]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 LT |
2 | /* |
3 | * linux/mm/vmscan.c | |
4 | * | |
5 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
6 | * | |
7 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
8 | * kswapd added: 7.1.96 sct | |
9 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
10 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
11 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
12 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
13 | */ | |
14 | ||
b1de0d13 MH |
15 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
16 | ||
1da177e4 | 17 | #include <linux/mm.h> |
5b3cc15a | 18 | #include <linux/sched/mm.h> |
1da177e4 | 19 | #include <linux/module.h> |
5a0e3ad6 | 20 | #include <linux/gfp.h> |
1da177e4 LT |
21 | #include <linux/kernel_stat.h> |
22 | #include <linux/swap.h> | |
23 | #include <linux/pagemap.h> | |
24 | #include <linux/init.h> | |
25 | #include <linux/highmem.h> | |
70ddf637 | 26 | #include <linux/vmpressure.h> |
e129b5c2 | 27 | #include <linux/vmstat.h> |
1da177e4 LT |
28 | #include <linux/file.h> |
29 | #include <linux/writeback.h> | |
30 | #include <linux/blkdev.h> | |
31 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
32 | buffer_heads_over_limit */ | |
33 | #include <linux/mm_inline.h> | |
1da177e4 LT |
34 | #include <linux/backing-dev.h> |
35 | #include <linux/rmap.h> | |
36 | #include <linux/topology.h> | |
37 | #include <linux/cpu.h> | |
38 | #include <linux/cpuset.h> | |
3e7d3449 | 39 | #include <linux/compaction.h> |
1da177e4 LT |
40 | #include <linux/notifier.h> |
41 | #include <linux/rwsem.h> | |
248a0301 | 42 | #include <linux/delay.h> |
3218ae14 | 43 | #include <linux/kthread.h> |
7dfb7103 | 44 | #include <linux/freezer.h> |
66e1707b | 45 | #include <linux/memcontrol.h> |
873b4771 | 46 | #include <linux/delayacct.h> |
af936a16 | 47 | #include <linux/sysctl.h> |
929bea7c | 48 | #include <linux/oom.h> |
268bb0ce | 49 | #include <linux/prefetch.h> |
b1de0d13 | 50 | #include <linux/printk.h> |
f9fe48be | 51 | #include <linux/dax.h> |
1da177e4 LT |
52 | |
53 | #include <asm/tlbflush.h> | |
54 | #include <asm/div64.h> | |
55 | ||
56 | #include <linux/swapops.h> | |
117aad1e | 57 | #include <linux/balloon_compaction.h> |
1da177e4 | 58 | |
0f8053a5 NP |
59 | #include "internal.h" |
60 | ||
33906bc5 MG |
61 | #define CREATE_TRACE_POINTS |
62 | #include <trace/events/vmscan.h> | |
63 | ||
1da177e4 | 64 | struct scan_control { |
22fba335 KM |
65 | /* How many pages shrink_list() should reclaim */ |
66 | unsigned long nr_to_reclaim; | |
67 | ||
1da177e4 | 68 | /* This context's GFP mask */ |
6daa0e28 | 69 | gfp_t gfp_mask; |
1da177e4 | 70 | |
ee814fe2 | 71 | /* Allocation order */ |
5ad333eb | 72 | int order; |
66e1707b | 73 | |
ee814fe2 JW |
74 | /* |
75 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
76 | * are scanned. | |
77 | */ | |
78 | nodemask_t *nodemask; | |
9e3b2f8c | 79 | |
f16015fb JW |
80 | /* |
81 | * The memory cgroup that hit its limit and as a result is the | |
82 | * primary target of this reclaim invocation. | |
83 | */ | |
84 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 85 | |
ee814fe2 JW |
86 | /* Scan (total_size >> priority) pages at once */ |
87 | int priority; | |
88 | ||
b2e18757 MG |
89 | /* The highest zone to isolate pages for reclaim from */ |
90 | enum zone_type reclaim_idx; | |
91 | ||
1276ad68 | 92 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
93 | unsigned int may_writepage:1; |
94 | ||
95 | /* Can mapped pages be reclaimed? */ | |
96 | unsigned int may_unmap:1; | |
97 | ||
98 | /* Can pages be swapped as part of reclaim? */ | |
99 | unsigned int may_swap:1; | |
100 | ||
d6622f63 YX |
101 | /* |
102 | * Cgroups are not reclaimed below their configured memory.low, | |
103 | * unless we threaten to OOM. If any cgroups are skipped due to | |
104 | * memory.low and nothing was reclaimed, go back for memory.low. | |
105 | */ | |
106 | unsigned int memcg_low_reclaim:1; | |
107 | unsigned int memcg_low_skipped:1; | |
241994ed | 108 | |
ee814fe2 JW |
109 | unsigned int hibernation_mode:1; |
110 | ||
111 | /* One of the zones is ready for compaction */ | |
112 | unsigned int compaction_ready:1; | |
113 | ||
114 | /* Incremented by the number of inactive pages that were scanned */ | |
115 | unsigned long nr_scanned; | |
116 | ||
117 | /* Number of pages freed so far during a call to shrink_zones() */ | |
118 | unsigned long nr_reclaimed; | |
1da177e4 LT |
119 | }; |
120 | ||
1da177e4 LT |
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; | |
d0480be4 WSH |
153 | /* |
154 | * The total number of pages which are beyond the high watermark within all | |
155 | * zones. | |
156 | */ | |
157 | unsigned long vm_total_pages; | |
1da177e4 LT |
158 | |
159 | static LIST_HEAD(shrinker_list); | |
160 | static DECLARE_RWSEM(shrinker_rwsem); | |
161 | ||
c255a458 | 162 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
163 | static bool global_reclaim(struct scan_control *sc) |
164 | { | |
f16015fb | 165 | return !sc->target_mem_cgroup; |
89b5fae5 | 166 | } |
97c9341f TH |
167 | |
168 | /** | |
169 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
170 | * @sc: scan_control in question | |
171 | * | |
172 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
173 | * completely broken with the legacy memcg and direct stalling in | |
174 | * shrink_page_list() is used for throttling instead, which lacks all the | |
175 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
176 | * allocation and configurability. | |
177 | * | |
178 | * This function tests whether the vmscan currently in progress can assume | |
179 | * that the normal dirty throttling mechanism is operational. | |
180 | */ | |
181 | static bool sane_reclaim(struct scan_control *sc) | |
182 | { | |
183 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
184 | ||
185 | if (!memcg) | |
186 | return true; | |
187 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 188 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
189 | return true; |
190 | #endif | |
191 | return false; | |
192 | } | |
91a45470 | 193 | #else |
89b5fae5 JW |
194 | static bool global_reclaim(struct scan_control *sc) |
195 | { | |
196 | return true; | |
197 | } | |
97c9341f TH |
198 | |
199 | static bool sane_reclaim(struct scan_control *sc) | |
200 | { | |
201 | return true; | |
202 | } | |
91a45470 KH |
203 | #endif |
204 | ||
5a1c84b4 MG |
205 | /* |
206 | * This misses isolated pages which are not accounted for to save counters. | |
207 | * As the data only determines if reclaim or compaction continues, it is | |
208 | * not expected that isolated pages will be a dominating factor. | |
209 | */ | |
210 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
211 | { | |
212 | unsigned long nr; | |
213 | ||
214 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
215 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
216 | if (get_nr_swap_pages() > 0) | |
217 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
218 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
219 | ||
220 | return nr; | |
221 | } | |
222 | ||
599d0c95 MG |
223 | unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat) |
224 | { | |
225 | unsigned long nr; | |
226 | ||
227 | nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) + | |
228 | node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) + | |
229 | node_page_state_snapshot(pgdat, NR_ISOLATED_FILE); | |
6e543d57 LD |
230 | |
231 | if (get_nr_swap_pages() > 0) | |
599d0c95 MG |
232 | nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) + |
233 | node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) + | |
234 | node_page_state_snapshot(pgdat, NR_ISOLATED_ANON); | |
6e543d57 LD |
235 | |
236 | return nr; | |
237 | } | |
238 | ||
fd538803 MH |
239 | /** |
240 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
241 | * @lruvec: lru vector | |
242 | * @lru: lru to use | |
243 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
244 | */ | |
245 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) | |
c9f299d9 | 246 | { |
fd538803 MH |
247 | unsigned long lru_size; |
248 | int zid; | |
249 | ||
c3c787e8 | 250 | if (!mem_cgroup_disabled()) |
fd538803 MH |
251 | lru_size = mem_cgroup_get_lru_size(lruvec, lru); |
252 | else | |
253 | lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); | |
a3d8e054 | 254 | |
fd538803 MH |
255 | for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { |
256 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; | |
257 | unsigned long size; | |
c9f299d9 | 258 | |
fd538803 MH |
259 | if (!managed_zone(zone)) |
260 | continue; | |
261 | ||
262 | if (!mem_cgroup_disabled()) | |
263 | size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); | |
264 | else | |
265 | size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], | |
266 | NR_ZONE_LRU_BASE + lru); | |
267 | lru_size -= min(size, lru_size); | |
268 | } | |
269 | ||
270 | return lru_size; | |
b4536f0c | 271 | |
b4536f0c MH |
272 | } |
273 | ||
1da177e4 | 274 | /* |
1d3d4437 | 275 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 276 | */ |
1d3d4437 | 277 | int register_shrinker(struct shrinker *shrinker) |
1da177e4 | 278 | { |
1d3d4437 GC |
279 | size_t size = sizeof(*shrinker->nr_deferred); |
280 | ||
1d3d4437 GC |
281 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
282 | size *= nr_node_ids; | |
283 | ||
284 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
285 | if (!shrinker->nr_deferred) | |
286 | return -ENOMEM; | |
287 | ||
8e1f936b RR |
288 | down_write(&shrinker_rwsem); |
289 | list_add_tail(&shrinker->list, &shrinker_list); | |
290 | up_write(&shrinker_rwsem); | |
1d3d4437 | 291 | return 0; |
1da177e4 | 292 | } |
8e1f936b | 293 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
294 | |
295 | /* | |
296 | * Remove one | |
297 | */ | |
8e1f936b | 298 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 299 | { |
7880fc54 TH |
300 | if (!shrinker->nr_deferred) |
301 | return; | |
1da177e4 LT |
302 | down_write(&shrinker_rwsem); |
303 | list_del(&shrinker->list); | |
304 | up_write(&shrinker_rwsem); | |
ae393321 | 305 | kfree(shrinker->nr_deferred); |
7880fc54 | 306 | shrinker->nr_deferred = NULL; |
1da177e4 | 307 | } |
8e1f936b | 308 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
309 | |
310 | #define SHRINK_BATCH 128 | |
1d3d4437 | 311 | |
cb731d6c VD |
312 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
313 | struct shrinker *shrinker, | |
314 | unsigned long nr_scanned, | |
315 | unsigned long nr_eligible) | |
1d3d4437 GC |
316 | { |
317 | unsigned long freed = 0; | |
318 | unsigned long long delta; | |
319 | long total_scan; | |
d5bc5fd3 | 320 | long freeable; |
1d3d4437 GC |
321 | long nr; |
322 | long new_nr; | |
323 | int nid = shrinkctl->nid; | |
324 | long batch_size = shrinker->batch ? shrinker->batch | |
325 | : SHRINK_BATCH; | |
5f33a080 | 326 | long scanned = 0, next_deferred; |
1d3d4437 | 327 | |
d5bc5fd3 VD |
328 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
329 | if (freeable == 0) | |
1d3d4437 GC |
330 | return 0; |
331 | ||
332 | /* | |
333 | * copy the current shrinker scan count into a local variable | |
334 | * and zero it so that other concurrent shrinker invocations | |
335 | * don't also do this scanning work. | |
336 | */ | |
337 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
338 | ||
339 | total_scan = nr; | |
6b4f7799 | 340 | delta = (4 * nr_scanned) / shrinker->seeks; |
d5bc5fd3 | 341 | delta *= freeable; |
6b4f7799 | 342 | do_div(delta, nr_eligible + 1); |
1d3d4437 GC |
343 | total_scan += delta; |
344 | if (total_scan < 0) { | |
8612c663 | 345 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
a0b02131 | 346 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 347 | total_scan = freeable; |
5f33a080 SL |
348 | next_deferred = nr; |
349 | } else | |
350 | next_deferred = total_scan; | |
1d3d4437 GC |
351 | |
352 | /* | |
353 | * We need to avoid excessive windup on filesystem shrinkers | |
354 | * due to large numbers of GFP_NOFS allocations causing the | |
355 | * shrinkers to return -1 all the time. This results in a large | |
356 | * nr being built up so when a shrink that can do some work | |
357 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 358 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
359 | * memory. |
360 | * | |
361 | * Hence only allow the shrinker to scan the entire cache when | |
362 | * a large delta change is calculated directly. | |
363 | */ | |
d5bc5fd3 VD |
364 | if (delta < freeable / 4) |
365 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
366 | |
367 | /* | |
368 | * Avoid risking looping forever due to too large nr value: | |
369 | * never try to free more than twice the estimate number of | |
370 | * freeable entries. | |
371 | */ | |
d5bc5fd3 VD |
372 | if (total_scan > freeable * 2) |
373 | total_scan = freeable * 2; | |
1d3d4437 GC |
374 | |
375 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
6b4f7799 JW |
376 | nr_scanned, nr_eligible, |
377 | freeable, delta, total_scan); | |
1d3d4437 | 378 | |
0b1fb40a VD |
379 | /* |
380 | * Normally, we should not scan less than batch_size objects in one | |
381 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
382 | * than batch_size objects in total and we are really tight on memory, | |
383 | * we will try to reclaim all available objects, otherwise we can end | |
384 | * up failing allocations although there are plenty of reclaimable | |
385 | * objects spread over several slabs with usage less than the | |
386 | * batch_size. | |
387 | * | |
388 | * We detect the "tight on memory" situations by looking at the total | |
389 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 390 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
391 | * scanning at high prio and therefore should try to reclaim as much as |
392 | * possible. | |
393 | */ | |
394 | while (total_scan >= batch_size || | |
d5bc5fd3 | 395 | total_scan >= freeable) { |
a0b02131 | 396 | unsigned long ret; |
0b1fb40a | 397 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 398 | |
0b1fb40a | 399 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 400 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
401 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
402 | if (ret == SHRINK_STOP) | |
403 | break; | |
404 | freed += ret; | |
1d3d4437 | 405 | |
d460acb5 CW |
406 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
407 | total_scan -= shrinkctl->nr_scanned; | |
408 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
409 | |
410 | cond_resched(); | |
411 | } | |
412 | ||
5f33a080 SL |
413 | if (next_deferred >= scanned) |
414 | next_deferred -= scanned; | |
415 | else | |
416 | next_deferred = 0; | |
1d3d4437 GC |
417 | /* |
418 | * move the unused scan count back into the shrinker in a | |
419 | * manner that handles concurrent updates. If we exhausted the | |
420 | * scan, there is no need to do an update. | |
421 | */ | |
5f33a080 SL |
422 | if (next_deferred > 0) |
423 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
424 | &shrinker->nr_deferred[nid]); |
425 | else | |
426 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
427 | ||
df9024a8 | 428 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 429 | return freed; |
1495f230 YH |
430 | } |
431 | ||
6b4f7799 | 432 | /** |
cb731d6c | 433 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
434 | * @gfp_mask: allocation context |
435 | * @nid: node whose slab caches to target | |
cb731d6c | 436 | * @memcg: memory cgroup whose slab caches to target |
6b4f7799 JW |
437 | * @nr_scanned: pressure numerator |
438 | * @nr_eligible: pressure denominator | |
1da177e4 | 439 | * |
6b4f7799 | 440 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 441 | * |
6b4f7799 JW |
442 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
443 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 444 | * |
cb731d6c VD |
445 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
446 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan | |
0fc9f58a VD |
447 | * objects from the memory cgroup specified. Otherwise, only unaware |
448 | * shrinkers are called. | |
cb731d6c | 449 | * |
6b4f7799 JW |
450 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of |
451 | * the available objects should be scanned. Page reclaim for example | |
452 | * passes the number of pages scanned and the number of pages on the | |
453 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned | |
454 | * when it encountered mapped pages. The ratio is further biased by | |
455 | * the ->seeks setting of the shrink function, which indicates the | |
456 | * cost to recreate an object relative to that of an LRU page. | |
b15e0905 | 457 | * |
6b4f7799 | 458 | * Returns the number of reclaimed slab objects. |
1da177e4 | 459 | */ |
cb731d6c VD |
460 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
461 | struct mem_cgroup *memcg, | |
462 | unsigned long nr_scanned, | |
463 | unsigned long nr_eligible) | |
1da177e4 LT |
464 | { |
465 | struct shrinker *shrinker; | |
24f7c6b9 | 466 | unsigned long freed = 0; |
1da177e4 | 467 | |
0fc9f58a | 468 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) |
cb731d6c VD |
469 | return 0; |
470 | ||
6b4f7799 JW |
471 | if (nr_scanned == 0) |
472 | nr_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 473 | |
f06590bd | 474 | if (!down_read_trylock(&shrinker_rwsem)) { |
24f7c6b9 DC |
475 | /* |
476 | * If we would return 0, our callers would understand that we | |
477 | * have nothing else to shrink and give up trying. By returning | |
478 | * 1 we keep it going and assume we'll be able to shrink next | |
479 | * time. | |
480 | */ | |
481 | freed = 1; | |
f06590bd MK |
482 | goto out; |
483 | } | |
1da177e4 LT |
484 | |
485 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
486 | struct shrink_control sc = { |
487 | .gfp_mask = gfp_mask, | |
488 | .nid = nid, | |
cb731d6c | 489 | .memcg = memcg, |
6b4f7799 | 490 | }; |
ec97097b | 491 | |
0fc9f58a VD |
492 | /* |
493 | * If kernel memory accounting is disabled, we ignore | |
494 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers | |
495 | * passing NULL for memcg. | |
496 | */ | |
497 | if (memcg_kmem_enabled() && | |
498 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
cb731d6c VD |
499 | continue; |
500 | ||
6b4f7799 JW |
501 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
502 | sc.nid = 0; | |
1da177e4 | 503 | |
cb731d6c | 504 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); |
204fd0bd MK |
505 | /* |
506 | * Bail out if someone want to register a new shrinker to | |
507 | * prevent the regsitration from being stalled for long periods | |
508 | * by parallel ongoing shrinking. | |
509 | */ | |
510 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
511 | freed = freed ? : 1; | |
512 | break; | |
513 | } | |
1da177e4 | 514 | } |
6b4f7799 | 515 | |
1da177e4 | 516 | up_read(&shrinker_rwsem); |
f06590bd MK |
517 | out: |
518 | cond_resched(); | |
24f7c6b9 | 519 | return freed; |
1da177e4 LT |
520 | } |
521 | ||
cb731d6c VD |
522 | void drop_slab_node(int nid) |
523 | { | |
524 | unsigned long freed; | |
525 | ||
526 | do { | |
527 | struct mem_cgroup *memcg = NULL; | |
528 | ||
529 | freed = 0; | |
530 | do { | |
531 | freed += shrink_slab(GFP_KERNEL, nid, memcg, | |
532 | 1000, 1000); | |
533 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
534 | } while (freed > 10); | |
535 | } | |
536 | ||
537 | void drop_slab(void) | |
538 | { | |
539 | int nid; | |
540 | ||
541 | for_each_online_node(nid) | |
542 | drop_slab_node(nid); | |
543 | } | |
544 | ||
1da177e4 LT |
545 | static inline int is_page_cache_freeable(struct page *page) |
546 | { | |
ceddc3a5 JW |
547 | /* |
548 | * A freeable page cache page is referenced only by the caller | |
549 | * that isolated the page, the page cache radix tree and | |
550 | * optional buffer heads at page->private. | |
551 | */ | |
bd4c82c2 HY |
552 | int radix_pins = PageTransHuge(page) && PageSwapCache(page) ? |
553 | HPAGE_PMD_NR : 1; | |
554 | return page_count(page) - page_has_private(page) == 1 + radix_pins; | |
1da177e4 LT |
555 | } |
556 | ||
703c2708 | 557 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
1da177e4 | 558 | { |
930d9152 | 559 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 560 | return 1; |
703c2708 | 561 | if (!inode_write_congested(inode)) |
1da177e4 | 562 | return 1; |
703c2708 | 563 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
