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> |
6dbf8323 | 52 | #include <linux/psi.h> |
1da177e4 LT |
53 | |
54 | #include <asm/tlbflush.h> | |
55 | #include <asm/div64.h> | |
56 | ||
57 | #include <linux/swapops.h> | |
117aad1e | 58 | #include <linux/balloon_compaction.h> |
1da177e4 | 59 | |
0f8053a5 NP |
60 | #include "internal.h" |
61 | ||
33906bc5 MG |
62 | #define CREATE_TRACE_POINTS |
63 | #include <trace/events/vmscan.h> | |
64 | ||
1da177e4 | 65 | struct scan_control { |
22fba335 KM |
66 | /* How many pages shrink_list() should reclaim */ |
67 | unsigned long nr_to_reclaim; | |
68 | ||
1da177e4 | 69 | /* This context's GFP mask */ |
6daa0e28 | 70 | gfp_t gfp_mask; |
1da177e4 | 71 | |
ee814fe2 | 72 | /* Allocation order */ |
5ad333eb | 73 | int order; |
66e1707b | 74 | |
ee814fe2 JW |
75 | /* |
76 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
77 | * are scanned. | |
78 | */ | |
79 | nodemask_t *nodemask; | |
9e3b2f8c | 80 | |
f16015fb JW |
81 | /* |
82 | * The memory cgroup that hit its limit and as a result is the | |
83 | * primary target of this reclaim invocation. | |
84 | */ | |
85 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 86 | |
ee814fe2 JW |
87 | /* Scan (total_size >> priority) pages at once */ |
88 | int priority; | |
89 | ||
b2e18757 MG |
90 | /* The highest zone to isolate pages for reclaim from */ |
91 | enum zone_type reclaim_idx; | |
92 | ||
1276ad68 | 93 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
94 | unsigned int may_writepage:1; |
95 | ||
96 | /* Can mapped pages be reclaimed? */ | |
97 | unsigned int may_unmap:1; | |
98 | ||
99 | /* Can pages be swapped as part of reclaim? */ | |
100 | unsigned int may_swap:1; | |
101 | ||
d6622f63 YX |
102 | /* |
103 | * Cgroups are not reclaimed below their configured memory.low, | |
104 | * unless we threaten to OOM. If any cgroups are skipped due to | |
105 | * memory.low and nothing was reclaimed, go back for memory.low. | |
106 | */ | |
107 | unsigned int memcg_low_reclaim:1; | |
108 | unsigned int memcg_low_skipped:1; | |
241994ed | 109 | |
ee814fe2 JW |
110 | unsigned int hibernation_mode:1; |
111 | ||
112 | /* One of the zones is ready for compaction */ | |
113 | unsigned int compaction_ready:1; | |
114 | ||
115 | /* Incremented by the number of inactive pages that were scanned */ | |
116 | unsigned long nr_scanned; | |
117 | ||
118 | /* Number of pages freed so far during a call to shrink_zones() */ | |
119 | unsigned long nr_reclaimed; | |
1da177e4 LT |
120 | }; |
121 | ||
1da177e4 LT |
122 | #ifdef ARCH_HAS_PREFETCH |
123 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
124 | do { \ | |
125 | if ((_page)->lru.prev != _base) { \ | |
126 | struct page *prev; \ | |
127 | \ | |
128 | prev = lru_to_page(&(_page->lru)); \ | |
129 | prefetch(&prev->_field); \ | |
130 | } \ | |
131 | } while (0) | |
132 | #else | |
133 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
134 | #endif | |
135 | ||
136 | #ifdef ARCH_HAS_PREFETCHW | |
137 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
138 | do { \ | |
139 | if ((_page)->lru.prev != _base) { \ | |
140 | struct page *prev; \ | |
141 | \ | |
142 | prev = lru_to_page(&(_page->lru)); \ | |
143 | prefetchw(&prev->_field); \ | |
144 | } \ | |
145 | } while (0) | |
146 | #else | |
147 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
148 | #endif | |
149 | ||
150 | /* | |
151 | * From 0 .. 100. Higher means more swappy. | |
152 | */ | |
153 | int vm_swappiness = 60; | |
d0480be4 WSH |
154 | /* |
155 | * The total number of pages which are beyond the high watermark within all | |
156 | * zones. | |
157 | */ | |
158 | unsigned long vm_total_pages; | |
1da177e4 LT |
159 | |
160 | static LIST_HEAD(shrinker_list); | |
161 | static DECLARE_RWSEM(shrinker_rwsem); | |
162 | ||
c255a458 | 163 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
164 | static bool global_reclaim(struct scan_control *sc) |
165 | { | |
f16015fb | 166 | return !sc->target_mem_cgroup; |
89b5fae5 | 167 | } |
97c9341f TH |
168 | |
169 | /** | |
170 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
171 | * @sc: scan_control in question | |
172 | * | |
173 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
174 | * completely broken with the legacy memcg and direct stalling in | |
175 | * shrink_page_list() is used for throttling instead, which lacks all the | |
176 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
177 | * allocation and configurability. | |
178 | * | |
179 | * This function tests whether the vmscan currently in progress can assume | |
180 | * that the normal dirty throttling mechanism is operational. | |
181 | */ | |
182 | static bool sane_reclaim(struct scan_control *sc) | |
183 | { | |
184 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
185 | ||
186 | if (!memcg) | |
187 | return true; | |
188 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 189 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
190 | return true; |
191 | #endif | |
192 | return false; | |
193 | } | |
91a45470 | 194 | #else |
89b5fae5 JW |
195 | static bool global_reclaim(struct scan_control *sc) |
196 | { | |
197 | return true; | |
198 | } | |
97c9341f TH |
199 | |
200 | static bool sane_reclaim(struct scan_control *sc) | |
201 | { | |
202 | return true; | |
203 | } | |
91a45470 KH |
204 | #endif |
205 | ||
5a1c84b4 MG |
206 | /* |
207 | * This misses isolated pages which are not accounted for to save counters. | |
208 | * As the data only determines if reclaim or compaction continues, it is | |
209 | * not expected that isolated pages will be a dominating factor. | |
210 | */ | |
211 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
212 | { | |
213 | unsigned long nr; | |
214 | ||
215 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
216 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
217 | if (get_nr_swap_pages() > 0) | |
218 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
219 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
220 | ||
221 | return nr; | |
222 | } | |
223 | ||
599d0c95 MG |
224 | unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat) |
225 | { | |
226 | unsigned long nr; | |
227 | ||
228 | nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) + | |
229 | node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) + | |
230 | node_page_state_snapshot(pgdat, NR_ISOLATED_FILE); | |
6e543d57 LD |
231 | |
232 | if (get_nr_swap_pages() > 0) | |
599d0c95 MG |
233 | nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) + |
234 | node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) + | |
235 | node_page_state_snapshot(pgdat, NR_ISOLATED_ANON); | |
6e543d57 LD |
236 | |
237 | return nr; | |
238 | } | |
239 | ||
fd538803 MH |
240 | /** |
241 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
242 | * @lruvec: lru vector | |
243 | * @lru: lru to use | |
244 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
245 | */ | |
246 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) | |
c9f299d9 | 247 | { |
fd538803 MH |
248 | unsigned long lru_size; |
249 | int zid; | |
250 | ||
c3c787e8 | 251 | if (!mem_cgroup_disabled()) |
fd538803 MH |
252 | lru_size = mem_cgroup_get_lru_size(lruvec, lru); |
253 | else | |
254 | lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); | |
a3d8e054 | 255 | |
fd538803 MH |
256 | for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { |
257 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; | |
258 | unsigned long size; | |
c9f299d9 | 259 | |
fd538803 MH |
260 | if (!managed_zone(zone)) |
261 | continue; | |
262 | ||
263 | if (!mem_cgroup_disabled()) | |
264 | size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); | |
265 | else | |
266 | size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], | |
267 | NR_ZONE_LRU_BASE + lru); | |
268 | lru_size -= min(size, lru_size); | |
269 | } | |
270 | ||
271 | return lru_size; | |
b4536f0c | 272 | |
b4536f0c MH |
273 | } |
274 | ||
1da177e4 | 275 | /* |
1d3d4437 | 276 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 277 | */ |
1d3d4437 | 278 | int register_shrinker(struct shrinker *shrinker) |
1da177e4 | 279 | { |
1d3d4437 GC |
280 | size_t size = sizeof(*shrinker->nr_deferred); |
281 | ||
1d3d4437 GC |
282 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
283 | size *= nr_node_ids; | |
284 | ||
285 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
286 | if (!shrinker->nr_deferred) | |
287 | return -ENOMEM; | |
288 | ||
8e1f936b RR |
289 | down_write(&shrinker_rwsem); |
290 | list_add_tail(&shrinker->list, &shrinker_list); | |
291 | up_write(&shrinker_rwsem); | |
1d3d4437 | 292 | return 0; |
1da177e4 | 293 | } |
8e1f936b | 294 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
295 | |
296 | /* | |
297 | * Remove one | |
298 | */ | |
8e1f936b | 299 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 300 | { |
7880fc54 TH |
301 | if (!shrinker->nr_deferred) |
302 | return; | |
1da177e4 LT |
303 | down_write(&shrinker_rwsem); |
304 | list_del(&shrinker->list); | |
305 | up_write(&shrinker_rwsem); | |
ae393321 | 306 | kfree(shrinker->nr_deferred); |
7880fc54 | 307 | shrinker->nr_deferred = NULL; |
1da177e4 | 308 | } |
8e1f936b | 309 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
310 | |
311 | #define SHRINK_BATCH 128 | |
1d3d4437 | 312 | |
cb731d6c VD |
313 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
314 | struct shrinker *shrinker, | |
315 | unsigned long nr_scanned, | |
316 | unsigned long nr_eligible) | |
1d3d4437 GC |
317 | { |
318 | unsigned long freed = 0; | |
319 | unsigned long long delta; | |
320 | long total_scan; | |
d5bc5fd3 | 321 | long freeable; |
1d3d4437 GC |
322 | long nr; |
323 | long new_nr; | |
324 | int nid = shrinkctl->nid; | |
325 | long batch_size = shrinker->batch ? shrinker->batch | |
326 | : SHRINK_BATCH; | |
5f33a080 | 327 | long scanned = 0, next_deferred; |
1d3d4437 | 328 | |
d5bc5fd3 VD |
329 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
330 | if (freeable == 0) | |
1d3d4437 GC |
331 | return 0; |
332 | ||
333 | /* | |
334 | * copy the current shrinker scan count into a local variable | |
335 | * and zero it so that other concurrent shrinker invocations | |
336 | * don't also do this scanning work. | |
337 | */ | |
338 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
339 | ||
340 | total_scan = nr; | |
6b4f7799 | 341 | delta = (4 * nr_scanned) / shrinker->seeks; |
d5bc5fd3 | 342 | delta *= freeable; |
6b4f7799 | 343 | do_div(delta, nr_eligible + 1); |
1d3d4437 GC |
344 | total_scan += delta; |
345 | if (total_scan < 0) { | |
8612c663 | 346 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
a0b02131 | 347 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 348 | total_scan = freeable; |
5f33a080 SL |
349 | next_deferred = nr; |
350 | } else | |
351 | next_deferred = total_scan; | |
1d3d4437 GC |
352 | |
353 | /* | |
354 | * We need to avoid excessive windup on filesystem shrinkers | |
355 | * due to large numbers of GFP_NOFS allocations causing the | |
356 | * shrinkers to return -1 all the time. This results in a large | |
357 | * nr being built up so when a shrink that can do some work | |
358 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 359 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
360 | * memory. |
361 | * | |
362 | * Hence only allow the shrinker to scan the entire cache when | |
363 | * a large delta change is calculated directly. | |
364 | */ | |
d5bc5fd3 VD |
365 | if (delta < freeable / 4) |
366 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
367 | |
368 | /* | |
369 | * Avoid risking looping forever due to too large nr value: | |
370 | * never try to free more than twice the estimate number of | |
371 | * freeable entries. | |
372 | */ | |
d5bc5fd3 VD |
373 | if (total_scan > freeable * 2) |
374 | total_scan = freeable * 2; | |
1d3d4437 GC |
375 | |
376 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
6b4f7799 JW |
377 | nr_scanned, nr_eligible, |
378 | freeable, delta, total_scan); | |
1d3d4437 | 379 | |
0b1fb40a VD |
380 | /* |
381 | * Normally, we should not scan less than batch_size objects in one | |
382 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
383 | * than batch_size objects in total and we are really tight on memory, | |
384 | * we will try to reclaim all available objects, otherwise we can end | |
385 | * up failing allocations although there are plenty of reclaimable | |
386 | * objects spread over several slabs with usage less than the | |
387 | * batch_size. | |
388 | * | |
389 | * We detect the "tight on memory" situations by looking at the total | |
390 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 391 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
392 | * scanning at high prio and therefore should try to reclaim as much as |
393 | * possible. | |
394 | */ | |
395 | while (total_scan >= batch_size || | |
d5bc5fd3 | 396 | total_scan >= freeable) { |
a0b02131 | 397 | unsigned long ret; |
0b1fb40a | 398 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 399 | |
0b1fb40a | 400 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 401 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
402 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
403 | if (ret == SHRINK_STOP) | |
404 | break; | |
405 | freed += ret; | |
1d3d4437 | 406 | |
d460acb5 CW |
407 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
408 | total_scan -= shrinkctl->nr_scanned; | |
409 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
410 | |
411 | cond_resched(); | |
412 | } | |
413 | ||
5f33a080 SL |
414 | if (next_deferred >= scanned) |
415 | next_deferred -= scanned; | |
416 | else | |
417 | next_deferred = 0; | |
1d3d4437 GC |
418 | /* |
419 | * move the unused scan count back into the shrinker in a | |
420 | * manner that handles concurrent updates. If we exhausted the | |
421 | * scan, there is no need to do an update. | |
422 | */ | |
5f33a080 SL |
423 | if (next_deferred > 0) |
424 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
425 | &shrinker->nr_deferred[nid]); |
426 | else | |
427 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
428 | ||
df9024a8 | 429 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 430 | return freed; |
1495f230 YH |
431 | } |
432 | ||
6b4f7799 | 433 | /** |
cb731d6c | 434 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
435 | * @gfp_mask: allocation context |
436 | * @nid: node whose slab caches to target | |
cb731d6c | 437 | * @memcg: memory cgroup whose slab caches to target |
6b4f7799 JW |
438 | * @nr_scanned: pressure numerator |
439 | * @nr_eligible: pressure denominator | |
1da177e4 | 440 | * |
6b4f7799 | 441 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 442 | * |
6b4f7799 JW |
443 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
444 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 445 | * |
cb731d6c VD |
446 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
447 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan | |
0fc9f58a VD |
448 | * objects from the memory cgroup specified. Otherwise, only unaware |
449 | * shrinkers are called. | |
cb731d6c | 450 | * |
6b4f7799 JW |
451 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of |
452 | * the available objects should be scanned. Page reclaim for example | |
453 | * passes the number of pages scanned and the number of pages on the | |
454 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned | |
455 | * when it encountered mapped pages. The ratio is further biased by | |
456 | * the ->seeks setting of the shrink function, which indicates the | |
457 | * cost to recreate an object relative to that of an LRU page. | |
b15e0905 | 458 | * |
6b4f7799 | 459 | * Returns the number of reclaimed slab objects. |
1da177e4 | 460 | */ |
cb731d6c VD |
461 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
462 | struct mem_cgroup *memcg, | |
463 | unsigned long nr_scanned, | |
464 | unsigned long nr_eligible) | |
1da177e4 LT |
465 | { |
466 | struct shrinker *shrinker; | |
24f7c6b9 | 467 | unsigned long freed = 0; |
1da177e4 | 468 | |
0fc9f58a | 469 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) |
cb731d6c VD |
470 | return 0; |
471 | ||
6b4f7799 JW |
472 | if (nr_scanned == 0) |
473 | nr_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 474 | |
f06590bd | 475 | if (!down_read_trylock(&shrinker_rwsem)) { |
24f7c6b9 DC |
476 | /* |
477 | * If we would return 0, our callers would understand that we | |
478 | * have nothing else to shrink and give up trying. By returning | |
479 | * 1 we keep it going and assume we'll be able to shrink next | |
480 | * time. | |
481 | */ | |
482 | freed = 1; | |
f06590bd MK |
483 | goto out; |
484 | } | |
1da177e4 LT |
485 | |
486 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
487 | struct shrink_control sc = { |
488 | .gfp_mask = gfp_mask, | |
489 | .nid = nid, | |
cb731d6c | 490 | .memcg = memcg, |
6b4f7799 | 491 | }; |
ec97097b | 492 | |
0fc9f58a VD |
493 | /* |
494 | * If kernel memory accounting is disabled, we ignore | |
495 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers | |
496 | * passing NULL for memcg. | |
497 | */ | |
498 | if (memcg_kmem_enabled() && | |
499 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
cb731d6c VD |
500 | continue; |
501 | ||
6b4f7799 JW |
502 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
503 | sc.nid = 0; | |
1da177e4 | 504 | |
cb731d6c | 505 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); |
1da177e4 | 506 | } |
6b4f7799 | 507 | |
1da177e4 | 508 | up_read(&shrinker_rwsem); |
f06590bd MK |
509 | out: |
510 | cond_resched(); | |
24f7c6b9 | 511 | return freed; |
1da177e4 LT |
512 | } |
513 | ||
cb731d6c VD |
514 | void drop_slab_node(int nid) |
515 | { | |
516 | unsigned long freed; | |
517 | ||
518 | do { | |
519 | struct mem_cgroup *memcg = NULL; | |
520 | ||
521 | freed = 0; | |
522 | do { | |
523 | freed += shrink_slab(GFP_KERNEL, nid, memcg, | |
524 | 1000, 1000); | |
525 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
526 | } while (freed > 10); | |
527 | } | |
528 | ||
529 | void drop_slab(void) | |
530 | { | |
531 | int nid; | |
532 | ||
533 | for_each_online_node(nid) | |
534 | drop_slab_node(nid); | |
535 | } | |
536 | ||
1da177e4 LT |
537 | static inline int is_page_cache_freeable(struct page *page) |
538 | { | |
ceddc3a5 JW |
539 | /* |
540 | * A freeable page cache page is referenced only by the caller | |
541 | * that isolated the page, the page cache radix tree and | |
542 | * optional buffer heads at page->private. | |
543 | */ | |
bd4c82c2 HY |
544 | int radix_pins = PageTransHuge(page) && PageSwapCache(page) ? |
545 | HPAGE_PMD_NR : 1; | |
546 | return page_count(page) - page_has_private(page) == 1 + radix_pins; | |
1da177e4 LT |
547 | } |
548 | ||
703c2708 | 549 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
1da177e4 | 550 | { |
930d9152 | 551 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 552 | return 1; |
703c2708 | 553 | if (!inode_write_congested(inode)) |
1da177e4 | 554 | return 1; |
703c2708 | 555 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
556 | return 1; |
557 | return 0; | |
558 | } | |
559 | ||
560 | /* | |
561 | * We detected a synchronous write error writing a page out. Probably | |
562 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
563 | * fsync(), msync() or close(). | |
564 | * | |
565 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
566 | * prevents it from being freed up. But we have a ref on the page and once | |
567 | * that page is locked, the mapping is pinned. | |
568 | * | |
569 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
570 | * __GFP_FS. | |
571 | */ | |
572 | static void handle_write_error(struct address_space *mapping, | |
573 | struct page *page, int error) | |
574 | { | |
7eaceacc | 575 | lock_page(page); |
3e9f45bd GC |
576 | if (page_mapping(page) == mapping) |
577 | mapping_set_error(mapping, error); | |
1da177e4 LT |
578 | unlock_page(page); |
579 | } | |
580 | ||
04e62a29 CL |
581 | /* possible outcome of pageout() */ |
582 | typedef enum { | |
583 | /* failed to write page out, page is locked */ | |
584 | PAGE_KEEP, | |
585 | /* move page to the active list, page is locked */ | |
586 | PAGE_ACTIVATE, | |
587 | /* page has been sent to the disk successfully, page is unlocked */ | |
588 | PAGE_SUCCESS, | |
589 | /* page is clean and locked */ | |
590 | PAGE_CLEAN, | |
591 | } pageout_t; | |
592 | ||
1da177e4 | 593 | /* |
1742f19f AM |
594 | * pageout is called by shrink_page_list() for each dirty page. |
595 | * Calls ->writepage(). | |
1da177e4 | 596 | */ |
