]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/vmscan.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
5 | * | |
6 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
7 | * kswapd added: 7.1.96 sct | |
8 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
9 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
10 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
11 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
12 | */ | |
13 | ||
b1de0d13 MH |
14 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
15 | ||
1da177e4 | 16 | #include <linux/mm.h> |
5b3cc15a | 17 | #include <linux/sched/mm.h> |
1da177e4 | 18 | #include <linux/module.h> |
5a0e3ad6 | 19 | #include <linux/gfp.h> |
1da177e4 LT |
20 | #include <linux/kernel_stat.h> |
21 | #include <linux/swap.h> | |
22 | #include <linux/pagemap.h> | |
23 | #include <linux/init.h> | |
24 | #include <linux/highmem.h> | |
70ddf637 | 25 | #include <linux/vmpressure.h> |
e129b5c2 | 26 | #include <linux/vmstat.h> |
1da177e4 LT |
27 | #include <linux/file.h> |
28 | #include <linux/writeback.h> | |
29 | #include <linux/blkdev.h> | |
30 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
31 | buffer_heads_over_limit */ | |
32 | #include <linux/mm_inline.h> | |
1da177e4 LT |
33 | #include <linux/backing-dev.h> |
34 | #include <linux/rmap.h> | |
35 | #include <linux/topology.h> | |
36 | #include <linux/cpu.h> | |
37 | #include <linux/cpuset.h> | |
3e7d3449 | 38 | #include <linux/compaction.h> |
1da177e4 LT |
39 | #include <linux/notifier.h> |
40 | #include <linux/rwsem.h> | |
248a0301 | 41 | #include <linux/delay.h> |
3218ae14 | 42 | #include <linux/kthread.h> |
7dfb7103 | 43 | #include <linux/freezer.h> |
66e1707b | 44 | #include <linux/memcontrol.h> |
873b4771 | 45 | #include <linux/delayacct.h> |
af936a16 | 46 | #include <linux/sysctl.h> |
929bea7c | 47 | #include <linux/oom.h> |
268bb0ce | 48 | #include <linux/prefetch.h> |
b1de0d13 | 49 | #include <linux/printk.h> |
f9fe48be | 50 | #include <linux/dax.h> |
1da177e4 LT |
51 | |
52 | #include <asm/tlbflush.h> | |
53 | #include <asm/div64.h> | |
54 | ||
55 | #include <linux/swapops.h> | |
117aad1e | 56 | #include <linux/balloon_compaction.h> |
1da177e4 | 57 | |
0f8053a5 NP |
58 | #include "internal.h" |
59 | ||
33906bc5 MG |
60 | #define CREATE_TRACE_POINTS |
61 | #include <trace/events/vmscan.h> | |
62 | ||
1da177e4 | 63 | struct scan_control { |
22fba335 KM |
64 | /* How many pages shrink_list() should reclaim */ |
65 | unsigned long nr_to_reclaim; | |
66 | ||
1da177e4 | 67 | /* This context's GFP mask */ |
6daa0e28 | 68 | gfp_t gfp_mask; |
1da177e4 | 69 | |
ee814fe2 | 70 | /* Allocation order */ |
5ad333eb | 71 | int order; |
66e1707b | 72 | |
ee814fe2 JW |
73 | /* |
74 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
75 | * are scanned. | |
76 | */ | |
77 | nodemask_t *nodemask; | |
9e3b2f8c | 78 | |
f16015fb JW |
79 | /* |
80 | * The memory cgroup that hit its limit and as a result is the | |
81 | * primary target of this reclaim invocation. | |
82 | */ | |
83 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 84 | |
ee814fe2 JW |
85 | /* Scan (total_size >> priority) pages at once */ |
86 | int priority; | |
87 | ||
b2e18757 MG |
88 | /* The highest zone to isolate pages for reclaim from */ |
89 | enum zone_type reclaim_idx; | |
90 | ||
1276ad68 | 91 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
92 | unsigned int may_writepage:1; |
93 | ||
94 | /* Can mapped pages be reclaimed? */ | |
95 | unsigned int may_unmap:1; | |
96 | ||
97 | /* Can pages be swapped as part of reclaim? */ | |
98 | unsigned int may_swap:1; | |
99 | ||
d6622f63 YX |
100 | /* |
101 | * Cgroups are not reclaimed below their configured memory.low, | |
102 | * unless we threaten to OOM. If any cgroups are skipped due to | |
103 | * memory.low and nothing was reclaimed, go back for memory.low. | |
104 | */ | |
105 | unsigned int memcg_low_reclaim:1; | |
106 | unsigned int memcg_low_skipped:1; | |
241994ed | 107 | |
ee814fe2 JW |
108 | unsigned int hibernation_mode:1; |
109 | ||
110 | /* One of the zones is ready for compaction */ | |
111 | unsigned int compaction_ready:1; | |
112 | ||
113 | /* Incremented by the number of inactive pages that were scanned */ | |
114 | unsigned long nr_scanned; | |
115 | ||
116 | /* Number of pages freed so far during a call to shrink_zones() */ | |
117 | unsigned long nr_reclaimed; | |
1da177e4 LT |
118 | }; |
119 | ||
1da177e4 LT |
120 | #ifdef ARCH_HAS_PREFETCH |
121 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
122 | do { \ | |
123 | if ((_page)->lru.prev != _base) { \ | |
124 | struct page *prev; \ | |
125 | \ | |
126 | prev = lru_to_page(&(_page->lru)); \ | |
127 | prefetch(&prev->_field); \ | |
128 | } \ | |
129 | } while (0) | |
130 | #else | |
131 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
132 | #endif | |
133 | ||
134 | #ifdef ARCH_HAS_PREFETCHW | |
135 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
136 | do { \ | |
137 | if ((_page)->lru.prev != _base) { \ | |
138 | struct page *prev; \ | |
139 | \ | |
140 | prev = lru_to_page(&(_page->lru)); \ | |
141 | prefetchw(&prev->_field); \ | |
142 | } \ | |
143 | } while (0) | |
144 | #else | |
145 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
146 | #endif | |
147 | ||
148 | /* | |
149 | * From 0 .. 100. Higher means more swappy. | |
150 | */ | |
151 | int vm_swappiness = 60; | |
d0480be4 WSH |
152 | /* |
153 | * The total number of pages which are beyond the high watermark within all | |
154 | * zones. | |
155 | */ | |
156 | unsigned long vm_total_pages; | |
1da177e4 LT |
157 | |
158 | static LIST_HEAD(shrinker_list); | |
159 | static DECLARE_RWSEM(shrinker_rwsem); | |
160 | ||
c255a458 | 161 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
162 | static bool global_reclaim(struct scan_control *sc) |
163 | { | |
f16015fb | 164 | return !sc->target_mem_cgroup; |
89b5fae5 | 165 | } |
97c9341f TH |
166 | |
167 | /** | |
168 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
169 | * @sc: scan_control in question | |
170 | * | |
171 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
172 | * completely broken with the legacy memcg and direct stalling in | |
173 | * shrink_page_list() is used for throttling instead, which lacks all the | |
174 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
175 | * allocation and configurability. | |
176 | * | |
177 | * This function tests whether the vmscan currently in progress can assume | |
178 | * that the normal dirty throttling mechanism is operational. | |
179 | */ | |
180 | static bool sane_reclaim(struct scan_control *sc) | |
181 | { | |
182 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
183 | ||
184 | if (!memcg) | |
185 | return true; | |
186 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 187 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
188 | return true; |
189 | #endif | |
190 | return false; | |
191 | } | |
91a45470 | 192 | #else |
89b5fae5 JW |
193 | static bool global_reclaim(struct scan_control *sc) |
194 | { | |
195 | return true; | |
196 | } | |
97c9341f TH |
197 | |
198 | static bool sane_reclaim(struct scan_control *sc) | |
199 | { | |
200 | return true; | |
201 | } | |
91a45470 KH |
202 | #endif |
203 | ||
5a1c84b4 MG |
204 | /* |
205 | * This misses isolated pages which are not accounted for to save counters. | |
206 | * As the data only determines if reclaim or compaction continues, it is | |
207 | * not expected that isolated pages will be a dominating factor. | |
208 | */ | |
209 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
210 | { | |
211 | unsigned long nr; | |
212 | ||
213 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
214 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
215 | if (get_nr_swap_pages() > 0) | |
216 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
217 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
218 | ||
219 | return nr; | |
220 | } | |
221 | ||
599d0c95 MG |
222 | unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat) |
223 | { | |
224 | unsigned long nr; | |
225 | ||
226 | nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) + | |
227 | node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) + | |
228 | node_page_state_snapshot(pgdat, NR_ISOLATED_FILE); | |
6e543d57 LD |
229 | |
230 | if (get_nr_swap_pages() > 0) | |
599d0c95 MG |
231 | nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) + |
232 | node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) + | |
233 | node_page_state_snapshot(pgdat, NR_ISOLATED_ANON); | |
6e543d57 LD |
234 | |
235 | return nr; | |
236 | } | |
237 | ||
fd538803 MH |
238 | /** |
239 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
240 | * @lruvec: lru vector | |
241 | * @lru: lru to use | |
242 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
243 | */ | |
244 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) | |
c9f299d9 | 245 | { |
fd538803 MH |
246 | unsigned long lru_size; |
247 | int zid; | |
248 | ||
c3c787e8 | 249 | if (!mem_cgroup_disabled()) |
fd538803 MH |
250 | lru_size = mem_cgroup_get_lru_size(lruvec, lru); |
251 | else | |
252 | lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); | |
a3d8e054 | 253 | |
fd538803 MH |
254 | for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { |
255 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; | |
256 | unsigned long size; | |
c9f299d9 | 257 | |
fd538803 MH |
258 | if (!managed_zone(zone)) |
259 | continue; | |
260 | ||
261 | if (!mem_cgroup_disabled()) | |
262 | size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); | |
263 | else | |
264 | size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], | |
265 | NR_ZONE_LRU_BASE + lru); | |
266 | lru_size -= min(size, lru_size); | |
267 | } | |
268 | ||
269 | return lru_size; | |
b4536f0c | 270 | |
b4536f0c MH |
271 | } |
272 | ||
1da177e4 | 273 | /* |
1d3d4437 | 274 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 275 | */ |
1d3d4437 | 276 | int register_shrinker(struct shrinker *shrinker) |
1da177e4 | 277 | { |
1d3d4437 GC |
278 | size_t size = sizeof(*shrinker->nr_deferred); |
279 | ||
1d3d4437 GC |
280 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
281 | size *= nr_node_ids; | |
282 | ||
283 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
284 | if (!shrinker->nr_deferred) | |
285 | return -ENOMEM; | |
286 | ||
8e1f936b RR |
287 | down_write(&shrinker_rwsem); |
288 | list_add_tail(&shrinker->list, &shrinker_list); | |
289 | up_write(&shrinker_rwsem); | |
1d3d4437 | 290 | return 0; |
1da177e4 | 291 | } |
8e1f936b | 292 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
293 | |
294 | /* | |
295 | * Remove one | |
296 | */ | |
8e1f936b | 297 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
298 | { |
299 | down_write(&shrinker_rwsem); | |
300 | list_del(&shrinker->list); | |
301 | up_write(&shrinker_rwsem); | |
ae393321 | 302 | kfree(shrinker->nr_deferred); |
1da177e4 | 303 | } |
8e1f936b | 304 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
305 | |
306 | #define SHRINK_BATCH 128 | |
1d3d4437 | 307 | |
cb731d6c VD |
308 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
309 | struct shrinker *shrinker, | |
310 | unsigned long nr_scanned, | |
311 | unsigned long nr_eligible) | |
1d3d4437 GC |
312 | { |
313 | unsigned long freed = 0; | |
314 | unsigned long long delta; | |
315 | long total_scan; | |
d5bc5fd3 | 316 | long freeable; |
1d3d4437 GC |
317 | long nr; |
318 | long new_nr; | |
319 | int nid = shrinkctl->nid; | |
320 | long batch_size = shrinker->batch ? shrinker->batch | |
321 | : SHRINK_BATCH; | |
5f33a080 | 322 | long scanned = 0, next_deferred; |
1d3d4437 | 323 | |
d5bc5fd3 VD |
324 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
325 | if (freeable == 0) | |
1d3d4437 GC |
326 | return 0; |
327 | ||
328 | /* | |
329 | * copy the current shrinker scan count into a local variable | |
330 | * and zero it so that other concurrent shrinker invocations | |
331 | * don't also do this scanning work. | |
332 | */ | |
333 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
334 | ||
335 | total_scan = nr; | |
6b4f7799 | 336 | delta = (4 * nr_scanned) / shrinker->seeks; |
d5bc5fd3 | 337 | delta *= freeable; |
6b4f7799 | 338 | do_div(delta, nr_eligible + 1); |
1d3d4437 GC |
339 | total_scan += delta; |
340 | if (total_scan < 0) { | |
8612c663 | 341 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
a0b02131 | 342 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 343 | total_scan = freeable; |
5f33a080 SL |
344 | next_deferred = nr; |
345 | } else | |
346 | next_deferred = total_scan; | |
1d3d4437 GC |
347 | |
348 | /* | |
349 | * We need to avoid excessive windup on filesystem shrinkers | |
350 | * due to large numbers of GFP_NOFS allocations causing the | |
351 | * shrinkers to return -1 all the time. This results in a large | |
352 | * nr being built up so when a shrink that can do some work | |
353 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 354 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
355 | * memory. |
356 | * | |
357 | * Hence only allow the shrinker to scan the entire cache when | |
358 | * a large delta change is calculated directly. | |
359 | */ | |
d5bc5fd3 VD |
360 | if (delta < freeable / 4) |
361 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
362 | |
363 | /* | |
364 | * Avoid risking looping forever due to too large nr value: | |
365 | * never try to free more than twice the estimate number of | |
366 | * freeable entries. | |
367 | */ | |
d5bc5fd3 VD |
368 | if (total_scan > freeable * 2) |
369 | total_scan = freeable * 2; | |
1d3d4437 GC |
370 | |
371 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
6b4f7799 JW |
372 | nr_scanned, nr_eligible, |
373 | freeable, delta, total_scan); | |
1d3d4437 | 374 | |
0b1fb40a VD |
375 | /* |
376 | * Normally, we should not scan less than batch_size objects in one | |
377 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
378 | * than batch_size objects in total and we are really tight on memory, | |
379 | * we will try to reclaim all available objects, otherwise we can end | |
380 | * up failing allocations although there are plenty of reclaimable | |
381 | * objects spread over several slabs with usage less than the | |
382 | * batch_size. | |
383 | * | |
384 | * We detect the "tight on memory" situations by looking at the total | |
385 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 386 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
387 | * scanning at high prio and therefore should try to reclaim as much as |
388 | * possible. | |
389 | */ | |
390 | while (total_scan >= batch_size || | |
d5bc5fd3 | 391 | total_scan >= freeable) { |
a0b02131 | 392 | unsigned long ret; |
0b1fb40a | 393 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 394 | |
0b1fb40a | 395 | shrinkctl->nr_to_scan = nr_to_scan; |
a0b02131 DC |
396 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
397 | if (ret == SHRINK_STOP) | |
398 | break; | |
399 | freed += ret; | |
1d3d4437 | 400 | |
0b1fb40a VD |
401 | count_vm_events(SLABS_SCANNED, nr_to_scan); |
402 | total_scan -= nr_to_scan; | |
5f33a080 | 403 | scanned += nr_to_scan; |
1d3d4437 GC |
404 | |
405 | cond_resched(); | |
406 | } | |
407 | ||
5f33a080 SL |
408 | if (next_deferred >= scanned) |
409 | next_deferred -= scanned; | |
410 | else | |
411 | next_deferred = 0; | |
1d3d4437 GC |
412 | /* |
413 | * move the unused scan count back into the shrinker in a | |
414 | * manner that handles concurrent updates. If we exhausted the | |
415 | * scan, there is no need to do an update. | |
416 | */ | |
5f33a080 SL |
417 | if (next_deferred > 0) |
418 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
419 | &shrinker->nr_deferred[nid]); |
420 | else | |
421 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
422 | ||
df9024a8 | 423 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 424 | return freed; |
1495f230 YH |
425 | } |
426 | ||
6b4f7799 | 427 | /** |
cb731d6c | 428 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
429 | * @gfp_mask: allocation context |
430 | * @nid: node whose slab caches to target | |
cb731d6c | 431 | * @memcg: memory cgroup whose slab caches to target |
6b4f7799 JW |
432 | * @nr_scanned: pressure numerator |
433 | * @nr_eligible: pressure denominator | |
1da177e4 | 434 | * |
6b4f7799 | 435 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 436 | * |
6b4f7799 JW |
437 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
438 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 439 | * |
cb731d6c VD |
440 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
441 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan | |
0fc9f58a VD |
442 | * objects from the memory cgroup specified. Otherwise, only unaware |
443 | * shrinkers are called. | |
cb731d6c | 444 | * |
6b4f7799 JW |
445 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of |
446 | * the available objects should be scanned. Page reclaim for example | |
447 | * passes the number of pages scanned and the number of pages on the | |
448 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned | |
449 | * when it encountered mapped pages. The ratio is further biased by | |
450 | * the ->seeks setting of the shrink function, which indicates the | |
451 | * cost to recreate an object relative to that of an LRU page. | |
b15e0905 | 452 | * |
6b4f7799 | 453 | * Returns the number of reclaimed slab objects. |
1da177e4 | 454 | */ |
cb731d6c VD |
455 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
456 | struct mem_cgroup *memcg, | |
457 | unsigned long nr_scanned, | |
458 | unsigned long nr_eligible) | |
1da177e4 LT |
459 | { |
460 | struct shrinker *shrinker; | |
24f7c6b9 | 461 | unsigned long freed = 0; |
1da177e4 | 462 | |
0fc9f58a | 463 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) |
cb731d6c VD |
464 | return 0; |
465 | ||
6b4f7799 JW |
466 | if (nr_scanned == 0) |
467 | nr_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 468 | |
f06590bd | 469 | if (!down_read_trylock(&shrinker_rwsem)) { |
24f7c6b9 DC |
470 | /* |
471 | * If we would return 0, our callers would understand that we | |
472 | * have nothing else to shrink and give up trying. By returning | |
473 | * 1 we keep it going and assume we'll be able to shrink next | |
474 | * time. | |
475 | */ | |
476 | freed = 1; | |
f06590bd MK |
477 | goto out; |
478 | } | |
1da177e4 LT |
479 | |
480 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
481 | struct shrink_control sc = { |
482 | .gfp_mask = gfp_mask, | |
483 | .nid = nid, | |
cb731d6c | 484 | .memcg = memcg, |
6b4f7799 | 485 | }; |
ec97097b | 486 | |
0fc9f58a VD |
487 | /* |
488 | * If kernel memory accounting is disabled, we ignore | |
489 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers | |
490 | * passing NULL for memcg. | |
491 | */ | |
492 | if (memcg_kmem_enabled() && | |
493 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
cb731d6c VD |
494 | continue; |
495 | ||
6b4f7799 JW |
496 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
497 | sc.nid = 0; | |
1da177e4 | 498 | |
cb731d6c | 499 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); |
1da177e4 | 500 | } |
6b4f7799 | 501 | |
1da177e4 | 502 | up_read(&shrinker_rwsem); |
f06590bd MK |
503 | out: |
504 | cond_resched(); | |
24f7c6b9 | 505 | return freed; |
1da177e4 LT |
506 | } |
507 | ||
cb731d6c VD |
508 | void drop_slab_node(int nid) |
509 | { | |
510 | unsigned long freed; | |
511 | ||
512 | do { | |
513 | struct mem_cgroup *memcg = NULL; | |
514 | ||
515 | freed = 0; | |
516 | do { | |
517 | freed += shrink_slab(GFP_KERNEL, nid, memcg, | |
518 | 1000, 1000); | |
519 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
520 | } while (freed > 10); | |
521 | } | |
522 | ||
523 | void drop_slab(void) | |
524 | { | |
525 | int nid; | |
526 | ||
527 | for_each_online_node(nid) | |
528 | drop_slab_node(nid); | |
529 | } | |
530 | ||
1da177e4 LT |