564 | return 1; |
565 | return 0; | |
566 | } | |
567 | ||
568 | /* | |
569 | * We detected a synchronous write error writing a page out. Probably | |
570 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
571 | * fsync(), msync() or close(). | |
572 | * | |
573 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
574 | * prevents it from being freed up. But we have a ref on the page and once | |
575 | * that page is locked, the mapping is pinned. | |
576 | * | |
577 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
578 | * __GFP_FS. | |
579 | */ | |
580 | static void handle_write_error(struct address_space *mapping, | |
581 | struct page *page, int error) | |
582 | { | |
7eaceacc | 583 | lock_page(page); |
3e9f45bd GC |
584 | if (page_mapping(page) == mapping) |
585 | mapping_set_error(mapping, error); | |
1da177e4 LT |
586 | unlock_page(page); |
587 | } | |
588 | ||
04e62a29 CL |
589 | /* possible outcome of pageout() */ |
590 | typedef enum { | |
591 | /* failed to write page out, page is locked */ | |
592 | PAGE_KEEP, | |
593 | /* move page to the active list, page is locked */ | |
594 | PAGE_ACTIVATE, | |
595 | /* page has been sent to the disk successfully, page is unlocked */ | |
596 | PAGE_SUCCESS, | |
597 | /* page is clean and locked */ | |
598 | PAGE_CLEAN, | |
599 | } pageout_t; | |
600 | ||
1da177e4 | 601 | /* |
1742f19f AM |
602 | * pageout is called by shrink_page_list() for each dirty page. |
603 | * Calls ->writepage(). | |
1da177e4 | 604 | */ |
c661b078 | 605 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 606 | struct scan_control *sc) |
1da177e4 LT |
607 | { |
608 | /* | |
609 | * If the page is dirty, only perform writeback if that write | |
610 | * will be non-blocking. To prevent this allocation from being | |
611 | * stalled by pagecache activity. But note that there may be | |
612 | * stalls if we need to run get_block(). We could test | |
613 | * PagePrivate for that. | |
614 | * | |
8174202b | 615 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
616 | * this page's queue, we can perform writeback even if that |
617 | * will block. | |
618 | * | |
619 | * If the page is swapcache, write it back even if that would | |
620 | * block, for some throttling. This happens by accident, because | |
621 | * swap_backing_dev_info is bust: it doesn't reflect the | |
622 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
623 | */ |
624 | if (!is_page_cache_freeable(page)) | |
625 | return PAGE_KEEP; | |
626 | if (!mapping) { | |
627 | /* | |
628 | * Some data journaling orphaned pages can have | |
629 | * page->mapping == NULL while being dirty with clean buffers. | |
630 | */ | |
266cf658 | 631 | if (page_has_private(page)) { |
1da177e4 LT |
632 | if (try_to_free_buffers(page)) { |
633 | ClearPageDirty(page); | |
b1de0d13 | 634 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
635 | return PAGE_CLEAN; |
636 | } | |
637 | } | |
638 | return PAGE_KEEP; | |
639 | } | |
640 | if (mapping->a_ops->writepage == NULL) | |
641 | return PAGE_ACTIVATE; | |
703c2708 | 642 | if (!may_write_to_inode(mapping->host, sc)) |
1da177e4 LT |
643 | return PAGE_KEEP; |
644 | ||
645 | if (clear_page_dirty_for_io(page)) { | |
646 | int res; | |
647 | struct writeback_control wbc = { | |
648 | .sync_mode = WB_SYNC_NONE, | |
649 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
650 | .range_start = 0, |
651 | .range_end = LLONG_MAX, | |
1da177e4 LT |
652 | .for_reclaim = 1, |
653 | }; | |
654 | ||
655 | SetPageReclaim(page); | |
656 | res = mapping->a_ops->writepage(page, &wbc); | |
657 | if (res < 0) | |
658 | handle_write_error(mapping, page, res); | |
994fc28c | 659 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
660 | ClearPageReclaim(page); |
661 | return PAGE_ACTIVATE; | |
662 | } | |
c661b078 | 663 | |
1da177e4 LT |
664 | if (!PageWriteback(page)) { |
665 | /* synchronous write or broken a_ops? */ | |
666 | ClearPageReclaim(page); | |
667 | } | |
3aa23851 | 668 | trace_mm_vmscan_writepage(page); |
c4a25635 | 669 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
670 | return PAGE_SUCCESS; |
671 | } | |
672 | ||
673 | return PAGE_CLEAN; | |
674 | } | |
675 | ||
a649fd92 | 676 | /* |
e286781d NP |
677 | * Same as remove_mapping, but if the page is removed from the mapping, it |
678 | * gets returned with a refcount of 0. | |
a649fd92 | 679 | */ |
a528910e JW |
680 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
681 | bool reclaimed) | |
49d2e9cc | 682 | { |
c4843a75 | 683 | unsigned long flags; |
bd4c82c2 | 684 | int refcount; |
c4843a75 | 685 | |
28e4d965 NP |
686 | BUG_ON(!PageLocked(page)); |
687 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 688 | |
c4843a75 | 689 | spin_lock_irqsave(&mapping->tree_lock, flags); |
49d2e9cc | 690 | /* |
0fd0e6b0 NP |
691 | * The non racy check for a busy page. |
692 | * | |
693 | * Must be careful with the order of the tests. When someone has | |
694 | * a ref to the page, it may be possible that they dirty it then | |
695 | * drop the reference. So if PageDirty is tested before page_count | |
696 | * here, then the following race may occur: | |
697 | * | |
698 | * get_user_pages(&page); | |
699 | * [user mapping goes away] | |
700 | * write_to(page); | |
701 | * !PageDirty(page) [good] | |
702 | * SetPageDirty(page); | |
703 | * put_page(page); | |
704 | * !page_count(page) [good, discard it] | |
705 | * | |
706 | * [oops, our write_to data is lost] | |
707 | * | |
708 | * Reversing the order of the tests ensures such a situation cannot | |
709 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 710 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
711 | * |
712 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
713 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 714 | */ |
bd4c82c2 HY |
715 | if (unlikely(PageTransHuge(page)) && PageSwapCache(page)) |
716 | refcount = 1 + HPAGE_PMD_NR; | |
717 | else | |
718 | refcount = 2; | |
719 | if (!page_ref_freeze(page, refcount)) | |
49d2e9cc | 720 | goto cannot_free; |
e286781d NP |
721 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
722 | if (unlikely(PageDirty(page))) { | |
bd4c82c2 | 723 | page_ref_unfreeze(page, refcount); |
49d2e9cc | 724 | goto cannot_free; |
e286781d | 725 | } |
49d2e9cc CL |
726 | |
727 | if (PageSwapCache(page)) { | |
728 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 729 | mem_cgroup_swapout(page, swap); |
49d2e9cc | 730 | __delete_from_swap_cache(page); |
c4843a75 | 731 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
75f6d6d2 | 732 | put_swap_page(page, swap); |
e286781d | 733 | } else { |
6072d13c | 734 | void (*freepage)(struct page *); |
a528910e | 735 | void *shadow = NULL; |
6072d13c LT |
736 | |
737 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
738 | /* |
739 | * Remember a shadow entry for reclaimed file cache in | |
740 | * order to detect refaults, thus thrashing, later on. | |
741 | * | |
742 | * But don't store shadows in an address space that is | |
743 | * already exiting. This is not just an optizimation, | |
744 | * inode reclaim needs to empty out the radix tree or | |
745 | * the nodes are lost. Don't plant shadows behind its | |
746 | * back. | |
f9fe48be RZ |
747 | * |
748 | * We also don't store shadows for DAX mappings because the | |
749 | * only page cache pages found in these are zero pages | |
750 | * covering holes, and because we don't want to mix DAX | |
751 | * exceptional entries and shadow exceptional entries in the | |
752 | * same page_tree. | |
a528910e JW |
753 | */ |
754 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 755 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
a528910e | 756 | shadow = workingset_eviction(mapping, page); |
62cccb8c | 757 | __delete_from_page_cache(page, shadow); |
c4843a75 | 758 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
6072d13c LT |
759 | |
760 | if (freepage != NULL) | |
761 | freepage(page); | |
49d2e9cc CL |
762 | } |
763 | ||
49d2e9cc CL |
764 | return 1; |
765 | ||
766 | cannot_free: | |
c4843a75 | 767 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
49d2e9cc CL |
768 | return 0; |
769 | } | |
770 | ||
e286781d NP |
771 | /* |
772 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
773 | * someone else has a ref on the page, abort and return 0. If it was | |
774 | * successfully detached, return 1. Assumes the caller has a single ref on | |
775 | * this page. | |
776 | */ | |
777 | int remove_mapping(struct address_space *mapping, struct page *page) | |
778 | { | |
a528910e | 779 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
780 | /* |
781 | * Unfreezing the refcount with 1 rather than 2 effectively | |
782 | * drops the pagecache ref for us without requiring another | |
783 | * atomic operation. | |
784 | */ | |
fe896d18 | 785 | page_ref_unfreeze(page, 1); |
e286781d NP |
786 | return 1; |
787 | } | |
788 | return 0; | |
789 | } | |
790 | ||
894bc310 LS |
791 | /** |
792 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
793 | * @page: page to be put back to appropriate lru list | |
794 | * | |
795 | * Add previously isolated @page to appropriate LRU list. | |
796 | * Page may still be unevictable for other reasons. | |
797 | * | |
798 | * lru_lock must not be held, interrupts must be enabled. | |
799 | */ | |
894bc310 LS |
800 | void putback_lru_page(struct page *page) |
801 | { | |
0ec3b74c | 802 | bool is_unevictable; |
bbfd28ee | 803 | int was_unevictable = PageUnevictable(page); |
894bc310 | 804 | |
309381fe | 805 | VM_BUG_ON_PAGE(PageLRU(page), page); |
894bc310 LS |
806 | |
807 | redo: | |
808 | ClearPageUnevictable(page); | |
809 | ||
39b5f29a | 810 | if (page_evictable(page)) { |
894bc310 LS |
811 | /* |
812 | * For evictable pages, we can use the cache. | |
813 | * In event of a race, worst case is we end up with an | |
814 | * unevictable page on [in]active list. | |
815 | * We know how to handle that. | |
816 | */ | |
0ec3b74c | 817 | is_unevictable = false; |
c53954a0 | 818 | lru_cache_add(page); |
894bc310 LS |
819 | } else { |
820 | /* | |
821 | * Put unevictable pages directly on zone's unevictable | |
822 | * list. | |
823 | */ | |
0ec3b74c | 824 | is_unevictable = true; |
894bc310 | 825 | add_page_to_unevictable_list(page); |
6a7b9548 | 826 | /* |
21ee9f39 MK |
827 | * When racing with an mlock or AS_UNEVICTABLE clearing |
828 | * (page is unlocked) make sure that if the other thread | |
829 | * does not observe our setting of PG_lru and fails | |
24513264 | 830 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 831 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
832 | * the page back to the evictable list. |
833 | * | |
21ee9f39 | 834 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
835 | */ |
836 | smp_mb(); | |
894bc310 | 837 | } |
894bc310 LS |
838 | |
839 | /* | |
840 | * page's status can change while we move it among lru. If an evictable | |
841 | * page is on unevictable list, it never be freed. To avoid that, | |
842 | * check after we added it to the list, again. | |
843 | */ | |
0ec3b74c | 844 | if (is_unevictable && page_evictable(page)) { |
894bc310 LS |
845 | if (!isolate_lru_page(page)) { |
846 | put_page(page); | |
847 | goto redo; | |
848 | } | |
849 | /* This means someone else dropped this page from LRU | |
850 | * So, it will be freed or putback to LRU again. There is | |
851 | * nothing to do here. | |
852 | */ | |
853 | } | |
854 | ||
0ec3b74c | 855 | if (was_unevictable && !is_unevictable) |
bbfd28ee | 856 | count_vm_event(UNEVICTABLE_PGRESCUED); |
0ec3b74c | 857 | else if (!was_unevictable && is_unevictable) |
bbfd28ee LS |
858 | count_vm_event(UNEVICTABLE_PGCULLED); |
859 | ||
894bc310 LS |
860 | put_page(page); /* drop ref from isolate */ |
861 | } | |
862 | ||
dfc8d636 JW |
863 | enum page_references { |
864 | PAGEREF_RECLAIM, | |
865 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 866 | PAGEREF_KEEP, |
dfc8d636 JW |
867 | PAGEREF_ACTIVATE, |
868 | }; | |
869 | ||
870 | static enum page_references page_check_references(struct page *page, | |
871 | struct scan_control *sc) | |
872 | { | |
64574746 | 873 | int referenced_ptes, referenced_page; |
dfc8d636 | 874 | unsigned long vm_flags; |
dfc8d636 | 875 | |
c3ac9a8a JW |
876 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
877 | &vm_flags); | |
64574746 | 878 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 879 | |
dfc8d636 JW |
880 | /* |
881 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
882 | * move the page to the unevictable list. | |
883 | */ | |
884 | if (vm_flags & VM_LOCKED) | |
885 | return PAGEREF_RECLAIM; | |
886 | ||
64574746 | 887 | if (referenced_ptes) { |
e4898273 | 888 | if (PageSwapBacked(page)) |
64574746 JW |
889 | return PAGEREF_ACTIVATE; |
890 | /* | |
891 | * All mapped pages start out with page table | |
892 | * references from the instantiating fault, so we need | |
893 | * to look twice if a mapped file page is used more | |
894 | * than once. | |
895 | * | |
896 | * Mark it and spare it for another trip around the | |
897 | * inactive list. Another page table reference will | |
898 | * lead to its activation. | |
899 | * | |
900 | * Note: the mark is set for activated pages as well | |
901 | * so that recently deactivated but used pages are | |
902 | * quickly recovered. | |
903 | */ | |
904 | SetPageReferenced(page); | |
905 | ||
34dbc67a | 906 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
907 | return PAGEREF_ACTIVATE; |
908 | ||
c909e993 KK |
909 | /* |
910 | * Activate file-backed executable pages after first usage. | |
911 | */ | |
912 | if (vm_flags & VM_EXEC) | |
913 | return PAGEREF_ACTIVATE; | |
914 | ||
64574746 JW |
915 | return PAGEREF_KEEP; |
916 | } | |
dfc8d636 JW |
917 | |
918 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 919 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
920 | return PAGEREF_RECLAIM_CLEAN; |
921 | ||
922 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
923 | } |
924 | ||
e2be15f6 MG |
925 | /* Check if a page is dirty or under writeback */ |
926 | static void page_check_dirty_writeback(struct page *page, | |
927 | bool *dirty, bool *writeback) | |
928 | { | |
b4597226 MG |
929 | struct address_space *mapping; |
930 | ||
e2be15f6 MG |
931 | /* |
932 | * Anonymous pages are not handled by flushers and must be written | |
933 | * from reclaim context. Do not stall reclaim based on them | |
934 | */ | |
802a3a92 SL |
935 | if (!page_is_file_cache(page) || |
936 | (PageAnon(page) && !PageSwapBacked(page))) { | |
e2be15f6 MG |
937 | *dirty = false; |
938 | *writeback = false; | |
939 | return; | |
940 | } | |
941 | ||
942 | /* By default assume that the page flags are accurate */ | |
943 | *dirty = PageDirty(page); | |
944 | *writeback = PageWriteback(page); | |
b4597226 MG |
945 | |
946 | /* Verify dirty/writeback state if the filesystem supports it */ | |
947 | if (!page_has_private(page)) | |
948 | return; | |
949 | ||
950 | mapping = page_mapping(page); | |
951 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
952 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
953 | } |
954 | ||
3c710c1a MH |
955 | struct reclaim_stat { |
956 | unsigned nr_dirty; | |
957 | unsigned nr_unqueued_dirty; | |
958 | unsigned nr_congested; | |
959 | unsigned nr_writeback; | |
960 | unsigned nr_immediate; | |
5bccd166 MH |
961 | unsigned nr_activate; |
962 | unsigned nr_ref_keep; | |
963 | unsigned nr_unmap_fail; | |
3c710c1a MH |
964 | }; |
965 | ||
1da177e4 | 966 | /* |
1742f19f | 967 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 968 | */ |
1742f19f | 969 | static unsigned long shrink_page_list(struct list_head *page_list, |
599d0c95 | 970 | struct pglist_data *pgdat, |
f84f6e2b | 971 | struct scan_control *sc, |
02c6de8d | 972 | enum ttu_flags ttu_flags, |
3c710c1a | 973 | struct reclaim_stat *stat, |
02c6de8d | 974 | bool force_reclaim) |
1da177e4 LT |
975 | { |
976 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 977 | LIST_HEAD(free_pages); |
1da177e4 | 978 | int pgactivate = 0; |
3c710c1a MH |
979 | unsigned nr_unqueued_dirty = 0; |
980 | unsigned nr_dirty = 0; | |
981 | unsigned nr_congested = 0; | |
982 | unsigned nr_reclaimed = 0; | |
983 | unsigned nr_writeback = 0; | |
984 | unsigned nr_immediate = 0; | |
5bccd166 MH |
985 | unsigned nr_ref_keep = 0; |
986 | unsigned nr_unmap_fail = 0; | |
1da177e4 LT |
987 | |
988 | cond_resched(); | |
989 | ||
1da177e4 LT |
990 | while (!list_empty(page_list)) { |
991 | struct address_space *mapping; | |
992 | struct page *page; | |
993 | int may_enter_fs; | |
02c6de8d | 994 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
e2be15f6 | 995 | bool dirty, writeback; |
1da177e4 LT |
996 | |
997 | cond_resched(); | |
998 | ||
999 | page = lru_to_page(page_list); | |
1000 | list_del(&page->lru); | |
1001 | ||
529ae9aa | 1002 | if (!trylock_page(page)) |
1da177e4 LT |
1003 | goto keep; |
1004 | ||
309381fe | 1005 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
1006 | |
1007 | sc->nr_scanned++; | |
80e43426 | 1008 | |
39b5f29a | 1009 | if (unlikely(!page_evictable(page))) |
ad6b6704 | 1010 | goto activate_locked; |
894bc310 | 1011 | |
a6dc60f8 | 1012 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
1013 | goto keep_locked; |
1014 | ||
1da177e4 | 1015 | /* Double the slab pressure for mapped and swapcache pages */ |
802a3a92 SL |
1016 | if ((page_mapped(page) || PageSwapCache(page)) && |
1017 | !(PageAnon(page) && !PageSwapBacked(page))) | |
1da177e4 LT |
1018 | sc->nr_scanned++; |
1019 | ||
c661b078 AW |
1020 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1021 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1022 | ||
e2be15f6 MG |
1023 | /* |
1024 | * The number of dirty pages determines if a zone is marked | |
1025 | * reclaim_congested which affects wait_iff_congested. kswapd | |
1026 | * will stall and start writing pages if the tail of the LRU | |
1027 | * is all dirty unqueued pages. | |
1028 | */ | |
1029 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1030 | if (dirty || writeback) | |
1031 | nr_dirty++; | |
1032 | ||
1033 | if (dirty && !writeback) | |
1034 | nr_unqueued_dirty++; | |
1035 | ||
d04e8acd MG |
1036 | /* |
1037 | * Treat this page as congested if the underlying BDI is or if | |
1038 | * pages are cycling through the LRU so quickly that the | |
1039 | * pages marked for immediate reclaim are making it to the | |
1040 | * end of the LRU a second time. | |
1041 | */ | |
e2be15f6 | 1042 | mapping = page_mapping(page); |
1da58ee2 | 1043 | if (((dirty || writeback) && mapping && |
703c2708 | 1044 | inode_write_congested(mapping->host)) || |
d04e8acd | 1045 | (writeback && PageReclaim(page))) |
e2be15f6 MG |
1046 | nr_congested++; |
1047 | ||
283aba9f MG |
1048 | /* |
1049 | * If a page at the tail of the LRU is under writeback, there | |
1050 | * are three cases to consider. | |
1051 | * | |
1052 | * 1) If reclaim is encountering an excessive number of pages | |
1053 | * under writeback and this page is both under writeback and | |
1054 | * PageReclaim then it indicates that pages are being queued | |
1055 | * for IO but are being recycled through the LRU before the | |
1056 | * IO can complete. Waiting on the page itself risks an | |
1057 | * indefinite stall if it is impossible to writeback the | |
1058 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1059 | * note that the LRU is being scanned too quickly and the |
1060 | * caller can stall after page list has been processed. | |
283aba9f | 1061 | * |
97c9341f | 1062 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1063 | * not marked for immediate reclaim, or the caller does not |
1064 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1065 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1066 | * reclaim and continue scanning. |
283aba9f | 1067 | * |
ecf5fc6e MH |
1068 | * Require may_enter_fs because we would wait on fs, which |
1069 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1070 | * enter reclaim, and deadlock if it waits on a page for |
1071 | * which it is needed to do the write (loop masks off | |
1072 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1073 | * would probably show more reasons. | |
1074 | * | |
7fadc820 | 1075 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1076 | * PageReclaim. memcg does not have any dirty pages |
1077 | * throttling so we could easily OOM just because too many | |
1078 | * pages are in writeback and there is nothing else to | |
