403 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
78dc583d | 1404 | |
35cd7815 | 1405 | while (unlikely(too_many_isolated(zone, file, sc))) { |
58355c78 | 1406 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1407 | |
1408 | /* We are about to die and free our memory. Return now. */ | |
1409 | if (fatal_signal_pending(current)) | |
1410 | return SWAP_CLUSTER_MAX; | |
1411 | } | |
1412 | ||
f3a310bc | 1413 | set_reclaim_mode(priority, sc, false); |
4356f21d | 1414 | |
1da177e4 | 1415 | lru_add_drain(); |
f80c0673 MK |
1416 | |
1417 | if (!sc->may_unmap) | |
61317289 | 1418 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1419 | if (!sc->may_writepage) |
61317289 | 1420 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1421 | |
1da177e4 | 1422 | spin_lock_irq(&zone->lru_lock); |
b35ea17b | 1423 | |
fe2c2a10 RR |
1424 | nr_taken = isolate_lru_pages(nr_to_scan, mz, &page_list, &nr_scanned, |
1425 | sc, isolate_mode, 0, file); | |
89b5fae5 | 1426 | if (global_reclaim(sc)) { |
e247dbce KM |
1427 | zone->pages_scanned += nr_scanned; |
1428 | if (current_is_kswapd()) | |
1429 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, | |
1430 | nr_scanned); | |
1431 | else | |
1432 | __count_zone_vm_events(PGSCAN_DIRECT, zone, | |
1433 | nr_scanned); | |
e247dbce | 1434 | } |
d563c050 | 1435 | spin_unlock_irq(&zone->lru_lock); |
b35ea17b | 1436 | |
d563c050 | 1437 | if (nr_taken == 0) |
66635629 | 1438 | return 0; |
5ad333eb | 1439 | |
3f79768f HD |
1440 | update_isolated_counts(mz, &page_list, &nr_anon, &nr_file); |
1441 | ||
f16015fb | 1442 | nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority, |
92df3a72 | 1443 | &nr_dirty, &nr_writeback); |
c661b078 | 1444 | |
e31f3698 WF |
1445 | /* Check if we should syncronously wait for writeback */ |
1446 | if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) { | |
f3a310bc | 1447 | set_reclaim_mode(priority, sc, true); |
f16015fb | 1448 | nr_reclaimed += shrink_page_list(&page_list, mz, sc, |
92df3a72 | 1449 | priority, &nr_dirty, &nr_writeback); |
e247dbce | 1450 | } |
b35ea17b | 1451 | |
3f79768f HD |
1452 | spin_lock_irq(&zone->lru_lock); |
1453 | ||
d563c050 HD |
1454 | reclaim_stat->recent_scanned[0] += nr_anon; |
1455 | reclaim_stat->recent_scanned[1] += nr_file; | |
1456 | ||
904249aa YH |
1457 | if (global_reclaim(sc)) { |
1458 | if (current_is_kswapd()) | |
1459 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1460 | nr_reclaimed); | |
1461 | else | |
1462 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1463 | nr_reclaimed); | |
1464 | } | |
a74609fa | 1465 | |
3f79768f HD |
1466 | putback_inactive_pages(mz, &page_list); |
1467 | ||
1468 | __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon); | |
1469 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file); | |
1470 | ||
1471 | spin_unlock_irq(&zone->lru_lock); | |
1472 | ||
1473 | free_hot_cold_page_list(&page_list, 1); | |
e11da5b4 | 1474 | |
92df3a72 MG |
1475 | /* |
1476 | * If reclaim is isolating dirty pages under writeback, it implies | |
1477 | * that the long-lived page allocation rate is exceeding the page | |
1478 | * laundering rate. Either the global limits are not being effective | |
1479 | * at throttling processes due to the page distribution throughout | |
1480 | * zones or there is heavy usage of a slow backing device. The | |
1481 | * only option is to throttle from reclaim context which is not ideal | |
1482 | * as there is no guarantee the dirtying process is throttled in the | |
1483 | * same way balance_dirty_pages() manages. | |
1484 | * | |
1485 | * This scales the number of dirty pages that must be under writeback | |
1486 | * before throttling depending on priority. It is a simple backoff | |
1487 | * function that has the most effect in the range DEF_PRIORITY to | |
1488 | * DEF_PRIORITY-2 which is the priority reclaim is considered to be | |
1489 | * in trouble and reclaim is considered to be in trouble. | |
1490 | * | |
1491 | * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle | |
1492 | * DEF_PRIORITY-1 50% must be PageWriteback | |
1493 | * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble | |
1494 | * ... | |
1495 | * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any | |
1496 | * isolated page is PageWriteback | |
1497 | */ | |
1498 | if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority))) | |
1499 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); | |
1500 | ||
e11da5b4 MG |
1501 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, |
1502 | zone_idx(zone), | |
1503 | nr_scanned, nr_reclaimed, | |
1504 | priority, | |
f3a310bc | 1505 | trace_shrink_flags(file, sc->reclaim_mode)); |
05ff5137 | 1506 | return nr_reclaimed; |
1da177e4 LT |
1507 | } |
1508 | ||
1509 | /* | |
1510 | * This moves pages from the active list to the inactive list. | |
1511 | * | |
1512 | * We move them the other way if the page is referenced by one or more | |
1513 | * processes, from rmap. | |
1514 | * | |
1515 | * If the pages are mostly unmapped, the processing is fast and it is | |
1516 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1517 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1518 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1519 | * this, so instead we remove the pages from the LRU while processing them. | |
1520 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1521 | * nobody will play with that bit on a non-LRU page. | |
1522 | * | |
1523 | * The downside is that we have to touch page->_count against each page. | |
1524 | * But we had to alter page->flags anyway. | |
1525 | */ | |
1cfb419b | 1526 | |
3eb4140f WF |
1527 | static void move_active_pages_to_lru(struct zone *zone, |
1528 | struct list_head *list, | |
2bcf8879 | 1529 | struct list_head *pages_to_free, |
3eb4140f WF |
1530 | enum lru_list lru) |
1531 | { | |
1532 | unsigned long pgmoved = 0; | |
3eb4140f WF |
1533 | struct page *page; |
1534 | ||
3eb4140f | 1535 | while (!list_empty(list)) { |
925b7673 JW |
1536 | struct lruvec *lruvec; |
1537 | ||
3eb4140f | 1538 | page = lru_to_page(list); |
3eb4140f WF |
1539 | |
1540 | VM_BUG_ON(PageLRU(page)); | |
1541 | SetPageLRU(page); | |
1542 | ||
925b7673 JW |
1543 | lruvec = mem_cgroup_lru_add_list(zone, page, lru); |
1544 | list_move(&page->lru, &lruvec->lists[lru]); | |
2c888cfb | 1545 | pgmoved += hpage_nr_pages(page); |
3eb4140f | 1546 | |
2bcf8879 HD |
1547 | if (put_page_testzero(page)) { |
1548 | __ClearPageLRU(page); | |
1549 | __ClearPageActive(page); | |
1550 | del_page_from_lru_list(zone, page, lru); | |
1551 | ||
1552 | if (unlikely(PageCompound(page))) { | |
1553 | spin_unlock_irq(&zone->lru_lock); | |
1554 | (*get_compound_page_dtor(page))(page); | |
1555 | spin_lock_irq(&zone->lru_lock); | |
1556 | } else | |
1557 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1558 | } |
1559 | } | |
1560 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1561 | if (!is_active_lru(lru)) | |
1562 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1563 | } | |
1cfb419b | 1564 | |
f626012d | 1565 | static void shrink_active_list(unsigned long nr_to_scan, |
f16015fb JW |
1566 | struct mem_cgroup_zone *mz, |
1567 | struct scan_control *sc, | |
1568 | int priority, int file) | |
1da177e4 | 1569 | { |
44c241f1 | 1570 | unsigned long nr_taken; |
f626012d | 1571 | unsigned long nr_scanned; |
6fe6b7e3 | 1572 | unsigned long vm_flags; |
1da177e4 | 1573 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1574 | LIST_HEAD(l_active); |
b69408e8 | 1575 | LIST_HEAD(l_inactive); |
1da177e4 | 1576 | struct page *page; |
f16015fb | 1577 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
44c241f1 | 1578 | unsigned long nr_rotated = 0; |
61317289 | 1579 | isolate_mode_t isolate_mode = ISOLATE_ACTIVE; |
f16015fb | 1580 | struct zone *zone = mz->zone; |
1da177e4 LT |
1581 | |
1582 | lru_add_drain(); | |
f80c0673 | 1583 | |
1480de03 KK |
1584 | reset_reclaim_mode(sc); |
1585 | ||
f80c0673 | 1586 | if (!sc->may_unmap) |
61317289 | 1587 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1588 | if (!sc->may_writepage) |
61317289 | 1589 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1590 | |
1da177e4 | 1591 | spin_lock_irq(&zone->lru_lock); |
925b7673 | 1592 | |
fe2c2a10 | 1593 | nr_taken = isolate_lru_pages(nr_to_scan, mz, &l_hold, &nr_scanned, sc, |
61317289 | 1594 | isolate_mode, 1, file); |
89b5fae5 | 1595 | if (global_reclaim(sc)) |
f626012d | 1596 | zone->pages_scanned += nr_scanned; |
89b5fae5 | 1597 | |
b7c46d15 | 1598 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1599 | |
f626012d | 1600 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); |
4f98a2fe | 1601 | if (file) |
44c241f1 | 1602 | __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken); |
4f98a2fe | 1603 | else |
44c241f1 | 1604 | __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken); |
a731286d | 1605 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
1da177e4 LT |
1606 | spin_unlock_irq(&zone->lru_lock); |
1607 | ||
1da177e4 LT |
1608 | while (!list_empty(&l_hold)) { |
1609 | cond_resched(); | |
1610 | page = lru_to_page(&l_hold); | |
1611 | list_del(&page->lru); | |
7e9cd484 | 1612 | |
894bc310 LS |
1613 | if (unlikely(!page_evictable(page, NULL))) { |
1614 | putback_lru_page(page); | |
1615 | continue; | |
1616 | } | |
1617 | ||
cc715d99 MG |
1618 | if (unlikely(buffer_heads_over_limit)) { |
1619 | if (page_has_private(page) && trylock_page(page)) { | |
1620 | if (page_has_private(page)) | |
1621 | try_to_release_page(page, 0); | |
1622 | unlock_page(page); | |
1623 | } | |
1624 | } | |
1625 | ||
f16015fb | 1626 | if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) { |
9992af10 | 1627 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1628 | /* |
1629 | * Identify referenced, file-backed active pages and | |
1630 | * give them one more trip around the active list. So | |
1631 | * that executable code get better chances to stay in | |
1632 | * memory under moderate memory pressure. Anon pages | |
1633 | * are not likely to be evicted by use-once streaming | |
1634 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1635 | * so we ignore them here. | |
1636 | */ | |
41e20983 | 1637 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1638 | list_add(&page->lru, &l_active); |
1639 | continue; | |
1640 | } | |
1641 | } | |
7e9cd484 | 1642 | |
5205e56e | 1643 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1644 | list_add(&page->lru, &l_inactive); |
1645 | } | |
1646 | ||
b555749a | 1647 | /* |
8cab4754 | 1648 | * Move pages back to the lru list. |
b555749a | 1649 | */ |
2a1dc509 | 1650 | spin_lock_irq(&zone->lru_lock); |
556adecb | 1651 | /* |
8cab4754 WF |
1652 | * Count referenced pages from currently used mappings as rotated, |
1653 | * even though only some of them are actually re-activated. This | |
1654 | * helps balance scan pressure between file and anonymous pages in | |
1655 | * get_scan_ratio. | |
7e9cd484 | 1656 | */ |
b7c46d15 | 1657 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1658 | |
2bcf8879 | 1659 | move_active_pages_to_lru(zone, &l_active, &l_hold, |