531 | static inline int is_page_cache_freeable(struct page *page) |
532 | { | |
ceddc3a5 JW |
533 | /* |
534 | * A freeable page cache page is referenced only by the caller | |
535 | * that isolated the page, the page cache radix tree and | |
536 | * optional buffer heads at page->private. | |
537 | */ | |
edcf4748 | 538 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
539 | } |
540 | ||
703c2708 | 541 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
1da177e4 | 542 | { |
930d9152 | 543 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 544 | return 1; |
703c2708 | 545 | if (!inode_write_congested(inode)) |
1da177e4 | 546 | return 1; |
703c2708 | 547 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
548 | return 1; |
549 | return 0; | |
550 | } | |
551 | ||
552 | /* | |
553 | * We detected a synchronous write error writing a page out. Probably | |
554 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
555 | * fsync(), msync() or close(). | |
556 | * | |
557 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
558 | * prevents it from being freed up. But we have a ref on the page and once | |
559 | * that page is locked, the mapping is pinned. | |
560 | * | |
561 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
562 | * __GFP_FS. | |
563 | */ | |
564 | static void handle_write_error(struct address_space *mapping, | |
565 | struct page *page, int error) | |
566 | { | |
7eaceacc | 567 | lock_page(page); |
3e9f45bd GC |
568 | if (page_mapping(page) == mapping) |
569 | mapping_set_error(mapping, error); | |
1da177e4 LT |
570 | unlock_page(page); |
571 | } | |
572 | ||
04e62a29 CL |
573 | /* possible outcome of pageout() */ |
574 | typedef enum { | |
575 | /* failed to write page out, page is locked */ | |
576 | PAGE_KEEP, | |
577 | /* move page to the active list, page is locked */ | |
578 | PAGE_ACTIVATE, | |
579 | /* page has been sent to the disk successfully, page is unlocked */ | |
580 | PAGE_SUCCESS, | |
581 | /* page is clean and locked */ | |
582 | PAGE_CLEAN, | |
583 | } pageout_t; | |
584 | ||
1da177e4 | 585 | /* |
1742f19f AM |
586 | * pageout is called by shrink_page_list() for each dirty page. |
587 | * Calls ->writepage(). | |
1da177e4 | 588 | */ |
c661b078 | 589 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 590 | struct scan_control *sc) |
1da177e4 LT |
591 | { |
592 | /* | |
593 | * If the page is dirty, only perform writeback if that write | |
594 | * will be non-blocking. To prevent this allocation from being | |
595 | * stalled by pagecache activity. But note that there may be | |
596 | * stalls if we need to run get_block(). We could test | |
597 | * PagePrivate for that. | |
598 | * | |
8174202b | 599 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
600 | * this page's queue, we can perform writeback even if that |
601 | * will block. | |
602 | * | |
603 | * If the page is swapcache, write it back even if that would | |
604 | * block, for some throttling. This happens by accident, because | |
605 | * swap_backing_dev_info is bust: it doesn't reflect the | |
606 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
607 | */ |
608 | if (!is_page_cache_freeable(page)) | |
609 | return PAGE_KEEP; | |
610 | if (!mapping) { | |
611 | /* | |
612 | * Some data journaling orphaned pages can have | |
613 | * page->mapping == NULL while being dirty with clean buffers. | |
614 | */ | |
266cf658 | 615 | if (page_has_private(page)) { |
1da177e4 LT |
616 | if (try_to_free_buffers(page)) { |
617 | ClearPageDirty(page); | |
b1de0d13 | 618 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
619 | return PAGE_CLEAN; |
620 | } | |
621 | } | |
622 | return PAGE_KEEP; | |
623 | } | |
624 | if (mapping->a_ops->writepage == NULL) | |
625 | return PAGE_ACTIVATE; | |
703c2708 | 626 | if (!may_write_to_inode(mapping->host, sc)) |
1da177e4 LT |
627 | return PAGE_KEEP; |
628 | ||
629 | if (clear_page_dirty_for_io(page)) { | |
630 | int res; | |
631 | struct writeback_control wbc = { | |
632 | .sync_mode = WB_SYNC_NONE, | |
633 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
634 | .range_start = 0, |
635 | .range_end = LLONG_MAX, | |
1da177e4 LT |
636 | .for_reclaim = 1, |
637 | }; | |
638 | ||
639 | SetPageReclaim(page); | |
640 | res = mapping->a_ops->writepage(page, &wbc); | |
641 | if (res < 0) | |
642 | handle_write_error(mapping, page, res); | |
994fc28c | 643 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
644 | ClearPageReclaim(page); |
645 | return PAGE_ACTIVATE; | |
646 | } | |
c661b078 | 647 | |
1da177e4 LT |
648 | if (!PageWriteback(page)) { |
649 | /* synchronous write or broken a_ops? */ | |
650 | ClearPageReclaim(page); | |
651 | } | |
3aa23851 | 652 | trace_mm_vmscan_writepage(page); |
c4a25635 | 653 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
654 | return PAGE_SUCCESS; |
655 | } | |
656 | ||
657 | return PAGE_CLEAN; | |
658 | } | |
659 | ||
a649fd92 | 660 | /* |
e286781d NP |
661 | * Same as remove_mapping, but if the page is removed from the mapping, it |
662 | * gets returned with a refcount of 0. | |
a649fd92 | 663 | */ |
a528910e JW |
664 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
665 | bool reclaimed) | |
49d2e9cc | 666 | { |
c4843a75 | 667 | unsigned long flags; |
c4843a75 | 668 | |
28e4d965 NP |
669 | BUG_ON(!PageLocked(page)); |
670 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 671 | |
c4843a75 | 672 | spin_lock_irqsave(&mapping->tree_lock, flags); |
49d2e9cc | 673 | /* |
0fd0e6b0 NP |
674 | * The non racy check for a busy page. |
675 | * | |
676 | * Must be careful with the order of the tests. When someone has | |
677 | * a ref to the page, it may be possible that they dirty it then | |
678 | * drop the reference. So if PageDirty is tested before page_count | |
679 | * here, then the following race may occur: | |
680 | * | |
681 | * get_user_pages(&page); | |
682 | * [user mapping goes away] | |
683 | * write_to(page); | |
684 | * !PageDirty(page) [good] | |
685 | * SetPageDirty(page); | |
686 | * put_page(page); | |
687 | * !page_count(page) [good, discard it] | |
688 | * | |
689 | * [oops, our write_to data is lost] | |
690 | * | |
691 | * Reversing the order of the tests ensures such a situation cannot | |
692 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 693 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
694 | * |
695 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
696 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 697 | */ |
fe896d18 | 698 | if (!page_ref_freeze(page, 2)) |
49d2e9cc | 699 | goto cannot_free; |
e286781d NP |
700 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
701 | if (unlikely(PageDirty(page))) { | |
fe896d18 | 702 | page_ref_unfreeze(page, 2); |
49d2e9cc | 703 | goto cannot_free; |
e286781d | 704 | } |
49d2e9cc CL |
705 | |
706 | if (PageSwapCache(page)) { | |
707 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 708 | mem_cgroup_swapout(page, swap); |
49d2e9cc | 709 | __delete_from_swap_cache(page); |
c4843a75 | 710 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
0a31bc97 | 711 | swapcache_free(swap); |
e286781d | 712 | } else { |
6072d13c | 713 | void (*freepage)(struct page *); |
a528910e | 714 | void *shadow = NULL; |
6072d13c LT |
715 | |
716 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
717 | /* |
718 | * Remember a shadow entry for reclaimed file cache in | |
719 | * order to detect refaults, thus thrashing, later on. | |
720 | * | |
721 | * But don't store shadows in an address space that is | |
722 | * already exiting. This is not just an optizimation, | |
723 | * inode reclaim needs to empty out the radix tree or | |
724 | * the nodes are lost. Don't plant shadows behind its | |
725 | * back. | |
f9fe48be RZ |
726 | * |
727 | * We also don't store shadows for DAX mappings because the | |
728 | * only page cache pages found in these are zero pages | |
729 | * covering holes, and because we don't want to mix DAX | |
730 | * exceptional entries and shadow exceptional entries in the | |
731 | * same page_tree. | |
a528910e JW |
732 | */ |
733 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 734 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
a528910e | 735 | shadow = workingset_eviction(mapping, page); |
62cccb8c | 736 | __delete_from_page_cache(page, shadow); |
c4843a75 | 737 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
6072d13c LT |
738 | |
739 | if (freepage != NULL) | |
740 | freepage(page); | |
49d2e9cc CL |
741 | } |
742 | ||
49d2e9cc CL |
743 | return 1; |
744 | ||
745 | cannot_free: | |
c4843a75 | 746 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
49d2e9cc CL |
747 | return 0; |
748 | } | |
749 | ||
e286781d NP |
750 | /* |
751 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
752 | * someone else has a ref on the page, abort and return 0. If it was | |
753 | * successfully detached, return 1. Assumes the caller has a single ref on | |
754 | * this page. | |
755 | */ | |
756 | int remove_mapping(struct address_space *mapping, struct page *page) | |
757 | { | |
a528910e | 758 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
759 | /* |
760 | * Unfreezing the refcount with 1 rather than 2 effectively | |
761 | * drops the pagecache ref for us without requiring another | |
762 | * atomic operation. | |
763 | */ | |
fe896d18 | 764 | page_ref_unfreeze(page, 1); |
e286781d NP |
765 | return 1; |
766 | } | |
767 | return 0; | |
768 | } | |
769 | ||
894bc310 LS |
770 | /** |
771 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
772 | * @page: page to be put back to appropriate lru list | |
773 | * | |
774 | * Add previously isolated @page to appropriate LRU list. | |
775 | * Page may still be unevictable for other reasons. | |
776 | * | |
777 | * lru_lock must not be held, interrupts must be enabled. | |
778 | */ | |
894bc310 LS |
779 | void putback_lru_page(struct page *page) |
780 | { | |
0ec3b74c | 781 | bool is_unevictable; |
bbfd28ee | 782 | int was_unevictable = PageUnevictable(page); |
894bc310 | 783 | |
309381fe | 784 | VM_BUG_ON_PAGE(PageLRU(page), page); |
894bc310 LS |
785 | |
786 | redo: | |
787 | ClearPageUnevictable(page); | |
788 | ||
39b5f29a | 789 | if (page_evictable(page)) { |
894bc310 LS |
790 | /* |
791 | * For evictable pages, we can use the cache. | |
792 | * In event of a race, worst case is we end up with an | |
793 | * unevictable page on [in]active list. | |
794 | * We know how to handle that. | |
795 | */ | |
0ec3b74c | 796 | is_unevictable = false; |
c53954a0 | 797 | lru_cache_add(page); |
894bc310 LS |
798 | } else { |
799 | /* | |
800 | * Put unevictable pages directly on zone's unevictable | |
801 | * list. | |
802 | */ | |
0ec3b74c | 803 | is_unevictable = true; |
894bc310 | 804 | add_page_to_unevictable_list(page); |
6a7b9548 | 805 | /* |
21ee9f39 MK |
806 | * When racing with an mlock or AS_UNEVICTABLE clearing |
807 | * (page is unlocked) make sure that if the other thread | |
808 | * does not observe our setting of PG_lru and fails | |
24513264 | 809 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 810 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
811 | * the page back to the evictable list. |
812 | * | |
21ee9f39 | 813 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
814 | */ |
815 | smp_mb(); | |
894bc310 | 816 | } |
894bc310 LS |
817 | |
818 | /* | |
819 | * page's status can change while we move it among lru. If an evictable | |
820 | * page is on unevictable list, it never be freed. To avoid that, | |
821 | * check after we added it to the list, again. | |
822 | */ | |
0ec3b74c | 823 | if (is_unevictable && page_evictable(page)) { |
894bc310 LS |
824 | if (!isolate_lru_page(page)) { |
825 | put_page(page); | |
826 | goto redo; | |
827 | } | |
828 | /* This means someone else dropped this page from LRU | |
829 | * So, it will be freed or putback to LRU again. There is | |
830 | * nothing to do here. | |
831 | */ | |
832 | } | |
833 | ||
0ec3b74c | 834 | if (was_unevictable && !is_unevictable) |
bbfd28ee | 835 | count_vm_event(UNEVICTABLE_PGRESCUED); |
0ec3b74c | 836 | else if (!was_unevictable && is_unevictable) |
bbfd28ee LS |
837 | count_vm_event(UNEVICTABLE_PGCULLED); |
838 | ||
894bc310 LS |
839 | put_page(page); /* drop ref from isolate */ |
840 | } | |
841 | ||
dfc8d636 JW |
842 | enum page_references { |
843 | PAGEREF_RECLAIM, | |
844 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 845 | PAGEREF_KEEP, |
dfc8d636 JW |
846 | PAGEREF_ACTIVATE, |
847 | }; | |
848 | ||
849 | static enum page_references page_check_references(struct page *page, | |
850 | struct scan_control *sc) | |
851 | { | |
64574746 | 852 | int referenced_ptes, referenced_page; |
dfc8d636 | 853 | unsigned long vm_flags; |
dfc8d636 | 854 | |
c3ac9a8a JW |
855 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
856 | &vm_flags); | |
64574746 | 857 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 858 | |
dfc8d636 JW |
859 | /* |
860 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
861 | * move the page to the unevictable list. | |
862 | */ | |
863 | if (vm_flags & VM_LOCKED) | |
864 | return PAGEREF_RECLAIM; | |
865 | ||
64574746 | 866 | if (referenced_ptes) { |
e4898273 | 867 | if (PageSwapBacked(page)) |
64574746 JW |
868 | return PAGEREF_ACTIVATE; |
869 | /* | |
870 | * All mapped pages start out with page table | |
871 | * references from the instantiating fault, so we need | |
872 | * to look twice if a mapped file page is used more | |
873 | * than once. | |
874 | * | |
875 | * Mark it and spare it for another trip around the | |
876 | * inactive list. Another page table reference will | |
877 | * lead to its activation. | |
878 | * | |
879 | * Note: the mark is set for activated pages as well | |
880 | * so that recently deactivated but used pages are | |
881 | * quickly recovered. | |
882 | */ | |
883 | SetPageReferenced(page); | |
884 | ||
34dbc67a | 885 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
886 | return PAGEREF_ACTIVATE; |
887 | ||
c909e993 KK |
888 | /* |
889 | * Activate file-backed executable pages after first usage. | |
890 | */ | |
891 | if (vm_flags & VM_EXEC) | |
892 | return PAGEREF_ACTIVATE; | |
893 | ||
64574746 JW |
894 | return PAGEREF_KEEP; |
895 | } | |
dfc8d636 JW |
896 | |
897 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 898 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
899 | return PAGEREF_RECLAIM_CLEAN; |
900 | ||
901 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
902 | } |
903 | ||
e2be15f6 MG |
904 | /* Check if a page is dirty or under writeback */ |
905 | static void page_check_dirty_writeback(struct page *page, | |
906 | bool *dirty, bool *writeback) | |
907 | { | |
b4597226 MG |
908 | struct address_space *mapping; |
909 | ||
e2be15f6 MG |
910 | /* |
911 | * Anonymous pages are not handled by flushers and must be written | |
912 | * from reclaim context. Do not stall reclaim based on them | |
913 | */ | |
802a3a92 SL |
914 | if (!page_is_file_cache(page) || |
915 | (PageAnon(page) && !PageSwapBacked(page))) { | |
e2be15f6 MG |
916 | *dirty = false; |
917 | *writeback = false; | |
918 | return; | |
919 | } | |
920 | ||
921 | /* By default assume that the page flags are accurate */ | |
922 | *dirty = PageDirty(page); | |
923 | *writeback = PageWriteback(page); | |
b4597226 MG |
924 | |
925 | /* Verify dirty/writeback state if the filesystem supports it */ | |
926 | if (!page_has_private(page)) | |
927 | return; | |
928 | ||
929 | mapping = page_mapping(page); | |
930 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
931 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
932 | } |
933 | ||
3c710c1a MH |
934 | struct reclaim_stat { |
935 | unsigned nr_dirty; | |
936 | unsigned nr_unqueued_dirty; | |
937 | unsigned nr_congested; | |
938 | unsigned nr_writeback; | |
939 | unsigned nr_immediate; | |
5bccd166 MH |
940 | unsigned nr_activate; |
941 | unsigned nr_ref_keep; | |
942 | unsigned nr_unmap_fail; | |
3c710c1a MH |
943 | }; |
944 | ||
1da177e4 | 945 | /* |
1742f19f | 946 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 947 | */ |
1742f19f | 948 | static unsigned long shrink_page_list(struct list_head *page_list, |
599d0c95 | 949 | struct pglist_data *pgdat, |
f84f6e2b | 950 | struct scan_control *sc, |
02c6de8d | 951 | enum ttu_flags ttu_flags, |
3c710c1a | 952 | struct reclaim_stat *stat, |
02c6de8d | 953 | bool force_reclaim) |
1da177e4 LT |
954 | { |
955 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 956 | LIST_HEAD(free_pages); |
1da177e4 | 957 | int pgactivate = 0; |
3c710c1a MH |
958 | unsigned nr_unqueued_dirty = 0; |
959 | unsigned nr_dirty = 0; | |
960 | unsigned nr_congested = 0; | |
961 | unsigned nr_reclaimed = 0; | |
962 | unsigned nr_writeback = 0; | |
963 | unsigned nr_immediate = 0; | |
5bccd166 MH |
964 | unsigned nr_ref_keep = 0; |
965 | unsigned nr_unmap_fail = 0; | |
1da177e4 LT |
966 | |
967 | cond_resched(); | |
968 | ||
1da177e4 LT |
969 | while (!list_empty(page_list)) { |
970 | struct address_space *mapping; | |
971 | struct page *page; | |
972 | int may_enter_fs; | |
02c6de8d | 973 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
e2be15f6 | 974 | bool dirty, writeback; |
1da177e4 LT |
975 | |
976 | cond_resched(); | |
977 | ||
978 | page = lru_to_page(page_list); | |
979 | list_del(&page->lru); | |
980 | ||
529ae9aa | 981 | if (!trylock_page(page)) |
1da177e4 LT |
982 | goto keep; |
983 | ||
309381fe | 984 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
985 | |
986 | sc->nr_scanned++; | |
80e43426 | 987 | |
39b5f29a | 988 | if (unlikely(!page_evictable(page))) |
ad6b6704 | 989 | goto activate_locked; |
894bc310 | 990 | |
a6dc60f8 | 991 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
992 | goto keep_locked; |
993 | ||
1da177e4 | 994 | /* Double the slab pressure for mapped and swapcache pages */ |
802a3a92 SL |
995 | if ((page_mapped(page) || PageSwapCache(page)) && |
996 | !(PageAnon(page) && !PageSwapBacked(page))) | |
1da177e4 LT |
997 | sc->nr_scanned++; |
998 | ||
c661b078 AW |
999 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1000 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1001 | ||
e2be15f6 MG |
1002 | /* |
1003 | * The number of dirty pages determines if a zone is marked | |
1004 | * reclaim_congested which affects wait_iff_congested. kswapd | |
1005 | * will stall and start writing pages if the tail of the LRU | |
1006 | * is all dirty unqueued pages. | |
1007 | */ | |
1008 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1009 | if (dirty || writeback) | |
1010 | nr_dirty++; | |
1011 | ||
1012 | if (dirty && !writeback) | |
1013 | nr_unqueued_dirty++; | |
1014 | ||
d04e8acd MG |
1015 | /* |
1016 | * Treat this page as congested if the underlying BDI is or if | |
1017 | * pages are cycling through the LRU so quickly that the | |
1018 | * pages marked for immediate reclaim are making it to the | |
1019 | * end of the LRU a second time. | |
1020 | */ | |
e2be15f6 | 1021 | mapping = page_mapping(page); |
1da58ee2 | 1022 | if (((dirty || writeback) && mapping && |
703c2708 | 1023 | inode_write_congested(mapping->host)) || |
d04e8acd | 1024 | (writeback && PageReclaim(page))) |
e2be15f6 MG |
1025 | nr_congested++; |
1026 | ||
283aba9f MG |
1027 | /* |
1028 | * If a page at the tail of the LRU is under writeback, there | |
1029 | * are three cases to consider. | |
1030 | * | |
1031 | * 1) If reclaim is encountering an excessive number of pages | |
1032 | * under writeback and this page is both under writeback and | |
1033 | * PageReclaim then it indicates that pages are being queued | |
1034 | * for IO but are being recycled through the LRU before the | |
1035 | * IO can complete. Waiting on the page itself risks an | |
1036 | * indefinite stall if it is impossible to writeback the | |
1037 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1038 | * note that the LRU is being scanned too quickly and the |
1039 | * caller can stall after page list has been processed. | |
283aba9f | 1040 | * |
