]>
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 | ||
241994ed JW |
100 | /* Can cgroups be reclaimed below their normal consumption range? */ |
101 | unsigned int may_thrash:1; | |
102 | ||
ee814fe2 JW |
103 | unsigned int hibernation_mode:1; |
104 | ||
105 | /* One of the zones is ready for compaction */ | |
106 | unsigned int compaction_ready:1; | |
107 | ||
108 | /* Incremented by the number of inactive pages that were scanned */ | |
109 | unsigned long nr_scanned; | |
110 | ||
111 | /* Number of pages freed so far during a call to shrink_zones() */ | |
112 | unsigned long nr_reclaimed; | |
1da177e4 LT |
113 | }; |
114 | ||
1da177e4 LT |
115 | #ifdef ARCH_HAS_PREFETCH |
116 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
117 | do { \ | |
118 | if ((_page)->lru.prev != _base) { \ | |
119 | struct page *prev; \ | |
120 | \ | |
121 | prev = lru_to_page(&(_page->lru)); \ | |
122 | prefetch(&prev->_field); \ | |
123 | } \ | |
124 | } while (0) | |
125 | #else | |
126 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
127 | #endif | |
128 | ||
129 | #ifdef ARCH_HAS_PREFETCHW | |
130 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
131 | do { \ | |
132 | if ((_page)->lru.prev != _base) { \ | |
133 | struct page *prev; \ | |
134 | \ | |
135 | prev = lru_to_page(&(_page->lru)); \ | |
136 | prefetchw(&prev->_field); \ | |
137 | } \ | |
138 | } while (0) | |
139 | #else | |
140 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
141 | #endif | |
142 | ||
143 | /* | |
144 | * From 0 .. 100. Higher means more swappy. | |
145 | */ | |
146 | int vm_swappiness = 60; | |
d0480be4 WSH |
147 | /* |
148 | * The total number of pages which are beyond the high watermark within all | |
149 | * zones. | |
150 | */ | |
151 | unsigned long vm_total_pages; | |
1da177e4 LT |
152 | |
153 | static LIST_HEAD(shrinker_list); | |
154 | static DECLARE_RWSEM(shrinker_rwsem); | |
155 | ||
c255a458 | 156 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
157 | static bool global_reclaim(struct scan_control *sc) |
158 | { | |
f16015fb | 159 | return !sc->target_mem_cgroup; |
89b5fae5 | 160 | } |
97c9341f TH |
161 | |
162 | /** | |
163 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
164 | * @sc: scan_control in question | |
165 | * | |
166 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
167 | * completely broken with the legacy memcg and direct stalling in | |
168 | * shrink_page_list() is used for throttling instead, which lacks all the | |
169 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
170 | * allocation and configurability. | |
171 | * | |
172 | * This function tests whether the vmscan currently in progress can assume | |
173 | * that the normal dirty throttling mechanism is operational. | |
174 | */ | |
175 | static bool sane_reclaim(struct scan_control *sc) | |
176 | { | |
177 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
178 | ||
179 | if (!memcg) | |
180 | return true; | |
181 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 182 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
183 | return true; |
184 | #endif | |
185 | return false; | |
186 | } | |
91a45470 | 187 | #else |
89b5fae5 JW |
188 | static bool global_reclaim(struct scan_control *sc) |
189 | { | |
190 | return true; | |
191 | } | |
97c9341f TH |
192 | |
193 | static bool sane_reclaim(struct scan_control *sc) | |
194 | { | |
195 | return true; | |
196 | } | |
91a45470 KH |
197 | #endif |
198 | ||
5a1c84b4 MG |
199 | /* |
200 | * This misses isolated pages which are not accounted for to save counters. | |
201 | * As the data only determines if reclaim or compaction continues, it is | |
202 | * not expected that isolated pages will be a dominating factor. | |
203 | */ | |
204 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
205 | { | |
206 | unsigned long nr; | |
207 | ||
208 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
209 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
210 | if (get_nr_swap_pages() > 0) | |
211 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
212 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
213 | ||
214 | return nr; | |
215 | } | |
216 | ||
599d0c95 MG |
217 | unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat) |
218 | { | |
219 | unsigned long nr; | |
220 | ||
221 | nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) + | |
222 | node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) + | |
223 | node_page_state_snapshot(pgdat, NR_ISOLATED_FILE); | |
6e543d57 LD |
224 | |
225 | if (get_nr_swap_pages() > 0) | |
599d0c95 MG |
226 | nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) + |
227 | node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) + | |
228 | node_page_state_snapshot(pgdat, NR_ISOLATED_ANON); | |
6e543d57 LD |
229 | |
230 | return nr; | |
231 | } | |
232 | ||
599d0c95 | 233 | bool pgdat_reclaimable(struct pglist_data *pgdat) |
6e543d57 | 234 | { |
599d0c95 MG |
235 | return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) < |
236 | pgdat_reclaimable_pages(pgdat) * 6; | |
6e543d57 LD |
237 | } |
238 | ||
fd538803 MH |
239 | /** |
240 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
241 | * @lruvec: lru vector | |
242 | * @lru: lru to use | |
243 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
244 | */ | |
245 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) | |
c9f299d9 | 246 | { |
fd538803 MH |
247 | unsigned long lru_size; |
248 | int zid; | |
249 | ||
c3c787e8 | 250 | if (!mem_cgroup_disabled()) |
fd538803 MH |
251 | lru_size = mem_cgroup_get_lru_size(lruvec, lru); |
252 | else | |
253 | lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); | |
a3d8e054 | 254 | |
fd538803 MH |
255 | for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { |
256 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; | |
257 | unsigned long size; | |
c9f299d9 | 258 | |
fd538803 MH |
259 | if (!managed_zone(zone)) |
260 | continue; | |
261 | ||
262 | if (!mem_cgroup_disabled()) | |
263 | size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); | |
264 | else | |
265 | size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], | |
266 | NR_ZONE_LRU_BASE + lru); | |
267 | lru_size -= min(size, lru_size); | |
268 | } | |
269 | ||
270 | return lru_size; | |
b4536f0c | 271 | |
b4536f0c MH |
272 | } |
273 | ||
1da177e4 | 274 | /* |
1d3d4437 | 275 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 276 | */ |
1d3d4437 | 277 | int register_shrinker(struct shrinker *shrinker) |
1da177e4 | 278 | { |
1d3d4437 GC |
279 | size_t size = sizeof(*shrinker->nr_deferred); |
280 | ||
1d3d4437 GC |
281 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
282 | size *= nr_node_ids; | |
283 | ||
284 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
285 | if (!shrinker->nr_deferred) | |
286 | return -ENOMEM; | |
287 | ||
8e1f936b RR |
288 | down_write(&shrinker_rwsem); |
289 | list_add_tail(&shrinker->list, &shrinker_list); | |
290 | up_write(&shrinker_rwsem); | |
1d3d4437 | 291 | return 0; |
1da177e4 | 292 | } |
8e1f936b | 293 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
294 | |
295 | /* | |
296 | * Remove one | |
297 | */ | |
8e1f936b | 298 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
299 | { |
300 | down_write(&shrinker_rwsem); | |
301 | list_del(&shrinker->list); | |
302 | up_write(&shrinker_rwsem); | |