3eb4140f | 1660 | LRU_ACTIVE + file * LRU_FILE); |
2bcf8879 | 1661 | move_active_pages_to_lru(zone, &l_inactive, &l_hold, |
3eb4140f | 1662 | LRU_BASE + file * LRU_FILE); |
a731286d | 1663 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
f8891e5e | 1664 | spin_unlock_irq(&zone->lru_lock); |
2bcf8879 HD |
1665 | |
1666 | free_hot_cold_page_list(&l_hold, 1); | |
1da177e4 LT |
1667 | } |
1668 | ||
74e3f3c3 | 1669 | #ifdef CONFIG_SWAP |
14797e23 | 1670 | static int inactive_anon_is_low_global(struct zone *zone) |
f89eb90e KM |
1671 | { |
1672 | unsigned long active, inactive; | |
1673 | ||
1674 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1675 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1676 | ||
1677 | if (inactive * zone->inactive_ratio < active) | |
1678 | return 1; | |
1679 | ||
1680 | return 0; | |
1681 | } | |
1682 | ||
14797e23 KM |
1683 | /** |
1684 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
1685 | * @zone: zone to check | |
1686 | * @sc: scan control of this context | |
1687 | * | |
1688 | * Returns true if the zone does not have enough inactive anon pages, | |
1689 | * meaning some active anon pages need to be deactivated. | |
1690 | */ | |
f16015fb | 1691 | static int inactive_anon_is_low(struct mem_cgroup_zone *mz) |
14797e23 | 1692 | { |
74e3f3c3 MK |
1693 | /* |
1694 | * If we don't have swap space, anonymous page deactivation | |
1695 | * is pointless. | |
1696 | */ | |
1697 | if (!total_swap_pages) | |
1698 | return 0; | |
1699 | ||
f16015fb JW |
1700 | if (!scanning_global_lru(mz)) |
1701 | return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup, | |
1702 | mz->zone); | |
1703 | ||
1704 | return inactive_anon_is_low_global(mz->zone); | |
14797e23 | 1705 | } |
74e3f3c3 | 1706 | #else |
f16015fb | 1707 | static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz) |
74e3f3c3 MK |
1708 | { |
1709 | return 0; | |
1710 | } | |
1711 | #endif | |
14797e23 | 1712 | |
56e49d21 RR |
1713 | static int inactive_file_is_low_global(struct zone *zone) |
1714 | { | |
1715 | unsigned long active, inactive; | |
1716 | ||
1717 | active = zone_page_state(zone, NR_ACTIVE_FILE); | |
1718 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1719 | ||
1720 | return (active > inactive); | |
1721 | } | |
1722 | ||
1723 | /** | |
1724 | * inactive_file_is_low - check if file pages need to be deactivated | |
f16015fb | 1725 | * @mz: memory cgroup and zone to check |
56e49d21 RR |
1726 | * |
1727 | * When the system is doing streaming IO, memory pressure here | |
1728 | * ensures that active file pages get deactivated, until more | |
1729 | * than half of the file pages are on the inactive list. | |
1730 | * | |
1731 | * Once we get to that situation, protect the system's working | |
1732 | * set from being evicted by disabling active file page aging. | |
1733 | * | |
1734 | * This uses a different ratio than the anonymous pages, because | |
1735 | * the page cache uses a use-once replacement algorithm. | |
1736 | */ | |
f16015fb | 1737 | static int inactive_file_is_low(struct mem_cgroup_zone *mz) |
56e49d21 | 1738 | { |
f16015fb JW |
1739 | if (!scanning_global_lru(mz)) |
1740 | return mem_cgroup_inactive_file_is_low(mz->mem_cgroup, | |
1741 | mz->zone); | |
56e49d21 | 1742 | |
f16015fb | 1743 | return inactive_file_is_low_global(mz->zone); |
56e49d21 RR |
1744 | } |
1745 | ||
f16015fb | 1746 | static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file) |
b39415b2 RR |
1747 | { |
1748 | if (file) | |
f16015fb | 1749 | return inactive_file_is_low(mz); |
b39415b2 | 1750 | else |
f16015fb | 1751 | return inactive_anon_is_low(mz); |
b39415b2 RR |
1752 | } |
1753 | ||
4f98a2fe | 1754 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
f16015fb JW |
1755 | struct mem_cgroup_zone *mz, |
1756 | struct scan_control *sc, int priority) | |
b69408e8 | 1757 | { |
4f98a2fe RR |
1758 | int file = is_file_lru(lru); |
1759 | ||
b39415b2 | 1760 | if (is_active_lru(lru)) { |
f16015fb JW |
1761 | if (inactive_list_is_low(mz, file)) |
1762 | shrink_active_list(nr_to_scan, mz, sc, priority, file); | |
556adecb RR |
1763 | return 0; |
1764 | } | |
1765 | ||
f16015fb | 1766 | return shrink_inactive_list(nr_to_scan, mz, sc, priority, file); |
4f98a2fe RR |
1767 | } |
1768 | ||
f16015fb JW |
1769 | static int vmscan_swappiness(struct mem_cgroup_zone *mz, |
1770 | struct scan_control *sc) | |
1f4c025b | 1771 | { |
89b5fae5 | 1772 | if (global_reclaim(sc)) |
1f4c025b | 1773 | return vm_swappiness; |
f16015fb | 1774 | return mem_cgroup_swappiness(mz->mem_cgroup); |
1f4c025b KH |
1775 | } |
1776 | ||
4f98a2fe RR |
1777 | /* |
1778 | * Determine how aggressively the anon and file LRU lists should be | |
1779 | * scanned. The relative value of each set of LRU lists is determined | |
1780 | * by looking at the fraction of the pages scanned we did rotate back | |
1781 | * onto the active list instead of evict. | |
1782 | * | |
76a33fc3 | 1783 | * nr[0] = anon pages to scan; nr[1] = file pages to scan |
4f98a2fe | 1784 | */ |
f16015fb JW |
1785 | static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc, |
1786 | unsigned long *nr, int priority) | |
4f98a2fe RR |
1787 | { |
1788 | unsigned long anon, file, free; | |
1789 | unsigned long anon_prio, file_prio; | |
1790 | unsigned long ap, fp; | |
f16015fb | 1791 | struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz); |
76a33fc3 | 1792 | u64 fraction[2], denominator; |
4111304d | 1793 | enum lru_list lru; |
76a33fc3 | 1794 | int noswap = 0; |
a4d3e9e7 | 1795 | bool force_scan = false; |
246e87a9 | 1796 | |
f11c0ca5 JW |
1797 | /* |
1798 | * If the zone or memcg is small, nr[l] can be 0. This | |
1799 | * results in no scanning on this priority and a potential | |
1800 | * priority drop. Global direct reclaim can go to the next | |
1801 | * zone and tends to have no problems. Global kswapd is for | |
1802 | * zone balancing and it needs to scan a minimum amount. When | |
1803 | * reclaiming for a memcg, a priority drop can cause high | |
1804 | * latencies, so it's better to scan a minimum amount there as | |
1805 | * well. | |
1806 | */ | |
b95a2f2d | 1807 | if (current_is_kswapd() && mz->zone->all_unreclaimable) |
a4d3e9e7 | 1808 | force_scan = true; |
89b5fae5 | 1809 | if (!global_reclaim(sc)) |
a4d3e9e7 | 1810 | force_scan = true; |
76a33fc3 SL |
1811 | |
1812 | /* If we have no swap space, do not bother scanning anon pages. */ | |
1813 | if (!sc->may_swap || (nr_swap_pages <= 0)) { | |
1814 | noswap = 1; | |
1815 | fraction[0] = 0; | |
1816 | fraction[1] = 1; | |
1817 | denominator = 1; | |
1818 | goto out; | |
1819 | } | |
4f98a2fe | 1820 | |
f16015fb JW |
1821 | anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) + |
1822 | zone_nr_lru_pages(mz, LRU_INACTIVE_ANON); | |
1823 | file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) + | |
1824 | zone_nr_lru_pages(mz, LRU_INACTIVE_FILE); | |
a4d3e9e7 | 1825 | |
89b5fae5 | 1826 | if (global_reclaim(sc)) { |
f16015fb | 1827 | free = zone_page_state(mz->zone, NR_FREE_PAGES); |
eeee9a8c KM |
1828 | /* If we have very few page cache pages, |
1829 | force-scan anon pages. */ | |
f16015fb | 1830 | if (unlikely(file + free <= high_wmark_pages(mz->zone))) { |
76a33fc3 SL |
1831 | fraction[0] = 1; |
1832 | fraction[1] = 0; | |
1833 | denominator = 1; | |
1834 | goto out; | |
eeee9a8c | 1835 | } |
4f98a2fe RR |
1836 | } |
1837 | ||
58c37f6e KM |
1838 | /* |
1839 | * With swappiness at 100, anonymous and file have the same priority. | |
1840 | * This scanning priority is essentially the inverse of IO cost. | |
1841 | */ | |
f16015fb JW |
1842 | anon_prio = vmscan_swappiness(mz, sc); |
1843 | file_prio = 200 - vmscan_swappiness(mz, sc); | |
58c37f6e | 1844 | |
4f98a2fe RR |
1845 | /* |
1846 | * OK, so we have swap space and a fair amount of page cache | |
1847 | * pages. We use the recently rotated / recently scanned | |
1848 | * ratios to determine how valuable each cache is. | |
1849 | * | |
1850 | * Because workloads change over time (and to avoid overflow) | |
1851 | * we keep these statistics as a floating average, which ends | |
1852 | * up weighing recent references more than old ones. | |
1853 | * | |
1854 | * anon in [0], file in [1] | |
1855 | */ | |
f16015fb | 1856 | spin_lock_irq(&mz->zone->lru_lock); |
6e901571 | 1857 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
1858 | reclaim_stat->recent_scanned[0] /= 2; |
1859 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
1860 | } |
1861 | ||
6e901571 | 1862 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
1863 | reclaim_stat->recent_scanned[1] /= 2; |
1864 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
1865 | } |
1866 | ||
4f98a2fe | 1867 | /* |
00d8089c RR |
1868 | * The amount of pressure on anon vs file pages is inversely |
1869 | * proportional to the fraction of recently scanned pages on | |
1870 | * each list that were recently referenced and in active use. | |
4f98a2fe | 1871 | */ |
6e901571 KM |
1872 | ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1); |
1873 | ap /= reclaim_stat->recent_rotated[0] + 1; | |
4f98a2fe | 1874 | |
6e901571 KM |
1875 | fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1); |
1876 | fp /= reclaim_stat->recent_rotated[1] + 1; | |
f16015fb | 1877 | spin_unlock_irq(&mz->zone->lru_lock); |
4f98a2fe | 1878 | |
76a33fc3 SL |
1879 | fraction[0] = ap; |
1880 | fraction[1] = fp; | |
1881 | denominator = ap + fp + 1; | |
1882 | out: | |
4111304d HD |
1883 | for_each_evictable_lru(lru) { |
1884 | int file = is_file_lru(lru); | |
76a33fc3 | 1885 | unsigned long scan; |
6e08a369 | 1886 | |
4111304d | 1887 | scan = zone_nr_lru_pages(mz, lru); |
76a33fc3 SL |
1888 | if (priority || noswap) { |
1889 | scan >>= priority; | |
f11c0ca5 JW |
1890 | if (!scan && force_scan) |
1891 | scan = SWAP_CLUSTER_MAX; | |
76a33fc3 SL |
1892 | scan = div64_u64(scan * fraction[file], denominator); |
1893 | } | |
4111304d | 1894 | nr[lru] = scan; |
76a33fc3 | 1895 | } |
6e08a369 | 1896 | } |
4f98a2fe | 1897 | |
3e7d3449 MG |
1898 | /* |
1899 | * Reclaim/compaction depends on a number of pages being freed. To avoid | |
1900 | * disruption to the system, a small number of order-0 pages continue to be | |
1901 | * rotated and reclaimed in the normal fashion. However, by the time we get | |
1902 | * back to the allocator and call try_to_compact_zone(), we ensure that | |
1903 | * there are enough free pages for it to be likely successful | |
1904 | */ | |
f16015fb | 1905 | static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz, |
3e7d3449 MG |
1906 | unsigned long nr_reclaimed, |
1907 | unsigned long nr_scanned, | |
1908 | struct scan_control *sc) | |
1909 | { | |
1910 | unsigned long pages_for_compaction; | |
1911 | unsigned long inactive_lru_pages; | |
1912 | ||
1913 | /* If not in reclaim/compaction mode, stop */ | |
f3a310bc | 1914 | if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION)) |
3e7d3449 MG |
1915 | return false; |
1916 | ||
2876592f MG |
1917 | /* Consider stopping depending on scan and reclaim activity */ |
1918 | if (sc->gfp_mask & __GFP_REPEAT) { | |
1919 | /* | |
1920 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