c661b078 | 597 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 598 | struct scan_control *sc) |
1da177e4 LT |
599 | { |
600 | /* | |
601 | * If the page is dirty, only perform writeback if that write | |
602 | * will be non-blocking. To prevent this allocation from being | |
603 | * stalled by pagecache activity. But note that there may be | |
604 | * stalls if we need to run get_block(). We could test | |
605 | * PagePrivate for that. | |
606 | * | |
8174202b | 607 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
608 | * this page's queue, we can perform writeback even if that |
609 | * will block. | |
610 | * | |
611 | * If the page is swapcache, write it back even if that would | |
612 | * block, for some throttling. This happens by accident, because | |
613 | * swap_backing_dev_info is bust: it doesn't reflect the | |
614 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
615 | */ |
616 | if (!is_page_cache_freeable(page)) | |
617 | return PAGE_KEEP; | |
618 | if (!mapping) { | |
619 | /* | |
620 | * Some data journaling orphaned pages can have | |
621 | * page->mapping == NULL while being dirty with clean buffers. | |
622 | */ | |
266cf658 | 623 | if (page_has_private(page)) { |
1da177e4 LT |
624 | if (try_to_free_buffers(page)) { |
625 | ClearPageDirty(page); | |
b1de0d13 | 626 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
627 | return PAGE_CLEAN; |
628 | } | |
629 | } | |
630 | return PAGE_KEEP; | |
631 | } | |
632 | if (mapping->a_ops->writepage == NULL) | |
633 | return PAGE_ACTIVATE; | |
703c2708 | 634 | if (!may_write_to_inode(mapping->host, sc)) |
1da177e4 LT |
635 | return PAGE_KEEP; |
636 | ||
637 | if (clear_page_dirty_for_io(page)) { | |
638 | int res; | |
639 | struct writeback_control wbc = { | |
640 | .sync_mode = WB_SYNC_NONE, | |
641 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
642 | .range_start = 0, |
643 | .range_end = LLONG_MAX, | |
1da177e4 LT |
644 | .for_reclaim = 1, |
645 | }; | |
646 | ||
647 | SetPageReclaim(page); | |
648 | res = mapping->a_ops->writepage(page, &wbc); | |
649 | if (res < 0) | |
650 | handle_write_error(mapping, page, res); | |
994fc28c | 651 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
652 | ClearPageReclaim(page); |
653 | return PAGE_ACTIVATE; | |
654 | } | |
c661b078 | 655 | |
1da177e4 LT |
656 | if (!PageWriteback(page)) { |
657 | /* synchronous write or broken a_ops? */ | |
658 | ClearPageReclaim(page); | |
659 | } | |
3aa23851 | 660 | trace_mm_vmscan_writepage(page); |
c4a25635 | 661 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
662 | return PAGE_SUCCESS; |
663 | } | |
664 | ||
665 | return PAGE_CLEAN; | |
666 | } | |
667 | ||
a649fd92 | 668 | /* |
e286781d NP |
669 | * Same as remove_mapping, but if the page is removed from the mapping, it |
670 | * gets returned with a refcount of 0. | |
a649fd92 | 671 | */ |
a528910e JW |
672 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
673 | bool reclaimed) | |
49d2e9cc | 674 | { |
c4843a75 | 675 | unsigned long flags; |
bd4c82c2 | 676 | int refcount; |
c4843a75 | 677 | |
28e4d965 NP |
678 | BUG_ON(!PageLocked(page)); |
679 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 680 | |
c4843a75 | 681 | spin_lock_irqsave(&mapping->tree_lock, flags); |
49d2e9cc | 682 | /* |
0fd0e6b0 NP |
683 | * The non racy check for a busy page. |
684 | * | |
685 | * Must be careful with the order of the tests. When someone has | |
686 | * a ref to the page, it may be possible that they dirty it then | |
687 | * drop the reference. So if PageDirty is tested before page_count | |
688 | * here, then the following race may occur: | |
689 | * | |
690 | * get_user_pages(&page); | |
691 | * [user mapping goes away] | |
692 | * write_to(page); | |
693 | * !PageDirty(page) [good] | |
694 | * SetPageDirty(page); | |
695 | * put_page(page); | |
696 | * !page_count(page) [good, discard it] | |
697 | * | |
698 | * [oops, our write_to data is lost] | |
699 | * | |
700 | * Reversing the order of the tests ensures such a situation cannot | |
701 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 702 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
703 | * |
704 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
705 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 706 | */ |
bd4c82c2 HY |
707 | if (unlikely(PageTransHuge(page)) && PageSwapCache(page)) |
708 | refcount = 1 + HPAGE_PMD_NR; | |
709 | else | |
710 | refcount = 2; | |
711 | if (!page_ref_freeze(page, refcount)) | |
49d2e9cc | 712 | goto cannot_free; |
e286781d NP |
713 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
714 | if (unlikely(PageDirty(page))) { | |
bd4c82c2 | 715 | page_ref_unfreeze(page, refcount); |
49d2e9cc | 716 | goto cannot_free; |
e286781d | 717 | } |
49d2e9cc CL |
718 | |
719 | if (PageSwapCache(page)) { | |
720 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 721 | mem_cgroup_swapout(page, swap); |
49d2e9cc | 722 | __delete_from_swap_cache(page); |
c4843a75 | 723 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
75f6d6d2 | 724 | put_swap_page(page, swap); |
e286781d | 725 | } else { |
6072d13c | 726 | void (*freepage)(struct page *); |
a528910e | 727 | void *shadow = NULL; |
6072d13c LT |
728 | |
729 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
730 | /* |
731 | * Remember a shadow entry for reclaimed file cache in | |
732 | * order to detect refaults, thus thrashing, later on. | |
733 | * | |
734 | * But don't store shadows in an address space that is | |
735 | * already exiting. This is not just an optizimation, | |
736 | * inode reclaim needs to empty out the radix tree or | |
737 | * the nodes are lost. Don't plant shadows behind its | |
738 | * back. | |
f9fe48be RZ |
739 | * |
740 | * We also don't store shadows for DAX mappings because the | |
741 | * only page cache pages found in these are zero pages | |
742 | * covering holes, and because we don't want to mix DAX | |
743 | * exceptional entries and shadow exceptional entries in the | |
744 | * same page_tree. | |
a528910e JW |
745 | */ |
746 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 747 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
a528910e | 748 | shadow = workingset_eviction(mapping, page); |
62cccb8c | 749 | __delete_from_page_cache(page, shadow); |
c4843a75 | 750 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
6072d13c LT |
751 | |
752 | if (freepage != NULL) | |
753 | freepage(page); | |
49d2e9cc CL |
754 | } |
755 | ||
49d2e9cc CL |
756 | return 1; |
757 | ||
758 | cannot_free: | |
c4843a75 | 759 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
49d2e9cc CL |
760 | return 0; |
761 | } | |
762 | ||
e286781d NP |
763 | /* |
764 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
765 | * someone else has a ref on the page, abort and return 0. If it was | |
766 | * successfully detached, return 1. Assumes the caller has a single ref on | |
767 | * this page. | |
768 | */ | |
769 | int remove_mapping(struct address_space *mapping, struct page *page) | |
770 | { | |
a528910e | 771 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
772 | /* |
773 | * Unfreezing the refcount with 1 rather than 2 effectively | |
774 | * drops the pagecache ref for us without requiring another | |
775 | * atomic operation. | |
776 | */ | |
fe896d18 | 777 | page_ref_unfreeze(page, 1); |
e286781d NP |
778 | return 1; |
779 | } | |
780 | return 0; | |
781 | } | |
782 | ||
894bc310 LS |
783 | /** |
784 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
785 | * @page: page to be put back to appropriate lru list | |
786 | * | |
787 | * Add previously isolated @page to appropriate LRU list. | |
788 | * Page may still be unevictable for other reasons. | |
789 | * | |
790 | * lru_lock must not be held, interrupts must be enabled. | |
791 | */ | |
894bc310 LS |
792 | void putback_lru_page(struct page *page) |
793 | { | |
0ec3b74c | 794 | bool is_unevictable; |
bbfd28ee | 795 | int was_unevictable = PageUnevictable(page); |
894bc310 | 796 | |
309381fe | 797 | VM_BUG_ON_PAGE(PageLRU(page), page); |
894bc310 LS |
798 | |
799 | redo: | |
800 | ClearPageUnevictable(page); | |
801 | ||
39b5f29a | 802 | if (page_evictable(page)) { |
894bc310 LS |
803 | /* |
804 | * For evictable pages, we can use the cache. | |
805 | * In event of a race, worst case is we end up with an | |
806 | * unevictable page on [in]active list. | |
807 | * We know how to handle that. | |
808 | */ | |
0ec3b74c | 809 | is_unevictable = false; |
c53954a0 | 810 | lru_cache_add(page); |
894bc310 LS |
811 | } else { |
812 | /* | |
813 | * Put unevictable pages directly on zone's unevictable | |
814 | * list. | |
815 | */ | |
0ec3b74c | 816 | is_unevictable = true; |
894bc310 | 817 | add_page_to_unevictable_list(page); |
6a7b9548 | 818 | /* |
21ee9f39 MK |
819 | * When racing with an mlock or AS_UNEVICTABLE clearing |
820 | * (page is unlocked) make sure that if the other thread | |
821 | * does not observe our setting of PG_lru and fails | |
24513264 | 822 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 823 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
824 | * the page back to the evictable list. |
825 | * | |
21ee9f39 | 826 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
827 | */ |
828 | smp_mb(); | |
894bc310 | 829 | } |
894bc310 LS |
830 | |
831 | /* | |
832 | * page's status can change while we move it among lru. If an evictable | |
833 | * page is on unevictable list, it never be freed. To avoid that, | |
834 | * check after we added it to the list, again. | |
835 | */ | |
0ec3b74c | 836 | if (is_unevictable && page_evictable(page)) { |
894bc310 LS |
837 | if (!isolate_lru_page(page)) { |
838 | put_page(page); | |
839 | goto redo; | |
840 | } | |
841 | /* This means someone else dropped this page from LRU | |
842 | * So, it will be freed or putback to LRU again. There is | |
843 | * nothing to do here. | |
844 | */ | |
845 | } | |
846 | ||
0ec3b74c | 847 | if (was_unevictable && !is_unevictable) |
bbfd28ee | 848 | count_vm_event(UNEVICTABLE_PGRESCUED); |
0ec3b74c | 849 | else if (!was_unevictable && is_unevictable) |
bbfd28ee LS |
850 | count_vm_event(UNEVICTABLE_PGCULLED); |
851 | ||
894bc310 LS |
852 | put_page(page); /* drop ref from isolate */ |
853 | } | |
854 | ||
dfc8d636 JW |
855 | enum page_references { |
856 | PAGEREF_RECLAIM, | |
857 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 858 | PAGEREF_KEEP, |
dfc8d636 JW |
859 | PAGEREF_ACTIVATE, |
860 | }; | |
861 | ||
862 | static enum page_references page_check_references(struct page *page, | |
863 | struct scan_control *sc) | |
864 | { | |
64574746 | 865 | int referenced_ptes, referenced_page; |
dfc8d636 | 866 | unsigned long vm_flags; |
dfc8d636 | 867 | |
c3ac9a8a JW |
868 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
869 | &vm_flags); | |
64574746 | 870 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 871 | |
dfc8d636 JW |
872 | /* |
873 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
874 | * move the page to the unevictable list. | |
875 | */ | |
876 | if (vm_flags & VM_LOCKED) | |
877 | return PAGEREF_RECLAIM; | |
878 | ||
64574746 | 879 | if (referenced_ptes) { |
e4898273 | 880 | if (PageSwapBacked(page)) |
64574746 JW |
881 | return PAGEREF_ACTIVATE; |
882 | /* | |
883 | * All mapped pages start out with page table | |
884 | * references from the instantiating fault, so we need | |
885 | * to look twice if a mapped file page is used more | |
886 | * than once. | |
887 | * | |
888 | * Mark it and spare it for another trip around the | |
889 | * inactive list. Another page table reference will | |
890 | * lead to its activation. | |
891 | * | |
892 | * Note: the mark is set for activated pages as well | |
893 | * so that recently deactivated but used pages are | |
894 | * quickly recovered. | |
895 | */ | |
896 | SetPageReferenced(page); | |
897 | ||
34dbc67a | 898 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
899 | return PAGEREF_ACTIVATE; |
900 | ||
c909e993 KK |
901 | /* |
902 | * Activate file-backed executable pages after first usage. | |
903 | */ | |
904 | if (vm_flags & VM_EXEC) | |
905 | return PAGEREF_ACTIVATE; | |
906 | ||
64574746 JW |
907 | return PAGEREF_KEEP; |
908 | } | |
dfc8d636 JW |
909 | |
910 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 911 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
912 | return PAGEREF_RECLAIM_CLEAN; |
913 | ||
914 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
915 | } |
916 | ||
e2be15f6 MG |
917 | /* Check if a page is dirty or under writeback */ |
918 | static void page_check_dirty_writeback(struct page *page, | |
919 | bool *dirty, bool *writeback) | |
920 | { | |
b4597226 MG |
921 | struct address_space *mapping; |
922 | ||
e2be15f6 MG |
923 | /* |
924 | * Anonymous pages are not handled by flushers and must be written | |
925 | * from reclaim context. Do not stall reclaim based on them | |
926 | */ | |
802a3a92 SL |
927 | if (!page_is_file_cache(page) || |
928 | (PageAnon(page) && !PageSwapBacked(page))) { | |
e2be15f6 MG |
929 | *dirty = false; |
930 | *writeback = false; | |
931 | return; | |
932 | } | |
933 | ||
934 | /* By default assume that the page flags are accurate */ | |
935 | *dirty = PageDirty(page); | |
936 | *writeback = PageWriteback(page); | |
b4597226 MG |
937 | |
938 | /* Verify dirty/writeback state if the filesystem supports it */ | |
939 | if (!page_has_private(page)) | |
940 | return; | |
941 | ||
942 | mapping = page_mapping(page); | |
943 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
944 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
945 | } |
946 | ||
3c710c1a MH |
947 | struct reclaim_stat { |
948 | unsigned nr_dirty; | |
949 | unsigned nr_unqueued_dirty; | |
950 | unsigned nr_congested; | |
951 | unsigned nr_writeback; | |
952 | unsigned nr_immediate; | |
5bccd166 MH |
953 | unsigned nr_activate; |
954 | unsigned nr_ref_keep; | |
955 | unsigned nr_unmap_fail; | |
3c710c1a MH |
956 | }; |
957 | ||
1da177e4 | 958 | /* |
1742f19f | 959 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 960 | */ |
1742f19f | 961 | static unsigned long shrink_page_list(struct list_head *page_list, |
599d0c95 | 962 | struct pglist_data *pgdat, |
f84f6e2b | 963 | struct scan_control *sc, |
02c6de8d | 964 | enum ttu_flags ttu_flags, |
3c710c1a | 965 | struct reclaim_stat *stat, |
a4c0d0f2 | 966 | bool skip_reference_check) |
1da177e4 LT |
967 | { |
968 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 969 | LIST_HEAD(free_pages); |
1da177e4 | 970 | int pgactivate = 0; |
3c710c1a MH |
971 | unsigned nr_unqueued_dirty = 0; |
972 | unsigned nr_dirty = 0; | |
973 | unsigned nr_congested = 0; | |
974 | unsigned nr_reclaimed = 0; | |
975 | unsigned nr_writeback = 0; | |
976 | unsigned nr_immediate = 0; | |
5bccd166 MH |
977 | unsigned nr_ref_keep = 0; |
978 | unsigned nr_unmap_fail = 0; | |
1da177e4 LT |
979 | |
980 | cond_resched(); | |
981 | ||
1da177e4 LT |
982 | while (!list_empty(page_list)) { |
983 | struct address_space *mapping; | |
984 | struct page *page; | |
985 | int may_enter_fs; | |
a4c0d0f2 | 986 | enum page_references references = PAGEREF_RECLAIM; |
e2be15f6 | 987 | bool dirty, writeback; |
1da177e4 LT |
988 | |
989 | cond_resched(); | |
990 | ||
991 | page = lru_to_page(page_list); | |
992 | list_del(&page->lru); | |
993 | ||
529ae9aa | 994 | if (!trylock_page(page)) |
1da177e4 LT |
995 | goto keep; |
996 | ||
309381fe | 997 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
998 | |
999 | sc->nr_scanned++; | |
80e43426 | 1000 | |
39b5f29a | 1001 | if (unlikely(!page_evictable(page))) |
ad6b6704 | 1002 | goto activate_locked; |
894bc310 | 1003 | |
a6dc60f8 | 1004 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
1005 | goto keep_locked; |
1006 | ||
1da177e4 | 1007 | /* Double the slab pressure for mapped and swapcache pages */ |
802a3a92 SL |
1008 | if ((page_mapped(page) || PageSwapCache(page)) && |
1009 | !(PageAnon(page) && !PageSwapBacked(page))) | |
1da177e4 LT |
1010 | sc->nr_scanned++; |
1011 | ||
c661b078 AW |
1012 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1013 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1014 | ||
e2be15f6 MG |
1015 | /* |
1016 | * The number of dirty pages determines if a zone is marked | |
1017 | * reclaim_congested which affects wait_iff_congested. kswapd | |
1018 | * will stall and start writing pages if the tail of the LRU | |
1019 | * is all dirty unqueued pages. | |
1020 | */ | |
1021 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1022 | if (dirty || writeback) | |
1023 | nr_dirty++; | |
1024 | ||
1025 | if (dirty && !writeback) | |
1026 | nr_unqueued_dirty++; | |
1027 | ||
d04e8acd MG |
1028 | /* |
1029 | * Treat this page as congested if the underlying BDI is or if | |
1030 | * pages are cycling through the LRU so quickly that the | |
1031 | * pages marked for immediate reclaim are making it to the | |
1032 | * end of the LRU a second time. | |
1033 | */ | |
e2be15f6 | 1034 | mapping = page_mapping(page); |
1da58ee2 | 1035 | if (((dirty || writeback) && mapping && |
703c2708 | 1036 | inode_write_congested(mapping->host)) || |
d04e8acd | 1037 | (writeback && PageReclaim(page))) |
e2be15f6 MG |
1038 | nr_congested++; |
1039 | ||
283aba9f MG |
1040 | /* |
1041 | * If a page at the tail of the LRU is under writeback, there | |
1042 | * are three cases to consider. | |
1043 | * | |
1044 | * 1) If reclaim is encountering an excessive number of pages | |
1045 | * under writeback and this page is both under writeback and | |
1046 | * PageReclaim then it indicates that pages are being queued | |
1047 | * for IO but are being recycled through the LRU before the | |
1048 | * IO can complete. Waiting on the page itself risks an | |
1049 | * indefinite stall if it is impossible to writeback the | |
1050 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1051 | * note that the LRU is being scanned too quickly and the |
1052 | * caller can stall after page list has been processed. | |
283aba9f | 1053 | * |
97c9341f | 1054 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1055 | * not marked for immediate reclaim, or the caller does not |
1056 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1057 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1058 | * reclaim and continue scanning. |
283aba9f | 1059 | * |
ecf5fc6e MH |
1060 | * Require may_enter_fs because we would wait on fs, which |
1061 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1062 | * enter reclaim, and deadlock if it waits on a page for |
1063 | * which it is needed to do the write (loop masks off | |
1064 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1065 | * would probably show more reasons. | |
1066 | * | |
7fadc820 | 1067 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1068 | * PageReclaim. memcg does not have any dirty pages |
1069 | * throttling so we could easily OOM just because too many | |
1070 | * pages are in writeback and there is nothing else to | |
1071 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1072 | * |
1073 | * In cases 1) and 2) we activate the pages to get them out of | |
1074 | * the way while we continue scanning for clean pages on the | |
1075 | * inactive list and refilling from the active list. The | |
1076 | * observation here is that waiting for disk writes is more | |
1077 | * expensive than potentially causing reloads down the line. | |
1078 | * Since they're marked for immediate reclaim, they won't put | |
1079 | * memory pressure on the cache working set any longer than it | |
1080 | * takes to write them to disk. | |
283aba9f | 1081 | */ |
c661b078 | 1082 | if (PageWriteback(page)) { |
283aba9f MG |
1083 | /* Case 1 above */ |
1084 | if (current_is_kswapd() && | |
1085 | PageReclaim(page) && | |
599d0c95 | 1086 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
b1a6f21e | 1087 | nr_immediate++; |
c55e8d03 | 1088 | goto activate_locked; |
283aba9f MG |
1089 | |
1090 | /* Case 2 above */ | |
97c9341f | 1091 | } else if (sane_reclaim(sc) || |
ecf5fc6e | 1092 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1093 | /* |
1094 | * This is slightly racy - end_page_writeback() | |
1095 | * might have just cleared PageReclaim, then | |
1096 | * setting PageReclaim here end up interpreted | |
1097 | * as PageReadahead - but that does not matter | |
1098 | * enough to care. What we do want is for this | |
1099 | * page to have PageReclaim set next time memcg | |
1100 | * reclaim reaches the tests above, so it will | |
1101 | * then wait_on_page_writeback() to avoid OOM; | |
1102 | * and it's also appropriate in global reclaim. | |
1103 | */ | |
1104 | SetPageReclaim(page); | |