ae393321 | 303 | kfree(shrinker->nr_deferred); |
1da177e4 | 304 | } |
8e1f936b | 305 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
306 | |
307 | #define SHRINK_BATCH 128 | |
1d3d4437 | 308 | |
cb731d6c VD |
309 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
310 | struct shrinker *shrinker, | |
311 | unsigned long nr_scanned, | |
312 | unsigned long nr_eligible) | |
1d3d4437 GC |
313 | { |
314 | unsigned long freed = 0; | |
315 | unsigned long long delta; | |
316 | long total_scan; | |
d5bc5fd3 | 317 | long freeable; |
1d3d4437 GC |
318 | long nr; |
319 | long new_nr; | |
320 | int nid = shrinkctl->nid; | |
321 | long batch_size = shrinker->batch ? shrinker->batch | |
322 | : SHRINK_BATCH; | |
5f33a080 | 323 | long scanned = 0, next_deferred; |
1d3d4437 | 324 | |
d5bc5fd3 VD |
325 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
326 | if (freeable == 0) | |
1d3d4437 GC |
327 | return 0; |
328 | ||
329 | /* | |
330 | * copy the current shrinker scan count into a local variable | |
331 | * and zero it so that other concurrent shrinker invocations | |
332 | * don't also do this scanning work. | |
333 | */ | |
334 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
335 | ||
336 | total_scan = nr; | |
6b4f7799 | 337 | delta = (4 * nr_scanned) / shrinker->seeks; |
d5bc5fd3 | 338 | delta *= freeable; |
6b4f7799 | 339 | do_div(delta, nr_eligible + 1); |
1d3d4437 GC |
340 | total_scan += delta; |
341 | if (total_scan < 0) { | |
8612c663 | 342 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
a0b02131 | 343 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 344 | total_scan = freeable; |
5f33a080 SL |
345 | next_deferred = nr; |
346 | } else | |
347 | next_deferred = total_scan; | |
1d3d4437 GC |
348 | |
349 | /* | |
350 | * We need to avoid excessive windup on filesystem shrinkers | |
351 | * due to large numbers of GFP_NOFS allocations causing the | |
352 | * shrinkers to return -1 all the time. This results in a large | |
353 | * nr being built up so when a shrink that can do some work | |
354 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 355 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
356 | * memory. |
357 | * | |
358 | * Hence only allow the shrinker to scan the entire cache when | |
359 | * a large delta change is calculated directly. | |
360 | */ | |
d5bc5fd3 VD |
361 | if (delta < freeable / 4) |
362 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
363 | |
364 | /* | |
365 | * Avoid risking looping forever due to too large nr value: | |
366 | * never try to free more than twice the estimate number of | |
367 | * freeable entries. | |
368 | */ | |
d5bc5fd3 VD |
369 | if (total_scan > freeable * 2) |
370 | total_scan = freeable * 2; | |
1d3d4437 GC |
371 | |
372 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
6b4f7799 JW |
373 | nr_scanned, nr_eligible, |
374 | freeable, delta, total_scan); | |
1d3d4437 | 375 | |
0b1fb40a VD |
376 | /* |
377 | * Normally, we should not scan less than batch_size objects in one | |
378 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
379 | * than batch_size objects in total and we are really tight on memory, | |
380 | * we will try to reclaim all available objects, otherwise we can end | |
381 | * up failing allocations although there are plenty of reclaimable | |
382 | * objects spread over several slabs with usage less than the | |
383 | * batch_size. | |
384 | * | |
385 | * We detect the "tight on memory" situations by looking at the total | |
386 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 387 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
388 | * scanning at high prio and therefore should try to reclaim as much as |
389 | * possible. | |
390 | */ | |
391 | while (total_scan >= batch_size || | |
d5bc5fd3 | 392 | total_scan >= freeable) { |
a0b02131 | 393 | unsigned long ret; |
0b1fb40a | 394 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 395 | |
0b1fb40a | 396 | shrinkctl->nr_to_scan = nr_to_scan; |
a0b02131 DC |
397 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
398 | if (ret == SHRINK_STOP) | |
399 | break; | |
400 | freed += ret; | |
1d3d4437 | 401 | |
0b1fb40a VD |
402 | count_vm_events(SLABS_SCANNED, nr_to_scan); |
403 | total_scan -= nr_to_scan; | |
5f33a080 | 404 | scanned += nr_to_scan; |
1d3d4437 GC |
405 | |
406 | cond_resched(); | |
407 | } | |
408 | ||
5f33a080 SL |
409 | if (next_deferred >= scanned) |
410 | next_deferred -= scanned; | |
411 | else | |
412 | next_deferred = 0; | |
1d3d4437 GC |
413 | /* |
414 | * move the unused scan count back into the shrinker in a | |
415 | * manner that handles concurrent updates. If we exhausted the | |
416 | * scan, there is no need to do an update. | |
417 | */ | |
5f33a080 SL |
418 | if (next_deferred > 0) |
419 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
420 | &shrinker->nr_deferred[nid]); |
421 | else | |
422 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
423 | ||
df9024a8 | 424 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 425 | return freed; |
1495f230 YH |
426 | } |
427 | ||
6b4f7799 | 428 | /** |
cb731d6c | 429 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
430 | * @gfp_mask: allocation context |
431 | * @nid: node whose slab caches to target | |
cb731d6c | 432 | * @memcg: memory cgroup whose slab caches to target |
6b4f7799 JW |
433 | * @nr_scanned: pressure numerator |
434 | * @nr_eligible: pressure denominator | |
1da177e4 | 435 | * |
6b4f7799 | 436 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 437 | * |
6b4f7799 JW |
438 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
439 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 440 | * |
cb731d6c VD |
441 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
442 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan | |
0fc9f58a VD |
443 | * objects from the memory cgroup specified. Otherwise, only unaware |
444 | * shrinkers are called. | |
cb731d6c | 445 | * |
6b4f7799 JW |
446 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of |
447 | * the available objects should be scanned. Page reclaim for example | |
448 | * passes the number of pages scanned and the number of pages on the | |
449 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned | |
450 | * when it encountered mapped pages. The ratio is further biased by | |
451 | * the ->seeks setting of the shrink function, which indicates the | |
452 | * cost to recreate an object relative to that of an LRU page. | |
b15e0905 | 453 | * |
6b4f7799 | 454 | * Returns the number of reclaimed slab objects. |
1da177e4 | 455 | */ |
cb731d6c VD |
456 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
457 | struct mem_cgroup *memcg, | |
458 | unsigned long nr_scanned, | |
459 | unsigned long nr_eligible) | |
1da177e4 LT |
460 | { |
461 | struct shrinker *shrinker; | |
24f7c6b9 | 462 | unsigned long freed = 0; |
1da177e4 | 463 | |
0fc9f58a | 464 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) |
cb731d6c VD |
465 | return 0; |
466 | ||
6b4f7799 JW |
467 | if (nr_scanned == 0) |
468 | nr_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 469 | |
f06590bd | 470 | if (!down_read_trylock(&shrinker_rwsem)) { |
24f7c6b9 DC |
471 | /* |
472 | * If we would return 0, our callers would understand that we | |
473 | * have nothing else to shrink and give up trying. By returning | |