1079 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1080 | * |
1081 | * In cases 1) and 2) we activate the pages to get them out of | |
1082 | * the way while we continue scanning for clean pages on the | |
1083 | * inactive list and refilling from the active list. The | |
1084 | * observation here is that waiting for disk writes is more | |
1085 | * expensive than potentially causing reloads down the line. | |
1086 | * Since they're marked for immediate reclaim, they won't put | |
1087 | * memory pressure on the cache working set any longer than it | |
1088 | * takes to write them to disk. | |
283aba9f | 1089 | */ |
c661b078 | 1090 | if (PageWriteback(page)) { |
283aba9f MG |
1091 | /* Case 1 above */ |
1092 | if (current_is_kswapd() && | |
1093 | PageReclaim(page) && | |
599d0c95 | 1094 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
b1a6f21e | 1095 | nr_immediate++; |
c55e8d03 | 1096 | goto activate_locked; |
283aba9f MG |
1097 | |
1098 | /* Case 2 above */ | |
97c9341f | 1099 | } else if (sane_reclaim(sc) || |
ecf5fc6e | 1100 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1101 | /* |
1102 | * This is slightly racy - end_page_writeback() | |
1103 | * might have just cleared PageReclaim, then | |
1104 | * setting PageReclaim here end up interpreted | |
1105 | * as PageReadahead - but that does not matter | |
1106 | * enough to care. What we do want is for this | |
1107 | * page to have PageReclaim set next time memcg | |
1108 | * reclaim reaches the tests above, so it will | |
1109 | * then wait_on_page_writeback() to avoid OOM; | |
1110 | * and it's also appropriate in global reclaim. | |
1111 | */ | |
1112 | SetPageReclaim(page); | |
e62e384e | 1113 | nr_writeback++; |
c55e8d03 | 1114 | goto activate_locked; |
283aba9f MG |
1115 | |
1116 | /* Case 3 above */ | |
1117 | } else { | |
7fadc820 | 1118 | unlock_page(page); |
283aba9f | 1119 | wait_on_page_writeback(page); |
7fadc820 HD |
1120 | /* then go back and try same page again */ |
1121 | list_add_tail(&page->lru, page_list); | |
1122 | continue; | |
e62e384e | 1123 | } |
c661b078 | 1124 | } |
1da177e4 | 1125 | |
02c6de8d MK |
1126 | if (!force_reclaim) |
1127 | references = page_check_references(page, sc); | |
1128 | ||
dfc8d636 JW |
1129 | switch (references) { |
1130 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1131 | goto activate_locked; |
64574746 | 1132 | case PAGEREF_KEEP: |
5bccd166 | 1133 | nr_ref_keep++; |
64574746 | 1134 | goto keep_locked; |
dfc8d636 JW |
1135 | case PAGEREF_RECLAIM: |
1136 | case PAGEREF_RECLAIM_CLEAN: | |
1137 | ; /* try to reclaim the page below */ | |
1138 | } | |
1da177e4 | 1139 | |
1da177e4 LT |
1140 | /* |
1141 | * Anonymous process memory has backing store? | |
1142 | * Try to allocate it some swap space here. | |
802a3a92 | 1143 | * Lazyfree page could be freed directly |
1da177e4 | 1144 | */ |
bd4c82c2 HY |
1145 | if (PageAnon(page) && PageSwapBacked(page)) { |
1146 | if (!PageSwapCache(page)) { | |
1147 | if (!(sc->gfp_mask & __GFP_IO)) | |
1148 | goto keep_locked; | |
1149 | if (PageTransHuge(page)) { | |
1150 | /* cannot split THP, skip it */ | |
1151 | if (!can_split_huge_page(page, NULL)) | |
1152 | goto activate_locked; | |
1153 | /* | |
1154 | * Split pages without a PMD map right | |
1155 | * away. Chances are some or all of the | |
1156 | * tail pages can be freed without IO. | |
1157 | */ | |
1158 | if (!compound_mapcount(page) && | |
1159 | split_huge_page_to_list(page, | |
1160 | page_list)) | |
1161 | goto activate_locked; | |
1162 | } | |
1163 | if (!add_to_swap(page)) { | |
1164 | if (!PageTransHuge(page)) | |
1165 | goto activate_locked; | |
1166 | /* Fallback to swap normal pages */ | |
1167 | if (split_huge_page_to_list(page, | |
1168 | page_list)) | |
1169 | goto activate_locked; | |
fe490cc0 HY |
1170 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1171 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1172 | #endif | |
bd4c82c2 HY |
1173 | if (!add_to_swap(page)) |
1174 | goto activate_locked; | |
1175 | } | |
0f074658 | 1176 | |
bd4c82c2 | 1177 | may_enter_fs = 1; |
1da177e4 | 1178 | |
bd4c82c2 HY |
1179 | /* Adding to swap updated mapping */ |
1180 | mapping = page_mapping(page); | |
1181 | } | |
7751b2da KS |
1182 | } else if (unlikely(PageTransHuge(page))) { |
1183 | /* Split file THP */ | |
1184 | if (split_huge_page_to_list(page, page_list)) | |
1185 | goto keep_locked; | |
e2be15f6 | 1186 | } |
1da177e4 LT |
1187 | |
1188 | /* | |
1189 | * The page is mapped into the page tables of one or more | |
1190 | * processes. Try to unmap it here. | |
1191 | */ | |
802a3a92 | 1192 | if (page_mapped(page)) { |
bd4c82c2 HY |
1193 | enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH; |
1194 | ||
1195 | if (unlikely(PageTransHuge(page))) | |
1196 | flags |= TTU_SPLIT_HUGE_PMD; | |
1197 | if (!try_to_unmap(page, flags)) { | |
5bccd166 | 1198 | nr_unmap_fail++; |
1da177e4 | 1199 | goto activate_locked; |
1da177e4 LT |
1200 | } |
1201 | } | |
1202 | ||
1203 | if (PageDirty(page)) { | |
ee72886d | 1204 | /* |
4eda4823 JW |
1205 | * Only kswapd can writeback filesystem pages |
1206 | * to avoid risk of stack overflow. But avoid | |
1207 | * injecting inefficient single-page IO into | |
1208 | * flusher writeback as much as possible: only | |
1209 | * write pages when we've encountered many | |
1210 | * dirty pages, and when we've already scanned | |
1211 | * the rest of the LRU for clean pages and see | |
1212 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1213 | */ |
f84f6e2b | 1214 | if (page_is_file_cache(page) && |
4eda4823 JW |
1215 | (!current_is_kswapd() || !PageReclaim(page) || |
1216 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1217 | /* |
1218 | * Immediately reclaim when written back. | |
1219 | * Similar in principal to deactivate_page() | |
1220 | * except we already have the page isolated | |
1221 | * and know it's dirty | |
1222 | */ | |
c4a25635 | 1223 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1224 | SetPageReclaim(page); |
1225 | ||
c55e8d03 | 1226 | goto activate_locked; |
ee72886d MG |
1227 | } |
1228 | ||
dfc8d636 | 1229 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1230 | goto keep_locked; |
4dd4b920 | 1231 | if (!may_enter_fs) |
1da177e4 | 1232 | goto keep_locked; |
52a8363e | 1233 | if (!sc->may_writepage) |
1da177e4 LT |
1234 | goto keep_locked; |
1235 | ||
d950c947 MG |
1236 | /* |
1237 | * Page is dirty. Flush the TLB if a writable entry | |
1238 | * potentially exists to avoid CPU writes after IO | |
1239 | * starts and then write it out here. | |
1240 | */ | |
1241 | try_to_unmap_flush_dirty(); | |
7d3579e8 | 1242 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1243 | case PAGE_KEEP: |
1244 | goto keep_locked; | |
1245 | case PAGE_ACTIVATE: | |
1246 | goto activate_locked; | |
1247 | case PAGE_SUCCESS: | |
7d3579e8 | 1248 | if (PageWriteback(page)) |
41ac1999 | 1249 | goto keep; |
7d3579e8 | 1250 | if (PageDirty(page)) |
1da177e4 | 1251 | goto keep; |
7d3579e8 | 1252 | |
1da177e4 LT |
1253 | /* |
1254 | * A synchronous write - probably a ramdisk. Go | |
1255 | * ahead and try to reclaim the page. | |
1256 | */ | |
529ae9aa | 1257 | if (!trylock_page(page)) |
1da177e4 LT |
1258 | goto keep; |
1259 | if (PageDirty(page) || PageWriteback(page)) | |
1260 | goto keep_locked; | |
1261 | mapping = page_mapping(page); | |
1262 | case PAGE_CLEAN: | |
1263 | ; /* try to free the page below */ | |
1264 | } | |
1265 | } | |
1266 | ||
1267 | /* | |
1268 | * If the page has buffers, try to free the buffer mappings | |
1269 | * associated with this page. If we succeed we try to free | |
1270 | * the page as well. | |
1271 | * | |
1272 | * We do this even if the page is PageDirty(). | |
1273 | * try_to_release_page() does not perform I/O, but it is | |
1274 | * possible for a page to have PageDirty set, but it is actually | |
1275 | * clean (all its buffers are clean). This happens if the | |
1276 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1277 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1278 | * try_to_release_page() will discover that cleanness and will |
1279 | * drop the buffers and mark the page clean - it can be freed. | |
1280 | * | |
1281 | * Rarely, pages can have buffers and no ->mapping. These are | |
1282 | * the pages which were not successfully invalidated in | |
1283 | * truncate_complete_page(). We try to drop those buffers here | |
1284 | * and if that worked, and the page is no longer mapped into | |
1285 | * process address space (page_count == 1) it can be freed. | |
1286 | * Otherwise, leave the page on the LRU so it is swappable. | |
1287 | */ | |
266cf658 | 1288 | if (page_has_private(page)) { |
1da177e4 LT |
1289 | if (!try_to_release_page(page, sc->gfp_mask)) |
1290 | goto activate_locked; | |
e286781d NP |
1291 | if (!mapping && page_count(page) == 1) { |
1292 | unlock_page(page); | |
1293 | if (put_page_testzero(page)) | |
1294 | goto free_it; | |
1295 | else { | |
1296 | /* | |
1297 | * rare race with speculative reference. | |
1298 | * the speculative reference will free | |
1299 | * this page shortly, so we may | |
1300 | * increment nr_reclaimed here (and | |
1301 | * leave it off the LRU). | |
1302 | */ | |
1303 | nr_reclaimed++; | |
1304 | continue; | |
1305 | } | |
1306 | } | |
1da177e4 LT |
1307 | } |
1308 | ||
802a3a92 SL |
1309 | if (PageAnon(page) && !PageSwapBacked(page)) { |
1310 | /* follow __remove_mapping for reference */ | |
1311 | if (!page_ref_freeze(page, 1)) | |
1312 | goto keep_locked; | |
1313 | if (PageDirty(page)) { | |
1314 | page_ref_unfreeze(page, 1); | |
1315 | goto keep_locked; | |
1316 | } | |
1da177e4 | 1317 | |
802a3a92 | 1318 | count_vm_event(PGLAZYFREED); |
2262185c | 1319 | count_memcg_page_event(page, PGLAZYFREED); |
802a3a92 SL |
1320 | } else if (!mapping || !__remove_mapping(mapping, page, true)) |
1321 | goto keep_locked; | |
a978d6f5 NP |
1322 | /* |
1323 | * At this point, we have no other references and there is | |
1324 | * no way to pick any more up (removed from LRU, removed | |
1325 | * from pagecache). Can use non-atomic bitops now (and | |
1326 | * we obviously don't have to worry about waking up a process | |
1327 | * waiting on the page lock, because there are no references. | |
1328 | */ | |
48c935ad | 1329 | __ClearPageLocked(page); |
e286781d | 1330 | free_it: |
05ff5137 | 1331 | nr_reclaimed++; |
abe4c3b5 MG |
1332 | |
1333 | /* | |
1334 | * Is there need to periodically free_page_list? It would | |
1335 | * appear not as the counts should be low | |
1336 | */ | |
bd4c82c2 HY |
1337 | if (unlikely(PageTransHuge(page))) { |
1338 | mem_cgroup_uncharge(page); | |
1339 | (*get_compound_page_dtor(page))(page); | |
1340 | } else | |
1341 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1342 | continue; |
1343 | ||
1344 | activate_locked: | |
68a22394 | 1345 | /* Not a candidate for swapping, so reclaim swap space. */ |
ad6b6704 MK |
1346 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
1347 | PageMlocked(page))) | |
a2c43eed | 1348 | try_to_free_swap(page); |
309381fe | 1349 | VM_BUG_ON_PAGE(PageActive(page), page); |
ad6b6704 MK |
1350 | if (!PageMlocked(page)) { |
1351 | SetPageActive(page); | |
1352 | pgactivate++; | |
2262185c | 1353 | count_memcg_page_event(page, PGACTIVATE); |
ad6b6704 | 1354 | } |
1da177e4 LT |
1355 | keep_locked: |
1356 | unlock_page(page); | |
1357 | keep: | |
1358 | list_add(&page->lru, &ret_pages); | |
309381fe | 1359 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1360 | } |
abe4c3b5 | 1361 | |
747db954 | 1362 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1363 | try_to_unmap_flush(); |
b745bc85 | 1364 | free_hot_cold_page_list(&free_pages, true); |
abe4c3b5 | 1365 | |
1da177e4 | 1366 | list_splice(&ret_pages, page_list); |
f8891e5e | 1367 | count_vm_events(PGACTIVATE, pgactivate); |
0a31bc97 | 1368 | |
3c710c1a MH |
1369 | if (stat) { |
1370 | stat->nr_dirty = nr_dirty; | |
1371 | stat->nr_congested = nr_congested; | |
1372 | stat->nr_unqueued_dirty = nr_unqueued_dirty; | |
1373 | stat->nr_writeback = nr_writeback; | |
1374 | stat->nr_immediate = nr_immediate; | |
5bccd166 MH |
1375 | stat->nr_activate = pgactivate; |
1376 | stat->nr_ref_keep = nr_ref_keep; | |
1377 | stat->nr_unmap_fail = nr_unmap_fail; | |
3c710c1a | 1378 | } |
05ff5137 | 1379 | return nr_reclaimed; |
1da177e4 LT |
1380 | } |
1381 | ||
02c6de8d MK |
1382 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1383 | struct list_head *page_list) | |
1384 | { | |
1385 | struct scan_control sc = { | |
1386 | .gfp_mask = GFP_KERNEL, | |
1387 | .priority = DEF_PRIORITY, | |
1388 | .may_unmap = 1, | |
1389 | }; | |
3c710c1a | 1390 | unsigned long ret; |
02c6de8d MK |
1391 | struct page *page, *next; |
1392 | LIST_HEAD(clean_pages); | |
1393 | ||
1394 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1395 | if (page_is_file_cache(page) && !PageDirty(page) && |
89062f03 | 1396 | !__PageMovable(page) && !PageUnevictable(page)) { |
02c6de8d MK |
1397 | ClearPageActive(page); |
1398 | list_move(&page->lru, &clean_pages); | |
1399 | } | |
1400 | } | |
1401 | ||
599d0c95 | 1402 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
a128ca71 | 1403 | TTU_IGNORE_ACCESS, NULL, true); |
02c6de8d | 1404 | list_splice(&clean_pages, page_list); |
599d0c95 | 1405 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1406 | return ret; |
1407 | } | |
1408 | ||
5ad333eb AW |
1409 | /* |
1410 | * Attempt to remove the specified page from its LRU. Only take this page | |
1411 | * if it is of the appropriate PageActive status. Pages which are being | |
1412 | * freed elsewhere are also ignored. | |
1413 | * | |
1414 | * page: page to consider | |
1415 | * mode: one of the LRU isolation modes defined above | |
1416 | * | |
1417 | * returns 0 on success, -ve errno on failure. | |
1418 | */ | |
f3fd4a61 | 1419 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1420 | { |
1421 | int ret = -EINVAL; | |
1422 | ||
1423 | /* Only take pages on the LRU. */ | |
1424 | if (!PageLRU(page)) | |
1425 | return ret; | |
1426 | ||
e46a2879 MK |
1427 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1428 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1429 | return ret; |
1430 | ||
5ad333eb | 1431 | ret = -EBUSY; |
08e552c6 | 1432 | |
c8244935 MG |
1433 | /* |
1434 | * To minimise LRU disruption, the caller can indicate that it only | |
1435 | * wants to isolate pages it will be able to operate on without | |
1436 | * blocking - clean pages for the most part. | |
1437 | * | |
c8244935 MG |
1438 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1439 | * that it is possible to migrate without blocking | |
1440 | */ | |
1276ad68 | 1441 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
1442 | /* All the caller can do on PageWriteback is block */ |
1443 | if (PageWriteback(page)) | |
1444 | return ret; | |
1445 | ||
1446 | if (PageDirty(page)) { | |
1447 | struct address_space *mapping; | |
1f9c87e2 | 1448 | bool migrate_dirty; |
c8244935 | 1449 | |
c8244935 MG |
1450 | /* |
1451 | * Only pages without mappings or that have a | |
1452 | * ->migratepage callback are possible to migrate | |
1f9c87e2 MG |
1453 | * without blocking. However, we can be racing with |
1454 | * truncation so it's necessary to lock the page | |
1455 | * to stabilise the mapping as truncation holds | |
1456 | * the page lock until after the page is removed | |
1457 | * from the page cache. | |
c8244935 | 1458 | */ |
1f9c87e2 MG |
1459 | if (!trylock_page(page)) |
1460 | return ret; | |
1461 | ||
c8244935 | 1462 | mapping = page_mapping(page); |
b968dd76 | 1463 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
1f9c87e2 MG |
1464 | unlock_page(page); |
1465 | if (!migrate_dirty) | |
c8244935 MG |
1466 | return ret; |
1467 | } | |
1468 | } | |
39deaf85 | 1469 | |
f80c0673 MK |
1470 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1471 | return ret; | |
1472 | ||
5ad333eb AW |
1473 | if (likely(get_page_unless_zero(page))) { |
1474 | /* | |
1475 | * Be careful not to clear PageLRU until after we're | |
1476 | * sure the page is not being freed elsewhere -- the | |
1477 | * page release code relies on it. | |
1478 | */ | |
1479 | ClearPageLRU(page); | |
1480 | ret = 0; | |
1481 | } | |
1482 | ||
1483 | return ret; | |
1484 | } | |
1485 | ||
7ee36a14 MG |
1486 | |
1487 | /* | |
1488 | * Update LRU sizes after isolating pages. The LRU size updates must | |
1489 | * be complete before mem_cgroup_update_lru_size due to a santity check. | |
1490 | */ | |
1491 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 1492 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 1493 | { |
7ee36a14 MG |
1494 | int zid; |
1495 | ||
7ee36a14 MG |
1496 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1497 | if (!nr_zone_taken[zid]) | |
1498 | continue; | |
1499 | ||
1500 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
7ee36a14 | 1501 | #ifdef CONFIG_MEMCG |
b4536f0c | 1502 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
7ee36a14 | 1503 | #endif |
b4536f0c MH |
1504 | } |
1505 | ||
7ee36a14 MG |
1506 | } |
1507 | ||
1da177e4 | 1508 | /* |
a52633d8 | 1509 | * zone_lru_lock is heavily contended. Some of the functions that |
1da177e4 LT |
1510 | * shrink the lists perform better by taking out a batch of pages |
1511 | * and working on them outside the LRU lock. | |
1512 | * | |
1513 | * For pagecache intensive workloads, this function is the hottest | |
1514 | * spot in the kernel (apart from copy_*_user functions). | |
1515 | * | |
1516 | * Appropriate locks must be held before calling this function. | |
1517 | * | |
791b48b6 | 1518 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 1519 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1520 | * @dst: The temp list to put pages on to. |
f626012d | 1521 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1522 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1523 | * @mode: One of the LRU isolation modes |
3cb99451 | 1524 | * @lru: LRU list id for isolating |
1da177e4 LT |
1525 | * |
1526 | * returns how many pages were moved onto *@dst. | |
1527 | */ | |
69e05944 | 1528 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1529 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1530 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1531 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1532 | { |
75b00af7 | 1533 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1534 | unsigned long nr_taken = 0; |
599d0c95 | 1535 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 1536 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 1537 | unsigned long skipped = 0; |
791b48b6 | 1538 | unsigned long scan, total_scan, nr_pages; |
b2e18757 | 1539 | LIST_HEAD(pages_skipped); |
1da177e4 | 1540 | |
791b48b6 MK |
1541 | scan = 0; |
1542 | for (total_scan = 0; | |
1543 | scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src); | |
1544 | total_scan++) { | |
5ad333eb | 1545 | struct page *page; |
5ad333eb | 1546 | |
1da177e4 LT |
1547 | page = lru_to_page(src); |
1548 | prefetchw_prev_lru_page(page, src, flags); | |
1549 | ||
309381fe | 1550 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1551 | |
b2e18757 MG |
1552 | if (page_zonenum(page) > sc->reclaim_idx) { |
1553 | list_move(&page->lru, &pages_skipped); | |
7cc30fcf | 1554 | nr_skipped[page_zonenum(page)]++; |
b2e18757 MG |
1555 | continue; |
1556 | } | |
1557 | ||
791b48b6 MK |
1558 | /* |
1559 | * Do not count skipped pages because that makes the function | |
1560 | * return with no isolated pages if the LRU mostly contains | |
1561 | * ineligible pages. This causes the VM to not reclaim any | |
1562 | * pages, triggering a premature OOM. | |
1563 | */ | |
1564 | scan++; | |
f3fd4a61 | 1565 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1566 | case 0: |
599d0c95 MG |
1567 | nr_pages = hpage_nr_pages(page); |
1568 | nr_taken += nr_pages; | |
1569 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1570 | list_move(&page->lru, dst); |
5ad333eb AW |
1571 | break; |
1572 | ||
1573 | case -EBUSY: | |
1574 | /* else it is being freed elsewhere */ | |
1575 | list_move(&page->lru, src); | |
1576 | continue; | |
46453a6e | 1577 | |
5ad333eb AW |
1578 | default: |
1579 | BUG(); | |
1580 | } | |
1da177e4 LT |
1581 | } |
1582 | ||
b2e18757 MG |
1583 | /* |
1584 | * Splice any skipped pages to the start of the LRU list. Note that | |
1585 | * this disrupts the LRU order when reclaiming for lower zones but | |