e62e384e | 1105 | nr_writeback++; |
c55e8d03 | 1106 | goto activate_locked; |
283aba9f MG |
1107 | |
1108 | /* Case 3 above */ | |
1109 | } else { | |
7fadc820 | 1110 | unlock_page(page); |
283aba9f | 1111 | wait_on_page_writeback(page); |
7fadc820 HD |
1112 | /* then go back and try same page again */ |
1113 | list_add_tail(&page->lru, page_list); | |
1114 | continue; | |
e62e384e | 1115 | } |
c661b078 | 1116 | } |
1da177e4 | 1117 | |
a4c0d0f2 | 1118 | if (!skip_reference_check) |
02c6de8d MK |
1119 | references = page_check_references(page, sc); |
1120 | ||
dfc8d636 JW |
1121 | switch (references) { |
1122 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1123 | goto activate_locked; |
64574746 | 1124 | case PAGEREF_KEEP: |
5bccd166 | 1125 | nr_ref_keep++; |
64574746 | 1126 | goto keep_locked; |
dfc8d636 JW |
1127 | case PAGEREF_RECLAIM: |
1128 | case PAGEREF_RECLAIM_CLEAN: | |
1129 | ; /* try to reclaim the page below */ | |
1130 | } | |
1da177e4 | 1131 | |
1da177e4 LT |
1132 | /* |
1133 | * Anonymous process memory has backing store? | |
1134 | * Try to allocate it some swap space here. | |
802a3a92 | 1135 | * Lazyfree page could be freed directly |
1da177e4 | 1136 | */ |
bd4c82c2 HY |
1137 | if (PageAnon(page) && PageSwapBacked(page)) { |
1138 | if (!PageSwapCache(page)) { | |
1139 | if (!(sc->gfp_mask & __GFP_IO)) | |
1140 | goto keep_locked; | |
1141 | if (PageTransHuge(page)) { | |
1142 | /* cannot split THP, skip it */ | |
1143 | if (!can_split_huge_page(page, NULL)) | |
1144 | goto activate_locked; | |
1145 | /* | |
1146 | * Split pages without a PMD map right | |
1147 | * away. Chances are some or all of the | |
1148 | * tail pages can be freed without IO. | |
1149 | */ | |
1150 | if (!compound_mapcount(page) && | |
1151 | split_huge_page_to_list(page, | |
1152 | page_list)) | |
1153 | goto activate_locked; | |
1154 | } | |
1155 | if (!add_to_swap(page)) { | |
1156 | if (!PageTransHuge(page)) | |
1157 | goto activate_locked; | |
1158 | /* Fallback to swap normal pages */ | |
1159 | if (split_huge_page_to_list(page, | |
1160 | page_list)) | |
1161 | goto activate_locked; | |
fe490cc0 HY |
1162 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1163 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1164 | #endif | |
bd4c82c2 HY |
1165 | if (!add_to_swap(page)) |
1166 | goto activate_locked; | |
1167 | } | |
0f074658 | 1168 | |
bd4c82c2 | 1169 | may_enter_fs = 1; |
1da177e4 | 1170 | |
bd4c82c2 HY |
1171 | /* Adding to swap updated mapping */ |
1172 | mapping = page_mapping(page); | |
1173 | } | |
7751b2da KS |
1174 | } else if (unlikely(PageTransHuge(page))) { |
1175 | /* Split file THP */ | |
1176 | if (split_huge_page_to_list(page, page_list)) | |
1177 | goto keep_locked; | |
e2be15f6 | 1178 | } |
1da177e4 LT |
1179 | |
1180 | /* | |
1181 | * The page is mapped into the page tables of one or more | |
1182 | * processes. Try to unmap it here. | |
1183 | */ | |
802a3a92 | 1184 | if (page_mapped(page)) { |
bd4c82c2 HY |
1185 | enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH; |
1186 | ||
1187 | if (unlikely(PageTransHuge(page))) | |
1188 | flags |= TTU_SPLIT_HUGE_PMD; | |
1189 | if (!try_to_unmap(page, flags)) { | |
5bccd166 | 1190 | nr_unmap_fail++; |
1da177e4 | 1191 | goto activate_locked; |
1da177e4 LT |
1192 | } |
1193 | } | |
1194 | ||
1195 | if (PageDirty(page)) { | |
ee72886d | 1196 | /* |
4eda4823 JW |
1197 | * Only kswapd can writeback filesystem pages |
1198 | * to avoid risk of stack overflow. But avoid | |
1199 | * injecting inefficient single-page IO into | |
1200 | * flusher writeback as much as possible: only | |
1201 | * write pages when we've encountered many | |
1202 | * dirty pages, and when we've already scanned | |
1203 | * the rest of the LRU for clean pages and see | |
1204 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1205 | */ |
f84f6e2b | 1206 | if (page_is_file_cache(page) && |
4eda4823 JW |
1207 | (!current_is_kswapd() || !PageReclaim(page) || |
1208 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1209 | /* |
1210 | * Immediately reclaim when written back. | |
1211 | * Similar in principal to deactivate_page() | |
1212 | * except we already have the page isolated | |
1213 | * and know it's dirty | |
1214 | */ | |
c4a25635 | 1215 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1216 | SetPageReclaim(page); |
1217 | ||
c55e8d03 | 1218 | goto activate_locked; |
ee72886d MG |
1219 | } |
1220 | ||
dfc8d636 | 1221 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1222 | goto keep_locked; |
4dd4b920 | 1223 | if (!may_enter_fs) |
1da177e4 | 1224 | goto keep_locked; |
52a8363e | 1225 | if (!sc->may_writepage) |
1da177e4 LT |
1226 | goto keep_locked; |
1227 | ||
d950c947 MG |
1228 | /* |
1229 | * Page is dirty. Flush the TLB if a writable entry | |
1230 | * potentially exists to avoid CPU writes after IO | |
1231 | * starts and then write it out here. | |
1232 | */ | |
1233 | try_to_unmap_flush_dirty(); | |
7d3579e8 | 1234 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1235 | case PAGE_KEEP: |
1236 | goto keep_locked; | |
1237 | case PAGE_ACTIVATE: | |
1238 | goto activate_locked; | |
1239 | case PAGE_SUCCESS: | |
7d3579e8 | 1240 | if (PageWriteback(page)) |
41ac1999 | 1241 | goto keep; |
7d3579e8 | 1242 | if (PageDirty(page)) |
1da177e4 | 1243 | goto keep; |
7d3579e8 | 1244 | |
1da177e4 LT |
1245 | /* |
1246 | * A synchronous write - probably a ramdisk. Go | |
1247 | * ahead and try to reclaim the page. | |
1248 | */ | |
529ae9aa | 1249 | if (!trylock_page(page)) |
1da177e4 LT |
1250 | goto keep; |
1251 | if (PageDirty(page) || PageWriteback(page)) | |
1252 | goto keep_locked; | |
1253 | mapping = page_mapping(page); | |
1254 | case PAGE_CLEAN: | |
1255 | ; /* try to free the page below */ | |
1256 | } | |
1257 | } | |
1258 | ||
1259 | /* | |
1260 | * If the page has buffers, try to free the buffer mappings | |
1261 | * associated with this page. If we succeed we try to free | |
1262 | * the page as well. | |
1263 | * | |
1264 | * We do this even if the page is PageDirty(). | |
1265 | * try_to_release_page() does not perform I/O, but it is | |
1266 | * possible for a page to have PageDirty set, but it is actually | |
1267 | * clean (all its buffers are clean). This happens if the | |
1268 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1269 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1270 | * try_to_release_page() will discover that cleanness and will |
1271 | * drop the buffers and mark the page clean - it can be freed. | |
1272 | * | |
1273 | * Rarely, pages can have buffers and no ->mapping. These are | |
1274 | * the pages which were not successfully invalidated in | |
1275 | * truncate_complete_page(). We try to drop those buffers here | |
1276 | * and if that worked, and the page is no longer mapped into | |
1277 | * process address space (page_count == 1) it can be freed. | |
1278 | * Otherwise, leave the page on the LRU so it is swappable. | |
1279 | */ | |
266cf658 | 1280 | if (page_has_private(page)) { |
1da177e4 LT |
1281 | if (!try_to_release_page(page, sc->gfp_mask)) |
1282 | goto activate_locked; | |
e286781d NP |
1283 | if (!mapping && page_count(page) == 1) { |
1284 | unlock_page(page); | |
1285 | if (put_page_testzero(page)) | |
1286 | goto free_it; | |
1287 | else { | |
1288 | /* | |
1289 | * rare race with speculative reference. | |
1290 | * the speculative reference will free | |
1291 | * this page shortly, so we may | |
1292 | * increment nr_reclaimed here (and | |
1293 | * leave it off the LRU). | |
1294 | */ | |
1295 | nr_reclaimed++; | |
1296 | continue; | |
1297 | } | |
1298 | } | |
1da177e4 LT |
1299 | } |
1300 | ||
802a3a92 SL |
1301 | if (PageAnon(page) && !PageSwapBacked(page)) { |
1302 | /* follow __remove_mapping for reference */ | |
1303 | if (!page_ref_freeze(page, 1)) | |
1304 | goto keep_locked; | |
1305 | if (PageDirty(page)) { | |
1306 | page_ref_unfreeze(page, 1); | |
1307 | goto keep_locked; | |
1308 | } | |
1da177e4 | 1309 | |
802a3a92 | 1310 | count_vm_event(PGLAZYFREED); |
2262185c | 1311 | count_memcg_page_event(page, PGLAZYFREED); |
802a3a92 SL |
1312 | } else if (!mapping || !__remove_mapping(mapping, page, true)) |
1313 | goto keep_locked; | |
a978d6f5 NP |
1314 | /* |
1315 | * At this point, we have no other references and there is | |
1316 | * no way to pick any more up (removed from LRU, removed | |
1317 | * from pagecache). Can use non-atomic bitops now (and | |
1318 | * we obviously don't have to worry about waking up a process | |
1319 | * waiting on the page lock, because there are no references. | |
1320 | */ | |
48c935ad | 1321 | __ClearPageLocked(page); |
e286781d | 1322 | free_it: |
05ff5137 | 1323 | nr_reclaimed++; |
abe4c3b5 MG |
1324 | |
1325 | /* | |
1326 | * Is there need to periodically free_page_list? It would | |
1327 | * appear not as the counts should be low | |
1328 | */ | |
bd4c82c2 HY |
1329 | if (unlikely(PageTransHuge(page))) { |
1330 | mem_cgroup_uncharge(page); | |
1331 | (*get_compound_page_dtor(page))(page); | |
1332 | } else | |
1333 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1334 | continue; |
1335 | ||
1336 | activate_locked: | |
68a22394 | 1337 | /* Not a candidate for swapping, so reclaim swap space. */ |
ad6b6704 MK |
1338 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
1339 | PageMlocked(page))) | |
a2c43eed | 1340 | try_to_free_swap(page); |
309381fe | 1341 | VM_BUG_ON_PAGE(PageActive(page), page); |
ad6b6704 MK |
1342 | if (!PageMlocked(page)) { |
1343 | SetPageActive(page); | |
1344 | pgactivate++; | |
2262185c | 1345 | count_memcg_page_event(page, PGACTIVATE); |
ad6b6704 | 1346 | } |
1da177e4 LT |
1347 | keep_locked: |
1348 | unlock_page(page); | |
1349 | keep: | |
1350 | list_add(&page->lru, &ret_pages); | |
309381fe | 1351 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1352 | } |
abe4c3b5 | 1353 | |
747db954 | 1354 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1355 | try_to_unmap_flush(); |
b745bc85 | 1356 | free_hot_cold_page_list(&free_pages, true); |
abe4c3b5 | 1357 | |
1da177e4 | 1358 | list_splice(&ret_pages, page_list); |
f8891e5e | 1359 | count_vm_events(PGACTIVATE, pgactivate); |
0a31bc97 | 1360 | |
3c710c1a MH |
1361 | if (stat) { |
1362 | stat->nr_dirty = nr_dirty; | |
1363 | stat->nr_congested = nr_congested; | |
1364 | stat->nr_unqueued_dirty = nr_unqueued_dirty; | |
1365 | stat->nr_writeback = nr_writeback; | |
1366 | stat->nr_immediate = nr_immediate; | |
5bccd166 MH |
1367 | stat->nr_activate = pgactivate; |
1368 | stat->nr_ref_keep = nr_ref_keep; | |
1369 | stat->nr_unmap_fail = nr_unmap_fail; | |
3c710c1a | 1370 | } |
05ff5137 | 1371 | return nr_reclaimed; |
1da177e4 LT |
1372 | } |
1373 | ||
02c6de8d MK |
1374 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1375 | struct list_head *page_list) | |
1376 | { | |
1377 | struct scan_control sc = { | |
1378 | .gfp_mask = GFP_KERNEL, | |
1379 | .priority = DEF_PRIORITY, | |
1380 | .may_unmap = 1, | |
1381 | }; | |
3c710c1a | 1382 | unsigned long ret; |
02c6de8d MK |
1383 | struct page *page, *next; |
1384 | LIST_HEAD(clean_pages); | |
1385 | ||
1386 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1387 | if (page_is_file_cache(page) && !PageDirty(page) && |
b1123ea6 | 1388 | !__PageMovable(page)) { |
02c6de8d MK |
1389 | ClearPageActive(page); |
1390 | list_move(&page->lru, &clean_pages); | |
1391 | } | |
1392 | } | |
1393 | ||
599d0c95 | 1394 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
a128ca71 | 1395 | TTU_IGNORE_ACCESS, NULL, true); |
02c6de8d | 1396 | list_splice(&clean_pages, page_list); |
599d0c95 | 1397 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1398 | return ret; |
1399 | } | |
1400 | ||
a4c0d0f2 MK |
1401 | #ifdef CONFIG_PROCESS_RECLAIM |
1402 | unsigned long reclaim_pages(struct list_head *page_list) | |
1403 | { | |
1404 | struct reclaim_stat dummy; | |
1405 | unsigned long nr_reclaimed; | |
1406 | struct page *page; | |
1407 | unsigned long nr_isolated[2] = {0, }; | |
1408 | struct pglist_data *pgdat = NULL; | |
1409 | struct scan_control sc = { | |
1410 | .gfp_mask = GFP_KERNEL, | |
1411 | .priority = DEF_PRIORITY, | |
1412 | .may_writepage = 1, | |
1413 | .may_unmap = 1, | |
1414 | .may_swap = 1, | |
1415 | }; | |
1416 | ||
1417 | if (list_empty(page_list)) | |
1418 | return 0; | |
1419 | ||
1420 | list_for_each_entry(page, page_list, lru) { | |
1421 | ClearPageActive(page); | |
1422 | if (pgdat == NULL) | |
1423 | pgdat = page_pgdat(page); | |
1424 | /* XXX: It could be multiple node in other config */ | |
1425 | WARN_ON_ONCE(pgdat != page_pgdat(page)); | |
1426 | if (!page_is_file_cache(page)) | |
1427 | nr_isolated[0]++; | |
1428 | else | |
1429 | nr_isolated[1]++; | |
1430 | } | |
1431 | ||
1432 | mod_node_page_state(pgdat, NR_ISOLATED_ANON, nr_isolated[0]); | |
1433 | mod_node_page_state(pgdat, NR_ISOLATED_FILE, nr_isolated[1]); | |
1434 | ||
1435 | nr_reclaimed = shrink_page_list(page_list, pgdat, &sc, | |
1436 | TTU_IGNORE_ACCESS, &dummy, true); | |
1437 | ||
1438 | while (!list_empty(page_list)) { | |
1439 | page = lru_to_page(page_list); | |
1440 | list_del(&page->lru); | |
1441 | putback_lru_page(page); | |
1442 | } | |
1443 | ||
1444 | mod_node_page_state(pgdat, NR_ISOLATED_ANON, -nr_isolated[0]); | |
1445 | mod_node_page_state(pgdat, NR_ISOLATED_FILE, -nr_isolated[1]); | |
1446 | ||
1447 | return nr_reclaimed; | |
1448 | } | |
1449 | #endif | |
1450 | ||
5ad333eb AW |
1451 | /* |
1452 | * Attempt to remove the specified page from its LRU. Only take this page | |
1453 | * if it is of the appropriate PageActive status. Pages which are being | |
1454 | * freed elsewhere are also ignored. | |
1455 | * | |
1456 | * page: page to consider | |
1457 | * mode: one of the LRU isolation modes defined above | |
1458 | * | |
1459 | * returns 0 on success, -ve errno on failure. | |
1460 | */ | |
f3fd4a61 | 1461 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1462 | { |
1463 | int ret = -EINVAL; | |
1464 | ||
1465 | /* Only take pages on the LRU. */ | |
1466 | if (!PageLRU(page)) | |
1467 | return ret; | |
1468 | ||
e46a2879 MK |
1469 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1470 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1471 | return ret; |
1472 | ||
5ad333eb | 1473 | ret = -EBUSY; |
08e552c6 | 1474 | |
c8244935 MG |
1475 | /* |
1476 | * To minimise LRU disruption, the caller can indicate that it only | |
1477 | * wants to isolate pages it will be able to operate on without | |
1478 | * blocking - clean pages for the most part. | |
1479 | * | |
c8244935 MG |
1480 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1481 | * that it is possible to migrate without blocking | |
1482 | */ | |
1276ad68 | 1483 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
1484 | /* All the caller can do on PageWriteback is block */ |
1485 | if (PageWriteback(page)) | |
1486 | return ret; | |
1487 | ||
1488 | if (PageDirty(page)) { | |
1489 | struct address_space *mapping; | |
1f9c87e2 | 1490 | bool migrate_dirty; |
c8244935 | 1491 | |
c8244935 MG |
1492 | /* |
1493 | * Only pages without mappings or that have a | |
1494 | * ->migratepage callback are possible to migrate | |
1f9c87e2 MG |
1495 | * without blocking. However, we can be racing with |
1496 | * truncation so it's necessary to lock the page | |
1497 | * to stabilise the mapping as truncation holds | |
1498 | * the page lock until after the page is removed | |
1499 | * from the page cache. | |
c8244935 | 1500 | */ |
1f9c87e2 MG |
1501 | if (!trylock_page(page)) |
1502 | return ret; | |
1503 | ||
c8244935 | 1504 | mapping = page_mapping(page); |
b968dd76 | 1505 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
1f9c87e2 MG |
1506 | unlock_page(page); |
1507 | if (!migrate_dirty) | |
c8244935 MG |
1508 | return ret; |
1509 | } | |
1510 | } | |
39deaf85 | 1511 | |
f80c0673 MK |
1512 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1513 | return ret; | |
1514 | ||
5ad333eb AW |
1515 | if (likely(get_page_unless_zero(page))) { |
1516 | /* | |
1517 | * Be careful not to clear PageLRU until after we're | |
1518 | * sure the page is not being freed elsewhere -- the | |
1519 | * page release code relies on it. | |
1520 | */ | |
1521 | ClearPageLRU(page); | |
1522 | ret = 0; | |
1523 | } | |
1524 | ||
1525 | return ret; | |
1526 | } | |
1527 | ||
7ee36a14 MG |
1528 | |
1529 | /* | |
1530 | * Update LRU sizes after isolating pages. The LRU size updates must | |
1531 | * be complete before mem_cgroup_update_lru_size due to a santity check. | |
1532 | */ | |
1533 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 1534 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 1535 | { |
7ee36a14 MG |
1536 | int zid; |
1537 | ||
7ee36a14 MG |
1538 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1539 | if (!nr_zone_taken[zid]) | |
1540 | continue; | |
1541 | ||
1542 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
7ee36a14 | 1543 | #ifdef CONFIG_MEMCG |
b4536f0c | 1544 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
7ee36a14 | 1545 | #endif |
b4536f0c MH |
1546 | } |
1547 | ||
7ee36a14 MG |
1548 | } |
1549 | ||
1da177e4 | 1550 | /* |
a52633d8 | 1551 | * zone_lru_lock is heavily contended. Some of the functions that |
1da177e4 LT |
1552 | * shrink the lists perform better by taking out a batch of pages |
1553 | * and working on them outside the LRU lock. | |
1554 | * | |
1555 | * For pagecache intensive workloads, this function is the hottest | |
1556 | * spot in the kernel (apart from copy_*_user functions). | |
1557 | * | |
1558 | * Appropriate locks must be held before calling this function. | |
1559 | * | |
791b48b6 | 1560 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 1561 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1562 | * @dst: The temp list to put pages on to. |
f626012d | 1563 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1564 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1565 | * @mode: One of the LRU isolation modes |
3cb99451 | 1566 | * @lru: LRU list id for isolating |
1da177e4 LT |
1567 | * |
1568 | * returns how many pages were moved onto *@dst. | |
1569 | */ | |
69e05944 | 1570 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1571 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1572 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1573 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1574 | { |
75b00af7 | 1575 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1576 | unsigned long nr_taken = 0; |
599d0c95 | 1577 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 1578 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 1579 | unsigned long skipped = 0; |
791b48b6 | 1580 | unsigned long scan, total_scan, nr_pages; |
b2e18757 | 1581 | LIST_HEAD(pages_skipped); |
1da177e4 | 1582 | |
791b48b6 MK |
1583 | scan = 0; |
1584 | for (total_scan = 0; | |
1585 | scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src); | |
1586 | total_scan++) { | |
5ad333eb | 1587 | struct page *page; |
5ad333eb | 1588 | |
1da177e4 LT |
1589 | page = lru_to_page(src); |
1590 | prefetchw_prev_lru_page(page, src, flags); | |
1591 | ||
309381fe | 1592 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1593 | |
b2e18757 MG |
1594 | if (page_zonenum(page) > sc->reclaim_idx) { |
1595 | list_move(&page->lru, &pages_skipped); | |
7cc30fcf | 1596 | nr_skipped[page_zonenum(page)]++; |
b2e18757 MG |
1597 | continue; |
1598 | } | |
1599 | ||
791b48b6 MK |
1600 | /* |
1601 | * Do not count skipped pages because that makes the function | |
1602 | * return with no isolated pages if the LRU mostly contains | |
1603 | * ineligible pages. This causes the VM to not reclaim any | |
1604 | * pages, triggering a premature OOM. | |
1605 | */ | |
1606 | scan++; | |
f3fd4a61 | 1607 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1608 | case 0: |
599d0c95 MG |
1609 | nr_pages = hpage_nr_pages(page); |
1610 | nr_taken += nr_pages; | |
1611 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1612 | list_move(&page->lru, dst); |
5ad333eb AW |
1613 | break; |
1614 | ||
1615 | case -EBUSY: | |
1616 | /* else it is being freed elsewhere */ | |
1617 | list_move(&page->lru, src); | |
1618 | continue; | |
46453a6e | 1619 | |
5ad333eb AW |
1620 | default: |
1621 | BUG(); | |
1622 | } | |
1da177e4 LT |
1623 | } |
1624 | ||
b2e18757 MG |
1625 | /* |
1626 | * Splice any skipped pages to the start of the LRU list. Note that | |
1627 | * this disrupts the LRU order when reclaiming for lower zones but | |
1628 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1629 | * scanning would soon rescan the same pages to skip and put the | |