1921 | * full LRU list has been scanned and we are still failing | |
1922 | * to reclaim pages. This full LRU scan is potentially | |
1923 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
1924 | */ | |
1925 | if (!nr_reclaimed && !nr_scanned) | |
1926 | return false; | |
1927 | } else { | |
1928 | /* | |
1929 | * For non-__GFP_REPEAT allocations which can presumably | |
1930 | * fail without consequence, stop if we failed to reclaim | |
1931 | * any pages from the last SWAP_CLUSTER_MAX number of | |
1932 | * pages that were scanned. This will return to the | |
1933 | * caller faster at the risk reclaim/compaction and | |
1934 | * the resulting allocation attempt fails | |
1935 | */ | |
1936 | if (!nr_reclaimed) | |
1937 | return false; | |
1938 | } | |
3e7d3449 MG |
1939 | |
1940 | /* | |
1941 | * If we have not reclaimed enough pages for compaction and the | |
1942 | * inactive lists are large enough, continue reclaiming | |
1943 | */ | |
1944 | pages_for_compaction = (2UL << sc->order); | |
f16015fb | 1945 | inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE); |
86cfd3a4 | 1946 | if (nr_swap_pages > 0) |
f16015fb | 1947 | inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON); |
3e7d3449 MG |
1948 | if (sc->nr_reclaimed < pages_for_compaction && |
1949 | inactive_lru_pages > pages_for_compaction) | |
1950 | return true; | |
1951 | ||
1952 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
f16015fb | 1953 | switch (compaction_suitable(mz->zone, sc->order)) { |
3e7d3449 MG |
1954 | case COMPACT_PARTIAL: |
1955 | case COMPACT_CONTINUE: | |
1956 | return false; | |
1957 | default: | |
1958 | return true; | |
1959 | } | |
1960 | } | |
1961 | ||
1da177e4 LT |
1962 | /* |
1963 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1964 | */ | |
f16015fb JW |
1965 | static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz, |
1966 | struct scan_control *sc) | |
1da177e4 | 1967 | { |
b69408e8 | 1968 | unsigned long nr[NR_LRU_LISTS]; |
8695949a | 1969 | unsigned long nr_to_scan; |
4111304d | 1970 | enum lru_list lru; |
f0fdc5e8 | 1971 | unsigned long nr_reclaimed, nr_scanned; |
22fba335 | 1972 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; |
3da367c3 | 1973 | struct blk_plug plug; |
e0f79b8f | 1974 | |
3e7d3449 MG |
1975 | restart: |
1976 | nr_reclaimed = 0; | |
f0fdc5e8 | 1977 | nr_scanned = sc->nr_scanned; |
f16015fb | 1978 | get_scan_count(mz, sc, nr, priority); |
1da177e4 | 1979 | |
3da367c3 | 1980 | blk_start_plug(&plug); |
556adecb RR |
1981 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || |
1982 | nr[LRU_INACTIVE_FILE]) { | |
4111304d HD |
1983 | for_each_evictable_lru(lru) { |
1984 | if (nr[lru]) { | |
ece74b2e | 1985 | nr_to_scan = min_t(unsigned long, |
4111304d HD |
1986 | nr[lru], SWAP_CLUSTER_MAX); |
1987 | nr[lru] -= nr_to_scan; | |
1da177e4 | 1988 | |
4111304d | 1989 | nr_reclaimed += shrink_list(lru, nr_to_scan, |
f16015fb | 1990 | mz, sc, priority); |
b69408e8 | 1991 | } |
1da177e4 | 1992 | } |
a79311c1 RR |
1993 | /* |
1994 | * On large memory systems, scan >> priority can become | |
1995 | * really large. This is fine for the starting priority; | |
1996 | * we want to put equal scanning pressure on each zone. | |
1997 | * However, if the VM has a harder time of freeing pages, | |
1998 | * with multiple processes reclaiming pages, the total | |
1999 | * freeing target can get unreasonably large. | |
2000 | */ | |
41c93088 | 2001 | if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY) |
a79311c1 | 2002 | break; |
1da177e4 | 2003 | } |
3da367c3 | 2004 | blk_finish_plug(&plug); |
3e7d3449 | 2005 | sc->nr_reclaimed += nr_reclaimed; |
01dbe5c9 | 2006 | |
556adecb RR |
2007 | /* |
2008 | * Even if we did not try to evict anon pages at all, we want to | |
2009 | * rebalance the anon lru active/inactive ratio. | |
2010 | */ | |
f16015fb JW |
2011 | if (inactive_anon_is_low(mz)) |
2012 | shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0); | |
556adecb | 2013 | |
3e7d3449 | 2014 | /* reclaim/compaction might need reclaim to continue */ |
f16015fb | 2015 | if (should_continue_reclaim(mz, nr_reclaimed, |
3e7d3449 MG |
2016 | sc->nr_scanned - nr_scanned, sc)) |
2017 | goto restart; | |
2018 | ||
232ea4d6 | 2019 | throttle_vm_writeout(sc->gfp_mask); |
1da177e4 LT |
2020 | } |
2021 | ||
f16015fb JW |
2022 | static void shrink_zone(int priority, struct zone *zone, |
2023 | struct scan_control *sc) | |
2024 | { | |
5660048c JW |
2025 | struct mem_cgroup *root = sc->target_mem_cgroup; |
2026 | struct mem_cgroup_reclaim_cookie reclaim = { | |
f16015fb | 2027 | .zone = zone, |
5660048c | 2028 | .priority = priority, |
f16015fb | 2029 | }; |
5660048c JW |
2030 | struct mem_cgroup *memcg; |
2031 | ||
5660048c JW |
2032 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2033 | do { | |
2034 | struct mem_cgroup_zone mz = { | |
2035 | .mem_cgroup = memcg, | |
2036 | .zone = zone, | |
2037 | }; | |
f16015fb | 2038 | |
5660048c JW |
2039 | shrink_mem_cgroup_zone(priority, &mz, sc); |
2040 | /* | |
2041 | * Limit reclaim has historically picked one memcg and | |
2042 | * scanned it with decreasing priority levels until | |
2043 | * nr_to_reclaim had been reclaimed. This priority | |
2044 | * cycle is thus over after a single memcg. | |
b95a2f2d JW |
2045 | * |
2046 | * Direct reclaim and kswapd, on the other hand, have | |
2047 | * to scan all memory cgroups to fulfill the overall | |
2048 | * scan target for the zone. | |
5660048c JW |
2049 | */ |
2050 | if (!global_reclaim(sc)) { | |
2051 | mem_cgroup_iter_break(root, memcg); | |
2052 | break; | |
2053 | } | |
2054 | memcg = mem_cgroup_iter(root, memcg, &reclaim); | |
2055 | } while (memcg); | |
f16015fb JW |
2056 | } |
2057 | ||
fe4b1b24 MG |
2058 | /* Returns true if compaction should go ahead for a high-order request */ |
2059 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
2060 | { | |
2061 | unsigned long balance_gap, watermark; | |
2062 | bool watermark_ok; | |
2063 | ||
2064 | /* Do not consider compaction for orders reclaim is meant to satisfy */ | |
2065 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
2066 | return false; | |
2067 | ||
2068 | /* | |
2069 | * Compaction takes time to run and there are potentially other | |
2070 | * callers using the pages just freed. Continue reclaiming until | |
2071 | * there is a buffer of free pages available to give compaction | |
2072 | * a reasonable chance of completing and allocating the page | |
2073 | */ | |
2074 | balance_gap = min(low_wmark_pages(zone), | |
2075 | (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2076 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
2077 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order); | |
2078 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); | |
2079 | ||
2080 | /* | |
2081 | * If compaction is deferred, reclaim up to a point where | |
2082 | * compaction will have a chance of success when re-enabled | |
2083 | */ | |
aff62249 | 2084 | if (compaction_deferred(zone, sc->order)) |
fe4b1b24 MG |
2085 | return watermark_ok; |
2086 | ||
2087 | /* If compaction is not ready to start, keep reclaiming */ | |
2088 | if (!compaction_suitable(zone, sc->order)) | |
2089 | return false; | |
2090 | ||
2091 | return watermark_ok; | |
2092 | } | |
2093 | ||
1da177e4 LT |
2094 | /* |
2095 | * This is the direct reclaim path, for page-allocating processes. We only | |
2096 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2097 | * request. | |
2098 | * | |
41858966 MG |
2099 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
2100 | * Because: | |
1da177e4 LT |
2101 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
2102 | * allocation or | |
41858966 MG |
2103 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
2104 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2105 | * zone defense algorithm. | |
1da177e4 | 2106 | * |
1da177e4 LT |
2107 | * If a zone is deemed to be full of pinned pages then just give it a light |
2108 | * scan then give up on it. | |
e0c23279 MG |
2109 | * |
2110 | * This function returns true if a zone is being reclaimed for a costly | |
fe4b1b24 | 2111 | * high-order allocation and compaction is ready to begin. This indicates to |
0cee34fd MG |
2112 | * the caller that it should consider retrying the allocation instead of |
2113 | * further reclaim. | |
1da177e4 | 2114 | */ |
e0c23279 | 2115 | static bool shrink_zones(int priority, struct zonelist *zonelist, |
05ff5137 | 2116 | struct scan_control *sc) |
1da177e4 | 2117 | { |
dd1a239f | 2118 | struct zoneref *z; |
54a6eb5c | 2119 | struct zone *zone; |
d149e3b2 YH |
2120 | unsigned long nr_soft_reclaimed; |
2121 | unsigned long nr_soft_scanned; | |
0cee34fd | 2122 | bool aborted_reclaim = false; |
1cfb419b | 2123 | |
cc715d99 MG |
2124 | /* |
2125 | * If the number of buffer_heads in the machine exceeds the maximum | |
2126 | * allowed level, force direct reclaim to scan the highmem zone as | |
2127 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2128 | */ | |
2129 | if (buffer_heads_over_limit) | |
2130 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2131 | ||
d4debc66 MG |
2132 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
2133 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
f3fe6512 | 2134 | if (!populated_zone(zone)) |
1da177e4 | 2135 | continue; |
1cfb419b KH |
2136 | /* |
2137 | * Take care memory controller reclaiming has small influence | |
2138 | * to global LRU. | |
2139 | */ | |
89b5fae5 | 2140 | if (global_reclaim(sc)) { |
1cfb419b KH |
2141 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2142 | continue; | |
93e4a89a | 2143 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
1cfb419b | 2144 | continue; /* Let kswapd poll it */ |
e0887c19 RR |
2145 | if (COMPACTION_BUILD) { |
2146 | /* | |
e0c23279 MG |
2147 | * If we already have plenty of memory free for |
2148 | * compaction in this zone, don't free any more. | |
2149 | * Even though compaction is invoked for any | |
2150 | * non-zero order, only frequent costly order | |
2151 | * reclamation is disruptive enough to become a | |
c7cfa37b CA |
2152 | * noticeable problem, like transparent huge |
2153 | * page allocations. | |
e0887c19 | 2154 | */ |
fe4b1b24 | 2155 | if (compaction_ready(zone, sc)) { |
0cee34fd | 2156 | aborted_reclaim = true; |
e0887c19 | 2157 | continue; |
e0c23279 | 2158 | } |
e0887c19 | 2159 | } |
ac34a1a3 KH |
2160 | /* |
2161 | * This steals pages from memory cgroups over softlimit | |
2162 | * and returns the number of reclaimed pages and | |
2163 | * scanned pages. This works for global memory pressure | |
2164 | * and balancing, not for a memcg's limit. | |
2165 | */ | |
2166 | nr_soft_scanned = 0; | |
2167 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2168 | sc->order, sc->gfp_mask, | |
2169 | &nr_soft_scanned); | |
2170 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2171 | sc->nr_scanned += nr_soft_scanned; | |
2172 | /* need some check for avoid more shrink_zone() */ | |
1cfb419b | 2173 | } |
408d8544 | 2174 | |
a79311c1 | 2175 | shrink_zone(priority, zone, sc); |