1586 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1587 | * scanning would soon rescan the same pages to skip and put the | |
1588 | * system at risk of premature OOM. | |
1589 | */ | |
7cc30fcf MG |
1590 | if (!list_empty(&pages_skipped)) { |
1591 | int zid; | |
1592 | ||
3db65812 | 1593 | list_splice(&pages_skipped, src); |
7cc30fcf MG |
1594 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1595 | if (!nr_skipped[zid]) | |
1596 | continue; | |
1597 | ||
1598 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 1599 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
1600 | } |
1601 | } | |
791b48b6 | 1602 | *nr_scanned = total_scan; |
1265e3a6 | 1603 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
791b48b6 | 1604 | total_scan, skipped, nr_taken, mode, lru); |
b4536f0c | 1605 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
1606 | return nr_taken; |
1607 | } | |
1608 | ||
62695a84 NP |
1609 | /** |
1610 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1611 | * @page: page to isolate from its LRU list | |
1612 | * | |
1613 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1614 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1615 | * | |
1616 | * Returns 0 if the page was removed from an LRU list. | |
1617 | * Returns -EBUSY if the page was not on an LRU list. | |
1618 | * | |
1619 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1620 | * the active list, it will have PageActive set. If it was found on |
1621 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1622 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1623 | * |
1624 | * The vmstat statistic corresponding to the list on which the page was | |
1625 | * found will be decremented. | |
1626 | * | |
1627 | * Restrictions: | |
1628 | * (1) Must be called with an elevated refcount on the page. This is a | |
1629 | * fundamentnal difference from isolate_lru_pages (which is called | |
1630 | * without a stable reference). | |
1631 | * (2) the lru_lock must not be held. | |
1632 | * (3) interrupts must be enabled. | |
1633 | */ | |
1634 | int isolate_lru_page(struct page *page) | |
1635 | { | |
1636 | int ret = -EBUSY; | |
1637 | ||
309381fe | 1638 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1639 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1640 | |
62695a84 NP |
1641 | if (PageLRU(page)) { |
1642 | struct zone *zone = page_zone(page); | |
fa9add64 | 1643 | struct lruvec *lruvec; |
62695a84 | 1644 | |
a52633d8 | 1645 | spin_lock_irq(zone_lru_lock(zone)); |
599d0c95 | 1646 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
0c917313 | 1647 | if (PageLRU(page)) { |
894bc310 | 1648 | int lru = page_lru(page); |
0c917313 | 1649 | get_page(page); |
62695a84 | 1650 | ClearPageLRU(page); |
fa9add64 HD |
1651 | del_page_from_lru_list(page, lruvec, lru); |
1652 | ret = 0; | |
62695a84 | 1653 | } |
a52633d8 | 1654 | spin_unlock_irq(zone_lru_lock(zone)); |
62695a84 NP |
1655 | } |
1656 | return ret; | |
1657 | } | |
1658 | ||
35cd7815 | 1659 | /* |
d37dd5dc FW |
1660 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1661 | * then get resheduled. When there are massive number of tasks doing page | |
1662 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1663 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1664 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1665 | */ |
599d0c95 | 1666 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1667 | struct scan_control *sc) |
1668 | { | |
1669 | unsigned long inactive, isolated; | |
1670 | ||
1671 | if (current_is_kswapd()) | |
1672 | return 0; | |
1673 | ||
97c9341f | 1674 | if (!sane_reclaim(sc)) |
35cd7815 RR |
1675 | return 0; |
1676 | ||
1677 | if (file) { | |
599d0c95 MG |
1678 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1679 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1680 | } else { |
599d0c95 MG |
1681 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1682 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1683 | } |
1684 | ||
3cf23841 FW |
1685 | /* |
1686 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1687 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1688 | * deadlock. | |
1689 | */ | |
d0164adc | 1690 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1691 | inactive >>= 3; |
1692 | ||
35cd7815 RR |
1693 | return isolated > inactive; |
1694 | } | |
1695 | ||
66635629 | 1696 | static noinline_for_stack void |
75b00af7 | 1697 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1698 | { |
27ac81d8 | 1699 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
599d0c95 | 1700 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3f79768f | 1701 | LIST_HEAD(pages_to_free); |
66635629 | 1702 | |
66635629 MG |
1703 | /* |
1704 | * Put back any unfreeable pages. | |
1705 | */ | |
66635629 | 1706 | while (!list_empty(page_list)) { |
3f79768f | 1707 | struct page *page = lru_to_page(page_list); |
66635629 | 1708 | int lru; |
3f79768f | 1709 | |
309381fe | 1710 | VM_BUG_ON_PAGE(PageLRU(page), page); |
66635629 | 1711 | list_del(&page->lru); |
39b5f29a | 1712 | if (unlikely(!page_evictable(page))) { |
599d0c95 | 1713 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1714 | putback_lru_page(page); |
599d0c95 | 1715 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1716 | continue; |
1717 | } | |
fa9add64 | 1718 | |
599d0c95 | 1719 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1720 | |
7a608572 | 1721 | SetPageLRU(page); |
66635629 | 1722 | lru = page_lru(page); |
fa9add64 HD |
1723 | add_page_to_lru_list(page, lruvec, lru); |
1724 | ||
66635629 MG |
1725 | if (is_active_lru(lru)) { |
1726 | int file = is_file_lru(lru); | |
9992af10 RR |
1727 | int numpages = hpage_nr_pages(page); |
1728 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1729 | } |
2bcf8879 HD |
1730 | if (put_page_testzero(page)) { |
1731 | __ClearPageLRU(page); | |
1732 | __ClearPageActive(page); | |
fa9add64 | 1733 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1734 | |
1735 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1736 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1737 | mem_cgroup_uncharge(page); |
2bcf8879 | 1738 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1739 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1740 | } else |
1741 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1742 | } |
1743 | } | |
66635629 | 1744 | |
3f79768f HD |
1745 | /* |
1746 | * To save our caller's stack, now use input list for pages to free. | |
1747 | */ | |
1748 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1749 | } |
1750 | ||
399ba0b9 N |
1751 | /* |
1752 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1753 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1754 | * In that case we should only throttle if the backing device it is | |
1755 | * writing to is congested. In other cases it is safe to throttle. | |
1756 | */ | |
1757 | static int current_may_throttle(void) | |
1758 | { | |
1759 | return !(current->flags & PF_LESS_THROTTLE) || | |
1760 | current->backing_dev_info == NULL || | |
1761 | bdi_write_congested(current->backing_dev_info); | |
1762 | } | |
1763 | ||
1da177e4 | 1764 | /* |
b2e18757 | 1765 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1766 | * of reclaimed pages |
1da177e4 | 1767 | */ |
66635629 | 1768 | static noinline_for_stack unsigned long |
1a93be0e | 1769 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1770 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1771 | { |
1772 | LIST_HEAD(page_list); | |
e247dbce | 1773 | unsigned long nr_scanned; |
05ff5137 | 1774 | unsigned long nr_reclaimed = 0; |
e247dbce | 1775 | unsigned long nr_taken; |
3c710c1a | 1776 | struct reclaim_stat stat = {}; |
f3fd4a61 | 1777 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1778 | int file = is_file_lru(lru); |
599d0c95 | 1779 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1780 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
db73ee0d | 1781 | bool stalled = false; |
78dc583d | 1782 | |
599d0c95 | 1783 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
1784 | if (stalled) |
1785 | return 0; | |
1786 | ||
1787 | /* wait a bit for the reclaimer. */ | |
1788 | msleep(100); | |
1789 | stalled = true; | |
35cd7815 RR |
1790 | |
1791 | /* We are about to die and free our memory. Return now. */ | |
1792 | if (fatal_signal_pending(current)) | |
1793 | return SWAP_CLUSTER_MAX; | |
1794 | } | |
1795 | ||
1da177e4 | 1796 | lru_add_drain(); |
f80c0673 MK |
1797 | |
1798 | if (!sc->may_unmap) | |
61317289 | 1799 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1800 | |
599d0c95 | 1801 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1802 | |
5dc35979 KK |
1803 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1804 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc | 1805 | |
599d0c95 | 1806 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1807 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1808 | |
2262185c RG |
1809 | if (current_is_kswapd()) { |
1810 | if (global_reclaim(sc)) | |
599d0c95 | 1811 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
2262185c RG |
1812 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD, |
1813 | nr_scanned); | |
1814 | } else { | |
1815 | if (global_reclaim(sc)) | |
599d0c95 | 1816 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
2262185c RG |
1817 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT, |
1818 | nr_scanned); | |
e247dbce | 1819 | } |
599d0c95 | 1820 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1821 | |
d563c050 | 1822 | if (nr_taken == 0) |
66635629 | 1823 | return 0; |
5ad333eb | 1824 | |
a128ca71 | 1825 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, |
3c710c1a | 1826 | &stat, false); |
c661b078 | 1827 | |
599d0c95 | 1828 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 1829 | |
2262185c RG |
1830 | if (current_is_kswapd()) { |
1831 | if (global_reclaim(sc)) | |
599d0c95 | 1832 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
2262185c RG |
1833 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD, |
1834 | nr_reclaimed); | |
1835 | } else { | |
1836 | if (global_reclaim(sc)) | |
599d0c95 | 1837 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
2262185c RG |
1838 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT, |
1839 | nr_reclaimed); | |
904249aa | 1840 | } |
a74609fa | 1841 | |
27ac81d8 | 1842 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1843 | |
599d0c95 | 1844 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 1845 | |
599d0c95 | 1846 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 1847 | |
747db954 | 1848 | mem_cgroup_uncharge_list(&page_list); |
b745bc85 | 1849 | free_hot_cold_page_list(&page_list, true); |
e11da5b4 | 1850 | |
92df3a72 MG |
1851 | /* |
1852 | * If reclaim is isolating dirty pages under writeback, it implies | |
1853 | * that the long-lived page allocation rate is exceeding the page | |
1854 | * laundering rate. Either the global limits are not being effective | |
1855 | * at throttling processes due to the page distribution throughout | |
1856 | * zones or there is heavy usage of a slow backing device. The | |
1857 | * only option is to throttle from reclaim context which is not ideal | |
1858 | * as there is no guarantee the dirtying process is throttled in the | |
1859 | * same way balance_dirty_pages() manages. | |
1860 | * | |
8e950282 MG |
1861 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1862 | * of pages under pages flagged for immediate reclaim and stall if any | |
1863 | * are encountered in the nr_immediate check below. | |
92df3a72 | 1864 | */ |
3c710c1a | 1865 | if (stat.nr_writeback && stat.nr_writeback == nr_taken) |
599d0c95 | 1866 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); |
92df3a72 | 1867 | |
b8b81518 AR |
1868 | /* |
1869 | * If dirty pages are scanned that are not queued for IO, it | |
1870 | * implies that flushers are not doing their job. This can | |
1871 | * happen when memory pressure pushes dirty pages to the end of | |
1872 | * the LRU before the dirty limits are breached and the dirty | |
1873 | * data has expired. It can also happen when the proportion of | |
1874 | * dirty pages grows not through writes but through memory | |
1875 | * pressure reclaiming all the clean cache. And in some cases, | |
1876 | * the flushers simply cannot keep up with the allocation | |
1877 | * rate. Nudge the flusher threads in case they are asleep. | |
1878 | */ | |
1879 | if (stat.nr_unqueued_dirty == nr_taken) | |
1880 | wakeup_flusher_threads(0, WB_REASON_VMSCAN); | |
1881 | ||
d43006d5 | 1882 | /* |
97c9341f TH |
1883 | * Legacy memcg will stall in page writeback so avoid forcibly |
1884 | * stalling here. | |
d43006d5 | 1885 | */ |
97c9341f | 1886 | if (sane_reclaim(sc)) { |
8e950282 MG |
1887 | /* |
1888 | * Tag a zone as congested if all the dirty pages scanned were | |
1889 | * backed by a congested BDI and wait_iff_congested will stall. | |
1890 | */ | |
3c710c1a | 1891 | if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested) |
599d0c95 | 1892 | set_bit(PGDAT_CONGESTED, &pgdat->flags); |
8e950282 | 1893 | |
b8b81518 AR |
1894 | /* Allow kswapd to start writing pages during reclaim. */ |
1895 | if (stat.nr_unqueued_dirty == nr_taken) | |
599d0c95 | 1896 | set_bit(PGDAT_DIRTY, &pgdat->flags); |
b1a6f21e MG |
1897 | |
1898 | /* | |
b738d764 LT |
1899 | * If kswapd scans pages marked marked for immediate |
1900 | * reclaim and under writeback (nr_immediate), it implies | |
1901 | * that pages are cycling through the LRU faster than | |
b1a6f21e MG |
1902 | * they are written so also forcibly stall. |
1903 | */ | |
3c710c1a | 1904 | if (stat.nr_immediate && current_may_throttle()) |
b1a6f21e | 1905 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
e2be15f6 | 1906 | } |
d43006d5 | 1907 | |
8e950282 MG |
1908 | /* |
1909 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1910 | * is congested. Allow kswapd to continue until it starts encountering | |
1911 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1912 | */ | |
399ba0b9 N |
1913 | if (!sc->hibernation_mode && !current_is_kswapd() && |
1914 | current_may_throttle()) | |
599d0c95 | 1915 | wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); |
8e950282 | 1916 | |
599d0c95 MG |
1917 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
1918 | nr_scanned, nr_reclaimed, | |
5bccd166 MH |
1919 | stat.nr_dirty, stat.nr_writeback, |
1920 | stat.nr_congested, stat.nr_immediate, | |
1921 | stat.nr_activate, stat.nr_ref_keep, | |
1922 | stat.nr_unmap_fail, | |
ba5e9579 | 1923 | sc->priority, file); |
05ff5137 | 1924 | return nr_reclaimed; |
1da177e4 LT |
1925 | } |
1926 | ||
1927 | /* | |
1928 | * This moves pages from the active list to the inactive list. | |
1929 | * | |
1930 | * We move them the other way if the page is referenced by one or more | |
1931 | * processes, from rmap. | |
1932 | * | |
1933 | * If the pages are mostly unmapped, the processing is fast and it is | |
a52633d8 | 1934 | * appropriate to hold zone_lru_lock across the whole operation. But if |
1da177e4 | 1935 | * the pages are mapped, the processing is slow (page_referenced()) so we |
a52633d8 | 1936 | * should drop zone_lru_lock around each page. It's impossible to balance |
1da177e4 LT |
1937 | * this, so instead we remove the pages from the LRU while processing them. |
1938 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1939 | * nobody will play with that bit on a non-LRU page. | |
1940 | * | |
0139aa7b | 1941 | * The downside is that we have to touch page->_refcount against each page. |
1da177e4 | 1942 | * But we had to alter page->flags anyway. |
9d998b4f MH |
1943 | * |
1944 | * Returns the number of pages moved to the given lru. | |
1da177e4 | 1945 | */ |
1cfb419b | 1946 | |
9d998b4f | 1947 | static unsigned move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1948 | struct list_head *list, |
2bcf8879 | 1949 | struct list_head *pages_to_free, |
3eb4140f WF |
1950 | enum lru_list lru) |
1951 | { | |
599d0c95 | 1952 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3eb4140f | 1953 | struct page *page; |
fa9add64 | 1954 | int nr_pages; |
9d998b4f | 1955 | int nr_moved = 0; |
3eb4140f | 1956 | |
3eb4140f WF |
1957 | while (!list_empty(list)) { |
1958 | page = lru_to_page(list); | |
599d0c95 | 1959 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
3eb4140f | 1960 | |
309381fe | 1961 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3eb4140f WF |
1962 | SetPageLRU(page); |
1963 | ||
fa9add64 | 1964 | nr_pages = hpage_nr_pages(page); |
599d0c95 | 1965 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
925b7673 | 1966 | list_move(&page->lru, &lruvec->lists[lru]); |
3eb4140f | 1967 | |
2bcf8879 HD |
1968 | if (put_page_testzero(page)) { |
1969 | __ClearPageLRU(page); | |
1970 | __ClearPageActive(page); | |
fa9add64 | 1971 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1972 | |
1973 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1974 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1975 | mem_cgroup_uncharge(page); |
2bcf8879 | 1976 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1977 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1978 | } else |
1979 | list_add(&page->lru, pages_to_free); | |
9d998b4f MH |
1980 | } else { |
1981 | nr_moved += nr_pages; | |
3eb4140f WF |
1982 | } |
1983 | } | |
9d5e6a9f | 1984 | |
2262185c | 1985 | if (!is_active_lru(lru)) { |
f0958906 | 1986 | __count_vm_events(PGDEACTIVATE, nr_moved); |
2262185c RG |
1987 | count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, |
1988 | nr_moved); | |
1989 | } | |
9d998b4f MH |
1990 | |
1991 | return nr_moved; | |
3eb4140f | 1992 | } |
1cfb419b | 1993 | |
f626012d | 1994 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1995 | struct lruvec *lruvec, |
f16015fb | 1996 | struct scan_control *sc, |
9e3b2f8c | 1997 | enum lru_list lru) |
1da177e4 | 1998 | { |
44c241f1 | 1999 | unsigned long nr_taken; |
f626012d | 2000 | unsigned long nr_scanned; |
6fe6b7e3 | 2001 | unsigned long vm_flags; |
1da177e4 | 2002 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 2003 | LIST_HEAD(l_active); |
b69408e8 | 2004 | LIST_HEAD(l_inactive); |
1da177e4 | 2005 | struct page *page; |
1a93be0e | 2006 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
9d998b4f MH |
2007 | unsigned nr_deactivate, nr_activate; |
2008 | unsigned nr_rotated = 0; | |
f3fd4a61 | 2009 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 2010 | int file = is_file_lru(lru); |
599d0c95 | 2011 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2012 | |
2013 | lru_add_drain(); | |
f80c0673 MK |
2014 | |
2015 | if (!sc->may_unmap) | |
61317289 | 2016 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 2017 | |
599d0c95 | 2018 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 2019 | |
5dc35979 KK |
2020 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
2021 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 2022 | |
599d0c95 | 2023 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 2024 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 2025 | |
599d0c95 | 2026 | __count_vm_events(PGREFILL, nr_scanned); |
2262185c | 2027 | count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2028 | |
599d0c95 | 2029 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 2030 | |
1da177e4 LT |
2031 | while (!list_empty(&l_hold)) { |
2032 | cond_resched(); | |
2033 | page = lru_to_page(&l_hold); | |
2034 | list_del(&page->lru); | |
7e9cd484 | 2035 | |
39b5f29a | 2036 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
2037 | putback_lru_page(page); |
2038 | continue; | |
2039 | } | |
2040 | ||
cc715d99 MG |
2041 | if (unlikely(buffer_heads_over_limit)) { |
2042 | if (page_has_private(page) && trylock_page(page)) { | |
2043 | if (page_has_private(page)) | |
2044 | try_to_release_page(page, 0); | |
2045 | unlock_page(page); | |
2046 | } | |
2047 | } | |
2048 | ||
c3ac9a8a JW |
2049 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
2050 | &vm_flags)) { | |
9992af10 | 2051 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
2052 | /* |
2053 | * Identify referenced, file-backed active pages and | |
2054 | * give them one more trip around the active list. So | |
2055 | * that executable code get better chances to stay in | |
2056 | * memory under moderate memory pressure. Anon pages | |
2057 | * are not likely to be evicted by use-once streaming | |
2058 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
2059 | * so we ignore them here. | |
2060 | */ | |
41e20983 | 2061 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