97c9341f | 1041 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1042 | * not marked for immediate reclaim, or the caller does not |
1043 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1044 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1045 | * reclaim and continue scanning. |
283aba9f | 1046 | * |
ecf5fc6e MH |
1047 | * Require may_enter_fs because we would wait on fs, which |
1048 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1049 | * enter reclaim, and deadlock if it waits on a page for |
1050 | * which it is needed to do the write (loop masks off | |
1051 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1052 | * would probably show more reasons. | |
1053 | * | |
7fadc820 | 1054 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1055 | * PageReclaim. memcg does not have any dirty pages |
1056 | * throttling so we could easily OOM just because too many | |
1057 | * pages are in writeback and there is nothing else to | |
1058 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1059 | * |
1060 | * In cases 1) and 2) we activate the pages to get them out of | |
1061 | * the way while we continue scanning for clean pages on the | |
1062 | * inactive list and refilling from the active list. The | |
1063 | * observation here is that waiting for disk writes is more | |
1064 | * expensive than potentially causing reloads down the line. | |
1065 | * Since they're marked for immediate reclaim, they won't put | |
1066 | * memory pressure on the cache working set any longer than it | |
1067 | * takes to write them to disk. | |
283aba9f | 1068 | */ |
c661b078 | 1069 | if (PageWriteback(page)) { |
283aba9f MG |
1070 | /* Case 1 above */ |
1071 | if (current_is_kswapd() && | |
1072 | PageReclaim(page) && | |
599d0c95 | 1073 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
b1a6f21e | 1074 | nr_immediate++; |
c55e8d03 | 1075 | goto activate_locked; |
283aba9f MG |
1076 | |
1077 | /* Case 2 above */ | |
97c9341f | 1078 | } else if (sane_reclaim(sc) || |
ecf5fc6e | 1079 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1080 | /* |
1081 | * This is slightly racy - end_page_writeback() | |
1082 | * might have just cleared PageReclaim, then | |
1083 | * setting PageReclaim here end up interpreted | |
1084 | * as PageReadahead - but that does not matter | |
1085 | * enough to care. What we do want is for this | |
1086 | * page to have PageReclaim set next time memcg | |
1087 | * reclaim reaches the tests above, so it will | |
1088 | * then wait_on_page_writeback() to avoid OOM; | |
1089 | * and it's also appropriate in global reclaim. | |
1090 | */ | |
1091 | SetPageReclaim(page); | |
e62e384e | 1092 | nr_writeback++; |
c55e8d03 | 1093 | goto activate_locked; |
283aba9f MG |
1094 | |
1095 | /* Case 3 above */ | |
1096 | } else { | |
7fadc820 | 1097 | unlock_page(page); |
283aba9f | 1098 | wait_on_page_writeback(page); |
7fadc820 HD |
1099 | /* then go back and try same page again */ |
1100 | list_add_tail(&page->lru, page_list); | |
1101 | continue; | |
e62e384e | 1102 | } |
c661b078 | 1103 | } |
1da177e4 | 1104 | |
02c6de8d MK |
1105 | if (!force_reclaim) |
1106 | references = page_check_references(page, sc); | |
1107 | ||
dfc8d636 JW |
1108 | switch (references) { |
1109 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1110 | goto activate_locked; |
64574746 | 1111 | case PAGEREF_KEEP: |
5bccd166 | 1112 | nr_ref_keep++; |
64574746 | 1113 | goto keep_locked; |
dfc8d636 JW |
1114 | case PAGEREF_RECLAIM: |
1115 | case PAGEREF_RECLAIM_CLEAN: | |
1116 | ; /* try to reclaim the page below */ | |
1117 | } | |
1da177e4 | 1118 | |
1da177e4 LT |
1119 | /* |
1120 | * Anonymous process memory has backing store? | |
1121 | * Try to allocate it some swap space here. | |
802a3a92 | 1122 | * Lazyfree page could be freed directly |
1da177e4 | 1123 | */ |
802a3a92 SL |
1124 | if (PageAnon(page) && PageSwapBacked(page) && |
1125 | !PageSwapCache(page)) { | |
63eb6b93 HD |
1126 | if (!(sc->gfp_mask & __GFP_IO)) |
1127 | goto keep_locked; | |
5bc7b8ac | 1128 | if (!add_to_swap(page, page_list)) |
1da177e4 | 1129 | goto activate_locked; |
63eb6b93 | 1130 | may_enter_fs = 1; |
1da177e4 | 1131 | |
e2be15f6 MG |
1132 | /* Adding to swap updated mapping */ |
1133 | mapping = page_mapping(page); | |
7751b2da KS |
1134 | } else if (unlikely(PageTransHuge(page))) { |
1135 | /* Split file THP */ | |
1136 | if (split_huge_page_to_list(page, page_list)) | |
1137 | goto keep_locked; | |
e2be15f6 | 1138 | } |
1da177e4 | 1139 | |
7751b2da KS |
1140 | VM_BUG_ON_PAGE(PageTransHuge(page), page); |
1141 | ||
1da177e4 LT |
1142 | /* |
1143 | * The page is mapped into the page tables of one or more | |
1144 | * processes. Try to unmap it here. | |
1145 | */ | |
802a3a92 | 1146 | if (page_mapped(page)) { |
666e5a40 | 1147 | if (!try_to_unmap(page, ttu_flags | TTU_BATCH_FLUSH)) { |
5bccd166 | 1148 | nr_unmap_fail++; |
1da177e4 | 1149 | goto activate_locked; |
1da177e4 LT |
1150 | } |
1151 | } | |
1152 | ||
1153 | if (PageDirty(page)) { | |
ee72886d | 1154 | /* |
4eda4823 JW |
1155 | * Only kswapd can writeback filesystem pages |
1156 | * to avoid risk of stack overflow. But avoid | |
1157 | * injecting inefficient single-page IO into | |
1158 | * flusher writeback as much as possible: only | |
1159 | * write pages when we've encountered many | |
1160 | * dirty pages, and when we've already scanned | |
1161 | * the rest of the LRU for clean pages and see | |
1162 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1163 | */ |
f84f6e2b | 1164 | if (page_is_file_cache(page) && |
4eda4823 JW |
1165 | (!current_is_kswapd() || !PageReclaim(page) || |
1166 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1167 | /* |
1168 | * Immediately reclaim when written back. | |
1169 | * Similar in principal to deactivate_page() | |
1170 | * except we already have the page isolated | |
1171 | * and know it's dirty | |
1172 | */ | |
c4a25635 | 1173 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1174 | SetPageReclaim(page); |
1175 | ||
c55e8d03 | 1176 | goto activate_locked; |
ee72886d MG |
1177 | } |
1178 | ||
dfc8d636 | 1179 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1180 | goto keep_locked; |
4dd4b920 | 1181 | if (!may_enter_fs) |
1da177e4 | 1182 | goto keep_locked; |
52a8363e | 1183 | if (!sc->may_writepage) |
1da177e4 LT |
1184 | goto keep_locked; |
1185 | ||
d950c947 MG |
1186 | /* |
1187 | * Page is dirty. Flush the TLB if a writable entry | |
1188 | * potentially exists to avoid CPU writes after IO | |
1189 | * starts and then write it out here. | |
1190 | */ | |
1191 | try_to_unmap_flush_dirty(); | |
7d3579e8 | 1192 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1193 | case PAGE_KEEP: |
1194 | goto keep_locked; | |
1195 | case PAGE_ACTIVATE: | |
1196 | goto activate_locked; | |
1197 | case PAGE_SUCCESS: | |
7d3579e8 | 1198 | if (PageWriteback(page)) |
41ac1999 | 1199 | goto keep; |
7d3579e8 | 1200 | if (PageDirty(page)) |
1da177e4 | 1201 | goto keep; |
7d3579e8 | 1202 | |
1da177e4 LT |
1203 | /* |
1204 | * A synchronous write - probably a ramdisk. Go | |
1205 | * ahead and try to reclaim the page. | |
1206 | */ | |
529ae9aa | 1207 | if (!trylock_page(page)) |
1da177e4 LT |
1208 | goto keep; |
1209 | if (PageDirty(page) || PageWriteback(page)) | |
1210 | goto keep_locked; | |
1211 | mapping = page_mapping(page); | |
1212 | case PAGE_CLEAN: | |
1213 | ; /* try to free the page below */ | |
1214 | } | |
1215 | } | |
1216 | ||
1217 | /* | |
1218 | * If the page has buffers, try to free the buffer mappings | |
1219 | * associated with this page. If we succeed we try to free | |
1220 | * the page as well. | |
1221 | * | |
1222 | * We do this even if the page is PageDirty(). | |
1223 | * try_to_release_page() does not perform I/O, but it is | |
1224 | * possible for a page to have PageDirty set, but it is actually | |
1225 | * clean (all its buffers are clean). This happens if the | |
1226 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1227 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1228 | * try_to_release_page() will discover that cleanness and will |
1229 | * drop the buffers and mark the page clean - it can be freed. | |
1230 | * | |
1231 | * Rarely, pages can have buffers and no ->mapping. These are | |
1232 | * the pages which were not successfully invalidated in | |
1233 | * truncate_complete_page(). We try to drop those buffers here | |
1234 | * and if that worked, and the page is no longer mapped into | |
1235 | * process address space (page_count == 1) it can be freed. | |
1236 | * Otherwise, leave the page on the LRU so it is swappable. | |
1237 | */ | |
266cf658 | 1238 | if (page_has_private(page)) { |
1da177e4 LT |
1239 | if (!try_to_release_page(page, sc->gfp_mask)) |
1240 | goto activate_locked; | |
e286781d NP |
1241 | if (!mapping && page_count(page) == 1) { |
1242 | unlock_page(page); | |
1243 | if (put_page_testzero(page)) | |
1244 | goto free_it; | |
1245 | else { | |
1246 | /* | |
1247 | * rare race with speculative reference. | |
1248 | * the speculative reference will free | |
1249 | * this page shortly, so we may | |
1250 | * increment nr_reclaimed here (and | |
1251 | * leave it off the LRU). | |
1252 | */ | |
1253 | nr_reclaimed++; | |
1254 | continue; | |
1255 | } | |
1256 | } | |
1da177e4 LT |
1257 | } |
1258 | ||
802a3a92 SL |
1259 | if (PageAnon(page) && !PageSwapBacked(page)) { |
1260 | /* follow __remove_mapping for reference */ | |
1261 | if (!page_ref_freeze(page, 1)) | |
1262 | goto keep_locked; | |
1263 | if (PageDirty(page)) { | |
1264 | page_ref_unfreeze(page, 1); | |
1265 | goto keep_locked; | |
1266 | } | |
1da177e4 | 1267 | |
802a3a92 SL |
1268 | count_vm_event(PGLAZYFREED); |
1269 | } else if (!mapping || !__remove_mapping(mapping, page, true)) | |
1270 | goto keep_locked; | |
a978d6f5 NP |
1271 | /* |
1272 | * At this point, we have no other references and there is | |
1273 | * no way to pick any more up (removed from LRU, removed | |
1274 | * from pagecache). Can use non-atomic bitops now (and | |
1275 | * we obviously don't have to worry about waking up a process | |
1276 | * waiting on the page lock, because there are no references. | |
1277 | */ | |
48c935ad | 1278 | __ClearPageLocked(page); |
e286781d | 1279 | free_it: |
05ff5137 | 1280 | nr_reclaimed++; |
abe4c3b5 MG |
1281 | |
1282 | /* | |
1283 | * Is there need to periodically free_page_list? It would | |
1284 | * appear not as the counts should be low | |
1285 | */ | |
1286 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1287 | continue; |
1288 | ||
1289 | activate_locked: | |
68a22394 | 1290 | /* Not a candidate for swapping, so reclaim swap space. */ |
ad6b6704 MK |
1291 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
1292 | PageMlocked(page))) | |
a2c43eed | 1293 | try_to_free_swap(page); |
309381fe | 1294 | VM_BUG_ON_PAGE(PageActive(page), page); |
ad6b6704 MK |
1295 | if (!PageMlocked(page)) { |
1296 | SetPageActive(page); | |
1297 | pgactivate++; | |
1298 | } | |
1da177e4 LT |
1299 | keep_locked: |
1300 | unlock_page(page); | |
1301 | keep: | |
1302 | list_add(&page->lru, &ret_pages); | |
309381fe | 1303 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1304 | } |
abe4c3b5 | 1305 | |
747db954 | 1306 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1307 | try_to_unmap_flush(); |
b745bc85 | 1308 | free_hot_cold_page_list(&free_pages, true); |
abe4c3b5 | 1309 | |
1da177e4 | 1310 | list_splice(&ret_pages, page_list); |
f8891e5e | 1311 | count_vm_events(PGACTIVATE, pgactivate); |
0a31bc97 | 1312 | |
3c710c1a MH |
1313 | if (stat) { |
1314 | stat->nr_dirty = nr_dirty; | |
1315 | stat->nr_congested = nr_congested; | |
1316 | stat->nr_unqueued_dirty = nr_unqueued_dirty; | |
1317 | stat->nr_writeback = nr_writeback; | |
1318 | stat->nr_immediate = nr_immediate; | |
5bccd166 MH |
1319 | stat->nr_activate = pgactivate; |
1320 | stat->nr_ref_keep = nr_ref_keep; | |
1321 | stat->nr_unmap_fail = nr_unmap_fail; | |
3c710c1a | 1322 | } |
05ff5137 | 1323 | return nr_reclaimed; |
1da177e4 LT |
1324 | } |
1325 | ||
02c6de8d MK |
1326 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1327 | struct list_head *page_list) | |
1328 | { | |
1329 | struct scan_control sc = { | |
1330 | .gfp_mask = GFP_KERNEL, | |
1331 | .priority = DEF_PRIORITY, | |
1332 | .may_unmap = 1, | |
1333 | }; | |
3c710c1a | 1334 | unsigned long ret; |
02c6de8d MK |
1335 | struct page *page, *next; |
1336 | LIST_HEAD(clean_pages); | |
1337 | ||
1338 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1339 | if (page_is_file_cache(page) && !PageDirty(page) && |
b1123ea6 | 1340 | !__PageMovable(page)) { |
02c6de8d MK |
1341 | ClearPageActive(page); |
1342 | list_move(&page->lru, &clean_pages); | |
1343 | } | |
1344 | } | |
1345 | ||
599d0c95 | 1346 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
a128ca71 | 1347 | TTU_IGNORE_ACCESS, NULL, true); |
02c6de8d | 1348 | list_splice(&clean_pages, page_list); |
599d0c95 | 1349 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1350 | return ret; |
1351 | } | |
1352 | ||
5ad333eb AW |
1353 | /* |
1354 | * Attempt to remove the specified page from its LRU. Only take this page | |
1355 | * if it is of the appropriate PageActive status. Pages which are being | |
1356 | * freed elsewhere are also ignored. | |
1357 | * | |
1358 | * page: page to consider | |
1359 | * mode: one of the LRU isolation modes defined above | |
1360 | * | |
1361 | * returns 0 on success, -ve errno on failure. | |
1362 | */ | |
f3fd4a61 | 1363 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1364 | { |
1365 | int ret = -EINVAL; | |
1366 | ||
1367 | /* Only take pages on the LRU. */ | |
1368 | if (!PageLRU(page)) | |
1369 | return ret; | |
1370 | ||
e46a2879 MK |
1371 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1372 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1373 | return ret; |
1374 | ||
5ad333eb | 1375 | ret = -EBUSY; |
08e552c6 | 1376 | |
c8244935 MG |
1377 | /* |
1378 | * To minimise LRU disruption, the caller can indicate that it only | |
1379 | * wants to isolate pages it will be able to operate on without | |
1380 | * blocking - clean pages for the most part. | |
1381 | * | |
c8244935 MG |
1382 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1383 | * that it is possible to migrate without blocking | |
1384 | */ | |
1276ad68 | 1385 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
1386 | /* All the caller can do on PageWriteback is block */ |
1387 | if (PageWriteback(page)) | |
1388 | return ret; | |
1389 | ||
1390 | if (PageDirty(page)) { | |
1391 | struct address_space *mapping; | |
1392 | ||
c8244935 MG |
1393 | /* |
1394 | * Only pages without mappings or that have a | |
1395 | * ->migratepage callback are possible to migrate | |
1396 | * without blocking | |
1397 | */ | |
1398 | mapping = page_mapping(page); | |
1399 | if (mapping && !mapping->a_ops->migratepage) | |
1400 | return ret; | |
1401 | } | |
1402 | } | |
39deaf85 | 1403 | |
f80c0673 MK |
1404 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1405 | return ret; | |
1406 | ||
5ad333eb AW |
1407 | if (likely(get_page_unless_zero(page))) { |
1408 | /* | |
1409 | * Be careful not to clear PageLRU until after we're | |
1410 | * sure the page is not being freed elsewhere -- the | |
1411 | * page release code relies on it. | |
1412 | */ | |
1413 | ClearPageLRU(page); | |
1414 | ret = 0; | |
1415 | } | |
1416 | ||
1417 | return ret; | |
1418 | } | |
1419 | ||
7ee36a14 MG |
1420 | |
1421 | /* | |
1422 | * Update LRU sizes after isolating pages. The LRU size updates must | |
1423 | * be complete before mem_cgroup_update_lru_size due to a santity check. | |
1424 | */ | |
1425 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 1426 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 1427 | { |
7ee36a14 MG |
1428 | int zid; |
1429 | ||
7ee36a14 MG |
1430 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1431 | if (!nr_zone_taken[zid]) | |
1432 | continue; | |
1433 | ||
1434 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
7ee36a14 | 1435 | #ifdef CONFIG_MEMCG |
b4536f0c | 1436 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
7ee36a14 | 1437 | #endif |
b4536f0c MH |
1438 | } |
1439 | ||
7ee36a14 MG |
1440 | } |
1441 | ||
1da177e4 | 1442 | /* |
a52633d8 | 1443 | * zone_lru_lock is heavily contended. Some of the functions that |
1da177e4 LT |
1444 | * shrink the lists perform better by taking out a batch of pages |
1445 | * and working on them outside the LRU lock. | |
1446 | * | |
1447 | * For pagecache intensive workloads, this function is the hottest | |
1448 | * spot in the kernel (apart from copy_*_user functions). | |
1449 | * | |
1450 | * Appropriate locks must be held before calling this function. | |
1451 | * | |
1452 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1453 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1454 | * @dst: The temp list to put pages on to. |
f626012d | 1455 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1456 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1457 | * @mode: One of the LRU isolation modes |
3cb99451 | 1458 | * @lru: LRU list id for isolating |
1da177e4 LT |
1459 | * |
1460 | * returns how many pages were moved onto *@dst. | |
1461 | */ | |
69e05944 | 1462 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1463 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1464 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1465 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1466 | { |
75b00af7 | 1467 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1468 | unsigned long nr_taken = 0; |
599d0c95 | 1469 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 1470 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 1471 | unsigned long skipped = 0; |
599d0c95 | 1472 | unsigned long scan, nr_pages; |
b2e18757 | 1473 | LIST_HEAD(pages_skipped); |
1da177e4 | 1474 | |
0b802f10 | 1475 | for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan && |
3db65812 | 1476 | !list_empty(src); scan++) { |
5ad333eb | 1477 | struct page *page; |
5ad333eb | 1478 | |
1da177e4 LT |
1479 | page = lru_to_page(src); |
1480 | prefetchw_prev_lru_page(page, src, flags); | |
1481 | ||
309381fe | 1482 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1483 | |
b2e18757 MG |
1484 | if (page_zonenum(page) > sc->reclaim_idx) { |
1485 | list_move(&page->lru, &pages_skipped); | |
7cc30fcf | 1486 | nr_skipped[page_zonenum(page)]++; |
b2e18757 MG |
1487 | continue; |
1488 | } | |
1489 | ||
f3fd4a61 | 1490 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1491 | case 0: |
599d0c95 MG |
1492 | nr_pages = hpage_nr_pages(page); |
1493 | nr_taken += nr_pages; | |
1494 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1495 | list_move(&page->lru, dst); |
5ad333eb AW |
1496 | break; |
1497 | ||
1498 | case -EBUSY: | |
1499 | /* else it is being freed elsewhere */ | |
1500 | list_move(&page->lru, src); | |
1501 | continue; | |
46453a6e | 1502 | |
5ad333eb AW |
1503 | default: |
1504 | BUG(); | |
1505 | } | |
1da177e4 LT |
1506 | } |
1507 | ||
b2e18757 MG |
1508 | /* |
1509 | * Splice any skipped pages to the start of the LRU list. Note that | |
1510 | * this disrupts the LRU order when reclaiming for lower zones but | |
1511 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1512 | * scanning would soon rescan the same pages to skip and put the | |
1513 | * system at risk of premature OOM. | |
1514 | */ | |
7cc30fcf MG |
1515 | if (!list_empty(&pages_skipped)) { |
1516 | int zid; | |
1517 | ||
3db65812 | 1518 | list_splice(&pages_skipped, src); |
7cc30fcf MG |
1519 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1520 | if (!nr_skipped[zid]) | |
1521 | continue; | |
1522 | ||
1523 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 1524 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
1525 | } |
1526 | } | |
3db65812 | 1527 | *nr_scanned = scan; |
1265e3a6 | 1528 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
32b3f297 | 1529 | scan, skipped, nr_taken, mode, lru); |