1da177e4 | 2176 | } |
e0c23279 | 2177 | |
0cee34fd | 2178 | return aborted_reclaim; |
d1908362 MK |
2179 | } |
2180 | ||
2181 | static bool zone_reclaimable(struct zone *zone) | |
2182 | { | |
2183 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
2184 | } | |
2185 | ||
929bea7c | 2186 | /* All zones in zonelist are unreclaimable? */ |
d1908362 MK |
2187 | static bool all_unreclaimable(struct zonelist *zonelist, |
2188 | struct scan_control *sc) | |
2189 | { | |
2190 | struct zoneref *z; | |
2191 | struct zone *zone; | |
d1908362 MK |
2192 | |
2193 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2194 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2195 | if (!populated_zone(zone)) | |
2196 | continue; | |
2197 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2198 | continue; | |
929bea7c KM |
2199 | if (!zone->all_unreclaimable) |
2200 | return false; | |
d1908362 MK |
2201 | } |
2202 | ||
929bea7c | 2203 | return true; |
1da177e4 | 2204 | } |
4f98a2fe | 2205 | |
1da177e4 LT |
2206 | /* |
2207 | * This is the main entry point to direct page reclaim. | |
2208 | * | |
2209 | * If a full scan of the inactive list fails to free enough memory then we | |
2210 | * are "out of memory" and something needs to be killed. | |
2211 | * | |
2212 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2213 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2214 | * caller can't do much about. We kick the writeback threads and take explicit |
2215 | * naps in the hope that some of these pages can be written. But if the | |
2216 | * allocating task holds filesystem locks which prevent writeout this might not | |
2217 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2218 | * |
2219 | * returns: 0, if no pages reclaimed | |
2220 | * else, the number of pages reclaimed | |
1da177e4 | 2221 | */ |
dac1d27b | 2222 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
a09ed5e0 YH |
2223 | struct scan_control *sc, |
2224 | struct shrink_control *shrink) | |
1da177e4 LT |
2225 | { |
2226 | int priority; | |
69e05944 | 2227 | unsigned long total_scanned = 0; |
1da177e4 | 2228 | struct reclaim_state *reclaim_state = current->reclaim_state; |
dd1a239f | 2229 | struct zoneref *z; |
54a6eb5c | 2230 | struct zone *zone; |
22fba335 | 2231 | unsigned long writeback_threshold; |
0cee34fd | 2232 | bool aborted_reclaim; |
1da177e4 | 2233 | |
873b4771 KK |
2234 | delayacct_freepages_start(); |
2235 | ||
89b5fae5 | 2236 | if (global_reclaim(sc)) |
1cfb419b | 2237 | count_vm_event(ALLOCSTALL); |
1da177e4 LT |
2238 | |
2239 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { | |
66e1707b | 2240 | sc->nr_scanned = 0; |
0cee34fd | 2241 | aborted_reclaim = shrink_zones(priority, zonelist, sc); |
e0c23279 | 2242 | |
66e1707b BS |
2243 | /* |
2244 | * Don't shrink slabs when reclaiming memory from | |
2245 | * over limit cgroups | |
2246 | */ | |
89b5fae5 | 2247 | if (global_reclaim(sc)) { |
c6a8a8c5 | 2248 | unsigned long lru_pages = 0; |
d4debc66 MG |
2249 | for_each_zone_zonelist(zone, z, zonelist, |
2250 | gfp_zone(sc->gfp_mask)) { | |
c6a8a8c5 KM |
2251 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2252 | continue; | |
2253 | ||
2254 | lru_pages += zone_reclaimable_pages(zone); | |
2255 | } | |
2256 | ||
1495f230 | 2257 | shrink_slab(shrink, sc->nr_scanned, lru_pages); |
91a45470 | 2258 | if (reclaim_state) { |
a79311c1 | 2259 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
91a45470 KH |
2260 | reclaim_state->reclaimed_slab = 0; |
2261 | } | |
1da177e4 | 2262 | } |
66e1707b | 2263 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2264 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
1da177e4 | 2265 | goto out; |
1da177e4 LT |
2266 | |
2267 | /* | |
2268 | * Try to write back as many pages as we just scanned. This | |
2269 | * tends to cause slow streaming writers to write data to the | |
2270 | * disk smoothly, at the dirtying rate, which is nice. But | |
2271 | * that's undesirable in laptop mode, where we *want* lumpy | |
2272 | * writeout. So in laptop mode, write out the whole world. | |
2273 | */ | |
22fba335 KM |
2274 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2275 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2276 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2277 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2278 | sc->may_writepage = 1; |
1da177e4 LT |
2279 | } |
2280 | ||
2281 | /* Take a nap, wait for some writeback to complete */ | |
7b51755c | 2282 | if (!sc->hibernation_mode && sc->nr_scanned && |
0e093d99 MG |
2283 | priority < DEF_PRIORITY - 2) { |
2284 | struct zone *preferred_zone; | |
2285 | ||
2286 | first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), | |
f33261d7 DR |
2287 | &cpuset_current_mems_allowed, |
2288 | &preferred_zone); | |
0e093d99 MG |
2289 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); |
2290 | } | |
1da177e4 | 2291 | } |
bb21c7ce | 2292 | |
1da177e4 | 2293 | out: |
873b4771 KK |
2294 | delayacct_freepages_end(); |
2295 | ||
bb21c7ce KM |
2296 | if (sc->nr_reclaimed) |
2297 | return sc->nr_reclaimed; | |
2298 | ||
929bea7c KM |
2299 | /* |
2300 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2301 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2302 | * check. | |
2303 | */ | |
2304 | if (oom_killer_disabled) | |
2305 | return 0; | |
2306 | ||
0cee34fd MG |
2307 | /* Aborted reclaim to try compaction? don't OOM, then */ |
2308 | if (aborted_reclaim) | |
7335084d MG |
2309 | return 1; |
2310 | ||
bb21c7ce | 2311 | /* top priority shrink_zones still had more to do? don't OOM, then */ |
89b5fae5 | 2312 | if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc)) |
bb21c7ce KM |
2313 | return 1; |
2314 | ||
2315 | return 0; | |
1da177e4 LT |
2316 | } |
2317 | ||
dac1d27b | 2318 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2319 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2320 | { |
33906bc5 | 2321 | unsigned long nr_reclaimed; |
66e1707b BS |
2322 | struct scan_control sc = { |
2323 | .gfp_mask = gfp_mask, | |
2324 | .may_writepage = !laptop_mode, | |
22fba335 | 2325 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
a6dc60f8 | 2326 | .may_unmap = 1, |
2e2e4259 | 2327 | .may_swap = 1, |
66e1707b | 2328 | .order = order, |
f16015fb | 2329 | .target_mem_cgroup = NULL, |
327c0e96 | 2330 | .nodemask = nodemask, |
66e1707b | 2331 | }; |
a09ed5e0 YH |
2332 | struct shrink_control shrink = { |
2333 | .gfp_mask = sc.gfp_mask, | |
2334 | }; | |
66e1707b | 2335 | |
33906bc5 MG |
2336 | trace_mm_vmscan_direct_reclaim_begin(order, |
2337 | sc.may_writepage, | |
2338 | gfp_mask); | |
2339 | ||
a09ed5e0 | 2340 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
33906bc5 MG |
2341 | |
2342 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2343 | ||
2344 | return nr_reclaimed; | |
66e1707b BS |
2345 | } |
2346 | ||
00f0b825 | 2347 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
66e1707b | 2348 | |
72835c86 | 2349 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, |
4e416953 | 2350 | gfp_t gfp_mask, bool noswap, |
0ae5e89c YH |
2351 | struct zone *zone, |
2352 | unsigned long *nr_scanned) | |
4e416953 BS |
2353 | { |
2354 | struct scan_control sc = { | |
0ae5e89c | 2355 | .nr_scanned = 0, |
b8f5c566 | 2356 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
4e416953 BS |
2357 | .may_writepage = !laptop_mode, |
2358 | .may_unmap = 1, | |
2359 | .may_swap = !noswap, | |
4e416953 | 2360 | .order = 0, |
72835c86 | 2361 | .target_mem_cgroup = memcg, |
4e416953 | 2362 | }; |
5660048c | 2363 | struct mem_cgroup_zone mz = { |
72835c86 | 2364 | .mem_cgroup = memcg, |
5660048c JW |
2365 | .zone = zone, |
2366 | }; | |
0ae5e89c | 2367 | |
4e416953 BS |
2368 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2369 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e KM |
2370 | |
2371 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(0, | |
2372 | sc.may_writepage, | |
2373 | sc.gfp_mask); | |
2374 | ||
4e416953 BS |
2375 | /* |
2376 | * NOTE: Although we can get the priority field, using it | |
2377 | * here is not a good idea, since it limits the pages we can scan. | |
2378 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2379 | * will pick up pages from other mem cgroup's as well. We hack | |
2380 | * the priority and make it zero. | |
2381 | */ | |
5660048c | 2382 | shrink_mem_cgroup_zone(0, &mz, &sc); |
bdce6d9e KM |
2383 | |
2384 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2385 | ||
0ae5e89c | 2386 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2387 | return sc.nr_reclaimed; |
2388 | } | |
2389 | ||
72835c86 | 2390 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
a7885eb8 | 2391 | gfp_t gfp_mask, |
185efc0f | 2392 | bool noswap) |
66e1707b | 2393 | { |
4e416953 | 2394 | struct zonelist *zonelist; |
bdce6d9e | 2395 | unsigned long nr_reclaimed; |
889976db | 2396 | int nid; |
66e1707b | 2397 | struct scan_control sc = { |
66e1707b | 2398 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2399 | .may_unmap = 1, |
2e2e4259 | 2400 | .may_swap = !noswap, |
22fba335 | 2401 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
66e1707b | 2402 | .order = 0, |
72835c86 | 2403 | .target_mem_cgroup = memcg, |
327c0e96 | 2404 | .nodemask = NULL, /* we don't care the placement */ |
a09ed5e0 YH |
2405 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2406 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2407 | }; | |
2408 | struct shrink_control shrink = { | |
2409 | .gfp_mask = sc.gfp_mask, | |
66e1707b | 2410 | }; |
66e1707b | 2411 | |
889976db YH |
2412 | /* |
2413 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2414 | * take care of from where we get pages. So the node where we start the | |
2415 | * scan does not need to be the current node. | |
2416 | */ | |
72835c86 | 2417 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2418 | |
2419 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2420 | |
2421 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2422 | sc.may_writepage, | |
2423 | sc.gfp_mask); | |
2424 | ||
a09ed5e0 | 2425 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
bdce6d9e KM |
2426 | |
2427 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2428 | ||
2429 | return nr_reclaimed; | |
66e1707b BS |
2430 | } |
2431 | #endif | |
2432 | ||
f16015fb JW |
2433 | static void age_active_anon(struct zone *zone, struct scan_control *sc, |
2434 | int priority) | |
2435 | { | |
b95a2f2d | 2436 | struct mem_cgroup *memcg; |
f16015fb | 2437 | |
b95a2f2d JW |
2438 | if (!total_swap_pages) |
2439 | return; | |
2440 | ||
2441 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2442 | do { | |
2443 | struct mem_cgroup_zone mz = { | |
2444 | .mem_cgroup = memcg, | |
2445 | .zone = zone, | |
2446 | }; | |
2447 | ||
2448 | if (inactive_anon_is_low(&mz)) | |
2449 | shrink_active_list(SWAP_CLUSTER_MAX, &mz, | |
2450 | sc, priority, 0); | |
2451 | ||
2452 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2453 | } while (memcg); | |
f16015fb JW |
2454 | } |
2455 | ||
1741c877 MG |
2456 | /* |
2457 | * pgdat_balanced is used when checking if a node is balanced for high-order | |