2062 | list_add(&page->lru, &l_active); |
2063 | continue; | |
2064 | } | |
2065 | } | |
7e9cd484 | 2066 | |
5205e56e | 2067 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
2068 | list_add(&page->lru, &l_inactive); |
2069 | } | |
2070 | ||
b555749a | 2071 | /* |
8cab4754 | 2072 | * Move pages back to the lru list. |
b555749a | 2073 | */ |
599d0c95 | 2074 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 2075 | /* |
8cab4754 WF |
2076 | * Count referenced pages from currently used mappings as rotated, |
2077 | * even though only some of them are actually re-activated. This | |
2078 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 2079 | * get_scan_count. |
7e9cd484 | 2080 | */ |
b7c46d15 | 2081 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 2082 | |
9d998b4f MH |
2083 | nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
2084 | nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
599d0c95 MG |
2085 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
2086 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 2087 | |
747db954 | 2088 | mem_cgroup_uncharge_list(&l_hold); |
b745bc85 | 2089 | free_hot_cold_page_list(&l_hold, true); |
9d998b4f MH |
2090 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2091 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2092 | } |
2093 | ||
59dc76b0 RR |
2094 | /* |
2095 | * The inactive anon list should be small enough that the VM never has | |
2096 | * to do too much work. | |
14797e23 | 2097 | * |
59dc76b0 RR |
2098 | * The inactive file list should be small enough to leave most memory |
2099 | * to the established workingset on the scan-resistant active list, | |
2100 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2101 | * |
59dc76b0 RR |
2102 | * Both inactive lists should also be large enough that each inactive |
2103 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2104 | * |
2a2e4885 JW |
2105 | * If that fails and refaulting is observed, the inactive list grows. |
2106 | * | |
59dc76b0 RR |
2107 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
2108 | * on this LRU, maintained by the pageout code. A zone->inactive_ratio | |
2109 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. | |
56e49d21 | 2110 | * |
59dc76b0 RR |
2111 | * total target max |
2112 | * memory ratio inactive | |
2113 | * ------------------------------------- | |
2114 | * 10MB 1 5MB | |
2115 | * 100MB 1 50MB | |
2116 | * 1GB 3 250MB | |
2117 | * 10GB 10 0.9GB | |
2118 | * 100GB 31 3GB | |
2119 | * 1TB 101 10GB | |
2120 | * 10TB 320 32GB | |
56e49d21 | 2121 | */ |
f8d1a311 | 2122 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
2fc44624 | 2123 | struct scan_control *sc, bool trace) |
56e49d21 | 2124 | { |
fd538803 | 2125 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; |
2a2e4885 JW |
2126 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
2127 | enum lru_list inactive_lru = file * LRU_FILE; | |
2128 | unsigned long inactive, active; | |
2129 | unsigned long inactive_ratio; | |
2130 | unsigned long refaults; | |
59dc76b0 | 2131 | unsigned long gb; |
e3790144 | 2132 | |
59dc76b0 RR |
2133 | /* |
2134 | * If we don't have swap space, anonymous page deactivation | |
2135 | * is pointless. | |
2136 | */ | |
2137 | if (!file && !total_swap_pages) | |
2138 | return false; | |
56e49d21 | 2139 | |
fd538803 MH |
2140 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
2141 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); | |
f8d1a311 | 2142 | |
2a2e4885 JW |
2143 | /* |
2144 | * When refaults are being observed, it means a new workingset | |
2145 | * is being established. Disable active list protection to get | |
2146 | * rid of the stale workingset quickly. | |
2147 | */ | |
aeceeab3 | 2148 | refaults = lruvec_page_state(lruvec, WORKINGSET_ACTIVATE); |
2fc44624 | 2149 | if (file && lruvec->refaults != refaults) { |
2a2e4885 JW |
2150 | inactive_ratio = 0; |
2151 | } else { | |
2152 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
2153 | if (gb) | |
2154 | inactive_ratio = int_sqrt(10 * gb); | |
2155 | else | |
2156 | inactive_ratio = 1; | |
2157 | } | |
59dc76b0 | 2158 | |
2fc44624 | 2159 | if (trace) |
2a2e4885 JW |
2160 | trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, |
2161 | lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, | |
2162 | lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, | |
2163 | inactive_ratio, file); | |
fd538803 | 2164 | |
59dc76b0 | 2165 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2166 | } |
2167 | ||
4f98a2fe | 2168 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
aeceeab3 | 2169 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 2170 | { |
b39415b2 | 2171 | if (is_active_lru(lru)) { |
aeceeab3 | 2172 | if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true)) |
1a93be0e | 2173 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
2174 | return 0; |
2175 | } | |
2176 | ||
1a93be0e | 2177 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
2178 | } |
2179 | ||
9a265114 JW |
2180 | enum scan_balance { |
2181 | SCAN_EQUAL, | |
2182 | SCAN_FRACT, | |
2183 | SCAN_ANON, | |
2184 | SCAN_FILE, | |
2185 | }; | |
2186 | ||
4f98a2fe RR |
2187 | /* |
2188 | * Determine how aggressively the anon and file LRU lists should be | |
2189 | * scanned. The relative value of each set of LRU lists is determined | |
2190 | * by looking at the fraction of the pages scanned we did rotate back | |
2191 | * onto the active list instead of evict. | |
2192 | * | |
be7bd59d WL |
2193 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2194 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2195 | */ |
33377678 | 2196 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
6b4f7799 JW |
2197 | struct scan_control *sc, unsigned long *nr, |
2198 | unsigned long *lru_pages) | |
4f98a2fe | 2199 | { |
33377678 | 2200 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
2201 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2202 | u64 fraction[2]; | |
2203 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 2204 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 2205 | unsigned long anon_prio, file_prio; |
9a265114 | 2206 | enum scan_balance scan_balance; |
0bf1457f | 2207 | unsigned long anon, file; |
4f98a2fe | 2208 | unsigned long ap, fp; |
4111304d | 2209 | enum lru_list lru; |
76a33fc3 SL |
2210 | |
2211 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2212 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2213 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2214 | goto out; |
2215 | } | |
4f98a2fe | 2216 | |
10316b31 JW |
2217 | /* |
2218 | * Global reclaim will swap to prevent OOM even with no | |
2219 | * swappiness, but memcg users want to use this knob to | |
2220 | * disable swapping for individual groups completely when | |
2221 | * using the memory controller's swap limit feature would be | |
2222 | * too expensive. | |
2223 | */ | |
02695175 | 2224 | if (!global_reclaim(sc) && !swappiness) { |
9a265114 | 2225 | scan_balance = SCAN_FILE; |
10316b31 JW |
2226 | goto out; |
2227 | } | |
2228 | ||
2229 | /* | |
2230 | * Do not apply any pressure balancing cleverness when the | |
2231 | * system is close to OOM, scan both anon and file equally | |
2232 | * (unless the swappiness setting disagrees with swapping). | |
2233 | */ | |
02695175 | 2234 | if (!sc->priority && swappiness) { |
9a265114 | 2235 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2236 | goto out; |
2237 | } | |
2238 | ||
62376251 JW |
2239 | /* |
2240 | * Prevent the reclaimer from falling into the cache trap: as | |
2241 | * cache pages start out inactive, every cache fault will tip | |
2242 | * the scan balance towards the file LRU. And as the file LRU | |
2243 | * shrinks, so does the window for rotation from references. | |
2244 | * This means we have a runaway feedback loop where a tiny | |
2245 | * thrashing file LRU becomes infinitely more attractive than | |
2246 | * anon pages. Try to detect this based on file LRU size. | |
2247 | */ | |
2248 | if (global_reclaim(sc)) { | |
599d0c95 MG |
2249 | unsigned long pgdatfile; |
2250 | unsigned long pgdatfree; | |
2251 | int z; | |
2252 | unsigned long total_high_wmark = 0; | |
2ab051e1 | 2253 | |
599d0c95 MG |
2254 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2255 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2256 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2257 | ||
2258 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2259 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2260 | if (!managed_zone(zone)) |
599d0c95 MG |
2261 | continue; |
2262 | ||
2263 | total_high_wmark += high_wmark_pages(zone); | |
2264 | } | |
62376251 | 2265 | |
599d0c95 | 2266 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
06226226 DR |
2267 | /* |
2268 | * Force SCAN_ANON if there are enough inactive | |
2269 | * anonymous pages on the LRU in eligible zones. | |
2270 | * Otherwise, the small LRU gets thrashed. | |
2271 | */ | |
aeceeab3 | 2272 | if (!inactive_list_is_low(lruvec, false, sc, false) && |
06226226 DR |
2273 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx) |
2274 | >> sc->priority) { | |
2275 | scan_balance = SCAN_ANON; | |
2276 | goto out; | |
2277 | } | |
62376251 JW |
2278 | } |
2279 | } | |
2280 | ||
7c5bd705 | 2281 | /* |
316bda0e VD |
2282 | * If there is enough inactive page cache, i.e. if the size of the |
2283 | * inactive list is greater than that of the active list *and* the | |
2284 | * inactive list actually has some pages to scan on this priority, we | |
2285 | * do not reclaim anything from the anonymous working set right now. | |
2286 | * Without the second condition we could end up never scanning an | |
2287 | * lruvec even if it has plenty of old anonymous pages unless the | |
2288 | * system is under heavy pressure. | |
7c5bd705 | 2289 | */ |
aeceeab3 | 2290 | if (!inactive_list_is_low(lruvec, true, sc, false) && |
71ab6cfe | 2291 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
9a265114 | 2292 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2293 | goto out; |
2294 | } | |
2295 | ||
9a265114 JW |
2296 | scan_balance = SCAN_FRACT; |
2297 | ||
58c37f6e KM |
2298 | /* |
2299 | * With swappiness at 100, anonymous and file have the same priority. | |
2300 | * This scanning priority is essentially the inverse of IO cost. | |
2301 | */ | |
02695175 | 2302 | anon_prio = swappiness; |
75b00af7 | 2303 | file_prio = 200 - anon_prio; |
58c37f6e | 2304 | |
4f98a2fe RR |
2305 | /* |
2306 | * OK, so we have swap space and a fair amount of page cache | |
2307 | * pages. We use the recently rotated / recently scanned | |
2308 | * ratios to determine how valuable each cache is. | |
2309 | * | |
2310 | * Because workloads change over time (and to avoid overflow) | |
2311 | * we keep these statistics as a floating average, which ends | |
2312 | * up weighing recent references more than old ones. | |
2313 | * | |
2314 | * anon in [0], file in [1] | |
2315 | */ | |
2ab051e1 | 2316 | |
fd538803 MH |
2317 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
2318 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); | |
2319 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + | |
2320 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); | |
2ab051e1 | 2321 | |
599d0c95 | 2322 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2323 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2324 | reclaim_stat->recent_scanned[0] /= 2; |
2325 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2326 | } |
2327 | ||
6e901571 | 2328 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2329 | reclaim_stat->recent_scanned[1] /= 2; |
2330 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2331 | } |
2332 | ||
4f98a2fe | 2333 | /* |
00d8089c RR |
2334 | * The amount of pressure on anon vs file pages is inversely |
2335 | * proportional to the fraction of recently scanned pages on | |
2336 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2337 | */ |
fe35004f | 2338 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2339 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2340 | |
fe35004f | 2341 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2342 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2343 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2344 | |
76a33fc3 SL |
2345 | fraction[0] = ap; |
2346 | fraction[1] = fp; | |
2347 | denominator = ap + fp + 1; | |
2348 | out: | |
688035f7 JW |
2349 | *lru_pages = 0; |
2350 | for_each_evictable_lru(lru) { | |
2351 | int file = is_file_lru(lru); | |
2352 | unsigned long size; | |
2353 | unsigned long scan; | |
6b4f7799 | 2354 | |
688035f7 JW |
2355 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); |
2356 | scan = size >> sc->priority; | |
2357 | /* | |
2358 | * If the cgroup's already been deleted, make sure to | |
2359 | * scrape out the remaining cache. | |
2360 | */ | |
2361 | if (!scan && !mem_cgroup_online(memcg)) | |
2362 | scan = min(size, SWAP_CLUSTER_MAX); | |
6b4f7799 | 2363 | |
688035f7 JW |
2364 | switch (scan_balance) { |
2365 | case SCAN_EQUAL: | |
2366 | /* Scan lists relative to size */ | |
2367 | break; | |
2368 | case SCAN_FRACT: | |
9a265114 | 2369 | /* |
688035f7 JW |
2370 | * Scan types proportional to swappiness and |
2371 | * their relative recent reclaim efficiency. | |
e3880aaa GS |
2372 | * Make sure we don't miss the last page on |
2373 | * the offlined memory cgroups because of a | |
2374 | * round-off error. | |
9a265114 | 2375 | */ |
e3880aaa GS |
2376 | scan = mem_cgroup_online(memcg) ? |
2377 | div64_u64(scan * fraction[file], denominator) : | |
2378 | DIV64_U64_ROUND_UP(scan * fraction[file], | |
36f93a2e | 2379 | denominator); |
688035f7 JW |
2380 | break; |
2381 | case SCAN_FILE: | |
2382 | case SCAN_ANON: | |
2383 | /* Scan one type exclusively */ | |
2384 | if ((scan_balance == SCAN_FILE) != file) { | |
2385 | size = 0; | |
2386 | scan = 0; | |
2387 | } | |
2388 | break; | |
2389 | default: | |
2390 | /* Look ma, no brain */ | |
2391 | BUG(); | |
9a265114 | 2392 | } |
688035f7 JW |
2393 | |
2394 | *lru_pages += size; | |
2395 | nr[lru] = scan; | |
76a33fc3 | 2396 | } |
6e08a369 | 2397 | } |
4f98a2fe | 2398 | |
9b4f98cd | 2399 | /* |
a9dd0a83 | 2400 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. |
9b4f98cd | 2401 | */ |
a9dd0a83 | 2402 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, |
33377678 | 2403 | struct scan_control *sc, unsigned long *lru_pages) |
9b4f98cd | 2404 | { |
ef8f2327 | 2405 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
9b4f98cd | 2406 | unsigned long nr[NR_LRU_LISTS]; |
e82e0561 | 2407 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2408 | unsigned long nr_to_scan; |
2409 | enum lru_list lru; | |
2410 | unsigned long nr_reclaimed = 0; | |
2411 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2412 | struct blk_plug plug; | |
1a501907 | 2413 | bool scan_adjusted; |
9b4f98cd | 2414 | |
33377678 | 2415 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
9b4f98cd | 2416 | |
e82e0561 MG |
2417 | /* Record the original scan target for proportional adjustments later */ |
2418 | memcpy(targets, nr, sizeof(nr)); | |
2419 | ||
1a501907 MG |
2420 | /* |
2421 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2422 | * event that can occur when there is little memory pressure e.g. | |
2423 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2424 | * when the requested number of pages are reclaimed when scanning at | |
2425 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2426 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2427 | * do a batch of work at once. For memcg reclaim one check is made to | |
2428 | * abort proportional reclaim if either the file or anon lru has already | |
2429 | * dropped to zero at the first pass. | |
2430 | */ | |
2431 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2432 | sc->priority == DEF_PRIORITY); | |
2433 | ||
9b4f98cd JW |
2434 | blk_start_plug(&plug); |
2435 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2436 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2437 | unsigned long nr_anon, nr_file, percentage; |
2438 | unsigned long nr_scanned; | |
2439 | ||
9b4f98cd JW |
2440 | for_each_evictable_lru(lru) { |
2441 | if (nr[lru]) { | |
2442 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2443 | nr[lru] -= nr_to_scan; | |
2444 | ||
2445 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
aeceeab3 | 2446 | lruvec, sc); |
9b4f98cd JW |
2447 | } |
2448 | } | |
e82e0561 | 2449 | |
bd041733 MH |
2450 | cond_resched(); |
2451 | ||
e82e0561 MG |
2452 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2453 | continue; | |
2454 | ||
e82e0561 MG |
2455 | /* |
2456 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2457 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2458 | * proportionally what was requested by get_scan_count(). We |
2459 | * stop reclaiming one LRU and reduce the amount scanning | |
2460 | * proportional to the original scan target. | |
2461 | */ | |
2462 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2463 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2464 | ||
1a501907 MG |
2465 | /* |
2466 | * It's just vindictive to attack the larger once the smaller | |
2467 | * has gone to zero. And given the way we stop scanning the | |
2468 | * smaller below, this makes sure that we only make one nudge | |
2469 | * towards proportionality once we've got nr_to_reclaim. | |
2470 | */ | |
2471 | if (!nr_file || !nr_anon) | |
2472 | break; | |
2473 | ||
e82e0561 MG |
2474 | if (nr_file > nr_anon) { |
2475 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2476 | targets[LRU_ACTIVE_ANON] + 1; | |
2477 | lru = LRU_BASE; | |
2478 | percentage = nr_anon * 100 / scan_target; | |
2479 | } else { | |
2480 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2481 | targets[LRU_ACTIVE_FILE] + 1; | |
2482 | lru = LRU_FILE; | |
2483 | percentage = nr_file * 100 / scan_target; | |
2484 | } | |
2485 | ||
2486 | /* Stop scanning the smaller of the LRU */ | |
2487 | nr[lru] = 0; | |
2488 | nr[lru + LRU_ACTIVE] = 0; | |
2489 | ||
2490 | /* | |
2491 | * Recalculate the other LRU scan count based on its original | |
2492 | * scan target and the percentage scanning already complete | |
2493 | */ | |
2494 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2495 | nr_scanned = targets[lru] - nr[lru]; | |
2496 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2497 | nr[lru] -= min(nr[lru], nr_scanned); | |
2498 | ||
2499 | lru += LRU_ACTIVE; | |
2500 | nr_scanned = targets[lru] - nr[lru]; | |
2501 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2502 | nr[lru] -= min(nr[lru], nr_scanned); | |
2503 | ||
2504 | scan_adjusted = true; | |
9b4f98cd JW |
2505 | } |
2506 | blk_finish_plug(&plug); | |
2507 | sc->nr_reclaimed += nr_reclaimed; | |
2508 | ||
2509 | /* | |
2510 | * Even if we did not try to evict anon pages at all, we want to | |
2511 | * rebalance the anon lru active/inactive ratio. | |
2512 | */ | |
aeceeab3 | 2513 | if (inactive_list_is_low(lruvec, false, sc, true)) |
9b4f98cd JW |
2514 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2515 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
2516 | } |
2517 | ||
23b9da55 | 2518 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2519 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2520 | { |
d84da3f9 | 2521 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2522 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2523 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2524 | return true; |
2525 | ||
2526 | return false; | |
2527 | } | |
2528 | ||
3e7d3449 | 2529 | /* |
23b9da55 MG |
2530 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2531 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2532 | * true if more pages should be reclaimed such that when the page allocator | |
2533 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2534 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2535 | */ |
a9dd0a83 | 2536 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 MG |
2537 | unsigned long nr_reclaimed, |
2538 | unsigned long nr_scanned, | |
2539 | struct scan_control *sc) | |
2540 | { | |
2541 | unsigned long pages_for_compaction; | |
2542 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2543 | int z; |
3e7d3449 MG |
2544 | |
2545 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2546 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2547 | return false; |