1630 | * system at risk of premature OOM. | |
1631 | */ | |
7cc30fcf MG |
1632 | if (!list_empty(&pages_skipped)) { |
1633 | int zid; | |
1634 | ||
3db65812 | 1635 | list_splice(&pages_skipped, src); |
7cc30fcf MG |
1636 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1637 | if (!nr_skipped[zid]) | |
1638 | continue; | |
1639 | ||
1640 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 1641 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
1642 | } |
1643 | } | |
791b48b6 | 1644 | *nr_scanned = total_scan; |
1265e3a6 | 1645 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
791b48b6 | 1646 | total_scan, skipped, nr_taken, mode, lru); |
b4536f0c | 1647 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
1648 | return nr_taken; |
1649 | } | |
1650 | ||
62695a84 NP |
1651 | /** |
1652 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1653 | * @page: page to isolate from its LRU list | |
1654 | * | |
1655 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1656 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1657 | * | |
1658 | * Returns 0 if the page was removed from an LRU list. | |
1659 | * Returns -EBUSY if the page was not on an LRU list. | |
1660 | * | |
1661 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1662 | * the active list, it will have PageActive set. If it was found on |
1663 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1664 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1665 | * |
1666 | * The vmstat statistic corresponding to the list on which the page was | |
1667 | * found will be decremented. | |
1668 | * | |
1669 | * Restrictions: | |
1670 | * (1) Must be called with an elevated refcount on the page. This is a | |
1671 | * fundamentnal difference from isolate_lru_pages (which is called | |
1672 | * without a stable reference). | |
1673 | * (2) the lru_lock must not be held. | |
1674 | * (3) interrupts must be enabled. | |
1675 | */ | |
1676 | int isolate_lru_page(struct page *page) | |
1677 | { | |
1678 | int ret = -EBUSY; | |
1679 | ||
309381fe | 1680 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1681 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1682 | |
62695a84 NP |
1683 | if (PageLRU(page)) { |
1684 | struct zone *zone = page_zone(page); | |
fa9add64 | 1685 | struct lruvec *lruvec; |
62695a84 | 1686 | |
a52633d8 | 1687 | spin_lock_irq(zone_lru_lock(zone)); |
599d0c95 | 1688 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
0c917313 | 1689 | if (PageLRU(page)) { |
894bc310 | 1690 | int lru = page_lru(page); |
0c917313 | 1691 | get_page(page); |
62695a84 | 1692 | ClearPageLRU(page); |
fa9add64 HD |
1693 | del_page_from_lru_list(page, lruvec, lru); |
1694 | ret = 0; | |
62695a84 | 1695 | } |
a52633d8 | 1696 | spin_unlock_irq(zone_lru_lock(zone)); |
62695a84 NP |
1697 | } |
1698 | return ret; | |
1699 | } | |
1700 | ||
35cd7815 | 1701 | /* |
d37dd5dc FW |
1702 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1703 | * then get resheduled. When there are massive number of tasks doing page | |
1704 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1705 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1706 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1707 | */ |
599d0c95 | 1708 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1709 | struct scan_control *sc) |
1710 | { | |
1711 | unsigned long inactive, isolated; | |
1712 | ||
1713 | if (current_is_kswapd()) | |
1714 | return 0; | |
1715 | ||
97c9341f | 1716 | if (!sane_reclaim(sc)) |
35cd7815 RR |
1717 | return 0; |
1718 | ||
1719 | if (file) { | |
599d0c95 MG |
1720 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1721 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1722 | } else { |
599d0c95 MG |
1723 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1724 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1725 | } |
1726 | ||
3cf23841 FW |
1727 | /* |
1728 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1729 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1730 | * deadlock. | |
1731 | */ | |
d0164adc | 1732 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1733 | inactive >>= 3; |
1734 | ||
35cd7815 RR |
1735 | return isolated > inactive; |
1736 | } | |
1737 | ||
66635629 | 1738 | static noinline_for_stack void |
75b00af7 | 1739 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1740 | { |
27ac81d8 | 1741 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
599d0c95 | 1742 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3f79768f | 1743 | LIST_HEAD(pages_to_free); |
66635629 | 1744 | |
66635629 MG |
1745 | /* |
1746 | * Put back any unfreeable pages. | |
1747 | */ | |
66635629 | 1748 | while (!list_empty(page_list)) { |
3f79768f | 1749 | struct page *page = lru_to_page(page_list); |
66635629 | 1750 | int lru; |
3f79768f | 1751 | |
309381fe | 1752 | VM_BUG_ON_PAGE(PageLRU(page), page); |
66635629 | 1753 | list_del(&page->lru); |
39b5f29a | 1754 | if (unlikely(!page_evictable(page))) { |
599d0c95 | 1755 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1756 | putback_lru_page(page); |
599d0c95 | 1757 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1758 | continue; |
1759 | } | |
fa9add64 | 1760 | |
599d0c95 | 1761 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1762 | |
7a608572 | 1763 | SetPageLRU(page); |
66635629 | 1764 | lru = page_lru(page); |
fa9add64 HD |
1765 | add_page_to_lru_list(page, lruvec, lru); |
1766 | ||
66635629 MG |
1767 | if (is_active_lru(lru)) { |
1768 | int file = is_file_lru(lru); | |
9992af10 RR |
1769 | int numpages = hpage_nr_pages(page); |
1770 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1771 | } |
2bcf8879 HD |
1772 | if (put_page_testzero(page)) { |
1773 | __ClearPageLRU(page); | |
1774 | __ClearPageActive(page); | |
fa9add64 | 1775 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1776 | |
1777 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1778 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1779 | mem_cgroup_uncharge(page); |
2bcf8879 | 1780 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1781 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1782 | } else |
1783 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1784 | } |
1785 | } | |
66635629 | 1786 | |
3f79768f HD |
1787 | /* |
1788 | * To save our caller's stack, now use input list for pages to free. | |
1789 | */ | |
1790 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1791 | } |
1792 | ||
399ba0b9 N |
1793 | /* |
1794 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1795 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1796 | * In that case we should only throttle if the backing device it is | |
1797 | * writing to is congested. In other cases it is safe to throttle. | |
1798 | */ | |
1799 | static int current_may_throttle(void) | |
1800 | { | |
1801 | return !(current->flags & PF_LESS_THROTTLE) || | |
1802 | current->backing_dev_info == NULL || | |
1803 | bdi_write_congested(current->backing_dev_info); | |
1804 | } | |
1805 | ||
1da177e4 | 1806 | /* |
b2e18757 | 1807 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1808 | * of reclaimed pages |
1da177e4 | 1809 | */ |
66635629 | 1810 | static noinline_for_stack unsigned long |
1a93be0e | 1811 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1812 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1813 | { |
1814 | LIST_HEAD(page_list); | |
e247dbce | 1815 | unsigned long nr_scanned; |
05ff5137 | 1816 | unsigned long nr_reclaimed = 0; |
e247dbce | 1817 | unsigned long nr_taken; |
3c710c1a | 1818 | struct reclaim_stat stat = {}; |
f3fd4a61 | 1819 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1820 | int file = is_file_lru(lru); |
599d0c95 | 1821 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1822 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
db73ee0d | 1823 | bool stalled = false; |
78dc583d | 1824 | |
599d0c95 | 1825 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
1826 | if (stalled) |
1827 | return 0; | |
1828 | ||
1829 | /* wait a bit for the reclaimer. */ | |
1830 | msleep(100); | |
1831 | stalled = true; | |
35cd7815 RR |
1832 | |
1833 | /* We are about to die and free our memory. Return now. */ | |
1834 | if (fatal_signal_pending(current)) | |
1835 | return SWAP_CLUSTER_MAX; | |
1836 | } | |
1837 | ||
1da177e4 | 1838 | lru_add_drain(); |
f80c0673 MK |
1839 | |
1840 | if (!sc->may_unmap) | |
61317289 | 1841 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1842 | |
599d0c95 | 1843 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1844 | |
5dc35979 KK |
1845 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1846 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc | 1847 | |
599d0c95 | 1848 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1849 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1850 | |
2262185c RG |
1851 | if (current_is_kswapd()) { |
1852 | if (global_reclaim(sc)) | |
599d0c95 | 1853 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
2262185c RG |
1854 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD, |
1855 | nr_scanned); | |
1856 | } else { | |
1857 | if (global_reclaim(sc)) | |
599d0c95 | 1858 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
2262185c RG |
1859 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT, |
1860 | nr_scanned); | |
e247dbce | 1861 | } |
599d0c95 | 1862 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1863 | |
d563c050 | 1864 | if (nr_taken == 0) |
66635629 | 1865 | return 0; |
5ad333eb | 1866 | |
a128ca71 | 1867 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, |
3c710c1a | 1868 | &stat, false); |
c661b078 | 1869 | |
599d0c95 | 1870 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 1871 | |
2262185c RG |
1872 | if (current_is_kswapd()) { |
1873 | if (global_reclaim(sc)) | |
599d0c95 | 1874 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
2262185c RG |
1875 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD, |
1876 | nr_reclaimed); | |
1877 | } else { | |
1878 | if (global_reclaim(sc)) | |
599d0c95 | 1879 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
2262185c RG |
1880 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT, |
1881 | nr_reclaimed); | |
904249aa | 1882 | } |
a74609fa | 1883 | |
27ac81d8 | 1884 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1885 | |
599d0c95 | 1886 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 1887 | |
599d0c95 | 1888 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 1889 | |
747db954 | 1890 | mem_cgroup_uncharge_list(&page_list); |
b745bc85 | 1891 | free_hot_cold_page_list(&page_list, true); |
e11da5b4 | 1892 | |
92df3a72 MG |
1893 | /* |
1894 | * If reclaim is isolating dirty pages under writeback, it implies | |
1895 | * that the long-lived page allocation rate is exceeding the page | |
1896 | * laundering rate. Either the global limits are not being effective | |
1897 | * at throttling processes due to the page distribution throughout | |
1898 | * zones or there is heavy usage of a slow backing device. The | |
1899 | * only option is to throttle from reclaim context which is not ideal | |
1900 | * as there is no guarantee the dirtying process is throttled in the | |
1901 | * same way balance_dirty_pages() manages. | |
1902 | * | |
8e950282 MG |
1903 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1904 | * of pages under pages flagged for immediate reclaim and stall if any | |
1905 | * are encountered in the nr_immediate check below. | |
92df3a72 | 1906 | */ |
3c710c1a | 1907 | if (stat.nr_writeback && stat.nr_writeback == nr_taken) |
599d0c95 | 1908 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); |
92df3a72 | 1909 | |
b8b81518 AR |
1910 | /* |
1911 | * If dirty pages are scanned that are not queued for IO, it | |
1912 | * implies that flushers are not doing their job. This can | |
1913 | * happen when memory pressure pushes dirty pages to the end of | |
1914 | * the LRU before the dirty limits are breached and the dirty | |
1915 | * data has expired. It can also happen when the proportion of | |
1916 | * dirty pages grows not through writes but through memory | |
1917 | * pressure reclaiming all the clean cache. And in some cases, | |
1918 | * the flushers simply cannot keep up with the allocation | |
1919 | * rate. Nudge the flusher threads in case they are asleep. | |
1920 | */ | |
1921 | if (stat.nr_unqueued_dirty == nr_taken) | |
1922 | wakeup_flusher_threads(0, WB_REASON_VMSCAN); | |
1923 | ||
d43006d5 | 1924 | /* |
97c9341f TH |
1925 | * Legacy memcg will stall in page writeback so avoid forcibly |
1926 | * stalling here. | |
d43006d5 | 1927 | */ |
97c9341f | 1928 | if (sane_reclaim(sc)) { |
8e950282 MG |
1929 | /* |
1930 | * Tag a zone as congested if all the dirty pages scanned were | |
1931 | * backed by a congested BDI and wait_iff_congested will stall. | |
1932 | */ | |
3c710c1a | 1933 | if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested) |
599d0c95 | 1934 | set_bit(PGDAT_CONGESTED, &pgdat->flags); |
8e950282 | 1935 | |
b8b81518 AR |
1936 | /* Allow kswapd to start writing pages during reclaim. */ |
1937 | if (stat.nr_unqueued_dirty == nr_taken) | |
599d0c95 | 1938 | set_bit(PGDAT_DIRTY, &pgdat->flags); |
b1a6f21e MG |
1939 | |
1940 | /* | |
b738d764 LT |
1941 | * If kswapd scans pages marked marked for immediate |
1942 | * reclaim and under writeback (nr_immediate), it implies | |
1943 | * that pages are cycling through the LRU faster than | |
b1a6f21e MG |
1944 | * they are written so also forcibly stall. |
1945 | */ | |
3c710c1a | 1946 | if (stat.nr_immediate && current_may_throttle()) |
b1a6f21e | 1947 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
e2be15f6 | 1948 | } |
d43006d5 | 1949 | |
8e950282 MG |
1950 | /* |
1951 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1952 | * is congested. Allow kswapd to continue until it starts encountering | |
1953 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1954 | */ | |
399ba0b9 N |
1955 | if (!sc->hibernation_mode && !current_is_kswapd() && |
1956 | current_may_throttle()) | |
599d0c95 | 1957 | wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); |
8e950282 | 1958 | |
599d0c95 MG |
1959 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
1960 | nr_scanned, nr_reclaimed, | |
5bccd166 MH |
1961 | stat.nr_dirty, stat.nr_writeback, |
1962 | stat.nr_congested, stat.nr_immediate, | |
1963 | stat.nr_activate, stat.nr_ref_keep, | |
1964 | stat.nr_unmap_fail, | |
ba5e9579 | 1965 | sc->priority, file); |
05ff5137 | 1966 | return nr_reclaimed; |
1da177e4 LT |
1967 | } |
1968 | ||
1969 | /* | |
1970 | * This moves pages from the active list to the inactive list. | |
1971 | * | |
1972 | * We move them the other way if the page is referenced by one or more | |
1973 | * processes, from rmap. | |
1974 | * | |
1975 | * If the pages are mostly unmapped, the processing is fast and it is | |
a52633d8 | 1976 | * appropriate to hold zone_lru_lock across the whole operation. But if |
1da177e4 | 1977 | * the pages are mapped, the processing is slow (page_referenced()) so we |
a52633d8 | 1978 | * should drop zone_lru_lock around each page. It's impossible to balance |
1da177e4 LT |
1979 | * this, so instead we remove the pages from the LRU while processing them. |
1980 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1981 | * nobody will play with that bit on a non-LRU page. | |
1982 | * | |
0139aa7b | 1983 | * The downside is that we have to touch page->_refcount against each page. |
1da177e4 | 1984 | * But we had to alter page->flags anyway. |
9d998b4f MH |
1985 | * |
1986 | * Returns the number of pages moved to the given lru. | |
1da177e4 | 1987 | */ |
1cfb419b | 1988 | |
9d998b4f | 1989 | static unsigned move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1990 | struct list_head *list, |
2bcf8879 | 1991 | struct list_head *pages_to_free, |
3eb4140f WF |
1992 | enum lru_list lru) |
1993 | { | |
599d0c95 | 1994 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3eb4140f | 1995 | struct page *page; |
fa9add64 | 1996 | int nr_pages; |
9d998b4f | 1997 | int nr_moved = 0; |
3eb4140f | 1998 | |
3eb4140f WF |
1999 | while (!list_empty(list)) { |
2000 | page = lru_to_page(list); | |
599d0c95 | 2001 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
3eb4140f | 2002 | |
309381fe | 2003 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3eb4140f WF |
2004 | SetPageLRU(page); |
2005 | ||
fa9add64 | 2006 | nr_pages = hpage_nr_pages(page); |
599d0c95 | 2007 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
925b7673 | 2008 | list_move(&page->lru, &lruvec->lists[lru]); |
3eb4140f | 2009 | |
2bcf8879 HD |
2010 | if (put_page_testzero(page)) { |
2011 | __ClearPageLRU(page); | |
2012 | __ClearPageActive(page); | |
fa9add64 | 2013 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
2014 | |
2015 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 2016 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 2017 | mem_cgroup_uncharge(page); |
2bcf8879 | 2018 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 2019 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
2020 | } else |
2021 | list_add(&page->lru, pages_to_free); | |
9d998b4f MH |
2022 | } else { |
2023 | nr_moved += nr_pages; | |
3eb4140f WF |
2024 | } |
2025 | } | |
9d5e6a9f | 2026 | |
2262185c | 2027 | if (!is_active_lru(lru)) { |
f0958906 | 2028 | __count_vm_events(PGDEACTIVATE, nr_moved); |
2262185c RG |
2029 | count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, |
2030 | nr_moved); | |
2031 | } | |
9d998b4f MH |
2032 | |
2033 | return nr_moved; | |
3eb4140f | 2034 | } |
1cfb419b | 2035 | |
f626012d | 2036 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 2037 | struct lruvec *lruvec, |
f16015fb | 2038 | struct scan_control *sc, |
9e3b2f8c | 2039 | enum lru_list lru) |
1da177e4 | 2040 | { |
44c241f1 | 2041 | unsigned long nr_taken; |
f626012d | 2042 | unsigned long nr_scanned; |
6fe6b7e3 | 2043 | unsigned long vm_flags; |
1da177e4 | 2044 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 2045 | LIST_HEAD(l_active); |
b69408e8 | 2046 | LIST_HEAD(l_inactive); |
1da177e4 | 2047 | struct page *page; |
1a93be0e | 2048 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
9d998b4f MH |
2049 | unsigned nr_deactivate, nr_activate; |
2050 | unsigned nr_rotated = 0; | |
f3fd4a61 | 2051 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 2052 | int file = is_file_lru(lru); |
599d0c95 | 2053 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2054 | |
2055 | lru_add_drain(); | |
f80c0673 MK |
2056 | |
2057 | if (!sc->may_unmap) | |
61317289 | 2058 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 2059 | |
599d0c95 | 2060 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 2061 | |
5dc35979 KK |
2062 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
2063 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 2064 | |
599d0c95 | 2065 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 2066 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 2067 | |
599d0c95 | 2068 | __count_vm_events(PGREFILL, nr_scanned); |
2262185c | 2069 | count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2070 | |
599d0c95 | 2071 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 2072 | |
1da177e4 LT |
2073 | while (!list_empty(&l_hold)) { |
2074 | cond_resched(); | |
2075 | page = lru_to_page(&l_hold); | |
2076 | list_del(&page->lru); | |
7e9cd484 | 2077 | |
39b5f29a | 2078 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
2079 | putback_lru_page(page); |
2080 | continue; | |
2081 | } | |
2082 | ||
cc715d99 MG |
2083 | if (unlikely(buffer_heads_over_limit)) { |
2084 | if (page_has_private(page) && trylock_page(page)) { | |
2085 | if (page_has_private(page)) | |
2086 | try_to_release_page(page, 0); | |
2087 | unlock_page(page); | |
2088 | } | |
2089 | } | |
2090 | ||
c3ac9a8a JW |
2091 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
2092 | &vm_flags)) { | |
9992af10 | 2093 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
2094 | /* |
2095 | * Identify referenced, file-backed active pages and | |
2096 | * give them one more trip around the active list. So | |
2097 | * that executable code get better chances to stay in | |
2098 | * memory under moderate memory pressure. Anon pages | |
2099 | * are not likely to be evicted by use-once streaming | |
2100 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
2101 | * so we ignore them here. | |
2102 | */ | |
41e20983 | 2103 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
2104 | list_add(&page->lru, &l_active); |
2105 | continue; | |
2106 | } | |
2107 | } | |
7e9cd484 | 2108 | |
5205e56e | 2109 | ClearPageActive(page); /* we are de-activating */ |