474 | * 1 we keep it going and assume we'll be able to shrink next | |
475 | * time. | |
476 | */ | |
477 | freed = 1; | |
f06590bd MK |
478 | goto out; |
479 | } | |
1da177e4 LT |
480 | |
481 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
482 | struct shrink_control sc = { |
483 | .gfp_mask = gfp_mask, | |
484 | .nid = nid, | |
cb731d6c | 485 | .memcg = memcg, |
6b4f7799 | 486 | }; |
ec97097b | 487 | |
0fc9f58a VD |
488 | /* |
489 | * If kernel memory accounting is disabled, we ignore | |
490 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers | |
491 | * passing NULL for memcg. | |
492 | */ | |
493 | if (memcg_kmem_enabled() && | |
494 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
cb731d6c VD |
495 | continue; |
496 | ||
6b4f7799 JW |
497 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
498 | sc.nid = 0; | |
1da177e4 | 499 | |
cb731d6c | 500 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); |
1da177e4 | 501 | } |
6b4f7799 | 502 | |
1da177e4 | 503 | up_read(&shrinker_rwsem); |
f06590bd MK |
504 | out: |
505 | cond_resched(); | |
24f7c6b9 | 506 | return freed; |
1da177e4 LT |
507 | } |
508 | ||
cb731d6c VD |
509 | void drop_slab_node(int nid) |
510 | { | |
511 | unsigned long freed; | |
512 | ||
513 | do { | |
514 | struct mem_cgroup *memcg = NULL; | |
515 | ||
516 | freed = 0; | |
517 | do { | |
518 | freed += shrink_slab(GFP_KERNEL, nid, memcg, | |
519 | 1000, 1000); | |
520 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
521 | } while (freed > 10); | |
522 | } | |
523 | ||
524 | void drop_slab(void) | |
525 | { | |
526 | int nid; | |
527 | ||
528 | for_each_online_node(nid) | |
529 | drop_slab_node(nid); | |
530 | } | |
531 | ||
1da177e4 LT |
532 | static inline int is_page_cache_freeable(struct page *page) |
533 | { | |
ceddc3a5 JW |
534 | /* |
535 | * A freeable page cache page is referenced only by the caller | |
536 | * that isolated the page, the page cache radix tree and | |
537 | * optional buffer heads at page->private. | |
538 | */ | |
edcf4748 | 539 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
540 | } |
541 | ||
703c2708 | 542 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
1da177e4 | 543 | { |
930d9152 | 544 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 545 | return 1; |
703c2708 | 546 | if (!inode_write_congested(inode)) |
1da177e4 | 547 | return 1; |
703c2708 | 548 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
549 | return 1; |
550 | return 0; | |
551 | } | |
552 | ||
553 | /* | |
554 | * We detected a synchronous write error writing a page out. Probably | |
555 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
556 | * fsync(), msync() or close(). | |
557 | * | |
558 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
559 | * prevents it from being freed up. But we have a ref on the page and once | |
560 | * that page is locked, the mapping is pinned. | |
561 | * | |
562 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
563 | * __GFP_FS. | |
564 | */ | |
565 | static void handle_write_error(struct address_space *mapping, | |
566 | struct page *page, int error) | |
567 | { | |
7eaceacc | 568 | lock_page(page); |
3e9f45bd GC |
569 | if (page_mapping(page) == mapping) |
570 | mapping_set_error(mapping, error); | |
1da177e4 LT |
571 | unlock_page(page); |
572 | } | |
573 | ||
04e62a29 CL |
574 | /* possible outcome of pageout() */ |
575 | typedef enum { | |
576 | /* failed to write page out, page is locked */ | |
577 | PAGE_KEEP, | |
578 | /* move page to the active list, page is locked */ | |
579 | PAGE_ACTIVATE, | |
580 | /* page has been sent to the disk successfully, page is unlocked */ | |
581 | PAGE_SUCCESS, | |
582 | /* page is clean and locked */ | |
583 | PAGE_CLEAN, | |
584 | } pageout_t; | |
585 | ||
1da177e4 | 586 | /* |
1742f19f AM |
587 | * pageout is called by shrink_page_list() for each dirty page. |
588 | * Calls ->writepage(). | |
1da177e4 | 589 | */ |
c661b078 | 590 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 591 | struct scan_control *sc) |
1da177e4 LT |
592 | { |
593 | /* | |
594 | * If the page is dirty, only perform writeback if that write | |
595 | * will be non-blocking. To prevent this allocation from being | |
596 | * stalled by pagecache activity. But note that there may be | |
597 | * stalls if we need to run get_block(). We could test | |
598 | * PagePrivate for that. | |
599 | * | |
8174202b | 600 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
601 | * this page's queue, we can perform writeback even if that |
602 | * will block. | |
603 | * | |
604 | * If the page is swapcache, write it back even if that would | |
605 | * block, for some throttling. This happens by accident, because | |
606 | * swap_backing_dev_info is bust: it doesn't reflect the | |
607 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
608 | */ |
609 | if (!is_page_cache_freeable(page)) | |
610 | return PAGE_KEEP; | |
611 | if (!mapping) { | |
612 | /* | |
613 | * Some data journaling orphaned pages can have | |
614 | * page->mapping == NULL while being dirty with clean buffers. | |
615 | */ | |
266cf658 | 616 | if (page_has_private(page)) { |
1da177e4 LT |
617 | if (try_to_free_buffers(page)) { |
618 | ClearPageDirty(page); | |
b1de0d13 | 619 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
620 | return PAGE_CLEAN; |
621 | } | |
622 | } | |
623 | return PAGE_KEEP; | |
624 | } | |
625 | if (mapping->a_ops->writepage == NULL) | |
626 | return PAGE_ACTIVATE; | |
703c2708 | 627 | if (!may_write_to_inode(mapping->host, sc)) |
1da177e4 LT |
628 | return PAGE_KEEP; |
629 | ||
630 | if (clear_page_dirty_for_io(page)) { | |
631 | int res; | |
632 | struct writeback_control wbc = { | |
633 | .sync_mode = WB_SYNC_NONE, | |
634 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
635 | .range_start = 0, |
636 | .range_end = LLONG_MAX, | |
1da177e4 LT |
637 | .for_reclaim = 1, |
638 | }; | |
639 | ||
640 | SetPageReclaim(page); | |
641 | res = mapping->a_ops->writepage(page, &wbc); | |
642 | if (res < 0) | |
643 | handle_write_error(mapping, page, res); | |
994fc28c | 644 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
645 | ClearPageReclaim(page); |
646 | return PAGE_ACTIVATE; | |
647 | } | |
c661b078 | 648 | |
1da177e4 LT |
649 | if (!PageWriteback(page)) { |
650 | /* synchronous write or broken a_ops? */ | |
651 | ClearPageReclaim(page); | |
652 | } | |
3aa23851 | 653 | trace_mm_vmscan_writepage(page); |
c4a25635 | 654 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
655 | return PAGE_SUCCESS; |
656 | } | |
657 | ||
658 | return PAGE_CLEAN; | |
659 | } | |
660 | ||
a649fd92 | 661 | /* |
e286781d NP |
662 | * Same as remove_mapping, but if the page is removed from the mapping, it |
663 | * gets returned with a refcount of 0. | |
a649fd92 | 664 | */ |
a528910e JW |
665 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
666 | bool reclaimed) | |
49d2e9cc | 667 | { |
c4843a75 | 668 | unsigned long flags; |
c4843a75 | 669 | |
28e4d965 NP |
670 | BUG_ON(!PageLocked(page)); |
671 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 672 | |
c4843a75 | 673 | spin_lock_irqsave(&mapping->tree_lock, flags); |
49d2e9cc | 674 | /* |
0fd0e6b0 NP |
675 | * The non racy check for a busy page. |
676 | * | |
677 | * Must be careful with the order of the tests. When someone has | |
678 | * a ref to the page, it may be possible that they dirty it then | |