2548 | ||
2876592f | 2549 | /* Consider stopping depending on scan and reclaim activity */ |
dcda9b04 | 2550 | if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) { |
2876592f | 2551 | /* |
dcda9b04 | 2552 | * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the |
2876592f MG |
2553 | * full LRU list has been scanned and we are still failing |
2554 | * to reclaim pages. This full LRU scan is potentially | |
dcda9b04 | 2555 | * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed |
2876592f MG |
2556 | */ |
2557 | if (!nr_reclaimed && !nr_scanned) | |
2558 | return false; | |
2559 | } else { | |
2560 | /* | |
dcda9b04 | 2561 | * For non-__GFP_RETRY_MAYFAIL allocations which can presumably |
2876592f MG |
2562 | * fail without consequence, stop if we failed to reclaim |
2563 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2564 | * pages that were scanned. This will return to the | |
2565 | * caller faster at the risk reclaim/compaction and | |
2566 | * the resulting allocation attempt fails | |
2567 | */ | |
2568 | if (!nr_reclaimed) | |
2569 | return false; | |
2570 | } | |
3e7d3449 MG |
2571 | |
2572 | /* | |
2573 | * If we have not reclaimed enough pages for compaction and the | |
2574 | * inactive lists are large enough, continue reclaiming | |
2575 | */ | |
9861a62c | 2576 | pages_for_compaction = compact_gap(sc->order); |
a9dd0a83 | 2577 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
ec8acf20 | 2578 | if (get_nr_swap_pages() > 0) |
a9dd0a83 | 2579 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
3e7d3449 MG |
2580 | if (sc->nr_reclaimed < pages_for_compaction && |
2581 | inactive_lru_pages > pages_for_compaction) | |
2582 | return true; | |
2583 | ||
2584 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
a9dd0a83 MG |
2585 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2586 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2587 | if (!managed_zone(zone)) |
a9dd0a83 MG |
2588 | continue; |
2589 | ||
2590 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 2591 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
2592 | case COMPACT_CONTINUE: |
2593 | return false; | |
2594 | default: | |
2595 | /* check next zone */ | |
2596 | ; | |
2597 | } | |
3e7d3449 | 2598 | } |
a9dd0a83 | 2599 | return true; |
3e7d3449 MG |
2600 | } |
2601 | ||
970a39a3 | 2602 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2603 | { |
cb731d6c | 2604 | struct reclaim_state *reclaim_state = current->reclaim_state; |
f0fdc5e8 | 2605 | unsigned long nr_reclaimed, nr_scanned; |
2344d7e4 | 2606 | bool reclaimable = false; |
1da177e4 | 2607 | |
9b4f98cd JW |
2608 | do { |
2609 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2610 | struct mem_cgroup_reclaim_cookie reclaim = { | |
ef8f2327 | 2611 | .pgdat = pgdat, |
9b4f98cd JW |
2612 | .priority = sc->priority, |
2613 | }; | |
a9dd0a83 | 2614 | unsigned long node_lru_pages = 0; |
694fbc0f | 2615 | struct mem_cgroup *memcg; |
3e7d3449 | 2616 | |
9b4f98cd JW |
2617 | nr_reclaimed = sc->nr_reclaimed; |
2618 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2619 | |
694fbc0f AM |
2620 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2621 | do { | |
6b4f7799 | 2622 | unsigned long lru_pages; |
8e8ae645 | 2623 | unsigned long reclaimed; |
cb731d6c | 2624 | unsigned long scanned; |
5660048c | 2625 | |
241994ed | 2626 | if (mem_cgroup_low(root, memcg)) { |
d6622f63 YX |
2627 | if (!sc->memcg_low_reclaim) { |
2628 | sc->memcg_low_skipped = 1; | |
241994ed | 2629 | continue; |
d6622f63 | 2630 | } |
31176c78 | 2631 | mem_cgroup_event(memcg, MEMCG_LOW); |
241994ed JW |
2632 | } |
2633 | ||
8e8ae645 | 2634 | reclaimed = sc->nr_reclaimed; |
cb731d6c | 2635 | scanned = sc->nr_scanned; |
f9be23d6 | 2636 | |
a9dd0a83 MG |
2637 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
2638 | node_lru_pages += lru_pages; | |
f16015fb | 2639 | |
b5afba29 | 2640 | if (memcg) |
a9dd0a83 | 2641 | shrink_slab(sc->gfp_mask, pgdat->node_id, |
cb731d6c VD |
2642 | memcg, sc->nr_scanned - scanned, |
2643 | lru_pages); | |
2644 | ||
8e8ae645 JW |
2645 | /* Record the group's reclaim efficiency */ |
2646 | vmpressure(sc->gfp_mask, memcg, false, | |
2647 | sc->nr_scanned - scanned, | |
2648 | sc->nr_reclaimed - reclaimed); | |
2649 | ||
9b4f98cd | 2650 | /* |
a394cb8e MH |
2651 | * Direct reclaim and kswapd have to scan all memory |
2652 | * cgroups to fulfill the overall scan target for the | |
a9dd0a83 | 2653 | * node. |
a394cb8e MH |
2654 | * |
2655 | * Limit reclaim, on the other hand, only cares about | |
2656 | * nr_to_reclaim pages to be reclaimed and it will | |
2657 | * retry with decreasing priority if one round over the | |
2658 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2659 | */ |
a394cb8e MH |
2660 | if (!global_reclaim(sc) && |
2661 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2662 | mem_cgroup_iter_break(root, memcg); |
2663 | break; | |
2664 | } | |
241994ed | 2665 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
70ddf637 | 2666 | |
6b4f7799 JW |
2667 | /* |
2668 | * Shrink the slab caches in the same proportion that | |
2669 | * the eligible LRU pages were scanned. | |
2670 | */ | |
b2e18757 | 2671 | if (global_reclaim(sc)) |
a9dd0a83 | 2672 | shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, |
cb731d6c | 2673 | sc->nr_scanned - nr_scanned, |
a9dd0a83 | 2674 | node_lru_pages); |
cb731d6c VD |
2675 | |
2676 | if (reclaim_state) { | |
2677 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2678 | reclaim_state->reclaimed_slab = 0; | |
6b4f7799 JW |
2679 | } |
2680 | ||
8e8ae645 JW |
2681 | /* Record the subtree's reclaim efficiency */ |
2682 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
70ddf637 AV |
2683 | sc->nr_scanned - nr_scanned, |
2684 | sc->nr_reclaimed - nr_reclaimed); | |
2685 | ||
2344d7e4 JW |
2686 | if (sc->nr_reclaimed - nr_reclaimed) |
2687 | reclaimable = true; | |
2688 | ||
a9dd0a83 | 2689 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
9b4f98cd | 2690 | sc->nr_scanned - nr_scanned, sc)); |
2344d7e4 | 2691 | |
c73322d0 JW |
2692 | /* |
2693 | * Kswapd gives up on balancing particular nodes after too | |
2694 | * many failures to reclaim anything from them and goes to | |
2695 | * sleep. On reclaim progress, reset the failure counter. A | |
2696 | * successful direct reclaim run will revive a dormant kswapd. | |
2697 | */ | |
2698 | if (reclaimable) | |
2699 | pgdat->kswapd_failures = 0; | |
2700 | ||
2344d7e4 | 2701 | return reclaimable; |
f16015fb JW |
2702 | } |
2703 | ||
53853e2d | 2704 | /* |
fdd4c614 VB |
2705 | * Returns true if compaction should go ahead for a costly-order request, or |
2706 | * the allocation would already succeed without compaction. Return false if we | |
2707 | * should reclaim first. | |
53853e2d | 2708 | */ |
4f588331 | 2709 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2710 | { |
31483b6a | 2711 | unsigned long watermark; |
fdd4c614 | 2712 | enum compact_result suitable; |
fe4b1b24 | 2713 | |
fdd4c614 VB |
2714 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2715 | if (suitable == COMPACT_SUCCESS) | |
2716 | /* Allocation should succeed already. Don't reclaim. */ | |
2717 | return true; | |
2718 | if (suitable == COMPACT_SKIPPED) | |
2719 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2720 | return false; | |
fe4b1b24 | 2721 | |
53853e2d | 2722 | /* |
fdd4c614 VB |
2723 | * Compaction is already possible, but it takes time to run and there |
2724 | * are potentially other callers using the pages just freed. So proceed | |
2725 | * with reclaim to make a buffer of free pages available to give | |
2726 | * compaction a reasonable chance of completing and allocating the page. | |
2727 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2728 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2729 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 2730 | */ |
fdd4c614 | 2731 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 2732 | |
fdd4c614 | 2733 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
2734 | } |
2735 | ||
1da177e4 LT |
2736 | /* |
2737 | * This is the direct reclaim path, for page-allocating processes. We only | |
2738 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2739 | * request. | |
2740 | * | |
1da177e4 LT |
2741 | * If a zone is deemed to be full of pinned pages then just give it a light |
2742 | * scan then give up on it. | |
2743 | */ | |
0a0337e0 | 2744 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2745 | { |
dd1a239f | 2746 | struct zoneref *z; |
54a6eb5c | 2747 | struct zone *zone; |
0608f43d AM |
2748 | unsigned long nr_soft_reclaimed; |
2749 | unsigned long nr_soft_scanned; | |
619d0d76 | 2750 | gfp_t orig_mask; |
79dafcdc | 2751 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2752 | |
cc715d99 MG |
2753 | /* |
2754 | * If the number of buffer_heads in the machine exceeds the maximum | |
2755 | * allowed level, force direct reclaim to scan the highmem zone as | |
2756 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2757 | */ | |
619d0d76 | 2758 | orig_mask = sc->gfp_mask; |
b2e18757 | 2759 | if (buffer_heads_over_limit) { |
cc715d99 | 2760 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2761 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2762 | } |
cc715d99 | 2763 | |
d4debc66 | 2764 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2765 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
2766 | /* |
2767 | * Take care memory controller reclaiming has small influence | |
2768 | * to global LRU. | |
2769 | */ | |
89b5fae5 | 2770 | if (global_reclaim(sc)) { |
344736f2 VD |
2771 | if (!cpuset_zone_allowed(zone, |
2772 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2773 | continue; |
65ec02cb | 2774 | |
0b06496a JW |
2775 | /* |
2776 | * If we already have plenty of memory free for | |
2777 | * compaction in this zone, don't free any more. | |
2778 | * Even though compaction is invoked for any | |
2779 | * non-zero order, only frequent costly order | |
2780 | * reclamation is disruptive enough to become a | |
2781 | * noticeable problem, like transparent huge | |
2782 | * page allocations. | |
2783 | */ | |
2784 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2785 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2786 | compaction_ready(zone, sc)) { |
0b06496a JW |
2787 | sc->compaction_ready = true; |
2788 | continue; | |
e0887c19 | 2789 | } |
0b06496a | 2790 | |
79dafcdc MG |
2791 | /* |
2792 | * Shrink each node in the zonelist once. If the | |
2793 | * zonelist is ordered by zone (not the default) then a | |
2794 | * node may be shrunk multiple times but in that case | |
2795 | * the user prefers lower zones being preserved. | |
2796 | */ | |
2797 | if (zone->zone_pgdat == last_pgdat) | |
2798 | continue; | |
2799 | ||
0608f43d AM |
2800 | /* |
2801 | * This steals pages from memory cgroups over softlimit | |
2802 | * and returns the number of reclaimed pages and | |
2803 | * scanned pages. This works for global memory pressure | |
2804 | * and balancing, not for a memcg's limit. | |
2805 | */ | |
2806 | nr_soft_scanned = 0; | |
ef8f2327 | 2807 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2808 | sc->order, sc->gfp_mask, |
2809 | &nr_soft_scanned); | |
2810 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2811 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2812 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2813 | } |
408d8544 | 2814 | |
79dafcdc MG |
2815 | /* See comment about same check for global reclaim above */ |
2816 | if (zone->zone_pgdat == last_pgdat) | |
2817 | continue; | |
2818 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2819 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2820 | } |
e0c23279 | 2821 | |
619d0d76 WY |
2822 | /* |
2823 | * Restore to original mask to avoid the impact on the caller if we | |
2824 | * promoted it to __GFP_HIGHMEM. | |
2825 | */ | |
2826 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2827 | } |
4f98a2fe | 2828 | |
2a2e4885 JW |
2829 | static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) |
2830 | { | |
2831 | struct mem_cgroup *memcg; | |
2832 | ||
2833 | memcg = mem_cgroup_iter(root_memcg, NULL, NULL); | |
2834 | do { | |
2835 | unsigned long refaults; | |
2836 | struct lruvec *lruvec; | |
2837 | ||
2a2e4885 | 2838 | lruvec = mem_cgroup_lruvec(pgdat, memcg); |
aeceeab3 | 2839 | refaults = lruvec_page_state(lruvec, WORKINGSET_ACTIVATE); |
2a2e4885 JW |
2840 | lruvec->refaults = refaults; |
2841 | } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); | |
2842 | } | |
2843 | ||
1da177e4 LT |
2844 | /* |
2845 | * This is the main entry point to direct page reclaim. | |
2846 | * | |
2847 | * If a full scan of the inactive list fails to free enough memory then we | |
2848 | * are "out of memory" and something needs to be killed. | |
2849 | * | |
2850 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2851 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2852 | * caller can't do much about. We kick the writeback threads and take explicit |
2853 | * naps in the hope that some of these pages can be written. But if the | |
2854 | * allocating task holds filesystem locks which prevent writeout this might not | |
2855 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2856 | * |
2857 | * returns: 0, if no pages reclaimed | |
2858 | * else, the number of pages reclaimed | |
1da177e4 | 2859 | */ |
dac1d27b | 2860 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2861 | struct scan_control *sc) |
1da177e4 | 2862 | { |
241994ed | 2863 | int initial_priority = sc->priority; |
2a2e4885 JW |
2864 | pg_data_t *last_pgdat; |
2865 | struct zoneref *z; | |
2866 | struct zone *zone; | |
241994ed | 2867 | retry: |
873b4771 KK |
2868 | delayacct_freepages_start(); |
2869 | ||
89b5fae5 | 2870 | if (global_reclaim(sc)) |
7cc30fcf | 2871 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 2872 | |
9e3b2f8c | 2873 | do { |
70ddf637 AV |
2874 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2875 | sc->priority); | |
66e1707b | 2876 | sc->nr_scanned = 0; |
0a0337e0 | 2877 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 2878 | |
bb21c7ce | 2879 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
2880 | break; |
2881 | ||
2882 | if (sc->compaction_ready) | |
2883 | break; | |
1da177e4 | 2884 | |
0e50ce3b MK |
2885 | /* |
2886 | * If we're getting trouble reclaiming, start doing | |
2887 | * writepage even in laptop mode. | |
2888 | */ | |
2889 | if (sc->priority < DEF_PRIORITY - 2) | |
2890 | sc->may_writepage = 1; | |
0b06496a | 2891 | } while (--sc->priority >= 0); |
bb21c7ce | 2892 | |
2a2e4885 JW |
2893 | last_pgdat = NULL; |
2894 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
2895 | sc->nodemask) { | |
2896 | if (zone->zone_pgdat == last_pgdat) | |
2897 | continue; | |
2898 | last_pgdat = zone->zone_pgdat; | |
2899 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); | |
2900 | } | |
2901 | ||
873b4771 KK |
2902 | delayacct_freepages_end(); |
2903 | ||
bb21c7ce KM |
2904 | if (sc->nr_reclaimed) |
2905 | return sc->nr_reclaimed; | |
2906 | ||
0cee34fd | 2907 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 2908 | if (sc->compaction_ready) |
7335084d MG |
2909 | return 1; |
2910 | ||
241994ed | 2911 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 2912 | if (sc->memcg_low_skipped) { |
241994ed | 2913 | sc->priority = initial_priority; |
d6622f63 YX |
2914 | sc->memcg_low_reclaim = 1; |
2915 | sc->memcg_low_skipped = 0; | |
241994ed JW |
2916 | goto retry; |
2917 | } | |
2918 | ||
bb21c7ce | 2919 | return 0; |
1da177e4 LT |
2920 | } |
2921 | ||
c73322d0 | 2922 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
2923 | { |
2924 | struct zone *zone; | |
2925 | unsigned long pfmemalloc_reserve = 0; | |
2926 | unsigned long free_pages = 0; | |
2927 | int i; | |
2928 | bool wmark_ok; | |
2929 | ||
c73322d0 JW |
2930 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
2931 | return true; | |
2932 | ||
5515061d MG |
2933 | for (i = 0; i <= ZONE_NORMAL; i++) { |
2934 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
2935 | if (!managed_zone(zone)) |
2936 | continue; | |
2937 | ||
2938 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
2939 | continue; |
2940 | ||
5515061d MG |
2941 | pfmemalloc_reserve += min_wmark_pages(zone); |
2942 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2943 | } | |
2944 | ||
675becce MG |
2945 | /* If there are no reserves (unexpected config) then do not throttle */ |
2946 | if (!pfmemalloc_reserve) | |
2947 | return true; | |
2948 | ||
5515061d MG |
2949 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
2950 | ||
2951 | /* kswapd must be awake if processes are being throttled */ | |
2952 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
38087d9b | 2953 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
5515061d MG |
2954 | (enum zone_type)ZONE_NORMAL); |
2955 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2956 | } | |
2957 | ||
2958 | return wmark_ok; | |
2959 | } | |
2960 | ||
2961 | /* | |
2962 | * Throttle direct reclaimers if backing storage is backed by the network | |
2963 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2964 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2965 | * when the low watermark is reached. |
2966 | * | |
2967 | * Returns true if a fatal signal was delivered during throttling. If this | |
2968 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2969 | */ |
50694c28 | 2970 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2971 | nodemask_t *nodemask) |
2972 | { | |
675becce | 2973 | struct zoneref *z; |
5515061d | 2974 | struct zone *zone; |
675becce | 2975 | pg_data_t *pgdat = NULL; |
5515061d MG |
2976 | |
2977 | /* | |
2978 | * Kernel threads should not be throttled as they may be indirectly | |
2979 | * responsible for cleaning pages necessary for reclaim to make forward | |
2980 | * progress. kjournald for example may enter direct reclaim while | |
2981 | * committing a transaction where throttling it could forcing other | |
2982 | * processes to block on log_wait_commit(). | |
2983 | */ | |
2984 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2985 | goto out; |
2986 | ||
2987 | /* | |
2988 | * If a fatal signal is pending, this process should not throttle. | |
2989 | * It should return quickly so it can exit and free its memory | |
2990 | */ | |
2991 | if (fatal_signal_pending(current)) | |
2992 | goto out; | |
5515061d | 2993 | |
675becce MG |
2994 | /* |
2995 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2996 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2997 | * GFP_KERNEL will be required for allocating network buffers when | |
2998 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2999 | * | |
3000 | * Throttling is based on the first usable node and throttled processes | |
3001 | * wait on a queue until kswapd makes progress and wakes them. There | |
3002 | * is an affinity then between processes waking up and where reclaim | |
3003 | * progress has been made assuming the process wakes on the same node. | |
3004 | * More importantly, processes running on remote nodes will not compete | |
3005 | * for remote pfmemalloc reserves and processes on different nodes | |
3006 | * should make reasonable progress. | |
3007 | */ | |
3008 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 3009 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
3010 | if (zone_idx(zone) > ZONE_NORMAL) |
3011 | continue; | |
3012 | ||
3013 | /* Throttle based on the first usable node */ | |
3014 | pgdat = zone->zone_pgdat; | |
c73322d0 | 3015 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
3016 | goto out; |
3017 | break; | |
3018 | } | |
3019 | ||
3020 | /* If no zone was usable by the allocation flags then do not throttle */ | |
3021 | if (!pgdat) | |
50694c28 | 3022 | goto out; |
5515061d | 3023 | |
68243e76 MG |
3024 | /* Account for the throttling */ |
3025 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
3026 | ||
5515061d MG |
3027 | /* |
3028 | * If the caller cannot enter the filesystem, it's possible that it | |
3029 | * is due to the caller holding an FS lock or performing a journal | |
3030 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
3031 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
3032 | * blocked waiting on the same lock. Instead, throttle for up to a | |
3033 | * second before continuing. | |
3034 | */ | |
3035 | if (!(gfp_mask & __GFP_FS)) { | |
3036 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