15c7f452 | 2110 | SetPageWorkingset(page); |
1da177e4 LT |
2111 | list_add(&page->lru, &l_inactive); |
2112 | } | |
2113 | ||
b555749a | 2114 | /* |
8cab4754 | 2115 | * Move pages back to the lru list. |
b555749a | 2116 | */ |
599d0c95 | 2117 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 2118 | /* |
8cab4754 WF |
2119 | * Count referenced pages from currently used mappings as rotated, |
2120 | * even though only some of them are actually re-activated. This | |
2121 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 2122 | * get_scan_count. |
7e9cd484 | 2123 | */ |
b7c46d15 | 2124 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 2125 | |
9d998b4f MH |
2126 | nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
2127 | nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
599d0c95 MG |
2128 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
2129 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 2130 | |
747db954 | 2131 | mem_cgroup_uncharge_list(&l_hold); |
b745bc85 | 2132 | free_hot_cold_page_list(&l_hold, true); |
9d998b4f MH |
2133 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2134 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2135 | } |
2136 | ||
59dc76b0 RR |
2137 | /* |
2138 | * The inactive anon list should be small enough that the VM never has | |
2139 | * to do too much work. | |
14797e23 | 2140 | * |
59dc76b0 RR |
2141 | * The inactive file list should be small enough to leave most memory |
2142 | * to the established workingset on the scan-resistant active list, | |
2143 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2144 | * |
59dc76b0 RR |
2145 | * Both inactive lists should also be large enough that each inactive |
2146 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2147 | * |
2a2e4885 JW |
2148 | * If that fails and refaulting is observed, the inactive list grows. |
2149 | * | |
59dc76b0 RR |
2150 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
2151 | * on this LRU, maintained by the pageout code. A zone->inactive_ratio | |
2152 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. | |
56e49d21 | 2153 | * |
59dc76b0 RR |
2154 | * total target max |
2155 | * memory ratio inactive | |
2156 | * ------------------------------------- | |
2157 | * 10MB 1 5MB | |
2158 | * 100MB 1 50MB | |
2159 | * 1GB 3 250MB | |
2160 | * 10GB 10 0.9GB | |
2161 | * 100GB 31 3GB | |
2162 | * 1TB 101 10GB | |
2163 | * 10TB 320 32GB | |
56e49d21 | 2164 | */ |
f8d1a311 | 2165 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
2a2e4885 JW |
2166 | struct mem_cgroup *memcg, |
2167 | struct scan_control *sc, bool actual_reclaim) | |
56e49d21 | 2168 | { |
fd538803 | 2169 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; |
2a2e4885 JW |
2170 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
2171 | enum lru_list inactive_lru = file * LRU_FILE; | |
2172 | unsigned long inactive, active; | |
2173 | unsigned long inactive_ratio; | |
2174 | unsigned long refaults; | |
59dc76b0 | 2175 | unsigned long gb; |
e3790144 | 2176 | |
59dc76b0 RR |
2177 | /* |
2178 | * If we don't have swap space, anonymous page deactivation | |
2179 | * is pointless. | |
2180 | */ | |
2181 | if (!file && !total_swap_pages) | |
2182 | return false; | |
56e49d21 | 2183 | |
fd538803 MH |
2184 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
2185 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); | |
f8d1a311 | 2186 | |
2a2e4885 | 2187 | if (memcg) |
ccda7f43 | 2188 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); |
b39415b2 | 2189 | else |
2a2e4885 JW |
2190 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); |
2191 | ||
2192 | /* | |
2193 | * When refaults are being observed, it means a new workingset | |
2194 | * is being established. Disable active list protection to get | |
2195 | * rid of the stale workingset quickly. | |
2196 | */ | |
2197 | if (file && actual_reclaim && lruvec->refaults != refaults) { | |
2198 | inactive_ratio = 0; | |
2199 | } else { | |
2200 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
2201 | if (gb) | |
2202 | inactive_ratio = int_sqrt(10 * gb); | |
2203 | else | |
2204 | inactive_ratio = 1; | |
2205 | } | |
59dc76b0 | 2206 | |
2a2e4885 JW |
2207 | if (actual_reclaim) |
2208 | trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, | |
2209 | lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, | |
2210 | lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, | |
2211 | inactive_ratio, file); | |
fd538803 | 2212 | |
59dc76b0 | 2213 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2214 | } |
2215 | ||
4f98a2fe | 2216 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
2a2e4885 JW |
2217 | struct lruvec *lruvec, struct mem_cgroup *memcg, |
2218 | struct scan_control *sc) | |
b69408e8 | 2219 | { |
b39415b2 | 2220 | if (is_active_lru(lru)) { |
2a2e4885 JW |
2221 | if (inactive_list_is_low(lruvec, is_file_lru(lru), |
2222 | memcg, sc, true)) | |
1a93be0e | 2223 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
2224 | return 0; |
2225 | } | |
2226 | ||
1a93be0e | 2227 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
2228 | } |
2229 | ||
9a265114 JW |
2230 | enum scan_balance { |
2231 | SCAN_EQUAL, | |
2232 | SCAN_FRACT, | |
2233 | SCAN_ANON, | |
2234 | SCAN_FILE, | |
2235 | }; | |
2236 | ||
4f98a2fe RR |
2237 | /* |
2238 | * Determine how aggressively the anon and file LRU lists should be | |
2239 | * scanned. The relative value of each set of LRU lists is determined | |
2240 | * by looking at the fraction of the pages scanned we did rotate back | |
2241 | * onto the active list instead of evict. | |
2242 | * | |
be7bd59d WL |
2243 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2244 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2245 | */ |
33377678 | 2246 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
6b4f7799 JW |
2247 | struct scan_control *sc, unsigned long *nr, |
2248 | unsigned long *lru_pages) | |
4f98a2fe | 2249 | { |
33377678 | 2250 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
2251 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2252 | u64 fraction[2]; | |
2253 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 2254 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 2255 | unsigned long anon_prio, file_prio; |
9a265114 | 2256 | enum scan_balance scan_balance; |
0bf1457f | 2257 | unsigned long anon, file; |
4f98a2fe | 2258 | unsigned long ap, fp; |
4111304d | 2259 | enum lru_list lru; |
76a33fc3 SL |
2260 | |
2261 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2262 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2263 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2264 | goto out; |
2265 | } | |
4f98a2fe | 2266 | |
10316b31 JW |
2267 | /* |
2268 | * Global reclaim will swap to prevent OOM even with no | |
2269 | * swappiness, but memcg users want to use this knob to | |
2270 | * disable swapping for individual groups completely when | |
2271 | * using the memory controller's swap limit feature would be | |
2272 | * too expensive. | |
2273 | */ | |
02695175 | 2274 | if (!global_reclaim(sc) && !swappiness) { |
9a265114 | 2275 | scan_balance = SCAN_FILE; |
10316b31 JW |
2276 | goto out; |
2277 | } | |
2278 | ||
2279 | /* | |
2280 | * Do not apply any pressure balancing cleverness when the | |
2281 | * system is close to OOM, scan both anon and file equally | |
2282 | * (unless the swappiness setting disagrees with swapping). | |
2283 | */ | |
02695175 | 2284 | if (!sc->priority && swappiness) { |
9a265114 | 2285 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2286 | goto out; |
2287 | } | |
2288 | ||
62376251 JW |
2289 | /* |
2290 | * Prevent the reclaimer from falling into the cache trap: as | |
2291 | * cache pages start out inactive, every cache fault will tip | |
2292 | * the scan balance towards the file LRU. And as the file LRU | |
2293 | * shrinks, so does the window for rotation from references. | |
2294 | * This means we have a runaway feedback loop where a tiny | |
2295 | * thrashing file LRU becomes infinitely more attractive than | |
2296 | * anon pages. Try to detect this based on file LRU size. | |
2297 | */ | |
2298 | if (global_reclaim(sc)) { | |
599d0c95 MG |
2299 | unsigned long pgdatfile; |
2300 | unsigned long pgdatfree; | |
2301 | int z; | |
2302 | unsigned long total_high_wmark = 0; | |
2ab051e1 | 2303 | |
599d0c95 MG |
2304 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2305 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2306 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2307 | ||
2308 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2309 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2310 | if (!managed_zone(zone)) |
599d0c95 MG |
2311 | continue; |
2312 | ||
2313 | total_high_wmark += high_wmark_pages(zone); | |
2314 | } | |
62376251 | 2315 | |
599d0c95 | 2316 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
06226226 DR |
2317 | /* |
2318 | * Force SCAN_ANON if there are enough inactive | |
2319 | * anonymous pages on the LRU in eligible zones. | |
2320 | * Otherwise, the small LRU gets thrashed. | |
2321 | */ | |
2322 | if (!inactive_list_is_low(lruvec, false, memcg, sc, false) && | |
2323 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx) | |
2324 | >> sc->priority) { | |
2325 | scan_balance = SCAN_ANON; | |
2326 | goto out; | |
2327 | } | |
62376251 JW |
2328 | } |
2329 | } | |
2330 | ||
7c5bd705 | 2331 | /* |
316bda0e VD |
2332 | * If there is enough inactive page cache, i.e. if the size of the |
2333 | * inactive list is greater than that of the active list *and* the | |
2334 | * inactive list actually has some pages to scan on this priority, we | |
2335 | * do not reclaim anything from the anonymous working set right now. | |
2336 | * Without the second condition we could end up never scanning an | |
2337 | * lruvec even if it has plenty of old anonymous pages unless the | |
2338 | * system is under heavy pressure. | |
7c5bd705 | 2339 | */ |
bb9d04e5 | 2340 | if (!IS_ENABLED(CONFIG_BALANCE_ANON_FILE_RECLAIM) && |
2341 | !inactive_list_is_low(lruvec, true, memcg, sc, false) && | |
71ab6cfe | 2342 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
9a265114 | 2343 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2344 | goto out; |
2345 | } | |
2346 | ||
9a265114 JW |
2347 | scan_balance = SCAN_FRACT; |
2348 | ||
58c37f6e KM |
2349 | /* |
2350 | * With swappiness at 100, anonymous and file have the same priority. | |
2351 | * This scanning priority is essentially the inverse of IO cost. | |
2352 | */ | |
02695175 | 2353 | anon_prio = swappiness; |
75b00af7 | 2354 | file_prio = 200 - anon_prio; |
58c37f6e | 2355 | |
4f98a2fe RR |
2356 | /* |
2357 | * OK, so we have swap space and a fair amount of page cache | |
2358 | * pages. We use the recently rotated / recently scanned | |
2359 | * ratios to determine how valuable each cache is. | |
2360 | * | |
2361 | * Because workloads change over time (and to avoid overflow) | |
2362 | * we keep these statistics as a floating average, which ends | |
2363 | * up weighing recent references more than old ones. | |
2364 | * | |
2365 | * anon in [0], file in [1] | |
2366 | */ | |
2ab051e1 | 2367 | |
fd538803 MH |
2368 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
2369 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); | |
2370 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + | |
2371 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); | |
2ab051e1 | 2372 | |
599d0c95 | 2373 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2374 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2375 | reclaim_stat->recent_scanned[0] /= 2; |
2376 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2377 | } |
2378 | ||
6e901571 | 2379 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2380 | reclaim_stat->recent_scanned[1] /= 2; |
2381 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2382 | } |
2383 | ||
4f98a2fe | 2384 | /* |
00d8089c RR |
2385 | * The amount of pressure on anon vs file pages is inversely |
2386 | * proportional to the fraction of recently scanned pages on | |
2387 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2388 | */ |
fe35004f | 2389 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2390 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2391 | |
fe35004f | 2392 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2393 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2394 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2395 | |
76a33fc3 SL |
2396 | fraction[0] = ap; |
2397 | fraction[1] = fp; | |
2398 | denominator = ap + fp + 1; | |
2399 | out: | |
688035f7 JW |
2400 | *lru_pages = 0; |
2401 | for_each_evictable_lru(lru) { | |
2402 | int file = is_file_lru(lru); | |
2403 | unsigned long size; | |
2404 | unsigned long scan; | |
6b4f7799 | 2405 | |
688035f7 JW |
2406 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); |
2407 | scan = size >> sc->priority; | |
2408 | /* | |
2409 | * If the cgroup's already been deleted, make sure to | |
2410 | * scrape out the remaining cache. | |
2411 | */ | |
2412 | if (!scan && !mem_cgroup_online(memcg)) | |
2413 | scan = min(size, SWAP_CLUSTER_MAX); | |
6b4f7799 | 2414 | |
688035f7 JW |
2415 | switch (scan_balance) { |
2416 | case SCAN_EQUAL: | |
2417 | /* Scan lists relative to size */ | |
2418 | break; | |
2419 | case SCAN_FRACT: | |
9a265114 | 2420 | /* |
688035f7 JW |
2421 | * Scan types proportional to swappiness and |
2422 | * their relative recent reclaim efficiency. | |
36f93a2e RG |
2423 | * Make sure we don't miss the last page |
2424 | * because of a round-off error. | |
9a265114 | 2425 | */ |
36f93a2e RG |
2426 | scan = DIV64_U64_ROUND_UP(scan * fraction[file], |
2427 | denominator); | |
688035f7 JW |
2428 | break; |
2429 | case SCAN_FILE: | |
2430 | case SCAN_ANON: | |
2431 | /* Scan one type exclusively */ | |
2432 | if ((scan_balance == SCAN_FILE) != file) { | |
2433 | size = 0; | |
2434 | scan = 0; | |
2435 | } | |
2436 | break; | |
2437 | default: | |
2438 | /* Look ma, no brain */ | |
2439 | BUG(); | |
9a265114 | 2440 | } |
688035f7 JW |
2441 | |
2442 | *lru_pages += size; | |
2443 | nr[lru] = scan; | |
76a33fc3 | 2444 | } |
6e08a369 | 2445 | } |
4f98a2fe | 2446 | |
9b4f98cd | 2447 | /* |
a9dd0a83 | 2448 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. |
9b4f98cd | 2449 | */ |
a9dd0a83 | 2450 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, |
33377678 | 2451 | struct scan_control *sc, unsigned long *lru_pages) |
9b4f98cd | 2452 | { |
ef8f2327 | 2453 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
9b4f98cd | 2454 | unsigned long nr[NR_LRU_LISTS]; |
e82e0561 | 2455 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2456 | unsigned long nr_to_scan; |
2457 | enum lru_list lru; | |
2458 | unsigned long nr_reclaimed = 0; | |
2459 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2460 | struct blk_plug plug; | |
1a501907 | 2461 | bool scan_adjusted; |
9b4f98cd | 2462 | |
33377678 | 2463 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
9b4f98cd | 2464 | |
e82e0561 MG |
2465 | /* Record the original scan target for proportional adjustments later */ |
2466 | memcpy(targets, nr, sizeof(nr)); | |
2467 | ||
1a501907 MG |
2468 | /* |
2469 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2470 | * event that can occur when there is little memory pressure e.g. | |
2471 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2472 | * when the requested number of pages are reclaimed when scanning at | |
2473 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2474 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2475 | * do a batch of work at once. For memcg reclaim one check is made to | |
2476 | * abort proportional reclaim if either the file or anon lru has already | |
2477 | * dropped to zero at the first pass. | |
2478 | */ | |
2479 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2480 | sc->priority == DEF_PRIORITY); | |
2481 | ||
9b4f98cd JW |
2482 | blk_start_plug(&plug); |
2483 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2484 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2485 | unsigned long nr_anon, nr_file, percentage; |
2486 | unsigned long nr_scanned; | |
2487 | ||
9b4f98cd JW |
2488 | for_each_evictable_lru(lru) { |
2489 | if (nr[lru]) { | |
2490 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2491 | nr[lru] -= nr_to_scan; | |
2492 | ||
2493 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2a2e4885 | 2494 | lruvec, memcg, sc); |
9b4f98cd JW |
2495 | } |
2496 | } | |
e82e0561 | 2497 | |
bd041733 MH |
2498 | cond_resched(); |
2499 | ||
e82e0561 MG |
2500 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2501 | continue; | |
2502 | ||
e82e0561 MG |
2503 | /* |
2504 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2505 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2506 | * proportionally what was requested by get_scan_count(). We |
2507 | * stop reclaiming one LRU and reduce the amount scanning | |
2508 | * proportional to the original scan target. | |
2509 | */ | |
2510 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2511 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2512 | ||
1a501907 MG |
2513 | /* |
2514 | * It's just vindictive to attack the larger once the smaller | |
2515 | * has gone to zero. And given the way we stop scanning the | |
2516 | * smaller below, this makes sure that we only make one nudge | |
2517 | * towards proportionality once we've got nr_to_reclaim. | |
2518 | */ | |
2519 | if (!nr_file || !nr_anon) | |
2520 | break; | |
2521 | ||
e82e0561 MG |
2522 | if (nr_file > nr_anon) { |
2523 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2524 | targets[LRU_ACTIVE_ANON] + 1; | |
2525 | lru = LRU_BASE; | |
2526 | percentage = nr_anon * 100 / scan_target; | |
2527 | } else { | |
2528 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2529 | targets[LRU_ACTIVE_FILE] + 1; | |
2530 | lru = LRU_FILE; | |
2531 | percentage = nr_file * 100 / scan_target; | |
2532 | } | |
2533 | ||
2534 | /* Stop scanning the smaller of the LRU */ | |
2535 | nr[lru] = 0; | |
2536 | nr[lru + LRU_ACTIVE] = 0; | |
2537 | ||
2538 | /* | |
2539 | * Recalculate the other LRU scan count based on its original | |
2540 | * scan target and the percentage scanning already complete | |
2541 | */ | |
2542 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2543 | nr_scanned = targets[lru] - nr[lru]; | |
2544 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2545 | nr[lru] -= min(nr[lru], nr_scanned); | |
2546 | ||
2547 | lru += LRU_ACTIVE; | |
2548 | nr_scanned = targets[lru] - nr[lru]; | |
2549 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2550 | nr[lru] -= min(nr[lru], nr_scanned); | |
2551 | ||
2552 | scan_adjusted = true; | |
9b4f98cd JW |
2553 | } |
2554 | blk_finish_plug(&plug); | |
2555 | sc->nr_reclaimed += nr_reclaimed; | |
2556 | ||
2557 | /* | |
2558 | * Even if we did not try to evict anon pages at all, we want to | |
2559 | * rebalance the anon lru active/inactive ratio. | |
2560 | */ | |
2a2e4885 | 2561 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) |
9b4f98cd JW |
2562 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2563 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
2564 | } |
2565 | ||
23b9da55 | 2566 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2567 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2568 | { |
d84da3f9 | 2569 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2570 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2571 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2572 | return true; |
2573 | ||
2574 | return false; | |
2575 | } | |
2576 | ||
3e7d3449 | 2577 | /* |
23b9da55 MG |
2578 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2579 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2580 | * true if more pages should be reclaimed such that when the page allocator | |
2581 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2582 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2583 | */ |
a9dd0a83 | 2584 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 MG |
2585 | unsigned long nr_reclaimed, |
2586 | unsigned long nr_scanned, | |
2587 | struct scan_control *sc) | |
2588 | { | |
2589 | unsigned long pages_for_compaction; | |
2590 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2591 | int z; |
3e7d3449 MG |
2592 | |
2593 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2594 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2595 | return false; |
2596 | ||
2876592f | 2597 | /* Consider stopping depending on scan and reclaim activity */ |
dcda9b04 | 2598 | if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) { |