679 | * drop the reference. So if PageDirty is tested before page_count | |
680 | * here, then the following race may occur: | |
681 | * | |
682 | * get_user_pages(&page); | |
683 | * [user mapping goes away] | |
684 | * write_to(page); | |
685 | * !PageDirty(page) [good] | |
686 | * SetPageDirty(page); | |
687 | * put_page(page); | |
688 | * !page_count(page) [good, discard it] | |
689 | * | |
690 | * [oops, our write_to data is lost] | |
691 | * | |
692 | * Reversing the order of the tests ensures such a situation cannot | |
693 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 694 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
695 | * |
696 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
697 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 698 | */ |
fe896d18 | 699 | if (!page_ref_freeze(page, 2)) |
49d2e9cc | 700 | goto cannot_free; |
e286781d NP |
701 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
702 | if (unlikely(PageDirty(page))) { | |
fe896d18 | 703 | page_ref_unfreeze(page, 2); |
49d2e9cc | 704 | goto cannot_free; |
e286781d | 705 | } |
49d2e9cc CL |
706 | |
707 | if (PageSwapCache(page)) { | |
708 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 709 | mem_cgroup_swapout(page, swap); |
49d2e9cc | 710 | __delete_from_swap_cache(page); |
c4843a75 | 711 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
0a31bc97 | 712 | swapcache_free(swap); |
e286781d | 713 | } else { |
6072d13c | 714 | void (*freepage)(struct page *); |
a528910e | 715 | void *shadow = NULL; |
6072d13c LT |
716 | |
717 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
718 | /* |
719 | * Remember a shadow entry for reclaimed file cache in | |
720 | * order to detect refaults, thus thrashing, later on. | |
721 | * | |
722 | * But don't store shadows in an address space that is | |
723 | * already exiting. This is not just an optizimation, | |
724 | * inode reclaim needs to empty out the radix tree or | |
725 | * the nodes are lost. Don't plant shadows behind its | |
726 | * back. | |
f9fe48be RZ |
727 | * |
728 | * We also don't store shadows for DAX mappings because the | |
729 | * only page cache pages found in these are zero pages | |
730 | * covering holes, and because we don't want to mix DAX | |
731 | * exceptional entries and shadow exceptional entries in the | |
732 | * same page_tree. | |
a528910e JW |
733 | */ |
734 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 735 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
a528910e | 736 | shadow = workingset_eviction(mapping, page); |
62cccb8c | 737 | __delete_from_page_cache(page, shadow); |
c4843a75 | 738 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
6072d13c LT |
739 | |
740 | if (freepage != NULL) | |
741 | freepage(page); | |
49d2e9cc CL |
742 | } |
743 | ||
49d2e9cc CL |
744 | return 1; |
745 | ||
746 | cannot_free: | |
c4843a75 | 747 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
49d2e9cc CL |
748 | return 0; |
749 | } | |
750 | ||
e286781d NP |
751 | /* |
752 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
753 | * someone else has a ref on the page, abort and return 0. If it was | |
754 | * successfully detached, return 1. Assumes the caller has a single ref on | |
755 | * this page. | |
756 | */ | |
757 | int remove_mapping(struct address_space *mapping, struct page *page) | |
758 | { | |
a528910e | 759 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
760 | /* |
761 | * Unfreezing the refcount with 1 rather than 2 effectively | |
762 | * drops the pagecache ref for us without requiring another | |
763 | * atomic operation. | |
764 | */ | |
fe896d18 | 765 | page_ref_unfreeze(page, 1); |
e286781d NP |
766 | return 1; |
767 | } | |
768 | return 0; | |
769 | } | |
770 | ||
894bc310 LS |
771 | /** |
772 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
773 | * @page: page to be put back to appropriate lru list | |
774 | * | |
775 | * Add previously isolated @page to appropriate LRU list. | |
776 | * Page may still be unevictable for other reasons. | |
777 | * | |
778 | * lru_lock must not be held, interrupts must be enabled. | |
779 | */ | |
894bc310 LS |
780 | void putback_lru_page(struct page *page) |
781 | { | |
0ec3b74c | 782 | bool is_unevictable; |
bbfd28ee | 783 | int was_unevictable = PageUnevictable(page); |
894bc310 | 784 | |
309381fe | 785 | VM_BUG_ON_PAGE(PageLRU(page), page); |
894bc310 LS |
786 | |
787 | redo: | |
788 | ClearPageUnevictable(page); | |
789 | ||
39b5f29a | 790 | if (page_evictable(page)) { |
894bc310 LS |
791 | /* |
792 | * For evictable pages, we can use the cache. | |
793 | * In event of a race, worst case is we end up with an | |
794 | * unevictable page on [in]active list. | |
795 | * We know how to handle that. | |
796 | */ | |
0ec3b74c | 797 | is_unevictable = false; |
c53954a0 | 798 | lru_cache_add(page); |
894bc310 LS |
799 | } else { |
800 | /* | |
801 | * Put unevictable pages directly on zone's unevictable | |
802 | * list. | |
803 | */ | |
0ec3b74c | 804 | is_unevictable = true; |
894bc310 | 805 | add_page_to_unevictable_list(page); |
6a7b9548 | 806 | /* |
21ee9f39 MK |
807 | * When racing with an mlock or AS_UNEVICTABLE clearing |
808 | * (page is unlocked) make sure that if the other thread | |
809 | * does not observe our setting of PG_lru and fails | |
24513264 | 810 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 811 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
812 | * the page back to the evictable list. |
813 | * | |
21ee9f39 | 814 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
815 | */ |
816 | smp_mb(); | |
894bc310 | 817 | } |
894bc310 LS |
818 | |
819 | /* | |
820 | * page's status can change while we move it among lru. If an evictable | |
821 | * page is on unevictable list, it never be freed. To avoid that, | |
822 | * check after we added it to the list, again. | |
823 | */ | |
0ec3b74c | 824 | if (is_unevictable && page_evictable(page)) { |
894bc310 LS |
825 | if (!isolate_lru_page(page)) { |
826 | put_page(page); | |
827 | goto redo; | |
828 | } | |
829 | /* This means someone else dropped this page from LRU | |
830 | * So, it will be freed or putback to LRU again. There is | |
831 | * nothing to do here. | |
832 | */ | |
833 | } | |
834 | ||
0ec3b74c | 835 | if (was_unevictable && !is_unevictable) |
bbfd28ee | 836 | count_vm_event(UNEVICTABLE_PGRESCUED); |
0ec3b74c | 837 | else if (!was_unevictable && is_unevictable) |
bbfd28ee LS |
838 | count_vm_event(UNEVICTABLE_PGCULLED); |
839 | ||
894bc310 LS |
840 | put_page(page); /* drop ref from isolate */ |
841 | } | |
842 | ||
dfc8d636 JW |
843 | enum page_references { |
844 | PAGEREF_RECLAIM, | |
845 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 846 | PAGEREF_KEEP, |
dfc8d636 JW |
847 | PAGEREF_ACTIVATE, |
848 | }; | |
849 | ||
850 | static enum page_references page_check_references(struct page *page, | |
851 | struct scan_control *sc) | |
852 | { | |
64574746 | 853 | int referenced_ptes, referenced_page; |
dfc8d636 | 854 | unsigned long vm_flags; |
dfc8d636 | 855 | |
c3ac9a8a JW |
856 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
857 | &vm_flags); | |
64574746 | 858 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 859 | |
dfc8d636 JW |
860 | /* |
861 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
862 | * move the page to the unevictable list. | |