b4536f0c | 1530 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
1531 | return nr_taken; |
1532 | } | |
1533 | ||
62695a84 NP |
1534 | /** |
1535 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1536 | * @page: page to isolate from its LRU list | |
1537 | * | |
1538 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1539 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1540 | * | |
1541 | * Returns 0 if the page was removed from an LRU list. | |
1542 | * Returns -EBUSY if the page was not on an LRU list. | |
1543 | * | |
1544 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1545 | * the active list, it will have PageActive set. If it was found on |
1546 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1547 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1548 | * |
1549 | * The vmstat statistic corresponding to the list on which the page was | |
1550 | * found will be decremented. | |
1551 | * | |
1552 | * Restrictions: | |
1553 | * (1) Must be called with an elevated refcount on the page. This is a | |
1554 | * fundamentnal difference from isolate_lru_pages (which is called | |
1555 | * without a stable reference). | |
1556 | * (2) the lru_lock must not be held. | |
1557 | * (3) interrupts must be enabled. | |
1558 | */ | |
1559 | int isolate_lru_page(struct page *page) | |
1560 | { | |
1561 | int ret = -EBUSY; | |
1562 | ||
309381fe | 1563 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1564 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1565 | |
62695a84 NP |
1566 | if (PageLRU(page)) { |
1567 | struct zone *zone = page_zone(page); | |
fa9add64 | 1568 | struct lruvec *lruvec; |
62695a84 | 1569 | |
a52633d8 | 1570 | spin_lock_irq(zone_lru_lock(zone)); |
599d0c95 | 1571 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
0c917313 | 1572 | if (PageLRU(page)) { |
894bc310 | 1573 | int lru = page_lru(page); |
0c917313 | 1574 | get_page(page); |
62695a84 | 1575 | ClearPageLRU(page); |
fa9add64 HD |
1576 | del_page_from_lru_list(page, lruvec, lru); |
1577 | ret = 0; | |
62695a84 | 1578 | } |
a52633d8 | 1579 | spin_unlock_irq(zone_lru_lock(zone)); |
62695a84 NP |
1580 | } |
1581 | return ret; | |
1582 | } | |
1583 | ||
35cd7815 | 1584 | /* |
d37dd5dc FW |
1585 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1586 | * then get resheduled. When there are massive number of tasks doing page | |
1587 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1588 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1589 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1590 | */ |
599d0c95 | 1591 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1592 | struct scan_control *sc) |
1593 | { | |
1594 | unsigned long inactive, isolated; | |
1595 | ||
1596 | if (current_is_kswapd()) | |
1597 | return 0; | |
1598 | ||
97c9341f | 1599 | if (!sane_reclaim(sc)) |
35cd7815 RR |
1600 | return 0; |
1601 | ||
1602 | if (file) { | |
599d0c95 MG |
1603 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1604 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1605 | } else { |
599d0c95 MG |
1606 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1607 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1608 | } |
1609 | ||
3cf23841 FW |
1610 | /* |
1611 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1612 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1613 | * deadlock. | |
1614 | */ | |
d0164adc | 1615 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1616 | inactive >>= 3; |
1617 | ||
35cd7815 RR |
1618 | return isolated > inactive; |
1619 | } | |
1620 | ||
66635629 | 1621 | static noinline_for_stack void |
75b00af7 | 1622 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1623 | { |
27ac81d8 | 1624 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
599d0c95 | 1625 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3f79768f | 1626 | LIST_HEAD(pages_to_free); |
66635629 | 1627 | |
66635629 MG |
1628 | /* |
1629 | * Put back any unfreeable pages. | |
1630 | */ | |
66635629 | 1631 | while (!list_empty(page_list)) { |
3f79768f | 1632 | struct page *page = lru_to_page(page_list); |
66635629 | 1633 | int lru; |
3f79768f | 1634 | |
309381fe | 1635 | VM_BUG_ON_PAGE(PageLRU(page), page); |
66635629 | 1636 | list_del(&page->lru); |
39b5f29a | 1637 | if (unlikely(!page_evictable(page))) { |
599d0c95 | 1638 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1639 | putback_lru_page(page); |
599d0c95 | 1640 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1641 | continue; |
1642 | } | |
fa9add64 | 1643 | |
599d0c95 | 1644 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1645 | |
7a608572 | 1646 | SetPageLRU(page); |
66635629 | 1647 | lru = page_lru(page); |
fa9add64 HD |
1648 | add_page_to_lru_list(page, lruvec, lru); |
1649 | ||
66635629 MG |
1650 | if (is_active_lru(lru)) { |
1651 | int file = is_file_lru(lru); | |
9992af10 RR |
1652 | int numpages = hpage_nr_pages(page); |
1653 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1654 | } |
2bcf8879 HD |
1655 | if (put_page_testzero(page)) { |
1656 | __ClearPageLRU(page); | |
1657 | __ClearPageActive(page); | |
fa9add64 | 1658 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1659 | |
1660 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1661 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1662 | mem_cgroup_uncharge(page); |
2bcf8879 | 1663 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1664 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1665 | } else |
1666 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1667 | } |
1668 | } | |
66635629 | 1669 | |
3f79768f HD |
1670 | /* |
1671 | * To save our caller's stack, now use input list for pages to free. | |
1672 | */ | |
1673 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1674 | } |
1675 | ||
399ba0b9 N |
1676 | /* |
1677 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1678 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1679 | * In that case we should only throttle if the backing device it is | |
1680 | * writing to is congested. In other cases it is safe to throttle. | |
1681 | */ | |
1682 | static int current_may_throttle(void) | |
1683 | { | |
1684 | return !(current->flags & PF_LESS_THROTTLE) || | |
1685 | current->backing_dev_info == NULL || | |
1686 | bdi_write_congested(current->backing_dev_info); | |
1687 | } | |
1688 | ||
1da177e4 | 1689 | /* |
b2e18757 | 1690 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1691 | * of reclaimed pages |
1da177e4 | 1692 | */ |
66635629 | 1693 | static noinline_for_stack unsigned long |
1a93be0e | 1694 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1695 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1696 | { |
1697 | LIST_HEAD(page_list); | |
e247dbce | 1698 | unsigned long nr_scanned; |
05ff5137 | 1699 | unsigned long nr_reclaimed = 0; |
e247dbce | 1700 | unsigned long nr_taken; |
3c710c1a | 1701 | struct reclaim_stat stat = {}; |
f3fd4a61 | 1702 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1703 | int file = is_file_lru(lru); |
599d0c95 | 1704 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1705 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
78dc583d | 1706 | |
599d0c95 | 1707 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
58355c78 | 1708 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1709 | |
1710 | /* We are about to die and free our memory. Return now. */ | |
1711 | if (fatal_signal_pending(current)) | |
1712 | return SWAP_CLUSTER_MAX; | |
1713 | } | |
1714 | ||
1da177e4 | 1715 | lru_add_drain(); |
f80c0673 MK |
1716 | |
1717 | if (!sc->may_unmap) | |
61317289 | 1718 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1719 | |
599d0c95 | 1720 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1721 | |
5dc35979 KK |
1722 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1723 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc | 1724 | |
599d0c95 | 1725 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1726 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1727 | |
89b5fae5 | 1728 | if (global_reclaim(sc)) { |
e247dbce | 1729 | if (current_is_kswapd()) |
599d0c95 | 1730 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
e247dbce | 1731 | else |
599d0c95 | 1732 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
e247dbce | 1733 | } |
599d0c95 | 1734 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1735 | |
d563c050 | 1736 | if (nr_taken == 0) |
66635629 | 1737 | return 0; |
5ad333eb | 1738 | |
a128ca71 | 1739 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, |
3c710c1a | 1740 | &stat, false); |
c661b078 | 1741 | |
599d0c95 | 1742 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 1743 | |
904249aa YH |
1744 | if (global_reclaim(sc)) { |
1745 | if (current_is_kswapd()) | |
599d0c95 | 1746 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
904249aa | 1747 | else |
599d0c95 | 1748 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
904249aa | 1749 | } |
a74609fa | 1750 | |
27ac81d8 | 1751 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1752 | |
599d0c95 | 1753 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 1754 | |
599d0c95 | 1755 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 1756 | |
747db954 | 1757 | mem_cgroup_uncharge_list(&page_list); |
b745bc85 | 1758 | free_hot_cold_page_list(&page_list, true); |
e11da5b4 | 1759 | |
92df3a72 MG |
1760 | /* |
1761 | * If reclaim is isolating dirty pages under writeback, it implies | |
1762 | * that the long-lived page allocation rate is exceeding the page | |
1763 | * laundering rate. Either the global limits are not being effective | |
1764 | * at throttling processes due to the page distribution throughout | |
1765 | * zones or there is heavy usage of a slow backing device. The | |
1766 | * only option is to throttle from reclaim context which is not ideal | |
1767 | * as there is no guarantee the dirtying process is throttled in the | |
1768 | * same way balance_dirty_pages() manages. | |
1769 | * | |
8e950282 MG |
1770 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1771 | * of pages under pages flagged for immediate reclaim and stall if any | |
1772 | * are encountered in the nr_immediate check below. | |
92df3a72 | 1773 | */ |
3c710c1a | 1774 | if (stat.nr_writeback && stat.nr_writeback == nr_taken) |
599d0c95 | 1775 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); |
92df3a72 | 1776 | |
d43006d5 | 1777 | /* |
97c9341f TH |
1778 | * Legacy memcg will stall in page writeback so avoid forcibly |
1779 | * stalling here. | |
d43006d5 | 1780 | */ |
97c9341f | 1781 | if (sane_reclaim(sc)) { |
8e950282 MG |
1782 | /* |
1783 | * Tag a zone as congested if all the dirty pages scanned were | |
1784 | * backed by a congested BDI and wait_iff_congested will stall. | |
1785 | */ | |
3c710c1a | 1786 | if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested) |
599d0c95 | 1787 | set_bit(PGDAT_CONGESTED, &pgdat->flags); |
8e950282 | 1788 | |
b1a6f21e MG |
1789 | /* |
1790 | * If dirty pages are scanned that are not queued for IO, it | |
726d061f JW |
1791 | * implies that flushers are not doing their job. This can |
1792 | * happen when memory pressure pushes dirty pages to the end of | |
1793 | * the LRU before the dirty limits are breached and the dirty | |
1794 | * data has expired. It can also happen when the proportion of | |
1795 | * dirty pages grows not through writes but through memory | |
1796 | * pressure reclaiming all the clean cache. And in some cases, | |
1797 | * the flushers simply cannot keep up with the allocation | |
1798 | * rate. Nudge the flusher threads in case they are asleep, but | |
1799 | * also allow kswapd to start writing pages during reclaim. | |
b1a6f21e | 1800 | */ |
726d061f JW |
1801 | if (stat.nr_unqueued_dirty == nr_taken) { |
1802 | wakeup_flusher_threads(0, WB_REASON_VMSCAN); | |
599d0c95 | 1803 | set_bit(PGDAT_DIRTY, &pgdat->flags); |
726d061f | 1804 | } |
b1a6f21e MG |
1805 | |
1806 | /* | |
b738d764 LT |
1807 | * If kswapd scans pages marked marked for immediate |
1808 | * reclaim and under writeback (nr_immediate), it implies | |
1809 | * that pages are cycling through the LRU faster than | |
b1a6f21e MG |
1810 | * they are written so also forcibly stall. |
1811 | */ | |
3c710c1a | 1812 | if (stat.nr_immediate && current_may_throttle()) |
b1a6f21e | 1813 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
e2be15f6 | 1814 | } |
d43006d5 | 1815 | |
8e950282 MG |
1816 | /* |
1817 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1818 | * is congested. Allow kswapd to continue until it starts encountering | |
1819 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1820 | */ | |
399ba0b9 N |
1821 | if (!sc->hibernation_mode && !current_is_kswapd() && |
1822 | current_may_throttle()) | |
599d0c95 | 1823 | wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); |
8e950282 | 1824 | |
599d0c95 MG |
1825 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
1826 | nr_scanned, nr_reclaimed, | |
5bccd166 MH |
1827 | stat.nr_dirty, stat.nr_writeback, |
1828 | stat.nr_congested, stat.nr_immediate, | |
1829 | stat.nr_activate, stat.nr_ref_keep, | |
1830 | stat.nr_unmap_fail, | |
ba5e9579 | 1831 | sc->priority, file); |
05ff5137 | 1832 | return nr_reclaimed; |
1da177e4 LT |
1833 | } |
1834 | ||
1835 | /* | |
1836 | * This moves pages from the active list to the inactive list. | |
1837 | * | |
1838 | * We move them the other way if the page is referenced by one or more | |
1839 | * processes, from rmap. | |
1840 | * | |
1841 | * If the pages are mostly unmapped, the processing is fast and it is | |
a52633d8 | 1842 | * appropriate to hold zone_lru_lock across the whole operation. But if |
1da177e4 | 1843 | * the pages are mapped, the processing is slow (page_referenced()) so we |
a52633d8 | 1844 | * should drop zone_lru_lock around each page. It's impossible to balance |
1da177e4 LT |
1845 | * this, so instead we remove the pages from the LRU while processing them. |
1846 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1847 | * nobody will play with that bit on a non-LRU page. | |
1848 | * | |
0139aa7b | 1849 | * The downside is that we have to touch page->_refcount against each page. |
1da177e4 | 1850 | * But we had to alter page->flags anyway. |
9d998b4f MH |
1851 | * |
1852 | * Returns the number of pages moved to the given lru. | |
1da177e4 | 1853 | */ |
1cfb419b | 1854 | |
9d998b4f | 1855 | static unsigned move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1856 | struct list_head *list, |
2bcf8879 | 1857 | struct list_head *pages_to_free, |
3eb4140f WF |
1858 | enum lru_list lru) |
1859 | { | |
599d0c95 | 1860 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3eb4140f | 1861 | struct page *page; |
fa9add64 | 1862 | int nr_pages; |
9d998b4f | 1863 | int nr_moved = 0; |
3eb4140f | 1864 | |
3eb4140f WF |
1865 | while (!list_empty(list)) { |
1866 | page = lru_to_page(list); | |
599d0c95 | 1867 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
3eb4140f | 1868 | |
309381fe | 1869 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3eb4140f WF |
1870 | SetPageLRU(page); |
1871 | ||
fa9add64 | 1872 | nr_pages = hpage_nr_pages(page); |
599d0c95 | 1873 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
925b7673 | 1874 | list_move(&page->lru, &lruvec->lists[lru]); |
3eb4140f | 1875 | |
2bcf8879 HD |
1876 | if (put_page_testzero(page)) { |
1877 | __ClearPageLRU(page); | |
1878 | __ClearPageActive(page); | |
fa9add64 | 1879 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1880 | |
1881 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1882 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1883 | mem_cgroup_uncharge(page); |
2bcf8879 | 1884 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1885 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1886 | } else |
1887 | list_add(&page->lru, pages_to_free); | |
9d998b4f MH |
1888 | } else { |
1889 | nr_moved += nr_pages; | |
3eb4140f WF |
1890 | } |
1891 | } | |
9d5e6a9f | 1892 | |
3eb4140f | 1893 | if (!is_active_lru(lru)) |
f0958906 | 1894 | __count_vm_events(PGDEACTIVATE, nr_moved); |
9d998b4f MH |
1895 | |
1896 | return nr_moved; | |
3eb4140f | 1897 | } |
1cfb419b | 1898 | |
f626012d | 1899 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1900 | struct lruvec *lruvec, |
f16015fb | 1901 | struct scan_control *sc, |
9e3b2f8c | 1902 | enum lru_list lru) |
1da177e4 | 1903 | { |
44c241f1 | 1904 | unsigned long nr_taken; |
f626012d | 1905 | unsigned long nr_scanned; |
6fe6b7e3 | 1906 | unsigned long vm_flags; |
1da177e4 | 1907 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1908 | LIST_HEAD(l_active); |
b69408e8 | 1909 | LIST_HEAD(l_inactive); |
1da177e4 | 1910 | struct page *page; |
1a93be0e | 1911 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
9d998b4f MH |
1912 | unsigned nr_deactivate, nr_activate; |
1913 | unsigned nr_rotated = 0; | |
f3fd4a61 | 1914 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1915 | int file = is_file_lru(lru); |
599d0c95 | 1916 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
1917 | |
1918 | lru_add_drain(); | |
f80c0673 MK |
1919 | |
1920 | if (!sc->may_unmap) | |
61317289 | 1921 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1922 | |
599d0c95 | 1923 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 1924 | |
5dc35979 KK |
1925 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1926 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1927 | |
599d0c95 | 1928 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 1929 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1930 | |
599d0c95 | 1931 | __count_vm_events(PGREFILL, nr_scanned); |
9d5e6a9f | 1932 | |
599d0c95 | 1933 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 1934 | |
1da177e4 LT |
1935 | while (!list_empty(&l_hold)) { |
1936 | cond_resched(); | |
1937 | page = lru_to_page(&l_hold); | |
1938 | list_del(&page->lru); | |
7e9cd484 | 1939 | |
39b5f29a | 1940 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
1941 | putback_lru_page(page); |
1942 | continue; | |
1943 | } | |
1944 | ||
cc715d99 MG |
1945 | if (unlikely(buffer_heads_over_limit)) { |
1946 | if (page_has_private(page) && trylock_page(page)) { | |
1947 | if (page_has_private(page)) | |
1948 | try_to_release_page(page, 0); | |
1949 | unlock_page(page); | |
1950 | } | |
1951 | } | |
1952 | ||
c3ac9a8a JW |
1953 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1954 | &vm_flags)) { | |
9992af10 | 1955 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1956 | /* |
1957 | * Identify referenced, file-backed active pages and | |
1958 | * give them one more trip around the active list. So | |
1959 | * that executable code get better chances to stay in | |
1960 | * memory under moderate memory pressure. Anon pages | |
1961 | * are not likely to be evicted by use-once streaming | |
1962 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1963 | * so we ignore them here. | |
1964 | */ | |
41e20983 | 1965 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1966 | list_add(&page->lru, &l_active); |
1967 | continue; | |
1968 | } | |
1969 | } | |
7e9cd484 | 1970 | |
5205e56e | 1971 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1972 | list_add(&page->lru, &l_inactive); |
1973 | } | |
1974 | ||
b555749a | 1975 | /* |
8cab4754 | 1976 | * Move pages back to the lru list. |
b555749a | 1977 | */ |
599d0c95 | 1978 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 1979 | /* |
8cab4754 WF |
1980 | * Count referenced pages from currently used mappings as rotated, |
1981 | * even though only some of them are actually re-activated. This | |
1982 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 1983 | * get_scan_count. |
7e9cd484 | 1984 | */ |
b7c46d15 | 1985 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1986 | |
9d998b4f MH |
1987 | nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
1988 | nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
599d0c95 MG |