c73322d0 | 3037 | allow_direct_reclaim(pgdat), HZ); |
50694c28 MG |
3038 | |
3039 | goto check_pending; | |
5515061d MG |
3040 | } |
3041 | ||
3042 | /* Throttle until kswapd wakes the process */ | |
3043 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
c73322d0 | 3044 | allow_direct_reclaim(pgdat)); |
50694c28 MG |
3045 | |
3046 | check_pending: | |
3047 | if (fatal_signal_pending(current)) | |
3048 | return true; | |
3049 | ||
3050 | out: | |
3051 | return false; | |
5515061d MG |
3052 | } |
3053 | ||
dac1d27b | 3054 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 3055 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 3056 | { |
33906bc5 | 3057 | unsigned long nr_reclaimed; |
66e1707b | 3058 | struct scan_control sc = { |
ee814fe2 | 3059 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 3060 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 3061 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
3062 | .order = order, |
3063 | .nodemask = nodemask, | |
3064 | .priority = DEF_PRIORITY, | |
66e1707b | 3065 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3066 | .may_unmap = 1, |
2e2e4259 | 3067 | .may_swap = 1, |
66e1707b BS |
3068 | }; |
3069 | ||
5515061d | 3070 | /* |
50694c28 MG |
3071 | * Do not enter reclaim if fatal signal was delivered while throttled. |
3072 | * 1 is returned so that the page allocator does not OOM kill at this | |
3073 | * point. | |
5515061d | 3074 | */ |
f2f43e56 | 3075 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
3076 | return 1; |
3077 | ||
33906bc5 MG |
3078 | trace_mm_vmscan_direct_reclaim_begin(order, |
3079 | sc.may_writepage, | |
f2f43e56 | 3080 | sc.gfp_mask, |
e5146b12 | 3081 | sc.reclaim_idx); |
33906bc5 | 3082 | |
3115cd91 | 3083 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
3084 | |
3085 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
3086 | ||
3087 | return nr_reclaimed; | |
66e1707b BS |
3088 | } |
3089 | ||
c255a458 | 3090 | #ifdef CONFIG_MEMCG |
66e1707b | 3091 | |
a9dd0a83 | 3092 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 3093 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 3094 | pg_data_t *pgdat, |
0ae5e89c | 3095 | unsigned long *nr_scanned) |
4e416953 BS |
3096 | { |
3097 | struct scan_control sc = { | |
b8f5c566 | 3098 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 3099 | .target_mem_cgroup = memcg, |
4e416953 BS |
3100 | .may_writepage = !laptop_mode, |
3101 | .may_unmap = 1, | |
b2e18757 | 3102 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 3103 | .may_swap = !noswap, |
4e416953 | 3104 | }; |
6b4f7799 | 3105 | unsigned long lru_pages; |
0ae5e89c | 3106 | |
4e416953 BS |
3107 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3108 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3109 | |
9e3b2f8c | 3110 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e | 3111 | sc.may_writepage, |
e5146b12 MG |
3112 | sc.gfp_mask, |
3113 | sc.reclaim_idx); | |
bdce6d9e | 3114 | |
4e416953 BS |
3115 | /* |
3116 | * NOTE: Although we can get the priority field, using it | |
3117 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3118 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3119 | * will pick up pages from other mem cgroup's as well. We hack |
3120 | * the priority and make it zero. | |
3121 | */ | |
ef8f2327 | 3122 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
bdce6d9e KM |
3123 | |
3124 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3125 | ||
0ae5e89c | 3126 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
3127 | return sc.nr_reclaimed; |
3128 | } | |
3129 | ||
72835c86 | 3130 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3131 | unsigned long nr_pages, |
a7885eb8 | 3132 | gfp_t gfp_mask, |
b70a2a21 | 3133 | bool may_swap) |
66e1707b | 3134 | { |
4e416953 | 3135 | struct zonelist *zonelist; |
bdce6d9e | 3136 | unsigned long nr_reclaimed; |
889976db | 3137 | int nid; |
499118e9 | 3138 | unsigned int noreclaim_flag; |
66e1707b | 3139 | struct scan_control sc = { |
b70a2a21 | 3140 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3141 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 3142 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 3143 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3144 | .target_mem_cgroup = memcg, |
3145 | .priority = DEF_PRIORITY, | |
3146 | .may_writepage = !laptop_mode, | |
3147 | .may_unmap = 1, | |
b70a2a21 | 3148 | .may_swap = may_swap, |
a09ed5e0 | 3149 | }; |
66e1707b | 3150 | |
889976db YH |
3151 | /* |
3152 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
3153 | * take care of from where we get pages. So the node where we start the | |
3154 | * scan does not need to be the current node. | |
3155 | */ | |
72835c86 | 3156 | nid = mem_cgroup_select_victim_node(memcg); |
889976db | 3157 | |
c9634cf0 | 3158 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; |
bdce6d9e KM |
3159 | |
3160 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
3161 | sc.may_writepage, | |
e5146b12 MG |
3162 | sc.gfp_mask, |
3163 | sc.reclaim_idx); | |
bdce6d9e | 3164 | |
499118e9 | 3165 | noreclaim_flag = memalloc_noreclaim_save(); |
3115cd91 | 3166 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
499118e9 | 3167 | memalloc_noreclaim_restore(noreclaim_flag); |
bdce6d9e KM |
3168 | |
3169 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
3170 | ||
3171 | return nr_reclaimed; | |
66e1707b BS |
3172 | } |
3173 | #endif | |
3174 | ||
1d82de61 | 3175 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3176 | struct scan_control *sc) |
f16015fb | 3177 | { |
b95a2f2d | 3178 | struct mem_cgroup *memcg; |
f16015fb | 3179 | |
b95a2f2d JW |
3180 | if (!total_swap_pages) |
3181 | return; | |
3182 | ||
3183 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
3184 | do { | |
ef8f2327 | 3185 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
b95a2f2d | 3186 | |
aeceeab3 | 3187 | if (inactive_list_is_low(lruvec, false, sc, true)) |
1a93be0e | 3188 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 3189 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
3190 | |
3191 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
3192 | } while (memcg); | |
f16015fb JW |
3193 | } |
3194 | ||
e716f2eb MG |
3195 | /* |
3196 | * Returns true if there is an eligible zone balanced for the request order | |
3197 | * and classzone_idx | |
3198 | */ | |
3199 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
60cefed4 | 3200 | { |
e716f2eb MG |
3201 | int i; |
3202 | unsigned long mark = -1; | |
3203 | struct zone *zone; | |
60cefed4 | 3204 | |
e716f2eb MG |
3205 | for (i = 0; i <= classzone_idx; i++) { |
3206 | zone = pgdat->node_zones + i; | |
6256c6b4 | 3207 | |
e716f2eb MG |
3208 | if (!managed_zone(zone)) |
3209 | continue; | |
3210 | ||
3211 | mark = high_wmark_pages(zone); | |
3212 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) | |
3213 | return true; | |
3214 | } | |
3215 | ||
3216 | /* | |
3217 | * If a node has no populated zone within classzone_idx, it does not | |
3218 | * need balancing by definition. This can happen if a zone-restricted | |
3219 | * allocation tries to wake a remote kswapd. | |
3220 | */ | |
3221 | if (mark == -1) | |
3222 | return true; | |
3223 | ||
3224 | return false; | |
60cefed4 JW |
3225 | } |
3226 | ||
631b6e08 MG |
3227 | /* Clear pgdat state for congested, dirty or under writeback. */ |
3228 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
3229 | { | |
3230 | clear_bit(PGDAT_CONGESTED, &pgdat->flags); | |
3231 | clear_bit(PGDAT_DIRTY, &pgdat->flags); | |
3232 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
3233 | } | |
3234 | ||
5515061d MG |
3235 | /* |
3236 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3237 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3238 | * | |
3239 | * Returns true if kswapd is ready to sleep | |
3240 | */ | |
d9f21d42 | 3241 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f50de2d3 | 3242 | { |
5515061d | 3243 | /* |
9e5e3661 | 3244 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 3245 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
3246 | * race between when kswapd checks the watermarks and a process gets |
3247 | * throttled. There is also a potential race if processes get | |
3248 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3249 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3250 | * the wake up checks. If kswapd is going to sleep, no process should | |
3251 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3252 | * the wake up is premature, processes will wake kswapd and get | |
3253 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3254 | * that here we are under prepare_to_wait(). | |
5515061d | 3255 | */ |
9e5e3661 VB |
3256 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3257 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3258 | |
c73322d0 JW |
3259 | /* Hopeless node, leave it to direct reclaim */ |
3260 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3261 | return true; | |
3262 | ||
e716f2eb MG |
3263 | if (pgdat_balanced(pgdat, order, classzone_idx)) { |
3264 | clear_pgdat_congested(pgdat); | |
3265 | return true; | |
1d82de61 MG |
3266 | } |
3267 | ||
333b0a45 | 3268 | return false; |
f50de2d3 MG |
3269 | } |
3270 | ||
75485363 | 3271 | /* |
1d82de61 MG |
3272 | * kswapd shrinks a node of pages that are at or below the highest usable |
3273 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3274 | * |
3275 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3276 | * reclaim or if the lack of progress was due to pages under writeback. |
3277 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3278 | */ |
1d82de61 | 3279 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3280 | struct scan_control *sc) |
75485363 | 3281 | { |
1d82de61 MG |
3282 | struct zone *zone; |
3283 | int z; | |
75485363 | 3284 | |
1d82de61 MG |
3285 | /* Reclaim a number of pages proportional to the number of zones */ |
3286 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3287 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 3288 | zone = pgdat->node_zones + z; |
6aa303de | 3289 | if (!managed_zone(zone)) |
1d82de61 | 3290 | continue; |
7c954f6d | 3291 | |
1d82de61 MG |
3292 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3293 | } | |
7c954f6d MG |
3294 | |
3295 | /* | |
1d82de61 MG |
3296 | * Historically care was taken to put equal pressure on all zones but |
3297 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3298 | */ |
970a39a3 | 3299 | shrink_node(pgdat, sc); |
283aba9f | 3300 | |
7c954f6d | 3301 | /* |
1d82de61 MG |
3302 | * Fragmentation may mean that the system cannot be rebalanced for |
3303 | * high-order allocations. If twice the allocation size has been | |
3304 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3305 | * excessive reclaim. Assume that a process requested a high-order | |
3306 | * can direct reclaim/compact. | |
7c954f6d | 3307 | */ |
9861a62c | 3308 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 3309 | sc->order = 0; |
7c954f6d | 3310 | |
b8e83b94 | 3311 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3312 | } |
3313 | ||
1da177e4 | 3314 | /* |
1d82de61 MG |
3315 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3316 | * that are eligible for use by the caller until at least one zone is | |
3317 | * balanced. | |
1da177e4 | 3318 | * |
1d82de61 | 3319 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3320 | * |
3321 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3322 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
1d82de61 MG |
3323 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone |
3324 | * or lower is eligible for reclaim until at least one usable zone is | |
3325 | * balanced. | |
1da177e4 | 3326 | */ |
accf6242 | 3327 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3328 | { |
1da177e4 | 3329 | int i; |
0608f43d AM |
3330 | unsigned long nr_soft_reclaimed; |
3331 | unsigned long nr_soft_scanned; | |
1d82de61 | 3332 | struct zone *zone; |
179e9639 AM |
3333 | struct scan_control sc = { |
3334 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3335 | .order = order, |
b8e83b94 | 3336 | .priority = DEF_PRIORITY, |
ee814fe2 | 3337 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3338 | .may_unmap = 1, |
2e2e4259 | 3339 | .may_swap = 1, |
179e9639 | 3340 | }; |
f8891e5e | 3341 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3342 | |
9e3b2f8c | 3343 | do { |
c73322d0 | 3344 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 MG |
3345 | bool raise_priority = true; |
3346 | ||
84c7a777 | 3347 | sc.reclaim_idx = classzone_idx; |
1da177e4 | 3348 | |
86c79f6b | 3349 | /* |
84c7a777 MG |
3350 | * If the number of buffer_heads exceeds the maximum allowed |
3351 | * then consider reclaiming from all zones. This has a dual | |
3352 | * purpose -- on 64-bit systems it is expected that | |
3353 | * buffer_heads are stripped during active rotation. On 32-bit | |
3354 | * systems, highmem pages can pin lowmem memory and shrinking | |
3355 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3356 | * go ahead if all eligible zones for the original allocation | |
3357 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3358 | */ |
3359 | if (buffer_heads_over_limit) { | |
3360 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3361 | zone = pgdat->node_zones + i; | |
6aa303de | 3362 | if (!managed_zone(zone)) |
86c79f6b | 3363 | continue; |
cc715d99 | 3364 | |
970a39a3 | 3365 | sc.reclaim_idx = i; |
e1dbeda6 | 3366 | break; |
1da177e4 | 3367 | } |
1da177e4 | 3368 | } |
dafcb73e | 3369 | |
86c79f6b | 3370 | /* |
e716f2eb MG |
3371 | * Only reclaim if there are no eligible zones. Note that |
3372 | * sc.reclaim_idx is not used as buffer_heads_over_limit may | |
3373 | * have adjusted it. | |
86c79f6b | 3374 | */ |
e716f2eb MG |
3375 | if (pgdat_balanced(pgdat, sc.order, classzone_idx)) |
3376 | goto out; | |
e1dbeda6 | 3377 | |
1d82de61 MG |
3378 | /* |
3379 | * Do some background aging of the anon list, to give | |
3380 | * pages a chance to be referenced before reclaiming. All | |
3381 | * pages are rotated regardless of classzone as this is | |
3382 | * about consistent aging. | |
3383 | */ | |
ef8f2327 | 3384 | age_active_anon(pgdat, &sc); |
1d82de61 | 3385 | |
b7ea3c41 MG |
3386 | /* |
3387 | * If we're getting trouble reclaiming, start doing writepage | |
3388 | * even in laptop mode. | |
3389 | */ | |
047d72c3 | 3390 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
3391 | sc.may_writepage = 1; |
3392 | ||
1d82de61 MG |
3393 | /* Call soft limit reclaim before calling shrink_node. */ |
3394 | sc.nr_scanned = 0; | |
3395 | nr_soft_scanned = 0; | |
ef8f2327 | 3396 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3397 | sc.gfp_mask, &nr_soft_scanned); |
3398 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3399 | ||
1da177e4 | 3400 | /* |
1d82de61 MG |
3401 | * There should be no need to raise the scanning priority if |
3402 | * enough pages are already being scanned that that high | |
3403 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3404 | */ |
970a39a3 | 3405 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3406 | raise_priority = false; |
5515061d MG |
3407 | |
3408 | /* | |
3409 | * If the low watermark is met there is no need for processes | |
3410 | * to be throttled on pfmemalloc_wait as they should not be | |
3411 | * able to safely make forward progress. Wake them | |
3412 | */ | |
3413 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 3414 | allow_direct_reclaim(pgdat)) |
cfc51155 | 3415 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3416 | |
b8e83b94 MG |
3417 | /* Check if kswapd should be suspending */ |
3418 | if (try_to_freeze() || kthread_should_stop()) | |
3419 | break; | |
8357376d | 3420 | |
73ce02e9 | 3421 | /* |
b8e83b94 MG |
3422 | * Raise priority if scanning rate is too low or there was no |
3423 | * progress in reclaiming pages | |
73ce02e9 | 3424 | */ |
c73322d0 JW |
3425 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
3426 | if (raise_priority || !nr_reclaimed) | |
b8e83b94 | 3427 | sc.priority--; |
1d82de61 | 3428 | } while (sc.priority >= 1); |
1da177e4 | 3429 | |
c73322d0 JW |
3430 | if (!sc.nr_reclaimed) |
3431 | pgdat->kswapd_failures++; | |
3432 | ||
b8e83b94 | 3433 | out: |
2a2e4885 | 3434 | snapshot_refaults(NULL, pgdat); |
0abdee2b | 3435 | /* |
1d82de61 MG |
3436 | * Return the order kswapd stopped reclaiming at as |
3437 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3438 | * entered the allocator slow path while kswapd was awake, order will | |
3439 | * remain at the higher level. | |
0abdee2b | 3440 | */ |
1d82de61 | 3441 | return sc.order; |
1da177e4 LT |
3442 | } |
3443 | ||
e716f2eb | 3444 | /* |
584810d3 SB |
3445 | * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be |
3446 | * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not | |
3447 | * a valid index then either kswapd runs for first time or kswapd couldn't sleep | |
3448 | * after previous reclaim attempt (node is still unbalanced). In that case | |
3449 | * return the zone index of the previous kswapd reclaim cycle. | |
e716f2eb MG |
3450 | */ |
3451 | static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, | |
584810d3 | 3452 | enum zone_type prev_classzone_idx) |
e716f2eb MG |
3453 | { |
3454 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
584810d3 SB |
3455 | return prev_classzone_idx; |
3456 | return pgdat->kswapd_classzone_idx; | |
e716f2eb MG |
3457 | } |
3458 | ||
38087d9b MG |
3459 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3460 | unsigned int classzone_idx) | |
f0bc0a60 KM |
3461 | { |
3462 | long remaining = 0; | |
3463 | DEFINE_WAIT(wait); | |
3464 | ||
3465 | if (freezing(current) || kthread_should_stop()) | |
3466 | return; | |
3467 | ||
3468 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3469 | ||
333b0a45 SG |
3470 | /* |
3471 | * Try to sleep for a short interval. Note that kcompactd will only be | |
3472 | * woken if it is possible to sleep for a short interval. This is | |
3473 | * deliberate on the assumption that if reclaim cannot keep an | |
3474 | * eligible zone balanced that it's also unlikely that compaction will | |
3475 | * succeed. | |
3476 | */ | |
d9f21d42 | 3477 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
fd901c95 VB |
3478 | /* |
3479 | * Compaction records what page blocks it recently failed to | |
3480 | * isolate pages from and skips them in the future scanning. | |
3481 | * When kswapd is going to sleep, it is reasonable to assume | |
3482 | * that pages and compaction may succeed so reset the cache. | |
3483 | */ | |
3484 | reset_isolation_suitable(pgdat); | |
3485 | ||
3486 | /* | |
3487 | * We have freed the memory, now we should compact it to make | |
3488 | * allocation of the requested order possible. | |
3489 | */ | |
38087d9b | 3490 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
fd901c95 | 3491 | |
f0bc0a60 | 3492 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3493 | |
3494 | /* | |
3495 | * If woken prematurely then reset kswapd_classzone_idx and | |
3496 | * order. The values will either be from a wakeup request or | |
3497 | * the previous request that slept prematurely. | |
3498 | */ | |
3499 | if (remaining) { | |
e716f2eb | 3500 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b MG |
3501 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); |
3502 | } | |
3503 | ||
f0bc0a60 KM |
3504 | finish_wait(&pgdat->kswapd_wait, &wait); |
3505 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3506 | } | |
3507 | ||
3508 | /* | |
3509 | * After a short sleep, check if it was a premature sleep. If not, then | |
3510 | * go fully to sleep until explicitly woken up. | |
3511 | */ | |
d9f21d42 MG |
3512 | if (!remaining && |
3513 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
f0bc0a60 KM |
3514 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3515 | ||
3516 | /* | |
3517 | * vmstat counters are not perfectly accurate and the estimated | |
3518 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3519 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3520 | * watermarks being breached while under pressure, we reduce the | |
3521 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3522 | * them before going back to sleep. | |
3523 | */ | |
3524 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3525 | |
3526 | if (!kthread_should_stop()) | |
3527 | schedule(); | |
3528 | ||
f0bc0a60 KM |