863 | */ | |
864 | if (vm_flags & VM_LOCKED) | |
865 | return PAGEREF_RECLAIM; | |
866 | ||
64574746 | 867 | if (referenced_ptes) { |
e4898273 | 868 | if (PageSwapBacked(page)) |
64574746 JW |
869 | return PAGEREF_ACTIVATE; |
870 | /* | |
871 | * All mapped pages start out with page table | |
872 | * references from the instantiating fault, so we need | |
873 | * to look twice if a mapped file page is used more | |
874 | * than once. | |
875 | * | |
876 | * Mark it and spare it for another trip around the | |
877 | * inactive list. Another page table reference will | |
878 | * lead to its activation. | |
879 | * | |
880 | * Note: the mark is set for activated pages as well | |
881 | * so that recently deactivated but used pages are | |
882 | * quickly recovered. | |
883 | */ | |
884 | SetPageReferenced(page); | |
885 | ||
34dbc67a | 886 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
887 | return PAGEREF_ACTIVATE; |
888 | ||
c909e993 KK |
889 | /* |
890 | * Activate file-backed executable pages after first usage. | |
891 | */ | |
892 | if (vm_flags & VM_EXEC) | |
893 | return PAGEREF_ACTIVATE; | |
894 | ||
64574746 JW |
895 | return PAGEREF_KEEP; |
896 | } | |
dfc8d636 JW |
897 | |
898 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 899 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
900 | return PAGEREF_RECLAIM_CLEAN; |
901 | ||
902 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
903 | } |
904 | ||
e2be15f6 MG |
905 | /* Check if a page is dirty or under writeback */ |
906 | static void page_check_dirty_writeback(struct page *page, | |
907 | bool *dirty, bool *writeback) | |
908 | { | |
b4597226 MG |
909 | struct address_space *mapping; |
910 | ||
e2be15f6 MG |
911 | /* |
912 | * Anonymous pages are not handled by flushers and must be written | |
913 | * from reclaim context. Do not stall reclaim based on them | |
914 | */ | |
915 | if (!page_is_file_cache(page)) { | |
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; |
854e9ed0 MK |
975 | bool lazyfree = false; |
976 | int ret = SWAP_SUCCESS; | |
1da177e4 LT |
977 | |
978 | cond_resched(); | |
979 | ||
980 | page = lru_to_page(page_list); | |
981 | list_del(&page->lru); | |
982 | ||
529ae9aa | 983 | if (!trylock_page(page)) |
1da177e4 LT |
984 | goto keep; |
985 | ||
309381fe | 986 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
987 | |
988 | sc->nr_scanned++; | |
80e43426 | 989 | |
39b5f29a | 990 | if (unlikely(!page_evictable(page))) |
b291f000 | 991 | goto cull_mlocked; |
894bc310 | 992 | |
a6dc60f8 | 993 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
994 | goto keep_locked; |
995 | ||
1da177e4 LT |
996 | /* Double the slab pressure for mapped and swapcache pages */ |
997 | if (page_mapped(page) || PageSwapCache(page)) | |
998 | sc->nr_scanned++; | |
999 | ||
c661b078 AW |
1000 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1001 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1002 | ||
e2be15f6 MG |
1003 | /* |
1004 | * The number of dirty pages determines if a zone is marked | |
1005 | * reclaim_congested which affects wait_iff_congested. kswapd | |
1006 | * will stall and start writing pages if the tail of the LRU | |
1007 | * is all dirty unqueued pages. | |
1008 | */ | |
1009 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1010 | if (dirty || writeback) | |
1011 | nr_dirty++; | |
1012 | ||
1013 | if (dirty && !writeback) | |
1014 | nr_unqueued_dirty++; | |
1015 | ||
d04e8acd MG |
1016 | /* |
1017 | * Treat this page as congested if the underlying BDI is or if | |
1018 | * pages are cycling through the LRU so quickly that the | |
1019 | * pages marked for immediate reclaim are making it to the | |
1020 | * end of the LRU a second time. | |
1021 | */ | |
e2be15f6 | 1022 | mapping = page_mapping(page); |
1da58ee2 | 1023 | if (((dirty || writeback) && mapping && |
703c2708 | 1024 | inode_write_congested(mapping->host)) || |
d04e8acd | 1025 | (writeback && PageReclaim(page))) |
e2be15f6 MG |
1026 | nr_congested++; |
1027 | ||
283aba9f MG |
1028 | /* |
1029 | * If a page at the tail of the LRU is under writeback, there | |
1030 | * are three cases to consider. | |
1031 | * | |
1032 | * 1) If reclaim is encountering an excessive number of pages | |
1033 | * under writeback and this page is both under writeback and | |
1034 | * PageReclaim then it indicates that pages are being queued | |
1035 | * for IO but are being recycled through the LRU before the | |
1036 | * IO can complete. Waiting on the page itself risks an | |
1037 | * indefinite stall if it is impossible to writeback the | |
1038 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1039 | * note that the LRU is being scanned too quickly and the |
1040 | * caller can stall after page list has been processed. | |
283aba9f | 1041 | * |
97c9341f | 1042 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1043 | * not marked for immediate reclaim, or the caller does not |
1044 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1045 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1046 | * reclaim and continue scanning. |
283aba9f | 1047 | * |
ecf5fc6e MH |
1048 | * Require may_enter_fs because we would wait on fs, which |
1049 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1050 | * enter reclaim, and deadlock if it waits on a page for |
1051 | * which it is needed to do the write (loop masks off | |
1052 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1053 | * would probably show more reasons. | |
1054 | * | |
7fadc820 | 1055 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1056 | * PageReclaim. memcg does not have any dirty pages |
1057 | * throttling so we could easily OOM just because too many | |
1058 | * pages are in writeback and there is nothing else to | |
1059 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1060 | * |
1061 | * In cases 1) and 2) we activate the pages to get them out of | |
1062 | * the way while we continue scanning for clean pages on the | |
1063 | * inactive list and refilling from the active list. The | |
1064 | * observation here is that waiting for disk writes is more | |
1065 | * expensive than potentially causing reloads down the line. | |
1066 | * Since they're marked for immediate reclaim, they won't put | |
1067 | * memory pressure on the cache working set any longer than it | |
1068 | * takes to write them to disk. | |
283aba9f | 1069 | */ |
c661b078 | 1070 | if (PageWriteback(page)) { |
283aba9f MG |
1071 | /* Case 1 above */ |
1072 | if (current_is_kswapd() && | |
1073 | PageReclaim(page) && | |
599d0c95 | 1074 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
b1a6f21e | 1075 | nr_immediate++; |
c55e8d03 | 1076 | goto activate_locked; |
283aba9f MG |
1077 | |
1078 | /* Case 2 above */ | |
97c9341f | 1079 | } else if (sane_reclaim(sc) || |
ecf5fc6e | 1080 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1081 | /* |
1082 | * This is slightly racy - end_page_writeback() | |
1083 | * might have just cleared PageReclaim, then | |
1084 | * setting PageReclaim here end up interpreted | |
1085 | * as PageReadahead - but that does not matter | |
1086 | * enough to care. What we do want is for this | |
1087 | * page to have PageReclaim set next time memcg | |
1088 | * reclaim reaches the tests above, so it will | |
1089 | * then wait_on_page_writeback() to avoid OOM; | |