2876592f | 2599 | /* |
dcda9b04 | 2600 | * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the |
2876592f MG |
2601 | * full LRU list has been scanned and we are still failing |
2602 | * to reclaim pages. This full LRU scan is potentially | |
dcda9b04 | 2603 | * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed |
2876592f MG |
2604 | */ |
2605 | if (!nr_reclaimed && !nr_scanned) | |
2606 | return false; | |
2607 | } else { | |
2608 | /* | |
dcda9b04 | 2609 | * For non-__GFP_RETRY_MAYFAIL allocations which can presumably |
2876592f MG |
2610 | * fail without consequence, stop if we failed to reclaim |
2611 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2612 | * pages that were scanned. This will return to the | |
2613 | * caller faster at the risk reclaim/compaction and | |
2614 | * the resulting allocation attempt fails | |
2615 | */ | |
2616 | if (!nr_reclaimed) | |
2617 | return false; | |
2618 | } | |
3e7d3449 MG |
2619 | |
2620 | /* | |
2621 | * If we have not reclaimed enough pages for compaction and the | |
2622 | * inactive lists are large enough, continue reclaiming | |
2623 | */ | |
9861a62c | 2624 | pages_for_compaction = compact_gap(sc->order); |
a9dd0a83 | 2625 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
ec8acf20 | 2626 | if (get_nr_swap_pages() > 0) |
a9dd0a83 | 2627 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
3e7d3449 MG |
2628 | if (sc->nr_reclaimed < pages_for_compaction && |
2629 | inactive_lru_pages > pages_for_compaction) | |
2630 | return true; | |
2631 | ||
2632 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
a9dd0a83 MG |
2633 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2634 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2635 | if (!managed_zone(zone)) |
a9dd0a83 MG |
2636 | continue; |
2637 | ||
2638 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 2639 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
2640 | case COMPACT_CONTINUE: |
2641 | return false; | |
2642 | default: | |
2643 | /* check next zone */ | |
2644 | ; | |
2645 | } | |
3e7d3449 | 2646 | } |
a9dd0a83 | 2647 | return true; |
3e7d3449 MG |
2648 | } |
2649 | ||
970a39a3 | 2650 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2651 | { |
cb731d6c | 2652 | struct reclaim_state *reclaim_state = current->reclaim_state; |
f0fdc5e8 | 2653 | unsigned long nr_reclaimed, nr_scanned; |
2344d7e4 | 2654 | bool reclaimable = false; |
1da177e4 | 2655 | |
9b4f98cd JW |
2656 | do { |
2657 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2658 | struct mem_cgroup_reclaim_cookie reclaim = { | |
ef8f2327 | 2659 | .pgdat = pgdat, |
9b4f98cd JW |
2660 | .priority = sc->priority, |
2661 | }; | |
a9dd0a83 | 2662 | unsigned long node_lru_pages = 0; |
694fbc0f | 2663 | struct mem_cgroup *memcg; |
3e7d3449 | 2664 | |
9b4f98cd JW |
2665 | nr_reclaimed = sc->nr_reclaimed; |
2666 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2667 | |
694fbc0f AM |
2668 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2669 | do { | |
6b4f7799 | 2670 | unsigned long lru_pages; |
8e8ae645 | 2671 | unsigned long reclaimed; |
cb731d6c | 2672 | unsigned long scanned; |
5660048c | 2673 | |
241994ed | 2674 | if (mem_cgroup_low(root, memcg)) { |
d6622f63 YX |
2675 | if (!sc->memcg_low_reclaim) { |
2676 | sc->memcg_low_skipped = 1; | |
241994ed | 2677 | continue; |
d6622f63 | 2678 | } |
31176c78 | 2679 | mem_cgroup_event(memcg, MEMCG_LOW); |
241994ed JW |
2680 | } |
2681 | ||
8e8ae645 | 2682 | reclaimed = sc->nr_reclaimed; |
cb731d6c | 2683 | scanned = sc->nr_scanned; |
f9be23d6 | 2684 | |
a9dd0a83 MG |
2685 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
2686 | node_lru_pages += lru_pages; | |
f16015fb | 2687 | |
b5afba29 | 2688 | if (memcg) |
a9dd0a83 | 2689 | shrink_slab(sc->gfp_mask, pgdat->node_id, |
cb731d6c VD |
2690 | memcg, sc->nr_scanned - scanned, |
2691 | lru_pages); | |
2692 | ||
8e8ae645 JW |
2693 | /* Record the group's reclaim efficiency */ |
2694 | vmpressure(sc->gfp_mask, memcg, false, | |
2695 | sc->nr_scanned - scanned, | |
2696 | sc->nr_reclaimed - reclaimed); | |
2697 | ||
9b4f98cd | 2698 | /* |
a394cb8e MH |
2699 | * Direct reclaim and kswapd have to scan all memory |
2700 | * cgroups to fulfill the overall scan target for the | |
a9dd0a83 | 2701 | * node. |
a394cb8e MH |
2702 | * |
2703 | * Limit reclaim, on the other hand, only cares about | |
2704 | * nr_to_reclaim pages to be reclaimed and it will | |
2705 | * retry with decreasing priority if one round over the | |
2706 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2707 | */ |
a394cb8e MH |
2708 | if (!global_reclaim(sc) && |
2709 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2710 | mem_cgroup_iter_break(root, memcg); |
2711 | break; | |
2712 | } | |
241994ed | 2713 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
70ddf637 | 2714 | |
6b4f7799 JW |
2715 | /* |
2716 | * Shrink the slab caches in the same proportion that | |
2717 | * the eligible LRU pages were scanned. | |
2718 | */ | |
b2e18757 | 2719 | if (global_reclaim(sc)) |
a9dd0a83 | 2720 | shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, |
cb731d6c | 2721 | sc->nr_scanned - nr_scanned, |
a9dd0a83 | 2722 | node_lru_pages); |
cb731d6c VD |
2723 | |
2724 | if (reclaim_state) { | |
2725 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2726 | reclaim_state->reclaimed_slab = 0; | |
6b4f7799 JW |
2727 | } |
2728 | ||
8e8ae645 JW |
2729 | /* Record the subtree's reclaim efficiency */ |
2730 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
70ddf637 AV |
2731 | sc->nr_scanned - nr_scanned, |
2732 | sc->nr_reclaimed - nr_reclaimed); | |
2733 | ||
2344d7e4 JW |
2734 | if (sc->nr_reclaimed - nr_reclaimed) |
2735 | reclaimable = true; | |
2736 | ||
a9dd0a83 | 2737 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
9b4f98cd | 2738 | sc->nr_scanned - nr_scanned, sc)); |
2344d7e4 | 2739 | |
c73322d0 JW |
2740 | /* |
2741 | * Kswapd gives up on balancing particular nodes after too | |
2742 | * many failures to reclaim anything from them and goes to | |
2743 | * sleep. On reclaim progress, reset the failure counter. A | |
2744 | * successful direct reclaim run will revive a dormant kswapd. | |
2745 | */ | |
2746 | if (reclaimable) | |
2747 | pgdat->kswapd_failures = 0; | |
2748 | ||
2344d7e4 | 2749 | return reclaimable; |
f16015fb JW |
2750 | } |
2751 | ||
53853e2d | 2752 | /* |
fdd4c614 VB |
2753 | * Returns true if compaction should go ahead for a costly-order request, or |
2754 | * the allocation would already succeed without compaction. Return false if we | |
2755 | * should reclaim first. | |
53853e2d | 2756 | */ |
4f588331 | 2757 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2758 | { |
31483b6a | 2759 | unsigned long watermark; |
fdd4c614 | 2760 | enum compact_result suitable; |
fe4b1b24 | 2761 | |
fdd4c614 VB |
2762 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2763 | if (suitable == COMPACT_SUCCESS) | |
2764 | /* Allocation should succeed already. Don't reclaim. */ | |
2765 | return true; | |
2766 | if (suitable == COMPACT_SKIPPED) | |
2767 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2768 | return false; | |
fe4b1b24 | 2769 | |
53853e2d | 2770 | /* |
fdd4c614 VB |
2771 | * Compaction is already possible, but it takes time to run and there |
2772 | * are potentially other callers using the pages just freed. So proceed | |
2773 | * with reclaim to make a buffer of free pages available to give | |
2774 | * compaction a reasonable chance of completing and allocating the page. | |
2775 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2776 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2777 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 2778 | */ |
fdd4c614 | 2779 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 2780 | |
fdd4c614 | 2781 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
2782 | } |
2783 | ||
1da177e4 LT |
2784 | /* |
2785 | * This is the direct reclaim path, for page-allocating processes. We only | |
2786 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2787 | * request. | |
2788 | * | |
1da177e4 LT |
2789 | * If a zone is deemed to be full of pinned pages then just give it a light |
2790 | * scan then give up on it. | |
2791 | */ | |
0a0337e0 | 2792 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2793 | { |
dd1a239f | 2794 | struct zoneref *z; |
54a6eb5c | 2795 | struct zone *zone; |
0608f43d AM |
2796 | unsigned long nr_soft_reclaimed; |
2797 | unsigned long nr_soft_scanned; | |
619d0d76 | 2798 | gfp_t orig_mask; |
79dafcdc | 2799 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2800 | |
cc715d99 MG |
2801 | /* |
2802 | * If the number of buffer_heads in the machine exceeds the maximum | |
2803 | * allowed level, force direct reclaim to scan the highmem zone as | |
2804 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2805 | */ | |
619d0d76 | 2806 | orig_mask = sc->gfp_mask; |
b2e18757 | 2807 | if (buffer_heads_over_limit) { |
cc715d99 | 2808 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2809 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2810 | } |
cc715d99 | 2811 | |
d4debc66 | 2812 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2813 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
2814 | /* |
2815 | * Take care memory controller reclaiming has small influence | |
2816 | * to global LRU. | |
2817 | */ | |
89b5fae5 | 2818 | if (global_reclaim(sc)) { |
344736f2 VD |
2819 | if (!cpuset_zone_allowed(zone, |
2820 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2821 | continue; |
65ec02cb | 2822 | |
0b06496a JW |
2823 | /* |
2824 | * If we already have plenty of memory free for | |
2825 | * compaction in this zone, don't free any more. | |
2826 | * Even though compaction is invoked for any | |
2827 | * non-zero order, only frequent costly order | |
2828 | * reclamation is disruptive enough to become a | |
2829 | * noticeable problem, like transparent huge | |
2830 | * page allocations. | |
2831 | */ | |
2832 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2833 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2834 | compaction_ready(zone, sc)) { |
0b06496a JW |
2835 | sc->compaction_ready = true; |
2836 | continue; | |
e0887c19 | 2837 | } |
0b06496a | 2838 | |
79dafcdc MG |
2839 | /* |
2840 | * Shrink each node in the zonelist once. If the | |
2841 | * zonelist is ordered by zone (not the default) then a | |
2842 | * node may be shrunk multiple times but in that case | |
2843 | * the user prefers lower zones being preserved. | |
2844 | */ | |
2845 | if (zone->zone_pgdat == last_pgdat) | |
2846 | continue; | |
2847 | ||
0608f43d AM |
2848 | /* |
2849 | * This steals pages from memory cgroups over softlimit | |
2850 | * and returns the number of reclaimed pages and | |
2851 | * scanned pages. This works for global memory pressure | |
2852 | * and balancing, not for a memcg's limit. | |
2853 | */ | |
2854 | nr_soft_scanned = 0; | |
ef8f2327 | 2855 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2856 | sc->order, sc->gfp_mask, |
2857 | &nr_soft_scanned); | |
2858 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2859 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2860 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2861 | } |
408d8544 | 2862 | |
79dafcdc MG |
2863 | /* See comment about same check for global reclaim above */ |
2864 | if (zone->zone_pgdat == last_pgdat) | |
2865 | continue; | |
2866 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2867 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2868 | } |
e0c23279 | 2869 | |
619d0d76 WY |
2870 | /* |
2871 | * Restore to original mask to avoid the impact on the caller if we | |
2872 | * promoted it to __GFP_HIGHMEM. | |
2873 | */ | |
2874 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2875 | } |
4f98a2fe | 2876 | |
2a2e4885 JW |
2877 | static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) |
2878 | { | |
2879 | struct mem_cgroup *memcg; | |
2880 | ||
2881 | memcg = mem_cgroup_iter(root_memcg, NULL, NULL); | |
2882 | do { | |
2883 | unsigned long refaults; | |
2884 | struct lruvec *lruvec; | |
2885 | ||
2886 | if (memcg) | |
ccda7f43 | 2887 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); |
2a2e4885 JW |
2888 | else |
2889 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); | |
2890 | ||
2891 | lruvec = mem_cgroup_lruvec(pgdat, memcg); | |
2892 | lruvec->refaults = refaults; | |
2893 | } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); | |
2894 | } | |
2895 | ||
1da177e4 LT |
2896 | /* |
2897 | * This is the main entry point to direct page reclaim. | |
2898 | * | |
2899 | * If a full scan of the inactive list fails to free enough memory then we | |
2900 | * are "out of memory" and something needs to be killed. | |
2901 | * | |
2902 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2903 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2904 | * caller can't do much about. We kick the writeback threads and take explicit |
2905 | * naps in the hope that some of these pages can be written. But if the | |
2906 | * allocating task holds filesystem locks which prevent writeout this might not | |
2907 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2908 | * |
2909 | * returns: 0, if no pages reclaimed | |
2910 | * else, the number of pages reclaimed | |
1da177e4 | 2911 | */ |
dac1d27b | 2912 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2913 | struct scan_control *sc) |
1da177e4 | 2914 | { |
241994ed | 2915 | int initial_priority = sc->priority; |
2a2e4885 JW |
2916 | pg_data_t *last_pgdat; |
2917 | struct zoneref *z; | |
2918 | struct zone *zone; | |
241994ed | 2919 | retry: |
873b4771 KK |
2920 | delayacct_freepages_start(); |
2921 | ||
89b5fae5 | 2922 | if (global_reclaim(sc)) |
7cc30fcf | 2923 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 2924 | |
9e3b2f8c | 2925 | do { |
70ddf637 AV |
2926 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2927 | sc->priority); | |
66e1707b | 2928 | sc->nr_scanned = 0; |
0a0337e0 | 2929 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 2930 | |
bb21c7ce | 2931 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
2932 | break; |
2933 | ||
2934 | if (sc->compaction_ready) | |
2935 | break; | |
1da177e4 | 2936 | |
0e50ce3b MK |
2937 | /* |
2938 | * If we're getting trouble reclaiming, start doing | |
2939 | * writepage even in laptop mode. | |
2940 | */ | |
2941 | if (sc->priority < DEF_PRIORITY - 2) | |
2942 | sc->may_writepage = 1; | |
0b06496a | 2943 | } while (--sc->priority >= 0); |
bb21c7ce | 2944 | |
2a2e4885 JW |
2945 | last_pgdat = NULL; |
2946 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
2947 | sc->nodemask) { | |
2948 | if (zone->zone_pgdat == last_pgdat) | |
2949 | continue; | |
2950 | last_pgdat = zone->zone_pgdat; | |
2951 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); | |
2952 | } | |
2953 | ||
873b4771 KK |
2954 | delayacct_freepages_end(); |
2955 | ||
bb21c7ce KM |
2956 | if (sc->nr_reclaimed) |
2957 | return sc->nr_reclaimed; | |
2958 | ||
0cee34fd | 2959 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 2960 | if (sc->compaction_ready) |
7335084d MG |
2961 | return 1; |
2962 | ||
241994ed | 2963 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 2964 | if (sc->memcg_low_skipped) { |
241994ed | 2965 | sc->priority = initial_priority; |
d6622f63 YX |
2966 | sc->memcg_low_reclaim = 1; |
2967 | sc->memcg_low_skipped = 0; | |
241994ed JW |
2968 | goto retry; |
2969 | } | |
2970 | ||
bb21c7ce | 2971 | return 0; |
1da177e4 LT |
2972 | } |
2973 | ||
c73322d0 | 2974 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
2975 | { |
2976 | struct zone *zone; | |
2977 | unsigned long pfmemalloc_reserve = 0; | |
2978 | unsigned long free_pages = 0; | |
2979 | int i; | |
2980 | bool wmark_ok; | |
2981 | ||
c73322d0 JW |
2982 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
2983 | return true; | |
2984 | ||
5515061d MG |
2985 | for (i = 0; i <= ZONE_NORMAL; i++) { |
2986 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
2987 | if (!managed_zone(zone)) |
2988 | continue; | |
2989 | ||
2990 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
2991 | continue; |
2992 | ||
5515061d MG |
2993 | pfmemalloc_reserve += min_wmark_pages(zone); |
2994 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2995 | } | |
2996 | ||
675becce MG |
2997 | /* If there are no reserves (unexpected config) then do not throttle */ |
2998 | if (!pfmemalloc_reserve) | |
2999 | return true; | |
3000 | ||
5515061d MG |
3001 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
3002 | ||
3003 | /* kswapd must be awake if processes are being throttled */ | |
3004 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
38087d9b | 3005 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
5515061d MG |
3006 | (enum zone_type)ZONE_NORMAL); |
3007 | wake_up_interruptible(&pgdat->kswapd_wait); | |
3008 | } | |
3009 | ||
3010 | return wmark_ok; | |
3011 | } | |
3012 | ||
3013 | /* | |
3014 | * Throttle direct reclaimers if backing storage is backed by the network | |
3015 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
3016 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
3017 | * when the low watermark is reached. |
3018 | * | |
3019 | * Returns true if a fatal signal was delivered during throttling. If this | |
3020 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 3021 | */ |
50694c28 | 3022 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
3023 | nodemask_t *nodemask) |
3024 | { | |
675becce | 3025 | struct zoneref *z; |
5515061d | 3026 | struct zone *zone; |
675becce | 3027 | pg_data_t *pgdat = NULL; |
5515061d MG |
3028 | |
3029 | /* | |
3030 | * Kernel threads should not be throttled as they may be indirectly | |
3031 | * responsible for cleaning pages necessary for reclaim to make forward | |
3032 | * progress. kjournald for example may enter direct reclaim while | |
3033 | * committing a transaction where throttling it could forcing other | |
3034 | * processes to block on log_wait_commit(). | |
3035 | */ | |
3036 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
3037 | goto out; |
3038 | ||
3039 | /* | |
3040 | * If a fatal signal is pending, this process should not throttle. | |
3041 | * It should return quickly so it can exit and free its memory | |
3042 | */ | |
3043 | if (fatal_signal_pending(current)) | |
3044 | goto out; | |
5515061d | 3045 | |
675becce MG |
3046 | /* |
3047 | * Check if the pfmemalloc reserves are ok by finding the first node | |
3048 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
3049 | * GFP_KERNEL will be required for allocating network buffers when | |
3050 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
3051 | * | |
3052 | * Throttling is based on the first usable node and throttled processes | |
3053 | * wait on a queue until kswapd makes progress and wakes them. There | |
3054 | * is an affinity then between processes waking up and where reclaim | |
3055 | * progress has been made assuming the process wakes on the same node. | |
3056 | * More importantly, processes running on remote nodes will not compete | |
3057 | * for remote pfmemalloc reserves and processes on different nodes | |
3058 | * should make reasonable progress. | |
3059 | */ | |
3060 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 3061 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
3062 | if (zone_idx(zone) > ZONE_NORMAL) |
3063 | continue; | |
3064 | ||
3065 | /* Throttle based on the first usable node */ | |
3066 | pgdat = zone->zone_pgdat; | |
c73322d0 | 3067 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
3068 | goto out; |
3069 | break; | |
3070 | } | |
3071 | ||
3072 | /* If no zone was usable by the allocation flags then do not throttle */ | |
3073 | if (!pgdat) | |
50694c28 | 3074 | goto out; |
5515061d | 3075 | |
68243e76 MG |
3076 | /* Account for the throttling */ |
3077 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
3078 | ||
5515061d MG |
3079 | /* |
3080 | * If the caller cannot enter the filesystem, it's possible that it | |
3081 | * is due to the caller holding an FS lock or performing a journal | |
3082 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
3083 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
3084 | * blocked waiting on the same lock. Instead, throttle for up to a | |
3085 | * second before continuing. | |
3086 | */ | |
3087 | if (!(gfp_mask & __GFP_FS)) { | |
3088 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
c73322d0 | 3089 | allow_direct_reclaim(pgdat), HZ); |
50694c28 MG |
3090 | |
3091 | goto check_pending; | |
5515061d MG |
3092 | } |