1989 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
1990 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 1991 | |
747db954 | 1992 | mem_cgroup_uncharge_list(&l_hold); |
b745bc85 | 1993 | free_hot_cold_page_list(&l_hold, true); |
9d998b4f MH |
1994 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
1995 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
1996 | } |
1997 | ||
59dc76b0 RR |
1998 | /* |
1999 | * The inactive anon list should be small enough that the VM never has | |
2000 | * to do too much work. | |
14797e23 | 2001 | * |
59dc76b0 RR |
2002 | * The inactive file list should be small enough to leave most memory |
2003 | * to the established workingset on the scan-resistant active list, | |
2004 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2005 | * |
59dc76b0 RR |
2006 | * Both inactive lists should also be large enough that each inactive |
2007 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2008 | * |
2a2e4885 JW |
2009 | * If that fails and refaulting is observed, the inactive list grows. |
2010 | * | |
59dc76b0 RR |
2011 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
2012 | * on this LRU, maintained by the pageout code. A zone->inactive_ratio | |
2013 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. | |
56e49d21 | 2014 | * |
59dc76b0 RR |
2015 | * total target max |
2016 | * memory ratio inactive | |
2017 | * ------------------------------------- | |
2018 | * 10MB 1 5MB | |
2019 | * 100MB 1 50MB | |
2020 | * 1GB 3 250MB | |
2021 | * 10GB 10 0.9GB | |
2022 | * 100GB 31 3GB | |
2023 | * 1TB 101 10GB | |
2024 | * 10TB 320 32GB | |
56e49d21 | 2025 | */ |
f8d1a311 | 2026 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
2a2e4885 JW |
2027 | struct mem_cgroup *memcg, |
2028 | struct scan_control *sc, bool actual_reclaim) | |
56e49d21 | 2029 | { |
fd538803 | 2030 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; |
2a2e4885 JW |
2031 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
2032 | enum lru_list inactive_lru = file * LRU_FILE; | |
2033 | unsigned long inactive, active; | |
2034 | unsigned long inactive_ratio; | |
2035 | unsigned long refaults; | |
59dc76b0 | 2036 | unsigned long gb; |
e3790144 | 2037 | |
59dc76b0 RR |
2038 | /* |
2039 | * If we don't have swap space, anonymous page deactivation | |
2040 | * is pointless. | |
2041 | */ | |
2042 | if (!file && !total_swap_pages) | |
2043 | return false; | |
56e49d21 | 2044 | |
fd538803 MH |
2045 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
2046 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); | |
f8d1a311 | 2047 | |
2a2e4885 | 2048 | if (memcg) |
ccda7f43 | 2049 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); |
b39415b2 | 2050 | else |
2a2e4885 JW |
2051 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); |
2052 | ||
2053 | /* | |
2054 | * When refaults are being observed, it means a new workingset | |
2055 | * is being established. Disable active list protection to get | |
2056 | * rid of the stale workingset quickly. | |
2057 | */ | |
2058 | if (file && actual_reclaim && lruvec->refaults != refaults) { | |
2059 | inactive_ratio = 0; | |
2060 | } else { | |
2061 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
2062 | if (gb) | |
2063 | inactive_ratio = int_sqrt(10 * gb); | |
2064 | else | |
2065 | inactive_ratio = 1; | |
2066 | } | |
59dc76b0 | 2067 | |
2a2e4885 JW |
2068 | if (actual_reclaim) |
2069 | trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, | |
2070 | lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, | |
2071 | lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, | |
2072 | inactive_ratio, file); | |
fd538803 | 2073 | |
59dc76b0 | 2074 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2075 | } |
2076 | ||
4f98a2fe | 2077 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
2a2e4885 JW |
2078 | struct lruvec *lruvec, struct mem_cgroup *memcg, |
2079 | struct scan_control *sc) | |
b69408e8 | 2080 | { |
b39415b2 | 2081 | if (is_active_lru(lru)) { |
2a2e4885 JW |
2082 | if (inactive_list_is_low(lruvec, is_file_lru(lru), |
2083 | memcg, sc, true)) | |
1a93be0e | 2084 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
2085 | return 0; |
2086 | } | |
2087 | ||
1a93be0e | 2088 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
2089 | } |
2090 | ||
9a265114 JW |
2091 | enum scan_balance { |
2092 | SCAN_EQUAL, | |
2093 | SCAN_FRACT, | |
2094 | SCAN_ANON, | |
2095 | SCAN_FILE, | |
2096 | }; | |
2097 | ||
4f98a2fe RR |
2098 | /* |
2099 | * Determine how aggressively the anon and file LRU lists should be | |
2100 | * scanned. The relative value of each set of LRU lists is determined | |
2101 | * by looking at the fraction of the pages scanned we did rotate back | |
2102 | * onto the active list instead of evict. | |
2103 | * | |
be7bd59d WL |
2104 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2105 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2106 | */ |
33377678 | 2107 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
6b4f7799 JW |
2108 | struct scan_control *sc, unsigned long *nr, |
2109 | unsigned long *lru_pages) | |
4f98a2fe | 2110 | { |
33377678 | 2111 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
2112 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2113 | u64 fraction[2]; | |
2114 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 2115 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 2116 | unsigned long anon_prio, file_prio; |
9a265114 | 2117 | enum scan_balance scan_balance; |
0bf1457f | 2118 | unsigned long anon, file; |
4f98a2fe | 2119 | unsigned long ap, fp; |
4111304d | 2120 | enum lru_list lru; |
76a33fc3 SL |
2121 | |
2122 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2123 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2124 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2125 | goto out; |
2126 | } | |
4f98a2fe | 2127 | |
10316b31 JW |
2128 | /* |
2129 | * Global reclaim will swap to prevent OOM even with no | |
2130 | * swappiness, but memcg users want to use this knob to | |
2131 | * disable swapping for individual groups completely when | |
2132 | * using the memory controller's swap limit feature would be | |
2133 | * too expensive. | |
2134 | */ | |
02695175 | 2135 | if (!global_reclaim(sc) && !swappiness) { |
9a265114 | 2136 | scan_balance = SCAN_FILE; |
10316b31 JW |
2137 | goto out; |
2138 | } | |
2139 | ||
2140 | /* | |
2141 | * Do not apply any pressure balancing cleverness when the | |
2142 | * system is close to OOM, scan both anon and file equally | |
2143 | * (unless the swappiness setting disagrees with swapping). | |
2144 | */ | |
02695175 | 2145 | if (!sc->priority && swappiness) { |
9a265114 | 2146 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2147 | goto out; |
2148 | } | |
2149 | ||
62376251 JW |
2150 | /* |
2151 | * Prevent the reclaimer from falling into the cache trap: as | |
2152 | * cache pages start out inactive, every cache fault will tip | |
2153 | * the scan balance towards the file LRU. And as the file LRU | |
2154 | * shrinks, so does the window for rotation from references. | |
2155 | * This means we have a runaway feedback loop where a tiny | |
2156 | * thrashing file LRU becomes infinitely more attractive than | |
2157 | * anon pages. Try to detect this based on file LRU size. | |
2158 | */ | |
2159 | if (global_reclaim(sc)) { | |
599d0c95 MG |
2160 | unsigned long pgdatfile; |
2161 | unsigned long pgdatfree; | |
2162 | int z; | |
2163 | unsigned long total_high_wmark = 0; | |
2ab051e1 | 2164 | |
599d0c95 MG |
2165 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2166 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2167 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2168 | ||
2169 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2170 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2171 | if (!managed_zone(zone)) |
599d0c95 MG |
2172 | continue; |
2173 | ||
2174 | total_high_wmark += high_wmark_pages(zone); | |
2175 | } | |
62376251 | 2176 | |
599d0c95 | 2177 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
62376251 JW |
2178 | scan_balance = SCAN_ANON; |
2179 | goto out; | |
2180 | } | |
2181 | } | |
2182 | ||
7c5bd705 | 2183 | /* |
316bda0e VD |
2184 | * If there is enough inactive page cache, i.e. if the size of the |
2185 | * inactive list is greater than that of the active list *and* the | |
2186 | * inactive list actually has some pages to scan on this priority, we | |
2187 | * do not reclaim anything from the anonymous working set right now. | |
2188 | * Without the second condition we could end up never scanning an | |
2189 | * lruvec even if it has plenty of old anonymous pages unless the | |
2190 | * system is under heavy pressure. | |
7c5bd705 | 2191 | */ |
2a2e4885 | 2192 | if (!inactive_list_is_low(lruvec, true, memcg, sc, false) && |
71ab6cfe | 2193 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
9a265114 | 2194 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2195 | goto out; |
2196 | } | |
2197 | ||
9a265114 JW |
2198 | scan_balance = SCAN_FRACT; |
2199 | ||
58c37f6e KM |
2200 | /* |
2201 | * With swappiness at 100, anonymous and file have the same priority. | |
2202 | * This scanning priority is essentially the inverse of IO cost. | |
2203 | */ | |
02695175 | 2204 | anon_prio = swappiness; |
75b00af7 | 2205 | file_prio = 200 - anon_prio; |
58c37f6e | 2206 | |
4f98a2fe RR |
2207 | /* |
2208 | * OK, so we have swap space and a fair amount of page cache | |
2209 | * pages. We use the recently rotated / recently scanned | |
2210 | * ratios to determine how valuable each cache is. | |
2211 | * | |
2212 | * Because workloads change over time (and to avoid overflow) | |
2213 | * we keep these statistics as a floating average, which ends | |
2214 | * up weighing recent references more than old ones. | |
2215 | * | |
2216 | * anon in [0], file in [1] | |
2217 | */ | |
2ab051e1 | 2218 | |
fd538803 MH |
2219 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
2220 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); | |
2221 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + | |
2222 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); | |
2ab051e1 | 2223 | |
599d0c95 | 2224 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2225 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2226 | reclaim_stat->recent_scanned[0] /= 2; |
2227 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2228 | } |
2229 | ||
6e901571 | 2230 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2231 | reclaim_stat->recent_scanned[1] /= 2; |
2232 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2233 | } |
2234 | ||
4f98a2fe | 2235 | /* |
00d8089c RR |
2236 | * The amount of pressure on anon vs file pages is inversely |
2237 | * proportional to the fraction of recently scanned pages on | |
2238 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2239 | */ |
fe35004f | 2240 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2241 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2242 | |
fe35004f | 2243 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2244 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2245 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2246 | |
76a33fc3 SL |
2247 | fraction[0] = ap; |
2248 | fraction[1] = fp; | |
2249 | denominator = ap + fp + 1; | |
2250 | out: | |
688035f7 JW |
2251 | *lru_pages = 0; |
2252 | for_each_evictable_lru(lru) { | |
2253 | int file = is_file_lru(lru); | |
2254 | unsigned long size; | |
2255 | unsigned long scan; | |
6b4f7799 | 2256 | |
688035f7 JW |
2257 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); |
2258 | scan = size >> sc->priority; | |
2259 | /* | |
2260 | * If the cgroup's already been deleted, make sure to | |
2261 | * scrape out the remaining cache. | |
2262 | */ | |
2263 | if (!scan && !mem_cgroup_online(memcg)) | |
2264 | scan = min(size, SWAP_CLUSTER_MAX); | |
6b4f7799 | 2265 | |
688035f7 JW |
2266 | switch (scan_balance) { |
2267 | case SCAN_EQUAL: | |
2268 | /* Scan lists relative to size */ | |
2269 | break; | |
2270 | case SCAN_FRACT: | |
9a265114 | 2271 | /* |
688035f7 JW |
2272 | * Scan types proportional to swappiness and |
2273 | * their relative recent reclaim efficiency. | |
9a265114 | 2274 | */ |
688035f7 JW |
2275 | scan = div64_u64(scan * fraction[file], |
2276 | denominator); | |
2277 | break; | |
2278 | case SCAN_FILE: | |
2279 | case SCAN_ANON: | |
2280 | /* Scan one type exclusively */ | |
2281 | if ((scan_balance == SCAN_FILE) != file) { | |
2282 | size = 0; | |
2283 | scan = 0; | |
2284 | } | |
2285 | break; | |
2286 | default: | |
2287 | /* Look ma, no brain */ | |
2288 | BUG(); | |
9a265114 | 2289 | } |
688035f7 JW |
2290 | |
2291 | *lru_pages += size; | |
2292 | nr[lru] = scan; | |
76a33fc3 | 2293 | } |
6e08a369 | 2294 | } |
4f98a2fe | 2295 | |
9b4f98cd | 2296 | /* |
a9dd0a83 | 2297 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. |
9b4f98cd | 2298 | */ |
a9dd0a83 | 2299 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, |
33377678 | 2300 | struct scan_control *sc, unsigned long *lru_pages) |
9b4f98cd | 2301 | { |
ef8f2327 | 2302 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
9b4f98cd | 2303 | unsigned long nr[NR_LRU_LISTS]; |
e82e0561 | 2304 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2305 | unsigned long nr_to_scan; |
2306 | enum lru_list lru; | |
2307 | unsigned long nr_reclaimed = 0; | |
2308 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2309 | struct blk_plug plug; | |
1a501907 | 2310 | bool scan_adjusted; |
9b4f98cd | 2311 | |
33377678 | 2312 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
9b4f98cd | 2313 | |
e82e0561 MG |
2314 | /* Record the original scan target for proportional adjustments later */ |
2315 | memcpy(targets, nr, sizeof(nr)); | |
2316 | ||
1a501907 MG |
2317 | /* |
2318 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2319 | * event that can occur when there is little memory pressure e.g. | |
2320 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2321 | * when the requested number of pages are reclaimed when scanning at | |
2322 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2323 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2324 | * do a batch of work at once. For memcg reclaim one check is made to | |
2325 | * abort proportional reclaim if either the file or anon lru has already | |
2326 | * dropped to zero at the first pass. | |
2327 | */ | |
2328 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2329 | sc->priority == DEF_PRIORITY); | |
2330 | ||
9b4f98cd JW |
2331 | blk_start_plug(&plug); |
2332 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2333 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2334 | unsigned long nr_anon, nr_file, percentage; |
2335 | unsigned long nr_scanned; | |
2336 | ||
9b4f98cd JW |
2337 | for_each_evictable_lru(lru) { |
2338 | if (nr[lru]) { | |
2339 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2340 | nr[lru] -= nr_to_scan; | |
2341 | ||
2342 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2a2e4885 | 2343 | lruvec, memcg, sc); |
9b4f98cd JW |
2344 | } |
2345 | } | |
e82e0561 | 2346 | |
bd041733 MH |
2347 | cond_resched(); |
2348 | ||
e82e0561 MG |
2349 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2350 | continue; | |
2351 | ||
e82e0561 MG |
2352 | /* |
2353 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2354 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2355 | * proportionally what was requested by get_scan_count(). We |
2356 | * stop reclaiming one LRU and reduce the amount scanning | |
2357 | * proportional to the original scan target. | |
2358 | */ | |
2359 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2360 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2361 | ||
1a501907 MG |
2362 | /* |
2363 | * It's just vindictive to attack the larger once the smaller | |
2364 | * has gone to zero. And given the way we stop scanning the | |
2365 | * smaller below, this makes sure that we only make one nudge | |
2366 | * towards proportionality once we've got nr_to_reclaim. | |
2367 | */ | |
2368 | if (!nr_file || !nr_anon) | |
2369 | break; | |
2370 | ||
e82e0561 MG |
2371 | if (nr_file > nr_anon) { |
2372 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2373 | targets[LRU_ACTIVE_ANON] + 1; | |
2374 | lru = LRU_BASE; | |
2375 | percentage = nr_anon * 100 / scan_target; | |
2376 | } else { | |
2377 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2378 | targets[LRU_ACTIVE_FILE] + 1; | |
2379 | lru = LRU_FILE; | |
2380 | percentage = nr_file * 100 / scan_target; | |
2381 | } | |
2382 | ||
2383 | /* Stop scanning the smaller of the LRU */ | |
2384 | nr[lru] = 0; | |
2385 | nr[lru + LRU_ACTIVE] = 0; | |
2386 | ||
2387 | /* | |
2388 | * Recalculate the other LRU scan count based on its original | |
2389 | * scan target and the percentage scanning already complete | |
2390 | */ | |
2391 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2392 | nr_scanned = targets[lru] - nr[lru]; | |
2393 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2394 | nr[lru] -= min(nr[lru], nr_scanned); | |
2395 | ||
2396 | lru += LRU_ACTIVE; | |
2397 | nr_scanned = targets[lru] - nr[lru]; | |
2398 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2399 | nr[lru] -= min(nr[lru], nr_scanned); | |
2400 | ||
2401 | scan_adjusted = true; | |
9b4f98cd JW |
2402 | } |
2403 | blk_finish_plug(&plug); | |
2404 | sc->nr_reclaimed += nr_reclaimed; | |
2405 | ||
2406 | /* | |
2407 | * Even if we did not try to evict anon pages at all, we want to | |
2408 | * rebalance the anon lru active/inactive ratio. | |
2409 | */ | |
2a2e4885 | 2410 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) |
9b4f98cd JW |
2411 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2412 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
2413 | } |
2414 | ||
23b9da55 | 2415 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2416 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2417 | { |
d84da3f9 | 2418 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2419 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2420 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2421 | return true; |
2422 | ||
2423 | return false; | |
2424 | } | |
2425 | ||
3e7d3449 | 2426 | /* |
23b9da55 MG |
2427 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2428 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2429 | * true if more pages should be reclaimed such that when the page allocator | |
2430 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2431 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2432 | */ |
a9dd0a83 | 2433 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 MG |
2434 | unsigned long nr_reclaimed, |
2435 | unsigned long nr_scanned, | |
2436 | struct scan_control *sc) | |
2437 | { | |
2438 | unsigned long pages_for_compaction; | |
2439 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2440 | int z; |
3e7d3449 MG |
2441 | |
2442 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2443 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2444 | return false; |
2445 | ||
2876592f MG |
2446 | /* Consider stopping depending on scan and reclaim activity */ |
2447 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2448 | /* | |
2449 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2450 | * full LRU list has been scanned and we are still failing | |
2451 | * to reclaim pages. This full LRU scan is potentially | |
2452 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2453 | */ | |
2454 | if (!nr_reclaimed && !nr_scanned) | |
2455 | return false; | |
2456 | } else { | |
2457 | /* | |
2458 | * For non-__GFP_REPEAT allocations which can presumably | |
2459 | * fail without consequence, stop if we failed to reclaim | |
2460 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2461 | * pages that were scanned. This will return to the | |
2462 | * caller faster at the risk reclaim/compaction and | |
2463 | * the resulting allocation attempt fails | |