2458 | * allocations. Only zones that meet watermarks and are in a zone allowed | |
2459 | * by the callers classzone_idx are added to balanced_pages. The total of | |
2460 | * balanced pages must be at least 25% of the zones allowed by classzone_idx | |
2461 | * for the node to be considered balanced. Forcing all zones to be balanced | |
2462 | * for high orders can cause excessive reclaim when there are imbalanced zones. | |
2463 | * The choice of 25% is due to | |
2464 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2465 | * reasonable sized machine | |
2466 | * o On all other machines, the top zone must be at least a reasonable | |
25985edc | 2467 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
1741c877 MG |
2468 | * would need to be at least 256M for it to be balance a whole node. |
2469 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2470 | * to balance a node on its own. These seemed like reasonable ratios. | |
2471 | */ | |
2472 | static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages, | |
2473 | int classzone_idx) | |
2474 | { | |
2475 | unsigned long present_pages = 0; | |
2476 | int i; | |
2477 | ||
2478 | for (i = 0; i <= classzone_idx; i++) | |
2479 | present_pages += pgdat->node_zones[i].present_pages; | |
2480 | ||
4746efde SL |
2481 | /* A special case here: if zone has no page, we think it's balanced */ |
2482 | return balanced_pages >= (present_pages >> 2); | |
1741c877 MG |
2483 | } |
2484 | ||
f50de2d3 | 2485 | /* is kswapd sleeping prematurely? */ |
dc83edd9 MG |
2486 | static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining, |
2487 | int classzone_idx) | |
f50de2d3 | 2488 | { |
bb3ab596 | 2489 | int i; |
1741c877 MG |
2490 | unsigned long balanced = 0; |
2491 | bool all_zones_ok = true; | |
f50de2d3 MG |
2492 | |
2493 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ | |
2494 | if (remaining) | |
dc83edd9 | 2495 | return true; |
f50de2d3 | 2496 | |
0abdee2b | 2497 | /* Check the watermark levels */ |
08951e54 | 2498 | for (i = 0; i <= classzone_idx; i++) { |
bb3ab596 KM |
2499 | struct zone *zone = pgdat->node_zones + i; |
2500 | ||
2501 | if (!populated_zone(zone)) | |
2502 | continue; | |
2503 | ||
355b09c4 MG |
2504 | /* |
2505 | * balance_pgdat() skips over all_unreclaimable after | |
2506 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2507 | * they must be considered balanced here as well if kswapd | |
2508 | * is to sleep | |
2509 | */ | |
2510 | if (zone->all_unreclaimable) { | |
2511 | balanced += zone->present_pages; | |
de3fab39 | 2512 | continue; |
355b09c4 | 2513 | } |
de3fab39 | 2514 | |
88f5acf8 | 2515 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone), |
da175d06 | 2516 | i, 0)) |
1741c877 MG |
2517 | all_zones_ok = false; |
2518 | else | |
2519 | balanced += zone->present_pages; | |
bb3ab596 | 2520 | } |
f50de2d3 | 2521 | |
1741c877 MG |
2522 | /* |
2523 | * For high-order requests, the balanced zones must contain at least | |
2524 | * 25% of the nodes pages for kswapd to sleep. For order-0, all zones | |
2525 | * must be balanced | |
2526 | */ | |
2527 | if (order) | |
afc7e326 | 2528 | return !pgdat_balanced(pgdat, balanced, classzone_idx); |
1741c877 MG |
2529 | else |
2530 | return !all_zones_ok; | |
f50de2d3 MG |
2531 | } |
2532 | ||
1da177e4 LT |
2533 | /* |
2534 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 2535 | * they are all at high_wmark_pages(zone). |
1da177e4 | 2536 | * |
0abdee2b | 2537 | * Returns the final order kswapd was reclaiming at |
1da177e4 LT |
2538 | * |
2539 | * There is special handling here for zones which are full of pinned pages. | |
2540 | * This can happen if the pages are all mlocked, or if they are all used by | |
2541 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
2542 | * What we do is to detect the case where all pages in the zone have been | |
2543 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
2544 | * dead and from now on, only perform a short scan. Basically we're polling | |
2545 | * the zone for when the problem goes away. | |
2546 | * | |
2547 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
2548 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
2549 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
2550 | * lower zones regardless of the number of free pages in the lower zones. This | |
2551 | * interoperates with the page allocator fallback scheme to ensure that aging | |
2552 | * of pages is balanced across the zones. | |
1da177e4 | 2553 | */ |
99504748 | 2554 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, |
dc83edd9 | 2555 | int *classzone_idx) |
1da177e4 | 2556 | { |
1da177e4 | 2557 | int all_zones_ok; |
1741c877 | 2558 | unsigned long balanced; |
1da177e4 LT |
2559 | int priority; |
2560 | int i; | |
99504748 | 2561 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
69e05944 | 2562 | unsigned long total_scanned; |
1da177e4 | 2563 | struct reclaim_state *reclaim_state = current->reclaim_state; |
0ae5e89c YH |
2564 | unsigned long nr_soft_reclaimed; |
2565 | unsigned long nr_soft_scanned; | |
179e9639 AM |
2566 | struct scan_control sc = { |
2567 | .gfp_mask = GFP_KERNEL, | |
a6dc60f8 | 2568 | .may_unmap = 1, |
2e2e4259 | 2569 | .may_swap = 1, |
22fba335 KM |
2570 | /* |
2571 | * kswapd doesn't want to be bailed out while reclaim. because | |
2572 | * we want to put equal scanning pressure on each zone. | |
2573 | */ | |
2574 | .nr_to_reclaim = ULONG_MAX, | |
5ad333eb | 2575 | .order = order, |
f16015fb | 2576 | .target_mem_cgroup = NULL, |
179e9639 | 2577 | }; |
a09ed5e0 YH |
2578 | struct shrink_control shrink = { |
2579 | .gfp_mask = sc.gfp_mask, | |
2580 | }; | |
1da177e4 LT |
2581 | loop_again: |
2582 | total_scanned = 0; | |
a79311c1 | 2583 | sc.nr_reclaimed = 0; |
c0bbbc73 | 2584 | sc.may_writepage = !laptop_mode; |
f8891e5e | 2585 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 2586 | |
1da177e4 | 2587 | for (priority = DEF_PRIORITY; priority >= 0; priority--) { |
1da177e4 | 2588 | unsigned long lru_pages = 0; |
bb3ab596 | 2589 | int has_under_min_watermark_zone = 0; |
1da177e4 LT |
2590 | |
2591 | all_zones_ok = 1; | |
1741c877 | 2592 | balanced = 0; |
1da177e4 | 2593 | |
d6277db4 RW |
2594 | /* |
2595 | * Scan in the highmem->dma direction for the highest | |
2596 | * zone which needs scanning | |
2597 | */ | |
2598 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
2599 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 2600 | |
d6277db4 RW |
2601 | if (!populated_zone(zone)) |
2602 | continue; | |
1da177e4 | 2603 | |
93e4a89a | 2604 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
d6277db4 | 2605 | continue; |
1da177e4 | 2606 | |
556adecb RR |
2607 | /* |
2608 | * Do some background aging of the anon list, to give | |
2609 | * pages a chance to be referenced before reclaiming. | |
2610 | */ | |
f16015fb | 2611 | age_active_anon(zone, &sc, priority); |
556adecb | 2612 | |
cc715d99 MG |
2613 | /* |
2614 | * If the number of buffer_heads in the machine | |
2615 | * exceeds the maximum allowed level and this node | |
2616 | * has a highmem zone, force kswapd to reclaim from | |
2617 | * it to relieve lowmem pressure. | |
2618 | */ | |
2619 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
2620 | end_zone = i; | |
2621 | break; | |
2622 | } | |
2623 | ||
88f5acf8 | 2624 | if (!zone_watermark_ok_safe(zone, order, |
41858966 | 2625 | high_wmark_pages(zone), 0, 0)) { |
d6277db4 | 2626 | end_zone = i; |
e1dbeda6 | 2627 | break; |
439423f6 SL |
2628 | } else { |
2629 | /* If balanced, clear the congested flag */ | |
2630 | zone_clear_flag(zone, ZONE_CONGESTED); | |
1da177e4 | 2631 | } |
1da177e4 | 2632 | } |
e1dbeda6 AM |
2633 | if (i < 0) |
2634 | goto out; | |
2635 | ||
1da177e4 LT |
2636 | for (i = 0; i <= end_zone; i++) { |
2637 | struct zone *zone = pgdat->node_zones + i; | |
2638 | ||
adea02a1 | 2639 | lru_pages += zone_reclaimable_pages(zone); |
1da177e4 LT |
2640 | } |
2641 | ||
2642 | /* | |
2643 | * Now scan the zone in the dma->highmem direction, stopping | |
2644 | * at the last zone which needs scanning. | |
2645 | * | |
2646 | * We do this because the page allocator works in the opposite | |
2647 | * direction. This prevents the page allocator from allocating | |
2648 | * pages behind kswapd's direction of progress, which would | |
2649 | * cause too much scanning of the lower zones. | |
2650 | */ | |
2651 | for (i = 0; i <= end_zone; i++) { | |
2652 | struct zone *zone = pgdat->node_zones + i; | |
fe2c2a10 | 2653 | int nr_slab, testorder; |
8afdcece | 2654 | unsigned long balance_gap; |
1da177e4 | 2655 | |
f3fe6512 | 2656 | if (!populated_zone(zone)) |
1da177e4 LT |
2657 | continue; |
2658 | ||
93e4a89a | 2659 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) |
1da177e4 LT |
2660 | continue; |
2661 | ||
1da177e4 | 2662 | sc.nr_scanned = 0; |
4e416953 | 2663 | |
0ae5e89c | 2664 | nr_soft_scanned = 0; |
4e416953 BS |
2665 | /* |
2666 | * Call soft limit reclaim before calling shrink_zone. | |
4e416953 | 2667 | */ |
0ae5e89c YH |
2668 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, |
2669 | order, sc.gfp_mask, | |
2670 | &nr_soft_scanned); | |
2671 | sc.nr_reclaimed += nr_soft_reclaimed; | |
2672 | total_scanned += nr_soft_scanned; | |
00918b6a | 2673 | |
32a4330d | 2674 | /* |
8afdcece MG |
2675 | * We put equal pressure on every zone, unless |
2676 | * one zone has way too many pages free | |
2677 | * already. The "too many pages" is defined | |
2678 | * as the high wmark plus a "gap" where the | |
2679 | * gap is either the low watermark or 1% | |
2680 | * of the zone, whichever is smaller. | |
32a4330d | 2681 | */ |
8afdcece MG |
2682 | balance_gap = min(low_wmark_pages(zone), |
2683 | (zone->present_pages + | |
2684 | KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / | |
2685 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
fe2c2a10 RR |
2686 | /* |
2687 | * Kswapd reclaims only single pages with compaction | |
2688 | * enabled. Trying too hard to reclaim until contiguous | |
2689 | * free pages have become available can hurt performance | |
2690 | * by evicting too much useful data from memory. | |
2691 | * Do not reclaim more than needed for compaction. | |
2692 | */ | |
2693 | testorder = order; | |
2694 | if (COMPACTION_BUILD && order && | |
2695 | compaction_suitable(zone, order) != | |
2696 | COMPACT_SKIPPED) | |
2697 | testorder = 0; | |
2698 | ||
cc715d99 | 2699 | if ((buffer_heads_over_limit && is_highmem_idx(i)) || |
643ac9fc | 2700 | !zone_watermark_ok_safe(zone, testorder, |
8afdcece | 2701 | high_wmark_pages(zone) + balance_gap, |
d7868dae | 2702 | end_zone, 0)) { |
a79311c1 | 2703 | shrink_zone(priority, zone, &sc); |
5a03b051 | 2704 | |
d7868dae MG |
2705 | reclaim_state->reclaimed_slab = 0; |
2706 | nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages); | |
2707 | sc.nr_reclaimed += reclaim_state->reclaimed_slab; | |
2708 | total_scanned += sc.nr_scanned; | |
2709 | ||
2710 | if (nr_slab == 0 && !zone_reclaimable(zone)) | |
2711 | zone->all_unreclaimable = 1; | |
2712 | } | |
2713 | ||
1da177e4 LT |
2714 | /* |