3093 | ||
3094 | /* Throttle until kswapd wakes the process */ | |
3095 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
c73322d0 | 3096 | allow_direct_reclaim(pgdat)); |
50694c28 MG |
3097 | |
3098 | check_pending: | |
3099 | if (fatal_signal_pending(current)) | |
3100 | return true; | |
3101 | ||
3102 | out: | |
3103 | return false; | |
5515061d MG |
3104 | } |
3105 | ||
dac1d27b | 3106 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 3107 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 3108 | { |
abd25eef | 3109 | ktime_t event_ts; |
33906bc5 | 3110 | unsigned long nr_reclaimed; |
66e1707b | 3111 | struct scan_control sc = { |
ee814fe2 | 3112 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 3113 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 3114 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
3115 | .order = order, |
3116 | .nodemask = nodemask, | |
3117 | .priority = DEF_PRIORITY, | |
66e1707b | 3118 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3119 | .may_unmap = 1, |
2e2e4259 | 3120 | .may_swap = 1, |
66e1707b BS |
3121 | }; |
3122 | ||
5515061d | 3123 | /* |
50694c28 MG |
3124 | * Do not enter reclaim if fatal signal was delivered while throttled. |
3125 | * 1 is returned so that the page allocator does not OOM kill at this | |
3126 | * point. | |
5515061d | 3127 | */ |
f2f43e56 | 3128 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
3129 | return 1; |
3130 | ||
abd25eef | 3131 | mm_event_start(&event_ts); |
33906bc5 MG |
3132 | trace_mm_vmscan_direct_reclaim_begin(order, |
3133 | sc.may_writepage, | |
f2f43e56 | 3134 | sc.gfp_mask, |
e5146b12 | 3135 | sc.reclaim_idx); |
33906bc5 | 3136 | |
3115cd91 | 3137 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
3138 | |
3139 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
abd25eef | 3140 | mm_event_end(MM_RECLAIM, event_ts); |
33906bc5 MG |
3141 | |
3142 | return nr_reclaimed; | |
66e1707b BS |
3143 | } |
3144 | ||
c255a458 | 3145 | #ifdef CONFIG_MEMCG |
66e1707b | 3146 | |
a9dd0a83 | 3147 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 3148 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 3149 | pg_data_t *pgdat, |
0ae5e89c | 3150 | unsigned long *nr_scanned) |
4e416953 BS |
3151 | { |
3152 | struct scan_control sc = { | |
b8f5c566 | 3153 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 3154 | .target_mem_cgroup = memcg, |
4e416953 BS |
3155 | .may_writepage = !laptop_mode, |
3156 | .may_unmap = 1, | |
b2e18757 | 3157 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 3158 | .may_swap = !noswap, |
4e416953 | 3159 | }; |
6b4f7799 | 3160 | unsigned long lru_pages; |
0ae5e89c | 3161 | |
4e416953 BS |
3162 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3163 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3164 | |
9e3b2f8c | 3165 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e | 3166 | sc.may_writepage, |
e5146b12 MG |
3167 | sc.gfp_mask, |
3168 | sc.reclaim_idx); | |
bdce6d9e | 3169 | |
4e416953 BS |
3170 | /* |
3171 | * NOTE: Although we can get the priority field, using it | |
3172 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3173 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3174 | * will pick up pages from other mem cgroup's as well. We hack |
3175 | * the priority and make it zero. | |
3176 | */ | |
ef8f2327 | 3177 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
bdce6d9e KM |
3178 | |
3179 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3180 | ||
0ae5e89c | 3181 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
3182 | return sc.nr_reclaimed; |
3183 | } | |
3184 | ||
72835c86 | 3185 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3186 | unsigned long nr_pages, |
a7885eb8 | 3187 | gfp_t gfp_mask, |
b70a2a21 | 3188 | bool may_swap) |
66e1707b | 3189 | { |
4e416953 | 3190 | struct zonelist *zonelist; |
bdce6d9e | 3191 | unsigned long nr_reclaimed; |
6dbf8323 | 3192 | unsigned long pflags; |
889976db | 3193 | int nid; |
499118e9 | 3194 | unsigned int noreclaim_flag; |
66e1707b | 3195 | struct scan_control sc = { |
b70a2a21 | 3196 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3197 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 3198 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 3199 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3200 | .target_mem_cgroup = memcg, |
3201 | .priority = DEF_PRIORITY, | |
3202 | .may_writepage = !laptop_mode, | |
3203 | .may_unmap = 1, | |
b70a2a21 | 3204 | .may_swap = may_swap, |
a09ed5e0 | 3205 | }; |
66e1707b | 3206 | |
889976db YH |
3207 | /* |
3208 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
3209 | * take care of from where we get pages. So the node where we start the | |
3210 | * scan does not need to be the current node. | |
3211 | */ | |
72835c86 | 3212 | nid = mem_cgroup_select_victim_node(memcg); |
889976db | 3213 | |
c9634cf0 | 3214 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; |
bdce6d9e KM |
3215 | |
3216 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
3217 | sc.may_writepage, | |
e5146b12 MG |
3218 | sc.gfp_mask, |
3219 | sc.reclaim_idx); | |
bdce6d9e | 3220 | |
6dbf8323 | 3221 | psi_memstall_enter(&pflags); |
499118e9 | 3222 | noreclaim_flag = memalloc_noreclaim_save(); |
6dbf8323 | 3223 | |
3115cd91 | 3224 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
6dbf8323 | 3225 | |
499118e9 | 3226 | memalloc_noreclaim_restore(noreclaim_flag); |
6dbf8323 | 3227 | psi_memstall_leave(&pflags); |
bdce6d9e KM |
3228 | |
3229 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
3230 | ||
3231 | return nr_reclaimed; | |
66e1707b BS |
3232 | } |
3233 | #endif | |
3234 | ||
1d82de61 | 3235 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3236 | struct scan_control *sc) |
f16015fb | 3237 | { |
b95a2f2d | 3238 | struct mem_cgroup *memcg; |
f16015fb | 3239 | |
b95a2f2d JW |
3240 | if (!total_swap_pages) |
3241 | return; | |
3242 | ||
3243 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
3244 | do { | |
ef8f2327 | 3245 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
b95a2f2d | 3246 | |
2a2e4885 | 3247 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) |
1a93be0e | 3248 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 3249 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
3250 | |
3251 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
3252 | } while (memcg); | |
f16015fb JW |
3253 | } |
3254 | ||
e716f2eb MG |
3255 | /* |
3256 | * Returns true if there is an eligible zone balanced for the request order | |
3257 | * and classzone_idx | |
3258 | */ | |
3259 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
60cefed4 | 3260 | { |
e716f2eb MG |
3261 | int i; |
3262 | unsigned long mark = -1; | |
3263 | struct zone *zone; | |
60cefed4 | 3264 | |
e716f2eb MG |
3265 | for (i = 0; i <= classzone_idx; i++) { |
3266 | zone = pgdat->node_zones + i; | |
6256c6b4 | 3267 | |
e716f2eb MG |
3268 | if (!managed_zone(zone)) |
3269 | continue; | |
3270 | ||
3271 | mark = high_wmark_pages(zone); | |
3272 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) | |
3273 | return true; | |
3274 | } | |
3275 | ||
3276 | /* | |
3277 | * If a node has no populated zone within classzone_idx, it does not | |
3278 | * need balancing by definition. This can happen if a zone-restricted | |
3279 | * allocation tries to wake a remote kswapd. | |
3280 | */ | |
3281 | if (mark == -1) | |
3282 | return true; | |
3283 | ||
3284 | return false; | |
60cefed4 JW |
3285 | } |
3286 | ||
631b6e08 MG |
3287 | /* Clear pgdat state for congested, dirty or under writeback. */ |
3288 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
3289 | { | |
3290 | clear_bit(PGDAT_CONGESTED, &pgdat->flags); | |
3291 | clear_bit(PGDAT_DIRTY, &pgdat->flags); | |
3292 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
3293 | } | |
3294 | ||
5515061d MG |
3295 | /* |
3296 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3297 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3298 | * | |
3299 | * Returns true if kswapd is ready to sleep | |
3300 | */ | |
d9f21d42 | 3301 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f50de2d3 | 3302 | { |
5515061d | 3303 | /* |
9e5e3661 | 3304 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 3305 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
3306 | * race between when kswapd checks the watermarks and a process gets |
3307 | * throttled. There is also a potential race if processes get | |
3308 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3309 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3310 | * the wake up checks. If kswapd is going to sleep, no process should | |
3311 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3312 | * the wake up is premature, processes will wake kswapd and get | |
3313 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3314 | * that here we are under prepare_to_wait(). | |
5515061d | 3315 | */ |
9e5e3661 VB |
3316 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3317 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3318 | |
c73322d0 JW |
3319 | /* Hopeless node, leave it to direct reclaim */ |
3320 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3321 | return true; | |
3322 | ||
e716f2eb MG |
3323 | if (pgdat_balanced(pgdat, order, classzone_idx)) { |
3324 | clear_pgdat_congested(pgdat); | |
3325 | return true; | |
1d82de61 MG |
3326 | } |
3327 | ||
333b0a45 | 3328 | return false; |
f50de2d3 MG |
3329 | } |
3330 | ||
75485363 | 3331 | /* |
1d82de61 MG |
3332 | * kswapd shrinks a node of pages that are at or below the highest usable |
3333 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3334 | * |
3335 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3336 | * reclaim or if the lack of progress was due to pages under writeback. |
3337 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3338 | */ |
1d82de61 | 3339 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3340 | struct scan_control *sc) |
75485363 | 3341 | { |
1d82de61 MG |
3342 | struct zone *zone; |
3343 | int z; | |
75485363 | 3344 | |
1d82de61 MG |
3345 | /* Reclaim a number of pages proportional to the number of zones */ |
3346 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3347 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 3348 | zone = pgdat->node_zones + z; |
6aa303de | 3349 | if (!managed_zone(zone)) |
1d82de61 | 3350 | continue; |
7c954f6d | 3351 | |
1d82de61 MG |
3352 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3353 | } | |
7c954f6d MG |
3354 | |
3355 | /* | |
1d82de61 MG |
3356 | * Historically care was taken to put equal pressure on all zones but |
3357 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3358 | */ |
970a39a3 | 3359 | shrink_node(pgdat, sc); |
283aba9f | 3360 | |
7c954f6d | 3361 | /* |
1d82de61 MG |
3362 | * Fragmentation may mean that the system cannot be rebalanced for |
3363 | * high-order allocations. If twice the allocation size has been | |
3364 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3365 | * excessive reclaim. Assume that a process requested a high-order | |
3366 | * can direct reclaim/compact. | |
7c954f6d | 3367 | */ |
9861a62c | 3368 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 3369 | sc->order = 0; |
7c954f6d | 3370 | |
b8e83b94 | 3371 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3372 | } |
3373 | ||
1da177e4 | 3374 | /* |
1d82de61 MG |
3375 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3376 | * that are eligible for use by the caller until at least one zone is | |
3377 | * balanced. | |
1da177e4 | 3378 | * |
1d82de61 | 3379 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3380 | * |
3381 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3382 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
1d82de61 MG |
3383 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone |
3384 | * or lower is eligible for reclaim until at least one usable zone is | |
3385 | * balanced. | |
1da177e4 | 3386 | */ |
accf6242 | 3387 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3388 | { |
1da177e4 | 3389 | int i; |
0608f43d AM |
3390 | unsigned long nr_soft_reclaimed; |
3391 | unsigned long nr_soft_scanned; | |
6dbf8323 | 3392 | unsigned long pflags; |
1d82de61 | 3393 | struct zone *zone; |
179e9639 AM |
3394 | struct scan_control sc = { |
3395 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3396 | .order = order, |
b8e83b94 | 3397 | .priority = DEF_PRIORITY, |
ee814fe2 | 3398 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3399 | .may_unmap = 1, |
2e2e4259 | 3400 | .may_swap = 1, |
179e9639 | 3401 | }; |
6dbf8323 | 3402 | psi_memstall_enter(&pflags); |
f8891e5e | 3403 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3404 | |
9e3b2f8c | 3405 | do { |
c73322d0 | 3406 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 MG |
3407 | bool raise_priority = true; |
3408 | ||
84c7a777 | 3409 | sc.reclaim_idx = classzone_idx; |
1da177e4 | 3410 | |
86c79f6b | 3411 | /* |
84c7a777 MG |
3412 | * If the number of buffer_heads exceeds the maximum allowed |
3413 | * then consider reclaiming from all zones. This has a dual | |
3414 | * purpose -- on 64-bit systems it is expected that | |
3415 | * buffer_heads are stripped during active rotation. On 32-bit | |
3416 | * systems, highmem pages can pin lowmem memory and shrinking | |
3417 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3418 | * go ahead if all eligible zones for the original allocation | |
3419 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3420 | */ |
3421 | if (buffer_heads_over_limit) { | |
3422 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3423 | zone = pgdat->node_zones + i; | |
6aa303de | 3424 | if (!managed_zone(zone)) |
86c79f6b | 3425 | continue; |
cc715d99 | 3426 | |
970a39a3 | 3427 | sc.reclaim_idx = i; |
e1dbeda6 | 3428 | break; |
1da177e4 | 3429 | } |
1da177e4 | 3430 | } |
dafcb73e | 3431 | |
86c79f6b | 3432 | /* |
e716f2eb MG |
3433 | * Only reclaim if there are no eligible zones. Note that |
3434 | * sc.reclaim_idx is not used as buffer_heads_over_limit may | |
3435 | * have adjusted it. | |
86c79f6b | 3436 | */ |
e716f2eb MG |
3437 | if (pgdat_balanced(pgdat, sc.order, classzone_idx)) |
3438 | goto out; | |
e1dbeda6 | 3439 | |
1d82de61 MG |
3440 | /* |
3441 | * Do some background aging of the anon list, to give | |
3442 | * pages a chance to be referenced before reclaiming. All | |
3443 | * pages are rotated regardless of classzone as this is | |
3444 | * about consistent aging. | |
3445 | */ | |
ef8f2327 | 3446 | age_active_anon(pgdat, &sc); |
1d82de61 | 3447 | |
b7ea3c41 MG |
3448 | /* |
3449 | * If we're getting trouble reclaiming, start doing writepage | |
3450 | * even in laptop mode. | |
3451 | */ | |
047d72c3 | 3452 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
3453 | sc.may_writepage = 1; |
3454 | ||
1d82de61 MG |
3455 | /* Call soft limit reclaim before calling shrink_node. */ |
3456 | sc.nr_scanned = 0; | |
3457 | nr_soft_scanned = 0; | |
ef8f2327 | 3458 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3459 | sc.gfp_mask, &nr_soft_scanned); |
3460 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3461 | ||
1da177e4 | 3462 | /* |
1d82de61 MG |
3463 | * There should be no need to raise the scanning priority if |
3464 | * enough pages are already being scanned that that high | |
3465 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3466 | */ |
970a39a3 | 3467 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3468 | raise_priority = false; |
5515061d MG |
3469 | |
3470 | /* | |
3471 | * If the low watermark is met there is no need for processes | |
3472 | * to be throttled on pfmemalloc_wait as they should not be | |
3473 | * able to safely make forward progress. Wake them | |
3474 | */ | |
3475 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 3476 | allow_direct_reclaim(pgdat)) |
cfc51155 | 3477 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3478 | |
b8e83b94 MG |
3479 | /* Check if kswapd should be suspending */ |
3480 | if (try_to_freeze() || kthread_should_stop()) | |
3481 | break; | |
8357376d | 3482 | |
73ce02e9 | 3483 | /* |
b8e83b94 MG |
3484 | * Raise priority if scanning rate is too low or there was no |
3485 | * progress in reclaiming pages | |
73ce02e9 | 3486 | */ |
c73322d0 JW |
3487 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
3488 | if (raise_priority || !nr_reclaimed) | |
b8e83b94 | 3489 | sc.priority--; |
1d82de61 | 3490 | } while (sc.priority >= 1); |
1da177e4 | 3491 | |
c73322d0 JW |
3492 | if (!sc.nr_reclaimed) |
3493 | pgdat->kswapd_failures++; | |
3494 | ||
b8e83b94 | 3495 | out: |
2a2e4885 | 3496 | snapshot_refaults(NULL, pgdat); |
6dbf8323 | 3497 | psi_memstall_leave(&pflags); |
0abdee2b | 3498 | /* |
1d82de61 MG |
3499 | * Return the order kswapd stopped reclaiming at as |
3500 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3501 | * entered the allocator slow path while kswapd was awake, order will | |
3502 | * remain at the higher level. | |
0abdee2b | 3503 | */ |
1d82de61 | 3504 | return sc.order; |
1da177e4 LT |
3505 | } |
3506 | ||
e716f2eb MG |
3507 | /* |
3508 | * pgdat->kswapd_classzone_idx is the highest zone index that a recent | |
3509 | * allocation request woke kswapd for. When kswapd has not woken recently, | |
3510 | * the value is MAX_NR_ZONES which is not a valid index. This compares a | |
3511 | * given classzone and returns it or the highest classzone index kswapd | |
3512 | * was recently woke for. | |
3513 | */ | |
3514 | static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, | |
3515 | enum zone_type classzone_idx) | |
3516 | { | |
3517 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
3518 | return classzone_idx; | |
3519 | ||
3520 | return max(pgdat->kswapd_classzone_idx, classzone_idx); | |
3521 | } | |
3522 | ||
38087d9b MG |
3523 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3524 | unsigned int classzone_idx) | |
f0bc0a60 KM |
3525 | { |
3526 | long remaining = 0; | |
3527 | DEFINE_WAIT(wait); | |
3528 | ||
3529 | if (freezing(current) || kthread_should_stop()) | |
3530 | return; | |
3531 | ||
3532 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3533 | ||
333b0a45 SG |
3534 | /* |
3535 | * Try to sleep for a short interval. Note that kcompactd will only be | |
3536 | * woken if it is possible to sleep for a short interval. This is | |
3537 | * deliberate on the assumption that if reclaim cannot keep an | |
3538 | * eligible zone balanced that it's also unlikely that compaction will | |
3539 | * succeed. | |
3540 | */ | |
d9f21d42 | 3541 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
fd901c95 VB |
3542 | /* |
3543 | * Compaction records what page blocks it recently failed to | |
3544 | * isolate pages from and skips them in the future scanning. | |
3545 | * When kswapd is going to sleep, it is reasonable to assume | |
3546 | * that pages and compaction may succeed so reset the cache. | |
3547 | */ | |
3548 | reset_isolation_suitable(pgdat); | |
3549 | ||
3550 | /* | |
3551 | * We have freed the memory, now we should compact it to make | |
3552 | * allocation of the requested order possible. | |
3553 | */ | |
38087d9b | 3554 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
fd901c95 | 3555 | |
f0bc0a60 | 3556 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3557 | |
3558 | /* | |
3559 | * If woken prematurely then reset kswapd_classzone_idx and | |
3560 | * order. The values will either be from a wakeup request or | |
3561 | * the previous request that slept prematurely. | |
3562 | */ | |
3563 | if (remaining) { | |
e716f2eb | 3564 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b MG |
3565 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); |
3566 | } | |
3567 | ||
f0bc0a60 KM |
3568 | finish_wait(&pgdat->kswapd_wait, &wait); |
3569 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3570 | } | |
3571 | ||
3572 | /* | |
3573 | * After a short sleep, check if it was a premature sleep. If not, then | |
3574 | * go fully to sleep until explicitly woken up. | |
3575 | */ | |
d9f21d42 MG |
3576 | if (!remaining && |
3577 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
f0bc0a60 KM |
3578 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3579 | ||
3580 | /* | |
3581 | * vmstat counters are not perfectly accurate and the estimated | |
3582 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3583 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3584 | * watermarks being breached while under pressure, we reduce the | |
3585 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3586 | * them before going back to sleep. | |
3587 | */ | |