2464 | */ | |
2465 | if (!nr_reclaimed) | |
2466 | return false; | |
2467 | } | |
3e7d3449 MG |
2468 | |
2469 | /* | |
2470 | * If we have not reclaimed enough pages for compaction and the | |
2471 | * inactive lists are large enough, continue reclaiming | |
2472 | */ | |
9861a62c | 2473 | pages_for_compaction = compact_gap(sc->order); |
a9dd0a83 | 2474 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
ec8acf20 | 2475 | if (get_nr_swap_pages() > 0) |
a9dd0a83 | 2476 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
3e7d3449 MG |
2477 | if (sc->nr_reclaimed < pages_for_compaction && |
2478 | inactive_lru_pages > pages_for_compaction) | |
2479 | return true; | |
2480 | ||
2481 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
a9dd0a83 MG |
2482 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2483 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2484 | if (!managed_zone(zone)) |
a9dd0a83 MG |
2485 | continue; |
2486 | ||
2487 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 2488 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
2489 | case COMPACT_CONTINUE: |
2490 | return false; | |
2491 | default: | |
2492 | /* check next zone */ | |
2493 | ; | |
2494 | } | |
3e7d3449 | 2495 | } |
a9dd0a83 | 2496 | return true; |
3e7d3449 MG |
2497 | } |
2498 | ||
970a39a3 | 2499 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2500 | { |
cb731d6c | 2501 | struct reclaim_state *reclaim_state = current->reclaim_state; |
f0fdc5e8 | 2502 | unsigned long nr_reclaimed, nr_scanned; |
2344d7e4 | 2503 | bool reclaimable = false; |
1da177e4 | 2504 | |
9b4f98cd JW |
2505 | do { |
2506 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2507 | struct mem_cgroup_reclaim_cookie reclaim = { | |
ef8f2327 | 2508 | .pgdat = pgdat, |
9b4f98cd JW |
2509 | .priority = sc->priority, |
2510 | }; | |
a9dd0a83 | 2511 | unsigned long node_lru_pages = 0; |
694fbc0f | 2512 | struct mem_cgroup *memcg; |
3e7d3449 | 2513 | |
9b4f98cd JW |
2514 | nr_reclaimed = sc->nr_reclaimed; |
2515 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2516 | |
694fbc0f AM |
2517 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2518 | do { | |
6b4f7799 | 2519 | unsigned long lru_pages; |
8e8ae645 | 2520 | unsigned long reclaimed; |
cb731d6c | 2521 | unsigned long scanned; |
5660048c | 2522 | |
241994ed | 2523 | if (mem_cgroup_low(root, memcg)) { |
d6622f63 YX |
2524 | if (!sc->memcg_low_reclaim) { |
2525 | sc->memcg_low_skipped = 1; | |
241994ed | 2526 | continue; |
d6622f63 | 2527 | } |
31176c78 | 2528 | mem_cgroup_event(memcg, MEMCG_LOW); |
241994ed JW |
2529 | } |
2530 | ||
8e8ae645 | 2531 | reclaimed = sc->nr_reclaimed; |
cb731d6c | 2532 | scanned = sc->nr_scanned; |
f9be23d6 | 2533 | |
a9dd0a83 MG |
2534 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
2535 | node_lru_pages += lru_pages; | |
f16015fb | 2536 | |
b5afba29 | 2537 | if (memcg) |
a9dd0a83 | 2538 | shrink_slab(sc->gfp_mask, pgdat->node_id, |
cb731d6c VD |
2539 | memcg, sc->nr_scanned - scanned, |
2540 | lru_pages); | |
2541 | ||
8e8ae645 JW |
2542 | /* Record the group's reclaim efficiency */ |
2543 | vmpressure(sc->gfp_mask, memcg, false, | |
2544 | sc->nr_scanned - scanned, | |
2545 | sc->nr_reclaimed - reclaimed); | |
2546 | ||
9b4f98cd | 2547 | /* |
a394cb8e MH |
2548 | * Direct reclaim and kswapd have to scan all memory |
2549 | * cgroups to fulfill the overall scan target for the | |
a9dd0a83 | 2550 | * node. |
a394cb8e MH |
2551 | * |
2552 | * Limit reclaim, on the other hand, only cares about | |
2553 | * nr_to_reclaim pages to be reclaimed and it will | |
2554 | * retry with decreasing priority if one round over the | |
2555 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2556 | */ |
a394cb8e MH |
2557 | if (!global_reclaim(sc) && |
2558 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2559 | mem_cgroup_iter_break(root, memcg); |
2560 | break; | |
2561 | } | |
241994ed | 2562 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
70ddf637 | 2563 | |
6b4f7799 JW |
2564 | /* |
2565 | * Shrink the slab caches in the same proportion that | |
2566 | * the eligible LRU pages were scanned. | |
2567 | */ | |
b2e18757 | 2568 | if (global_reclaim(sc)) |
a9dd0a83 | 2569 | shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, |
cb731d6c | 2570 | sc->nr_scanned - nr_scanned, |
a9dd0a83 | 2571 | node_lru_pages); |
cb731d6c VD |
2572 | |
2573 | if (reclaim_state) { | |
2574 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2575 | reclaim_state->reclaimed_slab = 0; | |
6b4f7799 JW |
2576 | } |
2577 | ||
8e8ae645 JW |
2578 | /* Record the subtree's reclaim efficiency */ |
2579 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
70ddf637 AV |
2580 | sc->nr_scanned - nr_scanned, |
2581 | sc->nr_reclaimed - nr_reclaimed); | |
2582 | ||
2344d7e4 JW |
2583 | if (sc->nr_reclaimed - nr_reclaimed) |
2584 | reclaimable = true; | |
2585 | ||
a9dd0a83 | 2586 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
9b4f98cd | 2587 | sc->nr_scanned - nr_scanned, sc)); |
2344d7e4 | 2588 | |
c73322d0 JW |
2589 | /* |
2590 | * Kswapd gives up on balancing particular nodes after too | |
2591 | * many failures to reclaim anything from them and goes to | |
2592 | * sleep. On reclaim progress, reset the failure counter. A | |
2593 | * successful direct reclaim run will revive a dormant kswapd. | |
2594 | */ | |
2595 | if (reclaimable) | |
2596 | pgdat->kswapd_failures = 0; | |
2597 | ||
2344d7e4 | 2598 | return reclaimable; |
f16015fb JW |
2599 | } |
2600 | ||
53853e2d | 2601 | /* |
fdd4c614 VB |
2602 | * Returns true if compaction should go ahead for a costly-order request, or |
2603 | * the allocation would already succeed without compaction. Return false if we | |
2604 | * should reclaim first. | |
53853e2d | 2605 | */ |
4f588331 | 2606 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2607 | { |
31483b6a | 2608 | unsigned long watermark; |
fdd4c614 | 2609 | enum compact_result suitable; |
fe4b1b24 | 2610 | |
fdd4c614 VB |
2611 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2612 | if (suitable == COMPACT_SUCCESS) | |
2613 | /* Allocation should succeed already. Don't reclaim. */ | |
2614 | return true; | |
2615 | if (suitable == COMPACT_SKIPPED) | |
2616 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2617 | return false; | |
fe4b1b24 | 2618 | |
53853e2d | 2619 | /* |
fdd4c614 VB |
2620 | * Compaction is already possible, but it takes time to run and there |
2621 | * are potentially other callers using the pages just freed. So proceed | |
2622 | * with reclaim to make a buffer of free pages available to give | |
2623 | * compaction a reasonable chance of completing and allocating the page. | |
2624 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2625 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2626 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 2627 | */ |
fdd4c614 | 2628 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 2629 | |
fdd4c614 | 2630 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
2631 | } |
2632 | ||
1da177e4 LT |
2633 | /* |
2634 | * This is the direct reclaim path, for page-allocating processes. We only | |
2635 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2636 | * request. | |
2637 | * | |
1da177e4 LT |
2638 | * If a zone is deemed to be full of pinned pages then just give it a light |
2639 | * scan then give up on it. | |
2640 | */ | |
0a0337e0 | 2641 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2642 | { |
dd1a239f | 2643 | struct zoneref *z; |
54a6eb5c | 2644 | struct zone *zone; |
0608f43d AM |
2645 | unsigned long nr_soft_reclaimed; |
2646 | unsigned long nr_soft_scanned; | |
619d0d76 | 2647 | gfp_t orig_mask; |
79dafcdc | 2648 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2649 | |
cc715d99 MG |
2650 | /* |
2651 | * If the number of buffer_heads in the machine exceeds the maximum | |
2652 | * allowed level, force direct reclaim to scan the highmem zone as | |
2653 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2654 | */ | |
619d0d76 | 2655 | orig_mask = sc->gfp_mask; |
b2e18757 | 2656 | if (buffer_heads_over_limit) { |
cc715d99 | 2657 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2658 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2659 | } |
cc715d99 | 2660 | |
d4debc66 | 2661 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2662 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
2663 | /* |
2664 | * Take care memory controller reclaiming has small influence | |
2665 | * to global LRU. | |
2666 | */ | |
89b5fae5 | 2667 | if (global_reclaim(sc)) { |
344736f2 VD |
2668 | if (!cpuset_zone_allowed(zone, |
2669 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2670 | continue; |
65ec02cb | 2671 | |
0b06496a JW |
2672 | /* |
2673 | * If we already have plenty of memory free for | |
2674 | * compaction in this zone, don't free any more. | |
2675 | * Even though compaction is invoked for any | |
2676 | * non-zero order, only frequent costly order | |
2677 | * reclamation is disruptive enough to become a | |
2678 | * noticeable problem, like transparent huge | |
2679 | * page allocations. | |
2680 | */ | |
2681 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2682 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2683 | compaction_ready(zone, sc)) { |
0b06496a JW |
2684 | sc->compaction_ready = true; |
2685 | continue; | |
e0887c19 | 2686 | } |
0b06496a | 2687 | |
79dafcdc MG |
2688 | /* |
2689 | * Shrink each node in the zonelist once. If the | |
2690 | * zonelist is ordered by zone (not the default) then a | |
2691 | * node may be shrunk multiple times but in that case | |
2692 | * the user prefers lower zones being preserved. | |
2693 | */ | |
2694 | if (zone->zone_pgdat == last_pgdat) | |
2695 | continue; | |
2696 | ||
0608f43d AM |
2697 | /* |
2698 | * This steals pages from memory cgroups over softlimit | |
2699 | * and returns the number of reclaimed pages and | |
2700 | * scanned pages. This works for global memory pressure | |
2701 | * and balancing, not for a memcg's limit. | |
2702 | */ | |
2703 | nr_soft_scanned = 0; | |
ef8f2327 | 2704 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2705 | sc->order, sc->gfp_mask, |
2706 | &nr_soft_scanned); | |
2707 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2708 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2709 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2710 | } |
408d8544 | 2711 | |
79dafcdc MG |
2712 | /* See comment about same check for global reclaim above */ |
2713 | if (zone->zone_pgdat == last_pgdat) | |
2714 | continue; | |
2715 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2716 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2717 | } |
e0c23279 | 2718 | |
619d0d76 WY |
2719 | /* |
2720 | * Restore to original mask to avoid the impact on the caller if we | |
2721 | * promoted it to __GFP_HIGHMEM. | |
2722 | */ | |
2723 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2724 | } |
4f98a2fe | 2725 | |
2a2e4885 JW |
2726 | static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) |
2727 | { | |
2728 | struct mem_cgroup *memcg; | |
2729 | ||
2730 | memcg = mem_cgroup_iter(root_memcg, NULL, NULL); | |
2731 | do { | |
2732 | unsigned long refaults; | |
2733 | struct lruvec *lruvec; | |
2734 | ||
2735 | if (memcg) | |
ccda7f43 | 2736 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); |
2a2e4885 JW |
2737 | else |
2738 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); | |
2739 | ||
2740 | lruvec = mem_cgroup_lruvec(pgdat, memcg); | |
2741 | lruvec->refaults = refaults; | |
2742 | } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); | |
2743 | } | |
2744 | ||
1da177e4 LT |
2745 | /* |
2746 | * This is the main entry point to direct page reclaim. | |
2747 | * | |
2748 | * If a full scan of the inactive list fails to free enough memory then we | |
2749 | * are "out of memory" and something needs to be killed. | |
2750 | * | |
2751 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2752 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2753 | * caller can't do much about. We kick the writeback threads and take explicit |
2754 | * naps in the hope that some of these pages can be written. But if the | |
2755 | * allocating task holds filesystem locks which prevent writeout this might not | |
2756 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2757 | * |
2758 | * returns: 0, if no pages reclaimed | |
2759 | * else, the number of pages reclaimed | |
1da177e4 | 2760 | */ |
dac1d27b | 2761 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2762 | struct scan_control *sc) |
1da177e4 | 2763 | { |
241994ed | 2764 | int initial_priority = sc->priority; |
2a2e4885 JW |
2765 | pg_data_t *last_pgdat; |
2766 | struct zoneref *z; | |
2767 | struct zone *zone; | |
241994ed | 2768 | retry: |
873b4771 KK |
2769 | delayacct_freepages_start(); |
2770 | ||
89b5fae5 | 2771 | if (global_reclaim(sc)) |
7cc30fcf | 2772 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 2773 | |
9e3b2f8c | 2774 | do { |
70ddf637 AV |
2775 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2776 | sc->priority); | |
66e1707b | 2777 | sc->nr_scanned = 0; |
0a0337e0 | 2778 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 2779 | |
bb21c7ce | 2780 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
2781 | break; |
2782 | ||
2783 | if (sc->compaction_ready) | |
2784 | break; | |
1da177e4 | 2785 | |
0e50ce3b MK |
2786 | /* |
2787 | * If we're getting trouble reclaiming, start doing | |
2788 | * writepage even in laptop mode. | |
2789 | */ | |
2790 | if (sc->priority < DEF_PRIORITY - 2) | |
2791 | sc->may_writepage = 1; | |
0b06496a | 2792 | } while (--sc->priority >= 0); |
bb21c7ce | 2793 | |
2a2e4885 JW |
2794 | last_pgdat = NULL; |
2795 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
2796 | sc->nodemask) { | |
2797 | if (zone->zone_pgdat == last_pgdat) | |
2798 | continue; | |
2799 | last_pgdat = zone->zone_pgdat; | |
2800 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); | |
2801 | } | |
2802 | ||
873b4771 KK |
2803 | delayacct_freepages_end(); |
2804 | ||
bb21c7ce KM |
2805 | if (sc->nr_reclaimed) |
2806 | return sc->nr_reclaimed; | |
2807 | ||
0cee34fd | 2808 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 2809 | if (sc->compaction_ready) |
7335084d MG |
2810 | return 1; |
2811 | ||
241994ed | 2812 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 2813 | if (sc->memcg_low_skipped) { |
241994ed | 2814 | sc->priority = initial_priority; |
d6622f63 YX |
2815 | sc->memcg_low_reclaim = 1; |
2816 | sc->memcg_low_skipped = 0; | |
241994ed JW |
2817 | goto retry; |
2818 | } | |
2819 | ||
bb21c7ce | 2820 | return 0; |
1da177e4 LT |
2821 | } |
2822 | ||
c73322d0 | 2823 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
2824 | { |
2825 | struct zone *zone; | |
2826 | unsigned long pfmemalloc_reserve = 0; | |
2827 | unsigned long free_pages = 0; | |
2828 | int i; | |
2829 | bool wmark_ok; | |
2830 | ||
c73322d0 JW |
2831 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
2832 | return true; | |
2833 | ||
5515061d MG |
2834 | for (i = 0; i <= ZONE_NORMAL; i++) { |
2835 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
2836 | if (!managed_zone(zone)) |
2837 | continue; | |
2838 | ||
2839 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
2840 | continue; |
2841 | ||
5515061d MG |
2842 | pfmemalloc_reserve += min_wmark_pages(zone); |
2843 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2844 | } | |
2845 | ||
675becce MG |
2846 | /* If there are no reserves (unexpected config) then do not throttle */ |
2847 | if (!pfmemalloc_reserve) | |
2848 | return true; | |
2849 | ||
5515061d MG |
2850 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
2851 | ||
2852 | /* kswapd must be awake if processes are being throttled */ | |
2853 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
38087d9b | 2854 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
5515061d MG |
2855 | (enum zone_type)ZONE_NORMAL); |
2856 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2857 | } | |
2858 | ||
2859 | return wmark_ok; | |
2860 | } | |
2861 | ||
2862 | /* | |
2863 | * Throttle direct reclaimers if backing storage is backed by the network | |
2864 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2865 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2866 | * when the low watermark is reached. |
2867 | * | |
2868 | * Returns true if a fatal signal was delivered during throttling. If this | |
2869 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2870 | */ |
50694c28 | 2871 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2872 | nodemask_t *nodemask) |
2873 | { | |
675becce | 2874 | struct zoneref *z; |
5515061d | 2875 | struct zone *zone; |
675becce | 2876 | pg_data_t *pgdat = NULL; |
5515061d MG |
2877 | |
2878 | /* | |
2879 | * Kernel threads should not be throttled as they may be indirectly | |
2880 | * responsible for cleaning pages necessary for reclaim to make forward | |
2881 | * progress. kjournald for example may enter direct reclaim while | |
2882 | * committing a transaction where throttling it could forcing other | |
2883 | * processes to block on log_wait_commit(). | |
2884 | */ | |
2885 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2886 | goto out; |
2887 | ||
2888 | /* | |
2889 | * If a fatal signal is pending, this process should not throttle. | |
2890 | * It should return quickly so it can exit and free its memory | |
2891 | */ | |
2892 | if (fatal_signal_pending(current)) | |
2893 | goto out; | |
5515061d | 2894 | |
675becce MG |
2895 | /* |
2896 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2897 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2898 | * GFP_KERNEL will be required for allocating network buffers when | |
2899 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2900 | * | |
2901 | * Throttling is based on the first usable node and throttled processes | |
2902 | * wait on a queue until kswapd makes progress and wakes them. There | |
2903 | * is an affinity then between processes waking up and where reclaim | |
2904 | * progress has been made assuming the process wakes on the same node. | |
2905 | * More importantly, processes running on remote nodes will not compete | |
2906 | * for remote pfmemalloc reserves and processes on different nodes | |
2907 | * should make reasonable progress. | |
2908 | */ | |
2909 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 2910 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
2911 | if (zone_idx(zone) > ZONE_NORMAL) |
2912 | continue; | |
2913 | ||
2914 | /* Throttle based on the first usable node */ | |
2915 | pgdat = zone->zone_pgdat; | |
c73322d0 | 2916 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
2917 | goto out; |
2918 | break; | |
2919 | } | |
2920 | ||
2921 | /* If no zone was usable by the allocation flags then do not throttle */ | |
2922 | if (!pgdat) | |
50694c28 | 2923 | goto out; |
5515061d | 2924 | |
68243e76 MG |
2925 | /* Account for the throttling */ |
2926 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2927 | ||
5515061d MG |
2928 | /* |
2929 | * If the caller cannot enter the filesystem, it's possible that it | |
2930 | * is due to the caller holding an FS lock or performing a journal | |
2931 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2932 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2933 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2934 | * second before continuing. | |
2935 | */ | |
2936 | if (!(gfp_mask & __GFP_FS)) { | |
2937 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
c73322d0 | 2938 | allow_direct_reclaim(pgdat), HZ); |
50694c28 MG |
2939 | |
2940 | goto check_pending; | |
5515061d MG |
2941 | } |
2942 | ||
2943 | /* Throttle until kswapd wakes the process */ | |
2944 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
c73322d0 | 2945 | allow_direct_reclaim(pgdat)); |
50694c28 MG |
2946 | |
2947 | check_pending: | |
2948 | if (fatal_signal_pending(current)) | |