2715 | * If we've done a decent amount of scanning and | |
2716 | * the reclaim ratio is low, start doing writepage | |
2717 | * even in laptop mode | |
2718 | */ | |
2719 | if (total_scanned > SWAP_CLUSTER_MAX * 2 && | |
a79311c1 | 2720 | total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) |
1da177e4 | 2721 | sc.may_writepage = 1; |
bb3ab596 | 2722 | |
215ddd66 MG |
2723 | if (zone->all_unreclaimable) { |
2724 | if (end_zone && end_zone == i) | |
2725 | end_zone--; | |
d7868dae | 2726 | continue; |
215ddd66 | 2727 | } |
d7868dae | 2728 | |
fe2c2a10 | 2729 | if (!zone_watermark_ok_safe(zone, testorder, |
45973d74 MK |
2730 | high_wmark_pages(zone), end_zone, 0)) { |
2731 | all_zones_ok = 0; | |
2732 | /* | |
2733 | * We are still under min water mark. This | |
2734 | * means that we have a GFP_ATOMIC allocation | |
2735 | * failure risk. Hurry up! | |
2736 | */ | |
88f5acf8 | 2737 | if (!zone_watermark_ok_safe(zone, order, |
45973d74 MK |
2738 | min_wmark_pages(zone), end_zone, 0)) |
2739 | has_under_min_watermark_zone = 1; | |
0e093d99 MG |
2740 | } else { |
2741 | /* | |
2742 | * If a zone reaches its high watermark, | |
2743 | * consider it to be no longer congested. It's | |
2744 | * possible there are dirty pages backed by | |
2745 | * congested BDIs but as pressure is relieved, | |
2746 | * spectulatively avoid congestion waits | |
2747 | */ | |
2748 | zone_clear_flag(zone, ZONE_CONGESTED); | |
dc83edd9 | 2749 | if (i <= *classzone_idx) |
1741c877 | 2750 | balanced += zone->present_pages; |
45973d74 | 2751 | } |
bb3ab596 | 2752 | |
1da177e4 | 2753 | } |
dc83edd9 | 2754 | if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx))) |
1da177e4 LT |
2755 | break; /* kswapd: all done */ |
2756 | /* | |
2757 | * OK, kswapd is getting into trouble. Take a nap, then take | |
2758 | * another pass across the zones. | |
2759 | */ | |
bb3ab596 KM |
2760 | if (total_scanned && (priority < DEF_PRIORITY - 2)) { |
2761 | if (has_under_min_watermark_zone) | |
2762 | count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT); | |
2763 | else | |
2764 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2765 | } | |
1da177e4 LT |
2766 | |
2767 | /* | |
2768 | * We do this so kswapd doesn't build up large priorities for | |
2769 | * example when it is freeing in parallel with allocators. It | |
2770 | * matches the direct reclaim path behaviour in terms of impact | |
2771 | * on zone->*_priority. | |
2772 | */ | |
a79311c1 | 2773 | if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) |
1da177e4 LT |
2774 | break; |
2775 | } | |
2776 | out: | |
99504748 MG |
2777 | |
2778 | /* | |
2779 | * order-0: All zones must meet high watermark for a balanced node | |
1741c877 MG |
2780 | * high-order: Balanced zones must make up at least 25% of the node |
2781 | * for the node to be balanced | |
99504748 | 2782 | */ |
dc83edd9 | 2783 | if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) { |
1da177e4 | 2784 | cond_resched(); |
8357376d RW |
2785 | |
2786 | try_to_freeze(); | |
2787 | ||
73ce02e9 KM |
2788 | /* |
2789 | * Fragmentation may mean that the system cannot be | |
2790 | * rebalanced for high-order allocations in all zones. | |
2791 | * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, | |
2792 | * it means the zones have been fully scanned and are still | |
2793 | * not balanced. For high-order allocations, there is | |
2794 | * little point trying all over again as kswapd may | |
2795 | * infinite loop. | |
2796 | * | |
2797 | * Instead, recheck all watermarks at order-0 as they | |
2798 | * are the most important. If watermarks are ok, kswapd will go | |
2799 | * back to sleep. High-order users can still perform direct | |
2800 | * reclaim if they wish. | |
2801 | */ | |
2802 | if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) | |
2803 | order = sc.order = 0; | |
2804 | ||
1da177e4 LT |
2805 | goto loop_again; |
2806 | } | |
2807 | ||
99504748 MG |
2808 | /* |
2809 | * If kswapd was reclaiming at a higher order, it has the option of | |
2810 | * sleeping without all zones being balanced. Before it does, it must | |
2811 | * ensure that the watermarks for order-0 on *all* zones are met and | |
2812 | * that the congestion flags are cleared. The congestion flag must | |
2813 | * be cleared as kswapd is the only mechanism that clears the flag | |
2814 | * and it is potentially going to sleep here. | |
2815 | */ | |
2816 | if (order) { | |
7be62de9 RR |
2817 | int zones_need_compaction = 1; |
2818 | ||
99504748 MG |
2819 | for (i = 0; i <= end_zone; i++) { |
2820 | struct zone *zone = pgdat->node_zones + i; | |
2821 | ||
2822 | if (!populated_zone(zone)) | |
2823 | continue; | |
2824 | ||
2825 | if (zone->all_unreclaimable && priority != DEF_PRIORITY) | |
2826 | continue; | |
2827 | ||
fe2c2a10 | 2828 | /* Would compaction fail due to lack of free memory? */ |
496b919b RR |
2829 | if (COMPACTION_BUILD && |
2830 | compaction_suitable(zone, order) == COMPACT_SKIPPED) | |
fe2c2a10 RR |
2831 | goto loop_again; |
2832 | ||
99504748 MG |
2833 | /* Confirm the zone is balanced for order-0 */ |
2834 | if (!zone_watermark_ok(zone, 0, | |
2835 | high_wmark_pages(zone), 0, 0)) { | |
2836 | order = sc.order = 0; | |
2837 | goto loop_again; | |
2838 | } | |
2839 | ||
7be62de9 RR |
2840 | /* Check if the memory needs to be defragmented. */ |
2841 | if (zone_watermark_ok(zone, order, | |
2842 | low_wmark_pages(zone), *classzone_idx, 0)) | |
2843 | zones_need_compaction = 0; | |
2844 | ||
99504748 MG |
2845 | /* If balanced, clear the congested flag */ |
2846 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2847 | } | |
7be62de9 RR |
2848 | |
2849 | if (zones_need_compaction) | |
2850 | compact_pgdat(pgdat, order); | |
99504748 MG |
2851 | } |
2852 | ||
0abdee2b MG |
2853 | /* |
2854 | * Return the order we were reclaiming at so sleeping_prematurely() | |
2855 | * makes a decision on the order we were last reclaiming at. However, | |
2856 | * if another caller entered the allocator slow path while kswapd | |
2857 | * was awake, order will remain at the higher level | |
2858 | */ | |
dc83edd9 | 2859 | *classzone_idx = end_zone; |
0abdee2b | 2860 | return order; |
1da177e4 LT |
2861 | } |
2862 | ||
dc83edd9 | 2863 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f0bc0a60 KM |
2864 | { |
2865 | long remaining = 0; | |
2866 | DEFINE_WAIT(wait); | |
2867 | ||
2868 | if (freezing(current) || kthread_should_stop()) | |
2869 | return; | |
2870 | ||
2871 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2872 | ||
2873 | /* Try to sleep for a short interval */ | |
dc83edd9 | 2874 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
2875 | remaining = schedule_timeout(HZ/10); |
2876 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2877 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
2878 | } | |
2879 | ||
2880 | /* | |
2881 | * After a short sleep, check if it was a premature sleep. If not, then | |
2882 | * go fully to sleep until explicitly woken up. | |
2883 | */ | |
dc83edd9 | 2884 | if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
2885 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
2886 | ||
2887 | /* | |
2888 | * vmstat counters are not perfectly accurate and the estimated | |
2889 | * value for counters such as NR_FREE_PAGES can deviate from the | |
2890 | * true value by nr_online_cpus * threshold. To avoid the zone | |
2891 | * watermarks being breached while under pressure, we reduce the | |
2892 | * per-cpu vmstat threshold while kswapd is awake and restore | |
2893 | * them before going back to sleep. | |
2894 | */ | |
2895 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
2896 | schedule(); | |
2897 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); | |
2898 | } else { | |
2899 | if (remaining) | |
2900 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
2901 | else | |
2902 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
2903 | } | |
2904 | finish_wait(&pgdat->kswapd_wait, &wait); | |
2905 | } | |
2906 | ||
1da177e4 LT |
2907 | /* |
2908 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 2909 | * from the init process. |
1da177e4 LT |
2910 | * |
2911 | * This basically trickles out pages so that we have _some_ | |
2912 | * free memory available even if there is no other activity | |
2913 | * that frees anything up. This is needed for things like routing | |
2914 | * etc, where we otherwise might have all activity going on in | |
2915 | * asynchronous contexts that cannot page things out. | |
2916 | * | |
2917 | * If there are applications that are active memory-allocators | |
2918 | * (most normal use), this basically shouldn't matter. | |
2919 | */ | |
2920 | static int kswapd(void *p) | |
2921 | { | |
215ddd66 | 2922 | unsigned long order, new_order; |
d2ebd0f6 | 2923 | unsigned balanced_order; |
215ddd66 | 2924 | int classzone_idx, new_classzone_idx; |
d2ebd0f6 | 2925 | int balanced_classzone_idx; |
1da177e4 LT |
2926 | pg_data_t *pgdat = (pg_data_t*)p; |
2927 | struct task_struct *tsk = current; | |
f0bc0a60 | 2928 | |
1da177e4 LT |
2929 | struct reclaim_state reclaim_state = { |
2930 | .reclaimed_slab = 0, | |
2931 | }; | |
a70f7302 | 2932 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 2933 | |
cf40bd16 NP |
2934 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
2935 | ||
174596a0 | 2936 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 2937 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
2938 | current->reclaim_state = &reclaim_state; |
2939 | ||
2940 | /* | |
2941 | * Tell the memory management that we're a "memory allocator", | |
2942 | * and that if we need more memory we should get access to it | |
2943 | * regardless (see "__alloc_pages()"). "kswapd" should | |
2944 | * never get caught in the normal page freeing logic. | |
2945 | * | |
2946 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
2947 | * you need a small amount of memory in order to be able to | |
2948 | * page out something else, and this flag essentially protects | |
2949 | * us from recursively trying to free more memory as we're | |
2950 | * trying to free the first piece of memory in the first place). | |
2951 | */ | |
930d9152 | 2952 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 2953 | set_freezable(); |
1da177e4 | 2954 | |
215ddd66 | 2955 | order = new_order = 0; |
d2ebd0f6 | 2956 | balanced_order = 0; |
215ddd66 | 2957 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; |
d2ebd0f6 | 2958 | balanced_classzone_idx = classzone_idx; |
1da177e4 | 2959 | for ( ; ; ) { |
8fe23e05 | 2960 | int ret; |
3e1d1d28 | 2961 | |
215ddd66 MG |
2962 | /* |
2963 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
2964 | * new request of a similar or harder type will succeed soon | |
2965 | * so consider going to sleep on the basis we reclaimed at | |
2966 | */ | |
d2ebd0f6 AS |
2967 | if (balanced_classzone_idx >= new_classzone_idx && |
2968 | balanced_order == new_order) { | |
215ddd66 MG |
2969 | new_order = pgdat->kswapd_max_order; |
2970 | new_classzone_idx = pgdat->classzone_idx; | |
2971 | pgdat->kswapd_max_order = 0; | |
2972 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