3529 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3530 | } else { | |
3531 | if (remaining) | |
3532 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3533 | else | |
3534 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3535 | } | |
3536 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3537 | } | |
3538 | ||
1da177e4 LT |
3539 | /* |
3540 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3541 | * from the init process. |
1da177e4 LT |
3542 | * |
3543 | * This basically trickles out pages so that we have _some_ | |
3544 | * free memory available even if there is no other activity | |
3545 | * that frees anything up. This is needed for things like routing | |
3546 | * etc, where we otherwise might have all activity going on in | |
3547 | * asynchronous contexts that cannot page things out. | |
3548 | * | |
3549 | * If there are applications that are active memory-allocators | |
3550 | * (most normal use), this basically shouldn't matter. | |
3551 | */ | |
3552 | static int kswapd(void *p) | |
3553 | { | |
e716f2eb MG |
3554 | unsigned int alloc_order, reclaim_order; |
3555 | unsigned int classzone_idx = MAX_NR_ZONES - 1; | |
1da177e4 LT |
3556 | pg_data_t *pgdat = (pg_data_t*)p; |
3557 | struct task_struct *tsk = current; | |
f0bc0a60 | 3558 | |
1da177e4 LT |
3559 | struct reclaim_state reclaim_state = { |
3560 | .reclaimed_slab = 0, | |
3561 | }; | |
a70f7302 | 3562 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3563 | |
174596a0 | 3564 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3565 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3566 | current->reclaim_state = &reclaim_state; |
3567 | ||
3568 | /* | |
3569 | * Tell the memory management that we're a "memory allocator", | |
3570 | * and that if we need more memory we should get access to it | |
3571 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3572 | * never get caught in the normal page freeing logic. | |
3573 | * | |
3574 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3575 | * you need a small amount of memory in order to be able to | |
3576 | * page out something else, and this flag essentially protects | |
3577 | * us from recursively trying to free more memory as we're | |
3578 | * trying to free the first piece of memory in the first place). | |
3579 | */ | |
930d9152 | 3580 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3581 | set_freezable(); |
1da177e4 | 3582 | |
e716f2eb MG |
3583 | pgdat->kswapd_order = 0; |
3584 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; | |
1da177e4 | 3585 | for ( ; ; ) { |
6f6313d4 | 3586 | bool ret; |
3e1d1d28 | 3587 | |
e716f2eb MG |
3588 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3589 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); | |
3590 | ||
38087d9b MG |
3591 | kswapd_try_sleep: |
3592 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3593 | classzone_idx); | |
215ddd66 | 3594 | |
38087d9b MG |
3595 | /* Read the new order and classzone_idx */ |
3596 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
584810d3 | 3597 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b | 3598 | pgdat->kswapd_order = 0; |
e716f2eb | 3599 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
1da177e4 | 3600 | |
8fe23e05 DR |
3601 | ret = try_to_freeze(); |
3602 | if (kthread_should_stop()) | |
3603 | break; | |
3604 | ||
3605 | /* | |
3606 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3607 | * after returning from the refrigerator | |
3608 | */ | |
38087d9b MG |
3609 | if (ret) |
3610 | continue; | |
3611 | ||
3612 | /* | |
3613 | * Reclaim begins at the requested order but if a high-order | |
3614 | * reclaim fails then kswapd falls back to reclaiming for | |
3615 | * order-0. If that happens, kswapd will consider sleeping | |
3616 | * for the order it finished reclaiming at (reclaim_order) | |
3617 | * but kcompactd is woken to compact for the original | |
3618 | * request (alloc_order). | |
3619 | */ | |
e5146b12 MG |
3620 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
3621 | alloc_order); | |
d92a8cfc | 3622 | fs_reclaim_acquire(GFP_KERNEL); |
38087d9b | 3623 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
d92a8cfc | 3624 | fs_reclaim_release(GFP_KERNEL); |
38087d9b MG |
3625 | if (reclaim_order < alloc_order) |
3626 | goto kswapd_try_sleep; | |
1da177e4 | 3627 | } |
b0a8cc58 | 3628 | |
71abdc15 | 3629 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
b0a8cc58 | 3630 | current->reclaim_state = NULL; |
71abdc15 | 3631 | |
1da177e4 LT |
3632 | return 0; |
3633 | } | |
3634 | ||
3635 | /* | |
3636 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3637 | */ | |
99504748 | 3638 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3639 | { |
3640 | pg_data_t *pgdat; | |
3641 | ||
6aa303de | 3642 | if (!managed_zone(zone)) |
1da177e4 LT |
3643 | return; |
3644 | ||
344736f2 | 3645 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) |
1da177e4 | 3646 | return; |
88f5acf8 | 3647 | pgdat = zone->zone_pgdat; |
584810d3 SB |
3648 | |
3649 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
3650 | pgdat->kswapd_classzone_idx = classzone_idx; | |
3651 | else | |
3652 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, | |
3653 | classzone_idx); | |
38087d9b | 3654 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
8d0986e2 | 3655 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3656 | return; |
e1a55637 | 3657 | |
c73322d0 JW |
3658 | /* Hopeless node, leave it to direct reclaim */ |
3659 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3660 | return; | |
3661 | ||
e716f2eb MG |
3662 | if (pgdat_balanced(pgdat, order, classzone_idx)) |
3663 | return; | |
88f5acf8 | 3664 | |
e716f2eb | 3665 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order); |
8d0986e2 | 3666 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3667 | } |
3668 | ||
c6f37f12 | 3669 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3670 | /* |
7b51755c | 3671 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3672 | * freed pages. |
3673 | * | |
3674 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3675 | * LRU order by reclaiming preferentially | |
3676 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3677 | */ |
7b51755c | 3678 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3679 | { |
d6277db4 | 3680 | struct reclaim_state reclaim_state; |
d6277db4 | 3681 | struct scan_control sc = { |
ee814fe2 | 3682 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3683 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 3684 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 3685 | .priority = DEF_PRIORITY, |
d6277db4 | 3686 | .may_writepage = 1, |
ee814fe2 JW |
3687 | .may_unmap = 1, |
3688 | .may_swap = 1, | |
7b51755c | 3689 | .hibernation_mode = 1, |
1da177e4 | 3690 | }; |
a09ed5e0 | 3691 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c KM |
3692 | struct task_struct *p = current; |
3693 | unsigned long nr_reclaimed; | |
499118e9 | 3694 | unsigned int noreclaim_flag; |
1da177e4 | 3695 | |
499118e9 | 3696 | noreclaim_flag = memalloc_noreclaim_save(); |
d92a8cfc | 3697 | fs_reclaim_acquire(sc.gfp_mask); |
7b51755c KM |
3698 | reclaim_state.reclaimed_slab = 0; |
3699 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3700 | |
3115cd91 | 3701 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3702 | |
7b51755c | 3703 | p->reclaim_state = NULL; |
d92a8cfc | 3704 | fs_reclaim_release(sc.gfp_mask); |
499118e9 | 3705 | memalloc_noreclaim_restore(noreclaim_flag); |
d6277db4 | 3706 | |
7b51755c | 3707 | return nr_reclaimed; |
1da177e4 | 3708 | } |
c6f37f12 | 3709 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3710 | |
1da177e4 LT |
3711 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3712 | not required for correctness. So if the last cpu in a node goes | |
3713 | away, we get changed to run anywhere: as the first one comes back, | |
3714 | restore their cpu bindings. */ | |
517bbed9 | 3715 | static int kswapd_cpu_online(unsigned int cpu) |
1da177e4 | 3716 | { |
58c0a4a7 | 3717 | int nid; |
1da177e4 | 3718 | |
517bbed9 SAS |
3719 | for_each_node_state(nid, N_MEMORY) { |
3720 | pg_data_t *pgdat = NODE_DATA(nid); | |
3721 | const struct cpumask *mask; | |
a70f7302 | 3722 | |
517bbed9 | 3723 | mask = cpumask_of_node(pgdat->node_id); |
c5f59f08 | 3724 | |
517bbed9 SAS |
3725 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3726 | /* One of our CPUs online: restore mask */ | |
3727 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
1da177e4 | 3728 | } |
517bbed9 | 3729 | return 0; |
1da177e4 | 3730 | } |
1da177e4 | 3731 | |
3218ae14 YG |
3732 | /* |
3733 | * This kswapd start function will be called by init and node-hot-add. | |
3734 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3735 | */ | |
3736 | int kswapd_run(int nid) | |
3737 | { | |
3738 | pg_data_t *pgdat = NODE_DATA(nid); | |
3739 | int ret = 0; | |
3740 | ||
3741 | if (pgdat->kswapd) | |
3742 | return 0; | |
3743 | ||
3744 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3745 | if (IS_ERR(pgdat->kswapd)) { | |
3746 | /* failure at boot is fatal */ | |
c6202adf | 3747 | BUG_ON(system_state < SYSTEM_RUNNING); |
d5dc0ad9 GS |
3748 | pr_err("Failed to start kswapd on node %d\n", nid); |
3749 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3750 | pgdat->kswapd = NULL; |
3218ae14 YG |
3751 | } |
3752 | return ret; | |
3753 | } | |
3754 | ||
8fe23e05 | 3755 | /* |
d8adde17 | 3756 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 3757 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
3758 | */ |
3759 | void kswapd_stop(int nid) | |
3760 | { | |
3761 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3762 | ||
d8adde17 | 3763 | if (kswapd) { |
8fe23e05 | 3764 | kthread_stop(kswapd); |
d8adde17 JL |
3765 | NODE_DATA(nid)->kswapd = NULL; |
3766 | } | |
8fe23e05 DR |
3767 | } |
3768 | ||
1da177e4 LT |
3769 | static int __init kswapd_init(void) |
3770 | { | |
517bbed9 | 3771 | int nid, ret; |
69e05944 | 3772 | |
1da177e4 | 3773 | swap_setup(); |
48fb2e24 | 3774 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3775 | kswapd_run(nid); |
517bbed9 SAS |
3776 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
3777 | "mm/vmscan:online", kswapd_cpu_online, | |
3778 | NULL); | |
3779 | WARN_ON(ret < 0); | |
1da177e4 LT |
3780 | return 0; |
3781 | } | |
3782 | ||
3783 | module_init(kswapd_init) | |
9eeff239 CL |
3784 | |
3785 | #ifdef CONFIG_NUMA | |
3786 | /* | |
a5f5f91d | 3787 | * Node reclaim mode |
9eeff239 | 3788 | * |
a5f5f91d | 3789 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 3790 | * the watermarks. |
9eeff239 | 3791 | */ |
a5f5f91d | 3792 | int node_reclaim_mode __read_mostly; |
9eeff239 | 3793 | |
1b2ffb78 | 3794 | #define RECLAIM_OFF 0 |
7d03431c | 3795 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 3796 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 3797 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 3798 | |
a92f7126 | 3799 | /* |
a5f5f91d | 3800 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
3801 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
3802 | * a zone. | |
3803 | */ | |
a5f5f91d | 3804 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 3805 | |
9614634f | 3806 | /* |
a5f5f91d | 3807 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
3808 | * occur. |
3809 | */ | |
3810 | int sysctl_min_unmapped_ratio = 1; | |
3811 | ||
0ff38490 CL |
3812 | /* |
3813 | * If the number of slab pages in a zone grows beyond this percentage then | |
3814 | * slab reclaim needs to occur. | |
3815 | */ | |
3816 | int sysctl_min_slab_ratio = 5; | |
3817 | ||
11fb9989 | 3818 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 3819 | { |
11fb9989 MG |
3820 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
3821 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
3822 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
3823 | |
3824 | /* | |
3825 | * It's possible for there to be more file mapped pages than | |
3826 | * accounted for by the pages on the file LRU lists because | |
3827 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3828 | */ | |
3829 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3830 | } | |
3831 | ||
3832 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 3833 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 3834 | { |
d031a157 AM |
3835 | unsigned long nr_pagecache_reclaimable; |
3836 | unsigned long delta = 0; | |
90afa5de MG |
3837 | |
3838 | /* | |
95bbc0c7 | 3839 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 3840 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 3841 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
3842 | * a better estimate |
3843 | */ | |
a5f5f91d MG |
3844 | if (node_reclaim_mode & RECLAIM_UNMAP) |
3845 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 3846 | else |
a5f5f91d | 3847 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
3848 | |
3849 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
3850 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
3851 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
3852 | |
3853 | /* Watch for any possible underflows due to delta */ | |
3854 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3855 | delta = nr_pagecache_reclaimable; | |
3856 | ||
3857 | return nr_pagecache_reclaimable - delta; | |
3858 | } | |
3859 | ||
9eeff239 | 3860 | /* |
a5f5f91d | 3861 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 3862 | */ |
a5f5f91d | 3863 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3864 | { |
7fb2d46d | 3865 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3866 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3867 | struct task_struct *p = current; |
3868 | struct reclaim_state reclaim_state; | |
499118e9 | 3869 | unsigned int noreclaim_flag; |
179e9639 | 3870 | struct scan_control sc = { |
62b726c1 | 3871 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 3872 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 3873 | .order = order, |
a5f5f91d MG |
3874 | .priority = NODE_RECLAIM_PRIORITY, |
3875 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
3876 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 3877 | .may_swap = 1, |
f2f43e56 | 3878 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 3879 | }; |
9eeff239 | 3880 | |
9eeff239 | 3881 | cond_resched(); |
d4f7796e | 3882 | /* |
95bbc0c7 | 3883 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 3884 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 3885 | * and RECLAIM_UNMAP. |
d4f7796e | 3886 | */ |
499118e9 VB |
3887 | noreclaim_flag = memalloc_noreclaim_save(); |
3888 | p->flags |= PF_SWAPWRITE; | |
d92a8cfc | 3889 | fs_reclaim_acquire(sc.gfp_mask); |
9eeff239 CL |
3890 | reclaim_state.reclaimed_slab = 0; |
3891 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3892 | |
a5f5f91d | 3893 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 CL |
3894 | /* |
3895 | * Free memory by calling shrink zone with increasing | |
3896 | * priorities until we have enough memory freed. | |
3897 | */ | |
0ff38490 | 3898 | do { |
970a39a3 | 3899 | shrink_node(pgdat, &sc); |
9e3b2f8c | 3900 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 3901 | } |
c84db23c | 3902 | |
9eeff239 | 3903 | p->reclaim_state = NULL; |
d92a8cfc | 3904 | fs_reclaim_release(gfp_mask); |
499118e9 VB |
3905 | current->flags &= ~PF_SWAPWRITE; |
3906 | memalloc_noreclaim_restore(noreclaim_flag); | |
a79311c1 | 3907 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3908 | } |
179e9639 | 3909 | |
a5f5f91d | 3910 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 3911 | { |
d773ed6b | 3912 | int ret; |
179e9639 AM |
3913 | |
3914 | /* | |
a5f5f91d | 3915 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 3916 | * slab pages if we are over the defined limits. |
34aa1330 | 3917 | * |
9614634f CL |
3918 | * A small portion of unmapped file backed pages is needed for |
3919 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
3920 | * thrown out if the node is overallocated. So we do not reclaim |
3921 | * if less than a specified percentage of the node is used by | |
9614634f | 3922 | * unmapped file backed pages. |
179e9639 | 3923 | */ |
a5f5f91d | 3924 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
385386cf | 3925 | node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) |
a5f5f91d | 3926 | return NODE_RECLAIM_FULL; |
179e9639 AM |
3927 | |
3928 | /* | |
d773ed6b | 3929 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3930 | */ |
d0164adc | 3931 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 3932 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
3933 | |
3934 | /* | |
a5f5f91d | 3935 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
3936 | * have associated processors. This will favor the local processor |
3937 | * over remote processors and spread off node memory allocations | |
3938 | * as wide as possible. | |
3939 | */ | |
a5f5f91d MG |
3940 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
3941 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 3942 | |
a5f5f91d MG |
3943 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
3944 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 3945 | |
a5f5f91d MG |
3946 | ret = __node_reclaim(pgdat, gfp_mask, order); |
3947 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 3948 | |
24cf7251 MG |
3949 | if (!ret) |
3950 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3951 | ||
d773ed6b | 3952 | return ret; |
179e9639 | 3953 | } |
9eeff239 | 3954 | #endif |
894bc310 | 3955 | |
894bc310 LS |
3956 | /* |
3957 | * page_evictable - test whether a page is evictable | |
3958 | * @page: the page to test | |
894bc310 LS |
3959 | * |
3960 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3961 | * lists vs unevictable list. |
894bc310 LS |
3962 | * |
3963 | * Reasons page might not be evictable: | |
ba9ddf49 | 3964 | * (1) page's mapping marked unevictable |
b291f000 | 3965 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3966 | * |
894bc310 | 3967 | */ |
39b5f29a | 3968 | int page_evictable(struct page *page) |
894bc310 | 3969 | { |
8d700626 HY |
3970 | int ret; |
3971 | ||
3972 | /* Prevent address_space of inode and swap cache from being freed */ | |
3973 | rcu_read_lock(); | |
3974 | ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); | |
3975 | rcu_read_unlock(); | |
3976 | return ret; | |
894bc310 | 3977 | } |
89e004ea | 3978 | |
85046579 | 3979 | #ifdef CONFIG_SHMEM |
89e004ea | 3980 | /** |
24513264 HD |
3981 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3982 | * @pages: array of pages to check | |
3983 | * @nr_pages: number of pages to check | |
89e004ea | 3984 | * |
24513264 | 3985 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3986 | * |
3987 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3988 | */ |
24513264 | 3989 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3990 | { |
925b7673 | 3991 | struct lruvec *lruvec; |
785b99fe | 3992 | struct pglist_data *pgdat = NULL; |
24513264 HD |
3993 | int pgscanned = 0; |
3994 | int pgrescued = 0; | |
3995 | int i; | |
89e004ea | 3996 | |
24513264 HD |
3997 | for (i = 0; i < nr_pages; i++) { |
3998 | struct page *page = pages[i]; | |
785b99fe | 3999 | struct pglist_data *pagepgdat = page_pgdat(page); |
89e004ea | 4000 | |
24513264 | 4001 | pgscanned++; |
785b99fe MG |
4002 | if (pagepgdat != pgdat) { |
4003 | if (pgdat) | |
4004 | spin_unlock_irq(&pgdat->lru_lock); | |
4005 | pgdat = pagepgdat; | |
4006 | spin_lock_irq(&pgdat->lru_lock); | |
24513264 | 4007 | } |
785b99fe | 4008 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
89e004ea | 4009 | |
24513264 HD |
4010 | if (!PageLRU(page) || !PageUnevictable(page)) |
4011 | continue; | |
89e004ea | 4012 | |
39b5f29a | 4013 | if (page_evictable(page)) { |
24513264 HD |
4014 | enum lru_list lru = page_lru_base_type(page); |
4015 | ||
309381fe | 4016 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 4017 | ClearPageUnevictable(page); |
fa9add64 HD |
4018 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
4019 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 4020 | pgrescued++; |
89e004ea | 4021 | } |
24513264 | 4022 | } |
89e004ea | 4023 | |
785b99fe | 4024 | if (pgdat) { |
24513264 HD |
4025 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
4026 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
785b99fe | 4027 | spin_unlock_irq(&pgdat->lru_lock); |
89e004ea | 4028 | } |
89e004ea | 4029 | } |
85046579 | 4030 | #endif /* CONFIG_SHMEM */ |