1090 | * and it's also appropriate in global reclaim. | |
1091 | */ | |
1092 | SetPageReclaim(page); | |
e62e384e | 1093 | nr_writeback++; |
c55e8d03 | 1094 | goto activate_locked; |
283aba9f MG |
1095 | |
1096 | /* Case 3 above */ | |
1097 | } else { | |
7fadc820 | 1098 | unlock_page(page); |
283aba9f | 1099 | wait_on_page_writeback(page); |
7fadc820 HD |
1100 | /* then go back and try same page again */ |
1101 | list_add_tail(&page->lru, page_list); | |
1102 | continue; | |
e62e384e | 1103 | } |
c661b078 | 1104 | } |
1da177e4 | 1105 | |
02c6de8d MK |
1106 | if (!force_reclaim) |
1107 | references = page_check_references(page, sc); | |
1108 | ||
dfc8d636 JW |
1109 | switch (references) { |
1110 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1111 | goto activate_locked; |
64574746 | 1112 | case PAGEREF_KEEP: |
5bccd166 | 1113 | nr_ref_keep++; |
64574746 | 1114 | goto keep_locked; |
dfc8d636 JW |
1115 | case PAGEREF_RECLAIM: |
1116 | case PAGEREF_RECLAIM_CLEAN: | |
1117 | ; /* try to reclaim the page below */ | |
1118 | } | |
1da177e4 | 1119 | |
1da177e4 LT |
1120 | /* |
1121 | * Anonymous process memory has backing store? | |
1122 | * Try to allocate it some swap space here. | |
1123 | */ | |
b291f000 | 1124 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
1125 | if (!(sc->gfp_mask & __GFP_IO)) |
1126 | goto keep_locked; | |
5bc7b8ac | 1127 | if (!add_to_swap(page, page_list)) |
1da177e4 | 1128 | goto activate_locked; |
854e9ed0 | 1129 | lazyfree = true; |
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 | */ | |
1146 | if (page_mapped(page) && mapping) { | |
854e9ed0 MK |
1147 | switch (ret = try_to_unmap(page, lazyfree ? |
1148 | (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) : | |
1149 | (ttu_flags | TTU_BATCH_FLUSH))) { | |
1da177e4 | 1150 | case SWAP_FAIL: |
5bccd166 | 1151 | nr_unmap_fail++; |
1da177e4 LT |
1152 | goto activate_locked; |
1153 | case SWAP_AGAIN: | |
1154 | goto keep_locked; | |
b291f000 NP |
1155 | case SWAP_MLOCK: |
1156 | goto cull_mlocked; | |
854e9ed0 MK |
1157 | case SWAP_LZFREE: |
1158 | goto lazyfree; | |
1da177e4 LT |
1159 | case SWAP_SUCCESS: |
1160 | ; /* try to free the page below */ | |
1161 | } | |
1162 | } | |
1163 | ||
1164 | if (PageDirty(page)) { | |
ee72886d | 1165 | /* |
4eda4823 JW |
1166 | * Only kswapd can writeback filesystem pages |
1167 | * to avoid risk of stack overflow. But avoid | |
1168 | * injecting inefficient single-page IO into | |
1169 | * flusher writeback as much as possible: only | |
1170 | * write pages when we've encountered many | |
1171 | * dirty pages, and when we've already scanned | |
1172 | * the rest of the LRU for clean pages and see | |
1173 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1174 | */ |
f84f6e2b | 1175 | if (page_is_file_cache(page) && |
4eda4823 JW |
1176 | (!current_is_kswapd() || !PageReclaim(page) || |
1177 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1178 | /* |
1179 | * Immediately reclaim when written back. | |
1180 | * Similar in principal to deactivate_page() | |
1181 | * except we already have the page isolated | |
1182 | * and know it's dirty | |
1183 | */ | |
c4a25635 | 1184 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1185 | SetPageReclaim(page); |
1186 | ||
c55e8d03 | 1187 | goto activate_locked; |
ee72886d MG |
1188 | } |
1189 | ||
dfc8d636 | 1190 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1191 | goto keep_locked; |
4dd4b920 | 1192 | if (!may_enter_fs) |
1da177e4 | 1193 | goto keep_locked; |
52a8363e | 1194 | if (!sc->may_writepage) |
1da177e4 LT |
1195 | goto keep_locked; |
1196 | ||
d950c947 MG |
1197 | /* |
1198 | * Page is dirty. Flush the TLB if a writable entry | |
1199 | * potentially exists to avoid CPU writes after IO | |
1200 | * starts and then write it out here. | |
1201 | */ | |
1202 | try_to_unmap_flush_dirty(); | |
7d3579e8 | 1203 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1204 | case PAGE_KEEP: |
1205 | goto keep_locked; | |
1206 | case PAGE_ACTIVATE: | |
1207 | goto activate_locked; | |
1208 | case PAGE_SUCCESS: | |
7d3579e8 | 1209 | if (PageWriteback(page)) |
41ac1999 | 1210 | goto keep; |
7d3579e8 | 1211 | if (PageDirty(page)) |
1da177e4 | 1212 | goto keep; |
7d3579e8 | 1213 | |
1da177e4 LT |
1214 | /* |
1215 | * A synchronous write - probably a ramdisk. Go | |
1216 | * ahead and try to reclaim the page. | |
1217 | */ | |
529ae9aa | 1218 | if (!trylock_page(page)) |
1da177e4 LT |
1219 | goto keep; |
1220 | if (PageDirty(page) || PageWriteback(page)) | |
1221 | goto keep_locked; | |
1222 | mapping = page_mapping(page); | |
1223 | case PAGE_CLEAN: | |
1224 | ; /* try to free the page below */ | |
1225 | } | |
1226 | } | |
1227 | ||
1228 | /* | |
1229 | * If the page has buffers, try to free the buffer mappings | |
1230 | * associated with this page. If we succeed we try to free | |
1231 | * the page as well. | |
1232 | * | |
1233 | * We do this even if the page is PageDirty(). | |
1234 | * try_to_release_page() does not perform I/O, but it is | |
1235 | * possible for a page to have PageDirty set, but it is actually | |
1236 | * clean (all its buffers are clean). This happens if the | |
1237 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1238 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1239 | * try_to_release_page() will discover that cleanness and will |
1240 | * drop the buffers and mark the page clean - it can be freed. | |
1241 | * | |
1242 | * Rarely, pages can have buffers and no ->mapping. These are | |
1243 | * the pages which were not successfully invalidated in | |
1244 | * truncate_complete_page(). We try to drop those buffers here | |
1245 | * and if that worked, and the page is no longer mapped into | |
1246 | * process address space (page_count == 1) it can be freed. | |
1247 | * Otherwise, leave the page on the LRU so it is swappable. | |
1248 | */ | |
266cf658 | 1249 | if (page_has_private(page)) { |
1da177e4 LT |
1250 | if (!try_to_release_page(page, sc->gfp_mask)) |
1251 | goto activate_locked; | |
e286781d NP |
1252 | if (!mapping && page_count(page) == 1) { |
1253 | unlock_page(page); | |
1254 | if (put_page_testzero(page)) | |
1255 | goto free_it; | |
1256 | else { | |
1257 | /* | |
1258 | * rare race with speculative reference. | |
1259 | * the speculative reference will free | |
1260 | * this page shortly, so we may | |
1261 | * increment nr_reclaimed here (and | |
1262 | * leave it off the LRU). | |
1263 | */ | |
1264 | nr_reclaimed++; | |
1265 | continue; | |
1266 | } | |
1267 | } | |
1da177e4 LT |
1268 | } |
1269 | ||
854e9ed0 | 1270 | lazyfree: |
a528910e | 1271 | if (!mapping || !__remove_mapping(mapping, page, true)) |
49d2e9cc | 1272 | goto keep_locked; |
1da177e4 | 1273 | |
a978d6f5 NP |
1274 | /* |
1275 | * At this point, we have no other references and there is | |
1276 | * no way to pick any more up (removed from LRU, removed | |
1277 | * from pagecache). Can use non-atomic bitops now (and | |
1278 | * we obviously don't have to worry about waking up a process | |
1279 | * waiting on the page lock, because there are no references. | |
1280 | */ | |
48c935ad | 1281 | __ClearPageLocked(page); |
e286781d | 1282 | free_it: |
854e9ed0 MK |
1283 | if (ret == SWAP_LZFREE) |
1284 | count_vm_event(PGLAZYFREED); | |
1285 | ||
05ff5137 | 1286 | nr_reclaimed++; |
abe4c3b5 MG |
1287 | |
1288 | /* | |