2973 | } | |
2974 | ||
99504748 | 2975 | if (order < new_order || classzone_idx > new_classzone_idx) { |
1da177e4 LT |
2976 | /* |
2977 | * Don't sleep if someone wants a larger 'order' | |
99504748 | 2978 | * allocation or has tigher zone constraints |
1da177e4 LT |
2979 | */ |
2980 | order = new_order; | |
99504748 | 2981 | classzone_idx = new_classzone_idx; |
1da177e4 | 2982 | } else { |
d2ebd0f6 AS |
2983 | kswapd_try_to_sleep(pgdat, balanced_order, |
2984 | balanced_classzone_idx); | |
1da177e4 | 2985 | order = pgdat->kswapd_max_order; |
99504748 | 2986 | classzone_idx = pgdat->classzone_idx; |
f0dfcde0 AS |
2987 | new_order = order; |
2988 | new_classzone_idx = classzone_idx; | |
4d40502e | 2989 | pgdat->kswapd_max_order = 0; |
215ddd66 | 2990 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
1da177e4 | 2991 | } |
1da177e4 | 2992 | |
8fe23e05 DR |
2993 | ret = try_to_freeze(); |
2994 | if (kthread_should_stop()) | |
2995 | break; | |
2996 | ||
2997 | /* | |
2998 | * We can speed up thawing tasks if we don't call balance_pgdat | |
2999 | * after returning from the refrigerator | |
3000 | */ | |
33906bc5 MG |
3001 | if (!ret) { |
3002 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
d2ebd0f6 AS |
3003 | balanced_classzone_idx = classzone_idx; |
3004 | balanced_order = balance_pgdat(pgdat, order, | |
3005 | &balanced_classzone_idx); | |
33906bc5 | 3006 | } |
1da177e4 LT |
3007 | } |
3008 | return 0; | |
3009 | } | |
3010 | ||
3011 | /* | |
3012 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3013 | */ | |
99504748 | 3014 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3015 | { |
3016 | pg_data_t *pgdat; | |
3017 | ||
f3fe6512 | 3018 | if (!populated_zone(zone)) |
1da177e4 LT |
3019 | return; |
3020 | ||
88f5acf8 | 3021 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1da177e4 | 3022 | return; |
88f5acf8 | 3023 | pgdat = zone->zone_pgdat; |
99504748 | 3024 | if (pgdat->kswapd_max_order < order) { |
1da177e4 | 3025 | pgdat->kswapd_max_order = order; |
99504748 MG |
3026 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
3027 | } | |
8d0986e2 | 3028 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3029 | return; |
88f5acf8 MG |
3030 | if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) |
3031 | return; | |
3032 | ||
3033 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3034 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3035 | } |
3036 | ||
adea02a1 WF |
3037 | /* |
3038 | * The reclaimable count would be mostly accurate. | |
3039 | * The less reclaimable pages may be | |
3040 | * - mlocked pages, which will be moved to unevictable list when encountered | |
3041 | * - mapped pages, which may require several travels to be reclaimed | |
3042 | * - dirty pages, which is not "instantly" reclaimable | |
3043 | */ | |
3044 | unsigned long global_reclaimable_pages(void) | |
4f98a2fe | 3045 | { |
adea02a1 WF |
3046 | int nr; |
3047 | ||
3048 | nr = global_page_state(NR_ACTIVE_FILE) + | |
3049 | global_page_state(NR_INACTIVE_FILE); | |
3050 | ||
3051 | if (nr_swap_pages > 0) | |
3052 | nr += global_page_state(NR_ACTIVE_ANON) + | |
3053 | global_page_state(NR_INACTIVE_ANON); | |
3054 | ||
3055 | return nr; | |
3056 | } | |
3057 | ||
3058 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
3059 | { | |
3060 | int nr; | |
3061 | ||
3062 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
3063 | zone_page_state(zone, NR_INACTIVE_FILE); | |
3064 | ||
3065 | if (nr_swap_pages > 0) | |
3066 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
3067 | zone_page_state(zone, NR_INACTIVE_ANON); | |
3068 | ||
3069 | return nr; | |
4f98a2fe RR |
3070 | } |
3071 | ||
c6f37f12 | 3072 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3073 | /* |
7b51755c | 3074 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3075 | * freed pages. |
3076 | * | |
3077 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3078 | * LRU order by reclaiming preferentially | |
3079 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3080 | */ |
7b51755c | 3081 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3082 | { |
d6277db4 | 3083 | struct reclaim_state reclaim_state; |
d6277db4 | 3084 | struct scan_control sc = { |
7b51755c KM |
3085 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
3086 | .may_swap = 1, | |
3087 | .may_unmap = 1, | |
d6277db4 | 3088 | .may_writepage = 1, |
7b51755c KM |
3089 | .nr_to_reclaim = nr_to_reclaim, |
3090 | .hibernation_mode = 1, | |
7b51755c | 3091 | .order = 0, |
1da177e4 | 3092 | }; |
a09ed5e0 YH |
3093 | struct shrink_control shrink = { |
3094 | .gfp_mask = sc.gfp_mask, | |
3095 | }; | |
3096 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
7b51755c KM |
3097 | struct task_struct *p = current; |
3098 | unsigned long nr_reclaimed; | |
1da177e4 | 3099 | |
7b51755c KM |
3100 | p->flags |= PF_MEMALLOC; |
3101 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3102 | reclaim_state.reclaimed_slab = 0; | |
3103 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3104 | |
a09ed5e0 | 3105 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
d979677c | 3106 | |
7b51755c KM |
3107 | p->reclaim_state = NULL; |
3108 | lockdep_clear_current_reclaim_state(); | |
3109 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3110 | |
7b51755c | 3111 | return nr_reclaimed; |
1da177e4 | 3112 | } |
c6f37f12 | 3113 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3114 | |
1da177e4 LT |
3115 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3116 | not required for correctness. So if the last cpu in a node goes | |
3117 | away, we get changed to run anywhere: as the first one comes back, | |
3118 | restore their cpu bindings. */ | |
9c7b216d | 3119 | static int __devinit cpu_callback(struct notifier_block *nfb, |
69e05944 | 3120 | unsigned long action, void *hcpu) |
1da177e4 | 3121 | { |
58c0a4a7 | 3122 | int nid; |
1da177e4 | 3123 | |
8bb78442 | 3124 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
58c0a4a7 | 3125 | for_each_node_state(nid, N_HIGH_MEMORY) { |
c5f59f08 | 3126 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3127 | const struct cpumask *mask; |
3128 | ||
3129 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3130 | |
3e597945 | 3131 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3132 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3133 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3134 | } |
3135 | } | |
3136 | return NOTIFY_OK; | |
3137 | } | |
1da177e4 | 3138 | |
3218ae14 YG |
3139 | /* |
3140 | * This kswapd start function will be called by init and node-hot-add. | |
3141 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3142 | */ | |
3143 | int kswapd_run(int nid) | |
3144 | { | |
3145 | pg_data_t *pgdat = NODE_DATA(nid); | |
3146 | int ret = 0; | |
3147 | ||
3148 | if (pgdat->kswapd) | |
3149 | return 0; | |
3150 | ||
3151 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3152 | if (IS_ERR(pgdat->kswapd)) { | |
3153 | /* failure at boot is fatal */ | |
3154 | BUG_ON(system_state == SYSTEM_BOOTING); | |
3155 | printk("Failed to start kswapd on node %d\n",nid); | |
3156 | ret = -1; | |
3157 | } | |
3158 | return ret; | |
3159 | } | |
3160 | ||
8fe23e05 DR |
3161 | /* |
3162 | * Called by memory hotplug when all memory in a node is offlined. | |
3163 | */ | |
3164 | void kswapd_stop(int nid) | |
3165 | { | |
3166 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3167 | ||
3168 | if (kswapd) | |
3169 | kthread_stop(kswapd); | |
3170 | } | |
3171 | ||
1da177e4 LT |
3172 | static int __init kswapd_init(void) |
3173 | { | |
3218ae14 | 3174 | int nid; |
69e05944 | 3175 | |
1da177e4 | 3176 | swap_setup(); |
9422ffba | 3177 | for_each_node_state(nid, N_HIGH_MEMORY) |
3218ae14 | 3178 | kswapd_run(nid); |
1da177e4 LT |
3179 | hotcpu_notifier(cpu_callback, 0); |
3180 | return 0; | |
3181 | } | |
3182 | ||
3183 | module_init(kswapd_init) | |
9eeff239 CL |
3184 | |
3185 | #ifdef CONFIG_NUMA | |
3186 | /* | |
3187 | * Zone reclaim mode | |
3188 | * | |
3189 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3190 | * the watermarks. | |
9eeff239 CL |
3191 | */ |
3192 | int zone_reclaim_mode __read_mostly; | |
3193 | ||
1b2ffb78 | 3194 | #define RECLAIM_OFF 0 |
7d03431c | 3195 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 CL |
3196 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
3197 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3198 | ||
a92f7126 CL |
3199 | /* |
3200 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3201 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3202 | * a zone. | |
3203 | */ | |
3204 | #define ZONE_RECLAIM_PRIORITY 4 | |
3205 | ||
9614634f CL |
3206 | /* |
3207 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3208 | * occur. | |
3209 | */ | |
3210 | int sysctl_min_unmapped_ratio = 1; | |
3211 | ||
0ff38490 CL |
3212 | /* |
3213 | * If the number of slab pages in a zone grows beyond this percentage then | |
3214 | * slab reclaim needs to occur. | |
3215 | */ | |
3216 | int sysctl_min_slab_ratio = 5; | |
3217 | ||
90afa5de MG |
3218 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
3219 | { | |
3220 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3221 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3222 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3223 | ||
3224 | /* | |
3225 | * It's possible for there to be more file mapped pages than | |
3226 | * accounted for by the pages on the file LRU lists because | |
3227 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3228 | */ | |
3229 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3230 | } | |
3231 | ||
3232 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3233 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3234 | { | |
3235 | long nr_pagecache_reclaimable; | |
3236 | long delta = 0; | |
3237 | ||
3238 | /* | |
3239 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3240 | * potentially reclaimable. Otherwise, we have to worry about | |
3241 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3242 | * a better estimate | |
3243 | */ | |
3244 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3245 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3246 | else | |
3247 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3248 | ||
3249 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3250 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3251 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3252 | ||
3253 | /* Watch for any possible underflows due to delta */ | |
3254 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3255 | delta = nr_pagecache_reclaimable; | |
3256 | ||
3257 | return nr_pagecache_reclaimable - delta; | |
3258 | } | |
3259 | ||
9eeff239 CL |
3260 | /* |
3261 | * Try to free up some pages from this zone through reclaim. | |
3262 | */ | |
179e9639 | 3263 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3264 | { |
7fb2d46d | 3265 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3266 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3267 | struct task_struct *p = current; |
3268 | struct reclaim_state reclaim_state; | |