2949 | return true; | |
2950 | ||
2951 | out: | |
2952 | return false; | |
5515061d MG |
2953 | } |
2954 | ||
dac1d27b | 2955 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2956 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2957 | { |
33906bc5 | 2958 | unsigned long nr_reclaimed; |
66e1707b | 2959 | struct scan_control sc = { |
ee814fe2 | 2960 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
7dea19f9 | 2961 | .gfp_mask = (gfp_mask = current_gfp_context(gfp_mask)), |
b2e18757 | 2962 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
2963 | .order = order, |
2964 | .nodemask = nodemask, | |
2965 | .priority = DEF_PRIORITY, | |
66e1707b | 2966 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2967 | .may_unmap = 1, |
2e2e4259 | 2968 | .may_swap = 1, |
66e1707b BS |
2969 | }; |
2970 | ||
5515061d | 2971 | /* |
50694c28 MG |
2972 | * Do not enter reclaim if fatal signal was delivered while throttled. |
2973 | * 1 is returned so that the page allocator does not OOM kill at this | |
2974 | * point. | |
5515061d | 2975 | */ |
50694c28 | 2976 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) |
5515061d MG |
2977 | return 1; |
2978 | ||
33906bc5 MG |
2979 | trace_mm_vmscan_direct_reclaim_begin(order, |
2980 | sc.may_writepage, | |
e5146b12 MG |
2981 | gfp_mask, |
2982 | sc.reclaim_idx); | |
33906bc5 | 2983 | |
3115cd91 | 2984 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
2985 | |
2986 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2987 | ||
2988 | return nr_reclaimed; | |
66e1707b BS |
2989 | } |
2990 | ||
c255a458 | 2991 | #ifdef CONFIG_MEMCG |
66e1707b | 2992 | |
a9dd0a83 | 2993 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 2994 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 2995 | pg_data_t *pgdat, |
0ae5e89c | 2996 | unsigned long *nr_scanned) |
4e416953 BS |
2997 | { |
2998 | struct scan_control sc = { | |
b8f5c566 | 2999 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 3000 | .target_mem_cgroup = memcg, |
4e416953 BS |
3001 | .may_writepage = !laptop_mode, |
3002 | .may_unmap = 1, | |
b2e18757 | 3003 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 3004 | .may_swap = !noswap, |
4e416953 | 3005 | }; |
6b4f7799 | 3006 | unsigned long lru_pages; |
0ae5e89c | 3007 | |
4e416953 BS |
3008 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3009 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3010 | |
9e3b2f8c | 3011 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e | 3012 | sc.may_writepage, |
e5146b12 MG |
3013 | sc.gfp_mask, |
3014 | sc.reclaim_idx); | |
bdce6d9e | 3015 | |
4e416953 BS |
3016 | /* |
3017 | * NOTE: Although we can get the priority field, using it | |
3018 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3019 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3020 | * will pick up pages from other mem cgroup's as well. We hack |
3021 | * the priority and make it zero. | |
3022 | */ | |
ef8f2327 | 3023 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
bdce6d9e KM |
3024 | |
3025 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3026 | ||
0ae5e89c | 3027 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
3028 | return sc.nr_reclaimed; |
3029 | } | |
3030 | ||
72835c86 | 3031 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3032 | unsigned long nr_pages, |
a7885eb8 | 3033 | gfp_t gfp_mask, |
b70a2a21 | 3034 | bool may_swap) |
66e1707b | 3035 | { |
4e416953 | 3036 | struct zonelist *zonelist; |
bdce6d9e | 3037 | unsigned long nr_reclaimed; |
889976db | 3038 | int nid; |
66e1707b | 3039 | struct scan_control sc = { |
b70a2a21 | 3040 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3041 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 3042 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 3043 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3044 | .target_mem_cgroup = memcg, |
3045 | .priority = DEF_PRIORITY, | |
3046 | .may_writepage = !laptop_mode, | |
3047 | .may_unmap = 1, | |
b70a2a21 | 3048 | .may_swap = may_swap, |
a09ed5e0 | 3049 | }; |
66e1707b | 3050 | |
889976db YH |
3051 | /* |
3052 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
3053 | * take care of from where we get pages. So the node where we start the | |
3054 | * scan does not need to be the current node. | |
3055 | */ | |
72835c86 | 3056 | nid = mem_cgroup_select_victim_node(memcg); |
889976db | 3057 | |
c9634cf0 | 3058 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; |
bdce6d9e KM |
3059 | |
3060 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
3061 | sc.may_writepage, | |
e5146b12 MG |
3062 | sc.gfp_mask, |
3063 | sc.reclaim_idx); | |
bdce6d9e | 3064 | |
89a28483 | 3065 | current->flags |= PF_MEMALLOC; |
3115cd91 | 3066 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
89a28483 | 3067 | current->flags &= ~PF_MEMALLOC; |
bdce6d9e KM |
3068 | |
3069 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
3070 | ||
3071 | return nr_reclaimed; | |
66e1707b BS |
3072 | } |
3073 | #endif | |
3074 | ||
1d82de61 | 3075 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3076 | struct scan_control *sc) |
f16015fb | 3077 | { |
b95a2f2d | 3078 | struct mem_cgroup *memcg; |
f16015fb | 3079 | |
b95a2f2d JW |
3080 | if (!total_swap_pages) |
3081 | return; | |
3082 | ||
3083 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
3084 | do { | |
ef8f2327 | 3085 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
b95a2f2d | 3086 | |
2a2e4885 | 3087 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) |
1a93be0e | 3088 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 3089 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
3090 | |
3091 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
3092 | } while (memcg); | |
f16015fb JW |
3093 | } |
3094 | ||
e716f2eb MG |
3095 | /* |
3096 | * Returns true if there is an eligible zone balanced for the request order | |
3097 | * and classzone_idx | |
3098 | */ | |
3099 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
60cefed4 | 3100 | { |
e716f2eb MG |
3101 | int i; |
3102 | unsigned long mark = -1; | |
3103 | struct zone *zone; | |
60cefed4 | 3104 | |
e716f2eb MG |
3105 | for (i = 0; i <= classzone_idx; i++) { |
3106 | zone = pgdat->node_zones + i; | |
6256c6b4 | 3107 | |
e716f2eb MG |
3108 | if (!managed_zone(zone)) |
3109 | continue; | |
3110 | ||
3111 | mark = high_wmark_pages(zone); | |
3112 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) | |
3113 | return true; | |
3114 | } | |
3115 | ||
3116 | /* | |
3117 | * If a node has no populated zone within classzone_idx, it does not | |
3118 | * need balancing by definition. This can happen if a zone-restricted | |
3119 | * allocation tries to wake a remote kswapd. | |
3120 | */ | |
3121 | if (mark == -1) | |
3122 | return true; | |
3123 | ||
3124 | return false; | |
60cefed4 JW |
3125 | } |
3126 | ||
631b6e08 MG |
3127 | /* Clear pgdat state for congested, dirty or under writeback. */ |
3128 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
3129 | { | |
3130 | clear_bit(PGDAT_CONGESTED, &pgdat->flags); | |
3131 | clear_bit(PGDAT_DIRTY, &pgdat->flags); | |
3132 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
3133 | } | |
3134 | ||
5515061d MG |
3135 | /* |
3136 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3137 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3138 | * | |
3139 | * Returns true if kswapd is ready to sleep | |
3140 | */ | |
d9f21d42 | 3141 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f50de2d3 | 3142 | { |
5515061d | 3143 | /* |
9e5e3661 | 3144 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 3145 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
3146 | * race between when kswapd checks the watermarks and a process gets |
3147 | * throttled. There is also a potential race if processes get | |
3148 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3149 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3150 | * the wake up checks. If kswapd is going to sleep, no process should | |
3151 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3152 | * the wake up is premature, processes will wake kswapd and get | |
3153 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3154 | * that here we are under prepare_to_wait(). | |
5515061d | 3155 | */ |
9e5e3661 VB |
3156 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3157 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3158 | |
c73322d0 JW |
3159 | /* Hopeless node, leave it to direct reclaim */ |
3160 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3161 | return true; | |
3162 | ||
e716f2eb MG |
3163 | if (pgdat_balanced(pgdat, order, classzone_idx)) { |
3164 | clear_pgdat_congested(pgdat); | |
3165 | return true; | |
1d82de61 MG |
3166 | } |
3167 | ||
333b0a45 | 3168 | return false; |
f50de2d3 MG |
3169 | } |
3170 | ||
75485363 | 3171 | /* |
1d82de61 MG |
3172 | * kswapd shrinks a node of pages that are at or below the highest usable |
3173 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3174 | * |
3175 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3176 | * reclaim or if the lack of progress was due to pages under writeback. |
3177 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3178 | */ |
1d82de61 | 3179 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3180 | struct scan_control *sc) |
75485363 | 3181 | { |
1d82de61 MG |
3182 | struct zone *zone; |
3183 | int z; | |
75485363 | 3184 | |
1d82de61 MG |
3185 | /* Reclaim a number of pages proportional to the number of zones */ |
3186 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3187 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 3188 | zone = pgdat->node_zones + z; |
6aa303de | 3189 | if (!managed_zone(zone)) |
1d82de61 | 3190 | continue; |
7c954f6d | 3191 | |
1d82de61 MG |
3192 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3193 | } | |
7c954f6d MG |
3194 | |
3195 | /* | |
1d82de61 MG |
3196 | * Historically care was taken to put equal pressure on all zones but |
3197 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3198 | */ |
970a39a3 | 3199 | shrink_node(pgdat, sc); |
283aba9f | 3200 | |
7c954f6d | 3201 | /* |
1d82de61 MG |
3202 | * Fragmentation may mean that the system cannot be rebalanced for |
3203 | * high-order allocations. If twice the allocation size has been | |
3204 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3205 | * excessive reclaim. Assume that a process requested a high-order | |
3206 | * can direct reclaim/compact. | |
7c954f6d | 3207 | */ |
9861a62c | 3208 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 3209 | sc->order = 0; |
7c954f6d | 3210 | |
b8e83b94 | 3211 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3212 | } |
3213 | ||
1da177e4 | 3214 | /* |
1d82de61 MG |
3215 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3216 | * that are eligible for use by the caller until at least one zone is | |
3217 | * balanced. | |
1da177e4 | 3218 | * |
1d82de61 | 3219 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3220 | * |
3221 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3222 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
1d82de61 MG |
3223 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone |
3224 | * or lower is eligible for reclaim until at least one usable zone is | |
3225 | * balanced. | |
1da177e4 | 3226 | */ |
accf6242 | 3227 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3228 | { |
1da177e4 | 3229 | int i; |
0608f43d AM |
3230 | unsigned long nr_soft_reclaimed; |
3231 | unsigned long nr_soft_scanned; | |
1d82de61 | 3232 | struct zone *zone; |
179e9639 AM |
3233 | struct scan_control sc = { |
3234 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3235 | .order = order, |
b8e83b94 | 3236 | .priority = DEF_PRIORITY, |
ee814fe2 | 3237 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3238 | .may_unmap = 1, |
2e2e4259 | 3239 | .may_swap = 1, |
179e9639 | 3240 | }; |
f8891e5e | 3241 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3242 | |
9e3b2f8c | 3243 | do { |
c73322d0 | 3244 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 MG |
3245 | bool raise_priority = true; |
3246 | ||
84c7a777 | 3247 | sc.reclaim_idx = classzone_idx; |
1da177e4 | 3248 | |
86c79f6b | 3249 | /* |
84c7a777 MG |
3250 | * If the number of buffer_heads exceeds the maximum allowed |
3251 | * then consider reclaiming from all zones. This has a dual | |
3252 | * purpose -- on 64-bit systems it is expected that | |
3253 | * buffer_heads are stripped during active rotation. On 32-bit | |
3254 | * systems, highmem pages can pin lowmem memory and shrinking | |
3255 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3256 | * go ahead if all eligible zones for the original allocation | |
3257 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3258 | */ |
3259 | if (buffer_heads_over_limit) { | |
3260 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3261 | zone = pgdat->node_zones + i; | |
6aa303de | 3262 | if (!managed_zone(zone)) |
86c79f6b | 3263 | continue; |
cc715d99 | 3264 | |
970a39a3 | 3265 | sc.reclaim_idx = i; |
e1dbeda6 | 3266 | break; |
1da177e4 | 3267 | } |
1da177e4 | 3268 | } |
dafcb73e | 3269 | |
86c79f6b | 3270 | /* |
e716f2eb MG |
3271 | * Only reclaim if there are no eligible zones. Note that |
3272 | * sc.reclaim_idx is not used as buffer_heads_over_limit may | |
3273 | * have adjusted it. | |
86c79f6b | 3274 | */ |
e716f2eb MG |
3275 | if (pgdat_balanced(pgdat, sc.order, classzone_idx)) |
3276 | goto out; | |
e1dbeda6 | 3277 | |
1d82de61 MG |
3278 | /* |
3279 | * Do some background aging of the anon list, to give | |
3280 | * pages a chance to be referenced before reclaiming. All | |
3281 | * pages are rotated regardless of classzone as this is | |
3282 | * about consistent aging. | |
3283 | */ | |
ef8f2327 | 3284 | age_active_anon(pgdat, &sc); |
1d82de61 | 3285 | |
b7ea3c41 MG |
3286 | /* |
3287 | * If we're getting trouble reclaiming, start doing writepage | |
3288 | * even in laptop mode. | |
3289 | */ | |
047d72c3 | 3290 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
3291 | sc.may_writepage = 1; |
3292 | ||
1d82de61 MG |
3293 | /* Call soft limit reclaim before calling shrink_node. */ |
3294 | sc.nr_scanned = 0; | |
3295 | nr_soft_scanned = 0; | |
ef8f2327 | 3296 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3297 | sc.gfp_mask, &nr_soft_scanned); |
3298 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3299 | ||
1da177e4 | 3300 | /* |
1d82de61 MG |
3301 | * There should be no need to raise the scanning priority if |
3302 | * enough pages are already being scanned that that high | |
3303 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3304 | */ |
970a39a3 | 3305 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3306 | raise_priority = false; |
5515061d MG |
3307 | |
3308 | /* | |
3309 | * If the low watermark is met there is no need for processes | |
3310 | * to be throttled on pfmemalloc_wait as they should not be | |
3311 | * able to safely make forward progress. Wake them | |
3312 | */ | |
3313 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 3314 | allow_direct_reclaim(pgdat)) |
cfc51155 | 3315 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3316 | |
b8e83b94 MG |
3317 | /* Check if kswapd should be suspending */ |
3318 | if (try_to_freeze() || kthread_should_stop()) | |
3319 | break; | |
8357376d | 3320 | |
73ce02e9 | 3321 | /* |
b8e83b94 MG |
3322 | * Raise priority if scanning rate is too low or there was no |
3323 | * progress in reclaiming pages | |
73ce02e9 | 3324 | */ |
c73322d0 JW |
3325 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
3326 | if (raise_priority || !nr_reclaimed) | |
b8e83b94 | 3327 | sc.priority--; |
1d82de61 | 3328 | } while (sc.priority >= 1); |
1da177e4 | 3329 | |
c73322d0 JW |
3330 | if (!sc.nr_reclaimed) |
3331 | pgdat->kswapd_failures++; | |
3332 | ||
b8e83b94 | 3333 | out: |
2a2e4885 | 3334 | snapshot_refaults(NULL, pgdat); |
0abdee2b | 3335 | /* |
1d82de61 MG |
3336 | * Return the order kswapd stopped reclaiming at as |
3337 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3338 | * entered the allocator slow path while kswapd was awake, order will | |
3339 | * remain at the higher level. | |
0abdee2b | 3340 | */ |
1d82de61 | 3341 | return sc.order; |
1da177e4 LT |
3342 | } |
3343 | ||
e716f2eb MG |
3344 | /* |
3345 | * pgdat->kswapd_classzone_idx is the highest zone index that a recent | |
3346 | * allocation request woke kswapd for. When kswapd has not woken recently, | |
3347 | * the value is MAX_NR_ZONES which is not a valid index. This compares a | |
3348 | * given classzone and returns it or the highest classzone index kswapd | |
3349 | * was recently woke for. | |
3350 | */ | |
3351 | static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, | |
3352 | enum zone_type classzone_idx) | |
3353 | { | |
3354 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
3355 | return classzone_idx; | |
3356 | ||
3357 | return max(pgdat->kswapd_classzone_idx, classzone_idx); | |
3358 | } | |
3359 | ||
38087d9b MG |
3360 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3361 | unsigned int classzone_idx) | |
f0bc0a60 KM |
3362 | { |
3363 | long remaining = 0; | |
3364 | DEFINE_WAIT(wait); | |
3365 | ||
3366 | if (freezing(current) || kthread_should_stop()) | |
3367 | return; | |
3368 | ||
3369 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3370 | ||
333b0a45 SG |
3371 | /* |
3372 | * Try to sleep for a short interval. Note that kcompactd will only be | |
3373 | * woken if it is possible to sleep for a short interval. This is | |
3374 | * deliberate on the assumption that if reclaim cannot keep an | |
3375 | * eligible zone balanced that it's also unlikely that compaction will | |
3376 | * succeed. | |
3377 | */ | |
d9f21d42 | 3378 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
fd901c95 VB |
3379 | /* |
3380 | * Compaction records what page blocks it recently failed to | |
3381 | * isolate pages from and skips them in the future scanning. | |
3382 | * When kswapd is going to sleep, it is reasonable to assume | |
3383 | * that pages and compaction may succeed so reset the cache. | |
3384 | */ | |
3385 | reset_isolation_suitable(pgdat); | |
3386 | ||
3387 | /* | |
3388 | * We have freed the memory, now we should compact it to make | |
3389 | * allocation of the requested order possible. | |
3390 | */ | |
38087d9b | 3391 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
fd901c95 | 3392 | |
f0bc0a60 | 3393 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3394 | |
3395 | /* | |
3396 | * If woken prematurely then reset kswapd_classzone_idx and | |
3397 | * order. The values will either be from a wakeup request or | |
3398 | * the previous request that slept prematurely. | |
3399 | */ | |
3400 | if (remaining) { | |
e716f2eb | 3401 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b MG |
3402 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); |
3403 | } | |
3404 | ||
f0bc0a60 KM |
3405 | finish_wait(&pgdat->kswapd_wait, &wait); |
3406 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3407 | } | |
3408 | ||
3409 | /* | |
3410 | * After a short sleep, check if it was a premature sleep. If not, then | |
3411 | * go fully to sleep until explicitly woken up. | |
3412 | */ | |
d9f21d42 MG |
3413 | if (!remaining && |
3414 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
f0bc0a60 KM |
3415 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3416 | ||
3417 | /* | |
3418 | * vmstat counters are not perfectly accurate and the estimated | |
3419 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3420 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3421 | * watermarks being breached while under pressure, we reduce the | |
3422 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3423 | * them before going back to sleep. | |
3424 | */ | |
3425 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3426 | |
3427 | if (!kthread_should_stop()) | |
3428 | schedule(); | |
3429 | ||
f0bc0a60 KM |