1289 | * Is there need to periodically free_page_list? It would | |
1290 | * appear not as the counts should be low | |
1291 | */ | |
1292 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1293 | continue; |
1294 | ||
b291f000 | 1295 | cull_mlocked: |
63d6c5ad HD |
1296 | if (PageSwapCache(page)) |
1297 | try_to_free_swap(page); | |
b291f000 | 1298 | unlock_page(page); |
c54839a7 | 1299 | list_add(&page->lru, &ret_pages); |
b291f000 NP |
1300 | continue; |
1301 | ||
1da177e4 | 1302 | activate_locked: |
68a22394 | 1303 | /* Not a candidate for swapping, so reclaim swap space. */ |
5ccc5aba | 1304 | if (PageSwapCache(page) && mem_cgroup_swap_full(page)) |
a2c43eed | 1305 | try_to_free_swap(page); |
309381fe | 1306 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
1307 | SetPageActive(page); |
1308 | pgactivate++; | |
1309 | keep_locked: | |
1310 | unlock_page(page); | |
1311 | keep: | |
1312 | list_add(&page->lru, &ret_pages); | |
309381fe | 1313 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1314 | } |
abe4c3b5 | 1315 | |
747db954 | 1316 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1317 | try_to_unmap_flush(); |
b745bc85 | 1318 | free_hot_cold_page_list(&free_pages, true); |
abe4c3b5 | 1319 | |
1da177e4 | 1320 | list_splice(&ret_pages, page_list); |
f8891e5e | 1321 | count_vm_events(PGACTIVATE, pgactivate); |
0a31bc97 | 1322 | |
3c710c1a MH |
1323 | if (stat) { |
1324 | stat->nr_dirty = nr_dirty; | |
1325 | stat->nr_congested = nr_congested; | |
1326 | stat->nr_unqueued_dirty = nr_unqueued_dirty; | |
1327 | stat->nr_writeback = nr_writeback; | |
1328 | stat->nr_immediate = nr_immediate; | |
5bccd166 MH |
1329 | stat->nr_activate = pgactivate; |
1330 | stat->nr_ref_keep = nr_ref_keep; | |
1331 | stat->nr_unmap_fail = nr_unmap_fail; | |
3c710c1a | 1332 | } |
05ff5137 | 1333 | return nr_reclaimed; |
1da177e4 LT |
1334 | } |
1335 | ||
02c6de8d MK |
1336 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1337 | struct list_head *page_list) | |
1338 | { | |
1339 | struct scan_control sc = { | |
1340 | .gfp_mask = GFP_KERNEL, | |
1341 | .priority = DEF_PRIORITY, | |
1342 | .may_unmap = 1, | |
1343 | }; | |
3c710c1a | 1344 | unsigned long ret; |
02c6de8d MK |
1345 | struct page *page, *next; |
1346 | LIST_HEAD(clean_pages); | |
1347 | ||
1348 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1349 | if (page_is_file_cache(page) && !PageDirty(page) && |
b1123ea6 | 1350 | !__PageMovable(page)) { |
02c6de8d MK |
1351 | ClearPageActive(page); |
1352 | list_move(&page->lru, &clean_pages); | |
1353 | } | |
1354 | } | |
1355 | ||
599d0c95 | 1356 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
3c710c1a | 1357 | TTU_UNMAP|TTU_IGNORE_ACCESS, NULL, true); |
02c6de8d | 1358 | list_splice(&clean_pages, page_list); |
599d0c95 | 1359 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1360 | return ret; |
1361 | } | |
1362 | ||
5ad333eb AW |
1363 | /* |
1364 | * Attempt to remove the specified page from its LRU. Only take this page | |
1365 | * if it is of the appropriate PageActive status. Pages which are being | |
1366 | * freed elsewhere are also ignored. | |
1367 | * | |
1368 | * page: page to consider | |
1369 | * mode: one of the LRU isolation modes defined above | |
1370 | * | |
1371 | * returns 0 on success, -ve errno on failure. | |
1372 | */ | |
f3fd4a61 | 1373 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1374 | { |
1375 | int ret = -EINVAL; | |
1376 | ||
1377 | /* Only take pages on the LRU. */ | |
1378 | if (!PageLRU(page)) | |
1379 | return ret; | |
1380 | ||
e46a2879 MK |
1381 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1382 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1383 | return ret; |
1384 | ||
5ad333eb | 1385 | ret = -EBUSY; |
08e552c6 | 1386 | |
c8244935 MG |
1387 | /* |
1388 | * To minimise LRU disruption, the caller can indicate that it only | |
1389 | * wants to isolate pages it will be able to operate on without | |
1390 | * blocking - clean pages for the most part. | |
1391 | * | |
c8244935 MG |
1392 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1393 | * that it is possible to migrate without blocking | |
1394 | */ | |
1276ad68 | 1395 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
1396 | /* All the caller can do on PageWriteback is block */ |
1397 | if (PageWriteback(page)) | |
1398 | return ret; | |
1399 | ||
1400 | if (PageDirty(page)) { | |
1401 | struct address_space *mapping; | |
1402 | ||
c8244935 MG |
1403 | /* |
1404 | * Only pages without mappings or that have a | |
1405 | * ->migratepage callback are possible to migrate | |
1406 | * without blocking | |
1407 | */ | |
1408 | mapping = page_mapping(page); | |
1409 | if (mapping && !mapping->a_ops->migratepage) | |
1410 | return ret; | |
1411 | } | |
1412 | } | |
39deaf85 | 1413 | |
f80c0673 MK |
1414 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1415 | return ret; | |
1416 | ||
5ad333eb AW |
1417 | if (likely(get_page_unless_zero(page))) { |
1418 | /* | |
1419 | * Be careful not to clear PageLRU until after we're | |
1420 | * sure the page is not being freed elsewhere -- the | |
1421 | * page release code relies on it. | |
1422 | */ | |
1423 | ClearPageLRU(page); | |
1424 | ret = 0; | |
1425 | } | |
1426 | ||
1427 | return ret; | |
1428 | } | |
1429 | ||
7ee36a14 MG |
1430 | |
1431 | /* | |
1432 | * Update LRU sizes after isolating pages. The LRU size updates must | |
1433 | * be complete before mem_cgroup_update_lru_size due to a santity check. | |
1434 | */ | |
1435 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 1436 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 1437 | { |
7ee36a14 MG |
1438 | int zid; |
1439 | ||
7ee36a14 MG |
1440 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1441 | if (!nr_zone_taken[zid]) | |
1442 | continue; | |
1443 | ||
1444 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
7ee36a14 | 1445 | #ifdef CONFIG_MEMCG |
b4536f0c | 1446 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
7ee36a14 | 1447 | #endif |
b4536f0c MH |
1448 | } |
1449 | ||
7ee36a14 MG |
1450 | } |
1451 | ||
1da177e4 | 1452 | /* |
a52633d8 | 1453 | * zone_lru_lock is heavily contended. Some of the functions that |
1da177e4 LT |
1454 | * shrink the lists perform better by taking out a batch of pages |
1455 | * and working on them outside the LRU lock. | |
1456 | * | |
1457 | * For pagecache intensive workloads, this function is the hottest | |
1458 | * spot in the kernel (apart from copy_*_user functions). | |
1459 | * | |
1460 | * Appropriate locks must be held before calling this function. | |
1461 | * | |
1462 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1463 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1464 | * @dst: The temp list to put pages on to. |
f626012d | 1465 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1466 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1467 | * @mode: One of the LRU isolation modes |
3cb99451 | 1468 | * @lru: LRU list id for isolating |
1da177e4 LT |
1469 | * |
1470 | * returns how many pages were moved onto *@dst. | |
1471 | */ | |
69e05944 | 1472 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1473 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1474 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1475 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1476 | { |