3588 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3589 | |
3590 | if (!kthread_should_stop()) | |
3591 | schedule(); | |
3592 | ||
f0bc0a60 KM |
3593 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3594 | } else { | |
3595 | if (remaining) | |
3596 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3597 | else | |
3598 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3599 | } | |
3600 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3601 | } | |
3602 | ||
1da177e4 LT |
3603 | /* |
3604 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3605 | * from the init process. |
1da177e4 LT |
3606 | * |
3607 | * This basically trickles out pages so that we have _some_ | |
3608 | * free memory available even if there is no other activity | |
3609 | * that frees anything up. This is needed for things like routing | |
3610 | * etc, where we otherwise might have all activity going on in | |
3611 | * asynchronous contexts that cannot page things out. | |
3612 | * | |
3613 | * If there are applications that are active memory-allocators | |
3614 | * (most normal use), this basically shouldn't matter. | |
3615 | */ | |
3616 | static int kswapd(void *p) | |
3617 | { | |
e716f2eb MG |
3618 | unsigned int alloc_order, reclaim_order; |
3619 | unsigned int classzone_idx = MAX_NR_ZONES - 1; | |
1da177e4 LT |
3620 | pg_data_t *pgdat = (pg_data_t*)p; |
3621 | struct task_struct *tsk = current; | |
f0bc0a60 | 3622 | |
1da177e4 LT |
3623 | struct reclaim_state reclaim_state = { |
3624 | .reclaimed_slab = 0, | |
3625 | }; | |
a70f7302 | 3626 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3627 | |
174596a0 | 3628 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3629 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3630 | current->reclaim_state = &reclaim_state; |
3631 | ||
3632 | /* | |
3633 | * Tell the memory management that we're a "memory allocator", | |
3634 | * and that if we need more memory we should get access to it | |
3635 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3636 | * never get caught in the normal page freeing logic. | |
3637 | * | |
3638 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3639 | * you need a small amount of memory in order to be able to | |
3640 | * page out something else, and this flag essentially protects | |
3641 | * us from recursively trying to free more memory as we're | |
3642 | * trying to free the first piece of memory in the first place). | |
3643 | */ | |
930d9152 | 3644 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3645 | set_freezable(); |
1da177e4 | 3646 | |
e716f2eb MG |
3647 | pgdat->kswapd_order = 0; |
3648 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; | |
1da177e4 | 3649 | for ( ; ; ) { |
6f6313d4 | 3650 | bool ret; |
3e1d1d28 | 3651 | |
abd25eef | 3652 | ktime_t event_ts; |
e716f2eb MG |
3653 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3654 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); | |
3655 | ||
38087d9b MG |
3656 | kswapd_try_sleep: |
3657 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3658 | classzone_idx); | |
215ddd66 | 3659 | |
38087d9b MG |
3660 | /* Read the new order and classzone_idx */ |
3661 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
e716f2eb | 3662 | classzone_idx = kswapd_classzone_idx(pgdat, 0); |
38087d9b | 3663 | pgdat->kswapd_order = 0; |
e716f2eb | 3664 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
1da177e4 | 3665 | |
8fe23e05 DR |
3666 | ret = try_to_freeze(); |
3667 | if (kthread_should_stop()) | |
3668 | break; | |
3669 | ||
3670 | /* | |
3671 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3672 | * after returning from the refrigerator | |
3673 | */ | |
38087d9b MG |
3674 | if (ret) |
3675 | continue; | |
3676 | ||
3677 | /* | |
3678 | * Reclaim begins at the requested order but if a high-order | |
3679 | * reclaim fails then kswapd falls back to reclaiming for | |
3680 | * order-0. If that happens, kswapd will consider sleeping | |
3681 | * for the order it finished reclaiming at (reclaim_order) | |
3682 | * but kcompactd is woken to compact for the original | |
3683 | * request (alloc_order). | |
3684 | */ | |
e5146b12 MG |
3685 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
3686 | alloc_order); | |
d92a8cfc | 3687 | fs_reclaim_acquire(GFP_KERNEL); |
abd25eef | 3688 | mm_event_start(&event_ts); |
38087d9b | 3689 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
abd25eef | 3690 | mm_event_end(MM_RECLAIM, event_ts); |
d92a8cfc | 3691 | fs_reclaim_release(GFP_KERNEL); |
38087d9b MG |
3692 | if (reclaim_order < alloc_order) |
3693 | goto kswapd_try_sleep; | |
1da177e4 | 3694 | } |
b0a8cc58 | 3695 | |
71abdc15 | 3696 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
b0a8cc58 | 3697 | current->reclaim_state = NULL; |
71abdc15 | 3698 | |
1da177e4 LT |
3699 | return 0; |
3700 | } | |
3701 | ||
3702 | /* | |
3703 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3704 | */ | |
99504748 | 3705 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3706 | { |
3707 | pg_data_t *pgdat; | |
3708 | ||
6aa303de | 3709 | if (!managed_zone(zone)) |
1da177e4 LT |
3710 | return; |
3711 | ||
344736f2 | 3712 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) |
1da177e4 | 3713 | return; |
88f5acf8 | 3714 | pgdat = zone->zone_pgdat; |
e716f2eb MG |
3715 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, |
3716 | classzone_idx); | |
38087d9b | 3717 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
8d0986e2 | 3718 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3719 | return; |
e1a55637 | 3720 | |
c73322d0 JW |
3721 | /* Hopeless node, leave it to direct reclaim */ |
3722 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3723 | return; | |
3724 | ||
e716f2eb MG |
3725 | if (pgdat_balanced(pgdat, order, classzone_idx)) |
3726 | return; | |
88f5acf8 | 3727 | |
e716f2eb | 3728 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order); |
8d0986e2 | 3729 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3730 | } |
3731 | ||
c6f37f12 | 3732 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3733 | /* |
7b51755c | 3734 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3735 | * freed pages. |
3736 | * | |
3737 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3738 | * LRU order by reclaiming preferentially | |
3739 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3740 | */ |
7b51755c | 3741 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3742 | { |
d6277db4 | 3743 | struct reclaim_state reclaim_state; |
d6277db4 | 3744 | struct scan_control sc = { |
ee814fe2 | 3745 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3746 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 3747 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 3748 | .priority = DEF_PRIORITY, |
d6277db4 | 3749 | .may_writepage = 1, |
ee814fe2 JW |
3750 | .may_unmap = 1, |
3751 | .may_swap = 1, | |
7b51755c | 3752 | .hibernation_mode = 1, |
1da177e4 | 3753 | }; |
a09ed5e0 | 3754 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c KM |
3755 | struct task_struct *p = current; |
3756 | unsigned long nr_reclaimed; | |
499118e9 | 3757 | unsigned int noreclaim_flag; |
1da177e4 | 3758 | |
499118e9 | 3759 | noreclaim_flag = memalloc_noreclaim_save(); |
d92a8cfc | 3760 | fs_reclaim_acquire(sc.gfp_mask); |
7b51755c KM |
3761 | reclaim_state.reclaimed_slab = 0; |
3762 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3763 | |
3115cd91 | 3764 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3765 | |
7b51755c | 3766 | p->reclaim_state = NULL; |
d92a8cfc | 3767 | fs_reclaim_release(sc.gfp_mask); |
499118e9 | 3768 | memalloc_noreclaim_restore(noreclaim_flag); |
d6277db4 | 3769 | |
7b51755c | 3770 | return nr_reclaimed; |
1da177e4 | 3771 | } |
c6f37f12 | 3772 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3773 | |
1da177e4 LT |
3774 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3775 | not required for correctness. So if the last cpu in a node goes | |
3776 | away, we get changed to run anywhere: as the first one comes back, | |
3777 | restore their cpu bindings. */ | |
517bbed9 | 3778 | static int kswapd_cpu_online(unsigned int cpu) |
1da177e4 | 3779 | { |
58c0a4a7 | 3780 | int nid; |
1da177e4 | 3781 | |
517bbed9 SAS |
3782 | for_each_node_state(nid, N_MEMORY) { |
3783 | pg_data_t *pgdat = NODE_DATA(nid); | |
3784 | const struct cpumask *mask; | |
a70f7302 | 3785 | |
517bbed9 | 3786 | mask = cpumask_of_node(pgdat->node_id); |
c5f59f08 | 3787 | |
517bbed9 SAS |
3788 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3789 | /* One of our CPUs online: restore mask */ | |
3790 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
1da177e4 | 3791 | } |
517bbed9 | 3792 | return 0; |
1da177e4 | 3793 | } |
1da177e4 | 3794 | |
3218ae14 YG |
3795 | /* |
3796 | * This kswapd start function will be called by init and node-hot-add. | |
3797 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3798 | */ | |
3799 | int kswapd_run(int nid) | |
3800 | { | |
3801 | pg_data_t *pgdat = NODE_DATA(nid); | |
3802 | int ret = 0; | |
3803 | ||
3804 | if (pgdat->kswapd) | |
3805 | return 0; | |
3806 | ||
3807 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3808 | if (IS_ERR(pgdat->kswapd)) { | |
3809 | /* failure at boot is fatal */ | |
c6202adf | 3810 | BUG_ON(system_state < SYSTEM_RUNNING); |
d5dc0ad9 GS |
3811 | pr_err("Failed to start kswapd on node %d\n", nid); |
3812 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3813 | pgdat->kswapd = NULL; |
3218ae14 YG |
3814 | } |
3815 | return ret; | |
3816 | } | |
3817 | ||
8fe23e05 | 3818 | /* |
d8adde17 | 3819 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 3820 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
3821 | */ |
3822 | void kswapd_stop(int nid) | |
3823 | { | |
3824 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3825 | ||
d8adde17 | 3826 | if (kswapd) { |
8fe23e05 | 3827 | kthread_stop(kswapd); |
d8adde17 JL |
3828 | NODE_DATA(nid)->kswapd = NULL; |
3829 | } | |
8fe23e05 DR |
3830 | } |
3831 | ||
1da177e4 LT |
3832 | static int __init kswapd_init(void) |
3833 | { | |
517bbed9 | 3834 | int nid, ret; |
69e05944 | 3835 | |
1da177e4 | 3836 | swap_setup(); |
48fb2e24 | 3837 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3838 | kswapd_run(nid); |
517bbed9 SAS |
3839 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
3840 | "mm/vmscan:online", kswapd_cpu_online, | |
3841 | NULL); | |
3842 | WARN_ON(ret < 0); | |
1da177e4 LT |
3843 | return 0; |
3844 | } | |
3845 | ||
3846 | module_init(kswapd_init) | |
9eeff239 CL |
3847 | |
3848 | #ifdef CONFIG_NUMA | |
3849 | /* | |
a5f5f91d | 3850 | * Node reclaim mode |
9eeff239 | 3851 | * |
a5f5f91d | 3852 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 3853 | * the watermarks. |
9eeff239 | 3854 | */ |
a5f5f91d | 3855 | int node_reclaim_mode __read_mostly; |
9eeff239 | 3856 | |
1b2ffb78 | 3857 | #define RECLAIM_OFF 0 |
7d03431c | 3858 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 3859 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 3860 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 3861 | |
a92f7126 | 3862 | /* |
a5f5f91d | 3863 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
3864 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
3865 | * a zone. | |
3866 | */ | |
a5f5f91d | 3867 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 3868 | |
9614634f | 3869 | /* |
a5f5f91d | 3870 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
3871 | * occur. |
3872 | */ | |
3873 | int sysctl_min_unmapped_ratio = 1; | |
3874 | ||
0ff38490 CL |
3875 | /* |
3876 | * If the number of slab pages in a zone grows beyond this percentage then | |
3877 | * slab reclaim needs to occur. | |
3878 | */ | |
3879 | int sysctl_min_slab_ratio = 5; | |
3880 | ||
11fb9989 | 3881 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 3882 | { |
11fb9989 MG |
3883 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
3884 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
3885 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
3886 | |
3887 | /* | |
3888 | * It's possible for there to be more file mapped pages than | |
3889 | * accounted for by the pages on the file LRU lists because | |
3890 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3891 | */ | |
3892 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3893 | } | |
3894 | ||
3895 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 3896 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 3897 | { |
d031a157 AM |
3898 | unsigned long nr_pagecache_reclaimable; |
3899 | unsigned long delta = 0; | |
90afa5de MG |
3900 | |
3901 | /* | |
95bbc0c7 | 3902 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 3903 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 3904 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
3905 | * a better estimate |
3906 | */ | |
a5f5f91d MG |
3907 | if (node_reclaim_mode & RECLAIM_UNMAP) |
3908 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 3909 | else |
a5f5f91d | 3910 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
3911 | |
3912 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
3913 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
3914 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
3915 | |
3916 | /* Watch for any possible underflows due to delta */ | |
3917 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3918 | delta = nr_pagecache_reclaimable; | |
3919 | ||
3920 | return nr_pagecache_reclaimable - delta; | |
3921 | } | |
3922 | ||
9eeff239 | 3923 | /* |
a5f5f91d | 3924 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 3925 | */ |
a5f5f91d | 3926 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3927 | { |
7fb2d46d | 3928 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3929 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3930 | struct task_struct *p = current; |
3931 | struct reclaim_state reclaim_state; | |
499118e9 | 3932 | unsigned int noreclaim_flag; |
179e9639 | 3933 | struct scan_control sc = { |
62b726c1 | 3934 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 3935 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 3936 | .order = order, |
a5f5f91d MG |
3937 | .priority = NODE_RECLAIM_PRIORITY, |
3938 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
3939 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 3940 | .may_swap = 1, |
f2f43e56 | 3941 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 3942 | }; |
9eeff239 | 3943 | |
9eeff239 | 3944 | cond_resched(); |
d4f7796e | 3945 | /* |
95bbc0c7 | 3946 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 3947 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 3948 | * and RECLAIM_UNMAP. |
d4f7796e | 3949 | */ |
499118e9 VB |
3950 | noreclaim_flag = memalloc_noreclaim_save(); |
3951 | p->flags |= PF_SWAPWRITE; | |
d92a8cfc | 3952 | fs_reclaim_acquire(sc.gfp_mask); |
9eeff239 CL |
3953 | reclaim_state.reclaimed_slab = 0; |
3954 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3955 | |
a5f5f91d | 3956 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 CL |
3957 | /* |
3958 | * Free memory by calling shrink zone with increasing | |
3959 | * priorities until we have enough memory freed. | |
3960 | */ | |
0ff38490 | 3961 | do { |
970a39a3 | 3962 | shrink_node(pgdat, &sc); |
9e3b2f8c | 3963 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 3964 | } |
c84db23c | 3965 | |
9eeff239 | 3966 | p->reclaim_state = NULL; |
d92a8cfc | 3967 | fs_reclaim_release(gfp_mask); |
499118e9 VB |
3968 | current->flags &= ~PF_SWAPWRITE; |
3969 | memalloc_noreclaim_restore(noreclaim_flag); | |
a79311c1 | 3970 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3971 | } |
179e9639 | 3972 | |
a5f5f91d | 3973 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 3974 | { |
d773ed6b | 3975 | int ret; |
179e9639 AM |
3976 | |
3977 | /* | |
a5f5f91d | 3978 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 3979 | * slab pages if we are over the defined limits. |
34aa1330 | 3980 | * |
9614634f CL |
3981 | * A small portion of unmapped file backed pages is needed for |
3982 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
3983 | * thrown out if the node is overallocated. So we do not reclaim |
3984 | * if less than a specified percentage of the node is used by | |
9614634f | 3985 | * unmapped file backed pages. |
179e9639 | 3986 | */ |
a5f5f91d | 3987 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
385386cf | 3988 | node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) |
a5f5f91d | 3989 | return NODE_RECLAIM_FULL; |
179e9639 AM |
3990 | |
3991 | /* | |
d773ed6b | 3992 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3993 | */ |
d0164adc | 3994 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 3995 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
3996 | |
3997 | /* | |
a5f5f91d | 3998 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
3999 | * have associated processors. This will favor the local processor |
4000 | * over remote processors and spread off node memory allocations | |
4001 | * as wide as possible. | |
4002 | */ | |
a5f5f91d MG |
4003 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
4004 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 4005 | |
a5f5f91d MG |
4006 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
4007 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 4008 | |
a5f5f91d MG |
4009 | ret = __node_reclaim(pgdat, gfp_mask, order); |
4010 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 4011 | |
24cf7251 MG |
4012 | if (!ret) |
4013 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
4014 | ||
d773ed6b | 4015 | return ret; |
179e9639 | 4016 | } |
9eeff239 | 4017 | #endif |
894bc310 | 4018 | |
894bc310 LS |
4019 | /* |
4020 | * page_evictable - test whether a page is evictable | |
4021 | * @page: the page to test | |
894bc310 LS |
4022 | * |
4023 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 4024 | * lists vs unevictable list. |
894bc310 LS |
4025 | * |
4026 | * Reasons page might not be evictable: | |
ba9ddf49 | 4027 | * (1) page's mapping marked unevictable |
b291f000 | 4028 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 4029 | * |
894bc310 | 4030 | */ |
39b5f29a | 4031 | int page_evictable(struct page *page) |
894bc310 | 4032 | { |
8d700626 HY |
4033 | int ret; |
4034 | ||
4035 | /* Prevent address_space of inode and swap cache from being freed */ | |
4036 | rcu_read_lock(); | |
4037 | ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); | |
4038 | rcu_read_unlock(); | |
4039 | return ret; | |
894bc310 | 4040 | } |
89e004ea | 4041 | |
85046579 | 4042 | #ifdef CONFIG_SHMEM |
89e004ea | 4043 | /** |
24513264 HD |
4044 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
4045 | * @pages: array of pages to check | |
4046 | * @nr_pages: number of pages to check | |
89e004ea | 4047 | * |
24513264 | 4048 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
4049 | * |
4050 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 4051 | */ |
24513264 | 4052 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 4053 | { |
925b7673 | 4054 | struct lruvec *lruvec; |
785b99fe | 4055 | struct pglist_data *pgdat = NULL; |
24513264 HD |
4056 | int pgscanned = 0; |
4057 | int pgrescued = 0; | |
4058 | int i; | |
89e004ea | 4059 | |
24513264 HD |
4060 | for (i = 0; i < nr_pages; i++) { |
4061 | struct page *page = pages[i]; | |
785b99fe | 4062 | struct pglist_data *pagepgdat = page_pgdat(page); |
89e004ea | 4063 | |
24513264 | 4064 | pgscanned++; |
785b99fe MG |
4065 | if (pagepgdat != pgdat) { |
4066 | if (pgdat) | |
4067 | spin_unlock_irq(&pgdat->lru_lock); | |
4068 | pgdat = pagepgdat; | |
4069 | spin_lock_irq(&pgdat->lru_lock); | |
24513264 | 4070 | } |
785b99fe | 4071 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
89e004ea | 4072 | |
24513264 HD |
4073 | if (!PageLRU(page) || !PageUnevictable(page)) |
4074 | continue; | |
89e004ea | 4075 | |
39b5f29a | 4076 | if (page_evictable(page)) { |
24513264 HD |
4077 | enum lru_list lru = page_lru_base_type(page); |
4078 | ||
309381fe | 4079 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 4080 | ClearPageUnevictable(page); |
fa9add64 HD |
4081 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
4082 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 4083 | pgrescued++; |
89e004ea | 4084 | } |
24513264 | 4085 | } |
89e004ea | 4086 | |
785b99fe | 4087 | if (pgdat) { |
24513264 HD |
4088 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
4089 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
785b99fe | 4090 | spin_unlock_irq(&pgdat->lru_lock); |
89e004ea | 4091 | } |
89e004ea | 4092 | } |
85046579 | 4093 | #endif /* CONFIG_SHMEM */ |