3430 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3431 | } else { | |
3432 | if (remaining) | |
3433 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3434 | else | |
3435 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3436 | } | |
3437 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3438 | } | |
3439 | ||
1da177e4 LT |
3440 | /* |
3441 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3442 | * from the init process. |
1da177e4 LT |
3443 | * |
3444 | * This basically trickles out pages so that we have _some_ | |
3445 | * free memory available even if there is no other activity | |
3446 | * that frees anything up. This is needed for things like routing | |
3447 | * etc, where we otherwise might have all activity going on in | |
3448 | * asynchronous contexts that cannot page things out. | |
3449 | * | |
3450 | * If there are applications that are active memory-allocators | |
3451 | * (most normal use), this basically shouldn't matter. | |
3452 | */ | |
3453 | static int kswapd(void *p) | |
3454 | { | |
e716f2eb MG |
3455 | unsigned int alloc_order, reclaim_order; |
3456 | unsigned int classzone_idx = MAX_NR_ZONES - 1; | |
1da177e4 LT |
3457 | pg_data_t *pgdat = (pg_data_t*)p; |
3458 | struct task_struct *tsk = current; | |
f0bc0a60 | 3459 | |
1da177e4 LT |
3460 | struct reclaim_state reclaim_state = { |
3461 | .reclaimed_slab = 0, | |
3462 | }; | |
a70f7302 | 3463 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3464 | |
cf40bd16 NP |
3465 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3466 | ||
174596a0 | 3467 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3468 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3469 | current->reclaim_state = &reclaim_state; |
3470 | ||
3471 | /* | |
3472 | * Tell the memory management that we're a "memory allocator", | |
3473 | * and that if we need more memory we should get access to it | |
3474 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3475 | * never get caught in the normal page freeing logic. | |
3476 | * | |
3477 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3478 | * you need a small amount of memory in order to be able to | |
3479 | * page out something else, and this flag essentially protects | |
3480 | * us from recursively trying to free more memory as we're | |
3481 | * trying to free the first piece of memory in the first place). | |
3482 | */ | |
930d9152 | 3483 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3484 | set_freezable(); |
1da177e4 | 3485 | |
e716f2eb MG |
3486 | pgdat->kswapd_order = 0; |
3487 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; | |
1da177e4 | 3488 | for ( ; ; ) { |
6f6313d4 | 3489 | bool ret; |
3e1d1d28 | 3490 | |
e716f2eb MG |
3491 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3492 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); | |
3493 | ||
38087d9b MG |
3494 | kswapd_try_sleep: |
3495 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3496 | classzone_idx); | |
215ddd66 | 3497 | |
38087d9b MG |
3498 | /* Read the new order and classzone_idx */ |
3499 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
e716f2eb | 3500 | classzone_idx = kswapd_classzone_idx(pgdat, 0); |
38087d9b | 3501 | pgdat->kswapd_order = 0; |
e716f2eb | 3502 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
1da177e4 | 3503 | |
8fe23e05 DR |
3504 | ret = try_to_freeze(); |
3505 | if (kthread_should_stop()) | |
3506 | break; | |
3507 | ||
3508 | /* | |
3509 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3510 | * after returning from the refrigerator | |
3511 | */ | |
38087d9b MG |
3512 | if (ret) |
3513 | continue; | |
3514 | ||
3515 | /* | |
3516 | * Reclaim begins at the requested order but if a high-order | |
3517 | * reclaim fails then kswapd falls back to reclaiming for | |
3518 | * order-0. If that happens, kswapd will consider sleeping | |
3519 | * for the order it finished reclaiming at (reclaim_order) | |
3520 | * but kcompactd is woken to compact for the original | |
3521 | * request (alloc_order). | |
3522 | */ | |
e5146b12 MG |
3523 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
3524 | alloc_order); | |
38087d9b MG |
3525 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
3526 | if (reclaim_order < alloc_order) | |
3527 | goto kswapd_try_sleep; | |
1da177e4 | 3528 | } |
b0a8cc58 | 3529 | |
71abdc15 | 3530 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
b0a8cc58 | 3531 | current->reclaim_state = NULL; |
71abdc15 JW |
3532 | lockdep_clear_current_reclaim_state(); |
3533 | ||
1da177e4 LT |
3534 | return 0; |
3535 | } | |
3536 | ||
3537 | /* | |
3538 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3539 | */ | |
99504748 | 3540 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3541 | { |
3542 | pg_data_t *pgdat; | |
3543 | ||
6aa303de | 3544 | if (!managed_zone(zone)) |
1da177e4 LT |
3545 | return; |
3546 | ||
344736f2 | 3547 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) |
1da177e4 | 3548 | return; |
88f5acf8 | 3549 | pgdat = zone->zone_pgdat; |
e716f2eb MG |
3550 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, |
3551 | classzone_idx); | |
38087d9b | 3552 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
8d0986e2 | 3553 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3554 | return; |
e1a55637 | 3555 | |
c73322d0 JW |
3556 | /* Hopeless node, leave it to direct reclaim */ |
3557 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3558 | return; | |
3559 | ||
e716f2eb MG |
3560 | if (pgdat_balanced(pgdat, order, classzone_idx)) |
3561 | return; | |
88f5acf8 | 3562 | |
e716f2eb | 3563 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order); |
8d0986e2 | 3564 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3565 | } |
3566 | ||
c6f37f12 | 3567 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3568 | /* |
7b51755c | 3569 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3570 | * freed pages. |
3571 | * | |
3572 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3573 | * LRU order by reclaiming preferentially | |
3574 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3575 | */ |
7b51755c | 3576 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3577 | { |
d6277db4 | 3578 | struct reclaim_state reclaim_state; |
d6277db4 | 3579 | struct scan_control sc = { |
ee814fe2 | 3580 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3581 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 3582 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 3583 | .priority = DEF_PRIORITY, |
d6277db4 | 3584 | .may_writepage = 1, |
ee814fe2 JW |
3585 | .may_unmap = 1, |
3586 | .may_swap = 1, | |
7b51755c | 3587 | .hibernation_mode = 1, |
1da177e4 | 3588 | }; |
a09ed5e0 | 3589 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c KM |
3590 | struct task_struct *p = current; |
3591 | unsigned long nr_reclaimed; | |
1da177e4 | 3592 | |
7b51755c KM |
3593 | p->flags |= PF_MEMALLOC; |
3594 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3595 | reclaim_state.reclaimed_slab = 0; | |
3596 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3597 | |
3115cd91 | 3598 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3599 | |
7b51755c KM |
3600 | p->reclaim_state = NULL; |
3601 | lockdep_clear_current_reclaim_state(); | |
3602 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3603 | |
7b51755c | 3604 | return nr_reclaimed; |
1da177e4 | 3605 | } |
c6f37f12 | 3606 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3607 | |
1da177e4 LT |
3608 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3609 | not required for correctness. So if the last cpu in a node goes | |
3610 | away, we get changed to run anywhere: as the first one comes back, | |
3611 | restore their cpu bindings. */ | |
517bbed9 | 3612 | static int kswapd_cpu_online(unsigned int cpu) |
1da177e4 | 3613 | { |
58c0a4a7 | 3614 | int nid; |
1da177e4 | 3615 | |
517bbed9 SAS |
3616 | for_each_node_state(nid, N_MEMORY) { |
3617 | pg_data_t *pgdat = NODE_DATA(nid); | |
3618 | const struct cpumask *mask; | |
a70f7302 | 3619 | |
517bbed9 | 3620 | mask = cpumask_of_node(pgdat->node_id); |
c5f59f08 | 3621 | |
517bbed9 SAS |
3622 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3623 | /* One of our CPUs online: restore mask */ | |
3624 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
1da177e4 | 3625 | } |
517bbed9 | 3626 | return 0; |
1da177e4 | 3627 | } |
1da177e4 | 3628 | |
3218ae14 YG |
3629 | /* |
3630 | * This kswapd start function will be called by init and node-hot-add. | |
3631 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3632 | */ | |
3633 | int kswapd_run(int nid) | |
3634 | { | |
3635 | pg_data_t *pgdat = NODE_DATA(nid); | |
3636 | int ret = 0; | |
3637 | ||
3638 | if (pgdat->kswapd) | |
3639 | return 0; | |
3640 | ||
3641 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3642 | if (IS_ERR(pgdat->kswapd)) { | |
3643 | /* failure at boot is fatal */ | |
3644 | BUG_ON(system_state == SYSTEM_BOOTING); | |
d5dc0ad9 GS |
3645 | pr_err("Failed to start kswapd on node %d\n", nid); |
3646 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3647 | pgdat->kswapd = NULL; |
3218ae14 YG |
3648 | } |
3649 | return ret; | |
3650 | } | |
3651 | ||
8fe23e05 | 3652 | /* |
d8adde17 | 3653 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 3654 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
3655 | */ |
3656 | void kswapd_stop(int nid) | |
3657 | { | |
3658 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3659 | ||
d8adde17 | 3660 | if (kswapd) { |
8fe23e05 | 3661 | kthread_stop(kswapd); |
d8adde17 JL |
3662 | NODE_DATA(nid)->kswapd = NULL; |
3663 | } | |
8fe23e05 DR |
3664 | } |
3665 | ||
1da177e4 LT |
3666 | static int __init kswapd_init(void) |
3667 | { | |
517bbed9 | 3668 | int nid, ret; |
69e05944 | 3669 | |
1da177e4 | 3670 | swap_setup(); |
48fb2e24 | 3671 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3672 | kswapd_run(nid); |
517bbed9 SAS |
3673 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
3674 | "mm/vmscan:online", kswapd_cpu_online, | |
3675 | NULL); | |
3676 | WARN_ON(ret < 0); | |
1da177e4 LT |
3677 | return 0; |
3678 | } | |
3679 | ||
3680 | module_init(kswapd_init) | |
9eeff239 CL |
3681 | |
3682 | #ifdef CONFIG_NUMA | |
3683 | /* | |
a5f5f91d | 3684 | * Node reclaim mode |
9eeff239 | 3685 | * |
a5f5f91d | 3686 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 3687 | * the watermarks. |
9eeff239 | 3688 | */ |
a5f5f91d | 3689 | int node_reclaim_mode __read_mostly; |
9eeff239 | 3690 | |
1b2ffb78 | 3691 | #define RECLAIM_OFF 0 |
7d03431c | 3692 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 3693 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 3694 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 3695 | |
a92f7126 | 3696 | /* |
a5f5f91d | 3697 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
3698 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
3699 | * a zone. | |
3700 | */ | |
a5f5f91d | 3701 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 3702 | |
9614634f | 3703 | /* |
a5f5f91d | 3704 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
3705 | * occur. |
3706 | */ | |
3707 | int sysctl_min_unmapped_ratio = 1; | |
3708 | ||
0ff38490 CL |
3709 | /* |
3710 | * If the number of slab pages in a zone grows beyond this percentage then | |
3711 | * slab reclaim needs to occur. | |
3712 | */ | |
3713 | int sysctl_min_slab_ratio = 5; | |
3714 | ||
11fb9989 | 3715 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 3716 | { |
11fb9989 MG |
3717 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
3718 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
3719 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
3720 | |
3721 | /* | |
3722 | * It's possible for there to be more file mapped pages than | |
3723 | * accounted for by the pages on the file LRU lists because | |
3724 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3725 | */ | |
3726 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3727 | } | |
3728 | ||
3729 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 3730 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 3731 | { |
d031a157 AM |
3732 | unsigned long nr_pagecache_reclaimable; |
3733 | unsigned long delta = 0; | |
90afa5de MG |
3734 | |
3735 | /* | |
95bbc0c7 | 3736 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 3737 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 3738 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
3739 | * a better estimate |
3740 | */ | |
a5f5f91d MG |
3741 | if (node_reclaim_mode & RECLAIM_UNMAP) |
3742 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 3743 | else |
a5f5f91d | 3744 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
3745 | |
3746 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
3747 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
3748 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
3749 | |
3750 | /* Watch for any possible underflows due to delta */ | |
3751 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3752 | delta = nr_pagecache_reclaimable; | |
3753 | ||
3754 | return nr_pagecache_reclaimable - delta; | |
3755 | } | |
3756 | ||
9eeff239 | 3757 | /* |
a5f5f91d | 3758 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 3759 | */ |
a5f5f91d | 3760 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3761 | { |
7fb2d46d | 3762 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3763 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3764 | struct task_struct *p = current; |
3765 | struct reclaim_state reclaim_state; | |
a5f5f91d | 3766 | int classzone_idx = gfp_zone(gfp_mask); |
179e9639 | 3767 | struct scan_control sc = { |
62b726c1 | 3768 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3769 | .gfp_mask = (gfp_mask = current_gfp_context(gfp_mask)), |
bd2f6199 | 3770 | .order = order, |
a5f5f91d MG |
3771 | .priority = NODE_RECLAIM_PRIORITY, |
3772 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
3773 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 3774 | .may_swap = 1, |
a5f5f91d | 3775 | .reclaim_idx = classzone_idx, |
179e9639 | 3776 | }; |
9eeff239 | 3777 | |
9eeff239 | 3778 | cond_resched(); |
d4f7796e | 3779 | /* |
95bbc0c7 | 3780 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 3781 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 3782 | * and RECLAIM_UNMAP. |
d4f7796e CL |
3783 | */ |
3784 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3785 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3786 | reclaim_state.reclaimed_slab = 0; |
3787 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3788 | |
a5f5f91d | 3789 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 CL |
3790 | /* |
3791 | * Free memory by calling shrink zone with increasing | |
3792 | * priorities until we have enough memory freed. | |
3793 | */ | |
0ff38490 | 3794 | do { |
970a39a3 | 3795 | shrink_node(pgdat, &sc); |
9e3b2f8c | 3796 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 3797 | } |
c84db23c | 3798 | |
9eeff239 | 3799 | p->reclaim_state = NULL; |
d4f7796e | 3800 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3801 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3802 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3803 | } |
179e9639 | 3804 | |
a5f5f91d | 3805 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 3806 | { |
d773ed6b | 3807 | int ret; |
179e9639 AM |
3808 | |
3809 | /* | |
a5f5f91d | 3810 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 3811 | * slab pages if we are over the defined limits. |
34aa1330 | 3812 | * |
9614634f CL |
3813 | * A small portion of unmapped file backed pages is needed for |
3814 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
3815 | * thrown out if the node is overallocated. So we do not reclaim |
3816 | * if less than a specified percentage of the node is used by | |
9614634f | 3817 | * unmapped file backed pages. |
179e9639 | 3818 | */ |
a5f5f91d MG |
3819 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
3820 | sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) | |
3821 | return NODE_RECLAIM_FULL; | |
179e9639 AM |
3822 | |
3823 | /* | |
d773ed6b | 3824 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3825 | */ |
d0164adc | 3826 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 3827 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
3828 | |
3829 | /* | |
a5f5f91d | 3830 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
3831 | * have associated processors. This will favor the local processor |
3832 | * over remote processors and spread off node memory allocations | |
3833 | * as wide as possible. | |
3834 | */ | |
a5f5f91d MG |
3835 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
3836 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 3837 | |
a5f5f91d MG |
3838 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
3839 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 3840 | |
a5f5f91d MG |
3841 | ret = __node_reclaim(pgdat, gfp_mask, order); |
3842 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 3843 | |
24cf7251 MG |
3844 | if (!ret) |
3845 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3846 | ||
d773ed6b | 3847 | return ret; |
179e9639 | 3848 | } |
9eeff239 | 3849 | #endif |
894bc310 | 3850 | |
894bc310 LS |
3851 | /* |
3852 | * page_evictable - test whether a page is evictable | |
3853 | * @page: the page to test | |
894bc310 LS |
3854 | * |
3855 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3856 | * lists vs unevictable list. |
894bc310 LS |
3857 | * |
3858 | * Reasons page might not be evictable: | |
ba9ddf49 | 3859 | * (1) page's mapping marked unevictable |
b291f000 | 3860 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3861 | * |
894bc310 | 3862 | */ |
39b5f29a | 3863 | int page_evictable(struct page *page) |
894bc310 | 3864 | { |
39b5f29a | 3865 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
894bc310 | 3866 | } |
89e004ea | 3867 | |
85046579 | 3868 | #ifdef CONFIG_SHMEM |
89e004ea | 3869 | /** |
24513264 HD |
3870 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3871 | * @pages: array of pages to check | |
3872 | * @nr_pages: number of pages to check | |
89e004ea | 3873 | * |
24513264 | 3874 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3875 | * |
3876 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3877 | */ |
24513264 | 3878 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3879 | { |
925b7673 | 3880 | struct lruvec *lruvec; |
785b99fe | 3881 | struct pglist_data *pgdat = NULL; |
24513264 HD |
3882 | int pgscanned = 0; |
3883 | int pgrescued = 0; | |
3884 | int i; | |
89e004ea | 3885 | |
24513264 HD |
3886 | for (i = 0; i < nr_pages; i++) { |
3887 | struct page *page = pages[i]; | |
785b99fe | 3888 | struct pglist_data *pagepgdat = page_pgdat(page); |
89e004ea | 3889 | |
24513264 | 3890 | pgscanned++; |
785b99fe MG |
3891 | if (pagepgdat != pgdat) { |
3892 | if (pgdat) | |
3893 | spin_unlock_irq(&pgdat->lru_lock); | |
3894 | pgdat = pagepgdat; | |
3895 | spin_lock_irq(&pgdat->lru_lock); | |
24513264 | 3896 | } |
785b99fe | 3897 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
89e004ea | 3898 | |
24513264 HD |
3899 | if (!PageLRU(page) || !PageUnevictable(page)) |
3900 | continue; | |
89e004ea | 3901 | |
39b5f29a | 3902 | if (page_evictable(page)) { |
24513264 HD |
3903 | enum lru_list lru = page_lru_base_type(page); |
3904 | ||
309381fe | 3905 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 3906 | ClearPageUnevictable(page); |
fa9add64 HD |
3907 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
3908 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 3909 | pgrescued++; |
89e004ea | 3910 | } |
24513264 | 3911 | } |
89e004ea | 3912 | |
785b99fe | 3913 | if (pgdat) { |
24513264 HD |
3914 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
3915 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
785b99fe | 3916 | spin_unlock_irq(&pgdat->lru_lock); |
89e004ea | 3917 | } |
89e004ea | 3918 | } |
85046579 | 3919 | #endif /* CONFIG_SHMEM */ |