8695949a | 3269 | int priority; |
179e9639 AM |
3270 | struct scan_control sc = { |
3271 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
a6dc60f8 | 3272 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), |
2e2e4259 | 3273 | .may_swap = 1, |
22fba335 KM |
3274 | .nr_to_reclaim = max_t(unsigned long, nr_pages, |
3275 | SWAP_CLUSTER_MAX), | |
179e9639 | 3276 | .gfp_mask = gfp_mask, |
bd2f6199 | 3277 | .order = order, |
179e9639 | 3278 | }; |
a09ed5e0 YH |
3279 | struct shrink_control shrink = { |
3280 | .gfp_mask = sc.gfp_mask, | |
3281 | }; | |
15748048 | 3282 | unsigned long nr_slab_pages0, nr_slab_pages1; |
9eeff239 | 3283 | |
9eeff239 | 3284 | cond_resched(); |
d4f7796e CL |
3285 | /* |
3286 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3287 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3288 | * and RECLAIM_SWAP. | |
3289 | */ | |
3290 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3291 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3292 | reclaim_state.reclaimed_slab = 0; |
3293 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3294 | |
90afa5de | 3295 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
3296 | /* |
3297 | * Free memory by calling shrink zone with increasing | |
3298 | * priorities until we have enough memory freed. | |
3299 | */ | |
3300 | priority = ZONE_RECLAIM_PRIORITY; | |
3301 | do { | |
a79311c1 | 3302 | shrink_zone(priority, zone, &sc); |
0ff38490 | 3303 | priority--; |
a79311c1 | 3304 | } while (priority >= 0 && sc.nr_reclaimed < nr_pages); |
0ff38490 | 3305 | } |
c84db23c | 3306 | |
15748048 KM |
3307 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3308 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
2a16e3f4 | 3309 | /* |
7fb2d46d | 3310 | * shrink_slab() does not currently allow us to determine how |
0ff38490 CL |
3311 | * many pages were freed in this zone. So we take the current |
3312 | * number of slab pages and shake the slab until it is reduced | |
3313 | * by the same nr_pages that we used for reclaiming unmapped | |
3314 | * pages. | |
2a16e3f4 | 3315 | * |
0ff38490 CL |
3316 | * Note that shrink_slab will free memory on all zones and may |
3317 | * take a long time. | |
2a16e3f4 | 3318 | */ |
4dc4b3d9 KM |
3319 | for (;;) { |
3320 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3321 | ||
3322 | /* No reclaimable slab or very low memory pressure */ | |
1495f230 | 3323 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) |
4dc4b3d9 KM |
3324 | break; |
3325 | ||
3326 | /* Freed enough memory */ | |
3327 | nr_slab_pages1 = zone_page_state(zone, | |
3328 | NR_SLAB_RECLAIMABLE); | |
3329 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3330 | break; | |
3331 | } | |
83e33a47 CL |
3332 | |
3333 | /* | |
3334 | * Update nr_reclaimed by the number of slab pages we | |
3335 | * reclaimed from this zone. | |
3336 | */ | |
15748048 KM |
3337 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3338 | if (nr_slab_pages1 < nr_slab_pages0) | |
3339 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
2a16e3f4 CL |
3340 | } |
3341 | ||
9eeff239 | 3342 | p->reclaim_state = NULL; |
d4f7796e | 3343 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3344 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3345 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3346 | } |
179e9639 AM |
3347 | |
3348 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3349 | { | |
179e9639 | 3350 | int node_id; |
d773ed6b | 3351 | int ret; |
179e9639 AM |
3352 | |
3353 | /* | |
0ff38490 CL |
3354 | * Zone reclaim reclaims unmapped file backed pages and |
3355 | * slab pages if we are over the defined limits. | |
34aa1330 | 3356 | * |
9614634f CL |
3357 | * A small portion of unmapped file backed pages is needed for |
3358 | * file I/O otherwise pages read by file I/O will be immediately | |
3359 | * thrown out if the zone is overallocated. So we do not reclaim | |
3360 | * if less than a specified percentage of the zone is used by | |
3361 | * unmapped file backed pages. | |
179e9639 | 3362 | */ |
90afa5de MG |
3363 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
3364 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 3365 | return ZONE_RECLAIM_FULL; |
179e9639 | 3366 | |
93e4a89a | 3367 | if (zone->all_unreclaimable) |
fa5e084e | 3368 | return ZONE_RECLAIM_FULL; |
d773ed6b | 3369 | |
179e9639 | 3370 | /* |
d773ed6b | 3371 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3372 | */ |
d773ed6b | 3373 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 3374 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
3375 | |
3376 | /* | |
3377 | * Only run zone reclaim on the local zone or on zones that do not | |
3378 | * have associated processors. This will favor the local processor | |
3379 | * over remote processors and spread off node memory allocations | |
3380 | * as wide as possible. | |
3381 | */ | |
89fa3024 | 3382 | node_id = zone_to_nid(zone); |
37c0708d | 3383 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 3384 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b DR |
3385 | |
3386 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
fa5e084e MG |
3387 | return ZONE_RECLAIM_NOSCAN; |
3388 | ||
d773ed6b DR |
3389 | ret = __zone_reclaim(zone, gfp_mask, order); |
3390 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3391 | ||
24cf7251 MG |
3392 | if (!ret) |
3393 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3394 | ||
d773ed6b | 3395 | return ret; |
179e9639 | 3396 | } |
9eeff239 | 3397 | #endif |
894bc310 | 3398 | |
894bc310 LS |
3399 | /* |
3400 | * page_evictable - test whether a page is evictable | |
3401 | * @page: the page to test | |
3402 | * @vma: the VMA in which the page is or will be mapped, may be NULL | |
3403 | * | |
3404 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
b291f000 NP |
3405 | * lists vs unevictable list. The vma argument is !NULL when called from the |
3406 | * fault path to determine how to instantate a new page. | |
894bc310 LS |
3407 | * |
3408 | * Reasons page might not be evictable: | |
ba9ddf49 | 3409 | * (1) page's mapping marked unevictable |
b291f000 | 3410 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3411 | * |
894bc310 LS |
3412 | */ |
3413 | int page_evictable(struct page *page, struct vm_area_struct *vma) | |
3414 | { | |
3415 | ||
ba9ddf49 LS |
3416 | if (mapping_unevictable(page_mapping(page))) |
3417 | return 0; | |
3418 | ||
b291f000 NP |
3419 | if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page))) |
3420 | return 0; | |
894bc310 LS |
3421 | |
3422 | return 1; | |
3423 | } | |
89e004ea | 3424 | |
85046579 | 3425 | #ifdef CONFIG_SHMEM |
89e004ea | 3426 | /** |
24513264 HD |
3427 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3428 | * @pages: array of pages to check | |
3429 | * @nr_pages: number of pages to check | |
89e004ea | 3430 | * |
24513264 | 3431 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3432 | * |
3433 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3434 | */ |
24513264 | 3435 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3436 | { |
925b7673 | 3437 | struct lruvec *lruvec; |
24513264 HD |
3438 | struct zone *zone = NULL; |
3439 | int pgscanned = 0; | |
3440 | int pgrescued = 0; | |
3441 | int i; | |
89e004ea | 3442 | |
24513264 HD |
3443 | for (i = 0; i < nr_pages; i++) { |
3444 | struct page *page = pages[i]; | |
3445 | struct zone *pagezone; | |
89e004ea | 3446 | |
24513264 HD |
3447 | pgscanned++; |
3448 | pagezone = page_zone(page); | |
3449 | if (pagezone != zone) { | |
3450 | if (zone) | |
3451 | spin_unlock_irq(&zone->lru_lock); | |
3452 | zone = pagezone; | |
3453 | spin_lock_irq(&zone->lru_lock); | |
3454 | } | |
89e004ea | 3455 | |
24513264 HD |
3456 | if (!PageLRU(page) || !PageUnevictable(page)) |
3457 | continue; | |
89e004ea | 3458 | |
24513264 HD |
3459 | if (page_evictable(page, NULL)) { |
3460 | enum lru_list lru = page_lru_base_type(page); | |
3461 | ||
3462 | VM_BUG_ON(PageActive(page)); | |
3463 | ClearPageUnevictable(page); | |
3464 | __dec_zone_state(zone, NR_UNEVICTABLE); | |
3465 | lruvec = mem_cgroup_lru_move_lists(zone, page, | |
3466 | LRU_UNEVICTABLE, lru); | |
3467 | list_move(&page->lru, &lruvec->lists[lru]); | |
3468 | __inc_zone_state(zone, NR_INACTIVE_ANON + lru); | |
3469 | pgrescued++; | |
89e004ea | 3470 | } |
24513264 | 3471 | } |
89e004ea | 3472 | |
24513264 HD |
3473 | if (zone) { |
3474 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3475 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3476 | spin_unlock_irq(&zone->lru_lock); | |
89e004ea | 3477 | } |
89e004ea | 3478 | } |
85046579 | 3479 | #endif /* CONFIG_SHMEM */ |
af936a16 | 3480 | |
264e56d8 | 3481 | static void warn_scan_unevictable_pages(void) |
af936a16 | 3482 | { |
264e56d8 | 3483 | printk_once(KERN_WARNING |
25bd91bd | 3484 | "%s: The scan_unevictable_pages sysctl/node-interface has been " |
264e56d8 | 3485 | "disabled for lack of a legitimate use case. If you have " |
25bd91bd KM |
3486 | "one, please send an email to linux-mm@kvack.org.\n", |
3487 | current->comm); | |
af936a16 LS |
3488 | } |
3489 | ||
3490 | /* | |
3491 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3492 | * all nodes' unevictable lists for evictable pages | |
3493 | */ | |
3494 | unsigned long scan_unevictable_pages; | |
3495 | ||
3496 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
8d65af78 | 3497 | void __user *buffer, |
af936a16 LS |
3498 | size_t *length, loff_t *ppos) |
3499 | { | |
264e56d8 | 3500 | warn_scan_unevictable_pages(); |
8d65af78 | 3501 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
af936a16 LS |
3502 | scan_unevictable_pages = 0; |
3503 | return 0; | |
3504 | } | |
3505 | ||
e4455abb | 3506 | #ifdef CONFIG_NUMA |
af936a16 LS |
3507 | /* |
3508 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3509 | * a specified node's per zone unevictable lists for evictable pages. | |
3510 | */ | |
3511 | ||
10fbcf4c KS |
3512 | static ssize_t read_scan_unevictable_node(struct device *dev, |
3513 | struct device_attribute *attr, | |
af936a16 LS |
3514 | char *buf) |
3515 | { | |
264e56d8 | 3516 | warn_scan_unevictable_pages(); |
af936a16 LS |
3517 | return sprintf(buf, "0\n"); /* always zero; should fit... */ |
3518 | } | |
3519 | ||
10fbcf4c KS |
3520 | static ssize_t write_scan_unevictable_node(struct device *dev, |
3521 | struct device_attribute *attr, | |
af936a16 LS |
3522 | const char *buf, size_t count) |
3523 | { | |
264e56d8 | 3524 | warn_scan_unevictable_pages(); |
af936a16 LS |
3525 | return 1; |
3526 | } | |
3527 | ||
3528 | ||
10fbcf4c | 3529 | static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, |
af936a16 LS |
3530 | read_scan_unevictable_node, |
3531 | write_scan_unevictable_node); | |
3532 | ||
3533 | int scan_unevictable_register_node(struct node *node) | |
3534 | { | |
10fbcf4c | 3535 | return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 LS |
3536 | } |
3537 | ||
3538 | void scan_unevictable_unregister_node(struct node *node) | |
3539 | { | |
10fbcf4c | 3540 | device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 | 3541 | } |
e4455abb | 3542 | #endif |