75b00af7 | 1477 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1478 | unsigned long nr_taken = 0; |
599d0c95 | 1479 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 1480 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
1265e3a6 | 1481 | unsigned long skipped = 0, total_skipped = 0; |
599d0c95 | 1482 | unsigned long scan, nr_pages; |
b2e18757 | 1483 | LIST_HEAD(pages_skipped); |
1da177e4 | 1484 | |
0b802f10 | 1485 | for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan && |
d7f05528 | 1486 | !list_empty(src);) { |
5ad333eb | 1487 | struct page *page; |
5ad333eb | 1488 | |
1da177e4 LT |
1489 | page = lru_to_page(src); |
1490 | prefetchw_prev_lru_page(page, src, flags); | |
1491 | ||
309381fe | 1492 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1493 | |
b2e18757 MG |
1494 | if (page_zonenum(page) > sc->reclaim_idx) { |
1495 | list_move(&page->lru, &pages_skipped); | |
7cc30fcf | 1496 | nr_skipped[page_zonenum(page)]++; |
b2e18757 MG |
1497 | continue; |
1498 | } | |
1499 | ||
d7f05528 MG |
1500 | /* |
1501 | * Account for scanned and skipped separetly to avoid the pgdat | |
1502 | * being prematurely marked unreclaimable by pgdat_reclaimable. | |
1503 | */ | |
1504 | scan++; | |
1505 | ||
f3fd4a61 | 1506 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1507 | case 0: |
599d0c95 MG |
1508 | nr_pages = hpage_nr_pages(page); |
1509 | nr_taken += nr_pages; | |
1510 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1511 | list_move(&page->lru, dst); |
5ad333eb AW |
1512 | break; |
1513 | ||
1514 | case -EBUSY: | |
1515 | /* else it is being freed elsewhere */ | |
1516 | list_move(&page->lru, src); | |
1517 | continue; | |
46453a6e | 1518 | |
5ad333eb AW |
1519 | default: |
1520 | BUG(); | |
1521 | } | |
1da177e4 LT |
1522 | } |
1523 | ||
b2e18757 MG |
1524 | /* |
1525 | * Splice any skipped pages to the start of the LRU list. Note that | |
1526 | * this disrupts the LRU order when reclaiming for lower zones but | |
1527 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1528 | * scanning would soon rescan the same pages to skip and put the | |
1529 | * system at risk of premature OOM. | |
1530 | */ | |
7cc30fcf MG |
1531 | if (!list_empty(&pages_skipped)) { |
1532 | int zid; | |
1533 | ||
7cc30fcf MG |
1534 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1535 | if (!nr_skipped[zid]) | |
1536 | continue; | |
1537 | ||
1538 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 1539 | skipped += nr_skipped[zid]; |
7cc30fcf | 1540 | } |
d7f05528 MG |
1541 | |
1542 | /* | |
1543 | * Account skipped pages as a partial scan as the pgdat may be | |
1544 | * close to unreclaimable. If the LRU list is empty, account | |
1545 | * skipped pages as a full scan. | |
1546 | */ | |
1265e3a6 | 1547 | total_skipped = list_empty(src) ? skipped : skipped >> 2; |
d7f05528 MG |
1548 | |
1549 | list_splice(&pages_skipped, src); | |
7cc30fcf | 1550 | } |
1265e3a6 MH |
1551 | *nr_scanned = scan + total_skipped; |
1552 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, | |
32b3f297 | 1553 | scan, skipped, nr_taken, mode, lru); |
b4536f0c | 1554 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
1555 | return nr_taken; |
1556 | } | |
1557 | ||
62695a84 NP |
1558 | /** |
1559 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1560 | * @page: page to isolate from its LRU list | |
1561 | * | |
1562 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1563 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1564 | * | |
1565 | * Returns 0 if the page was removed from an LRU list. | |
1566 | * Returns -EBUSY if the page was not on an LRU list. | |
1567 | * | |
1568 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1569 | * the active list, it will have PageActive set. If it was found on |
1570 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1571 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1572 | * |
1573 | * The vmstat statistic corresponding to the list on which the page was | |
1574 | * found will be decremented. | |
1575 | * | |
1576 | * Restrictions: | |
1577 | * (1) Must be called with an elevated refcount on the page. This is a | |
1578 | * fundamentnal difference from isolate_lru_pages (which is called | |
1579 | * without a stable reference). | |
1580 | * (2) the lru_lock must not be held. | |
1581 | * (3) interrupts must be enabled. | |
1582 | */ | |
1583 | int isolate_lru_page(struct page *page) | |
1584 | { | |
1585 | int ret = -EBUSY; | |
1586 | ||
309381fe | 1587 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1588 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1589 | |
62695a84 NP |
1590 | if (PageLRU(page)) { |
1591 | struct zone *zone = page_zone(page); | |
fa9add64 | 1592 | struct lruvec *lruvec; |
62695a84 | 1593 | |
a52633d8 | 1594 | spin_lock_irq(zone_lru_lock(zone)); |
599d0c95 | 1595 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
0c917313 | 1596 | if (PageLRU(page)) { |
894bc310 | 1597 | int lru = page_lru(page); |
0c917313 | 1598 | get_page(page); |
62695a84 | 1599 | ClearPageLRU(page); |
fa9add64 HD |
1600 | del_page_from_lru_list(page, lruvec, lru); |
1601 | ret = 0; | |
62695a84 | 1602 | } |
a52633d8 | 1603 | spin_unlock_irq(zone_lru_lock(zone)); |
62695a84 NP |
1604 | } |
1605 | return ret; | |
1606 | } | |
1607 | ||
35cd7815 | 1608 | /* |
d37dd5dc FW |
1609 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1610 | * then get resheduled. When there are massive number of tasks doing page | |
1611 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1612 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1613 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1614 | */ |
599d0c95 | 1615 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1616 | struct scan_control *sc) |
1617 | { | |
1618 | unsigned long inactive, isolated; | |
1619 | ||
1620 | if (current_is_kswapd()) | |
1621 | return 0; | |
1622 | ||
97c9341f | 1623 | if (!sane_reclaim(sc)) |
35cd7815 RR |
1624 | return 0; |
1625 | ||
1626 | if (file) { | |
599d0c95 MG |
1627 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1628 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1629 | } else { |
599d0c95 MG |
1630 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1631 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1632 | } |
1633 | ||
3cf23841 FW |
1634 | /* |
1635 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1636 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1637 | * deadlock. | |
1638 | */ | |
d0164adc | 1639 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1640 | inactive >>= 3; |
1641 | ||
35cd7815 RR |
1642 | return isolated > inactive; |
1643 | } | |
1644 | ||
66635629 | 1645 | static noinline_for_stack void |
75b00af7 | 1646 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1647 | { |
27ac81d8 | 1648 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
599d0c95 | 1649 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3f79768f | 1650 | LIST_HEAD(pages_to_free); |
66635629 | 1651 | |
66635629 MG |
1652 | /* |
1653 | * Put back any unfreeable pages. | |
1654 | */ | |
66635629 | 1655 | while (!list_empty(page_list)) { |
3f79768f | 1656 | struct page *page = lru_to_page(page_list); |
66635629 | 1657 | int lru; |
3f79768f | 1658 | |
309381fe | 1659 | VM_BUG_ON_PAGE(PageLRU(page), page); |
66635629 | 1660 | list_del(&page->lru); |
39b5f29a | 1661 | if (unlikely(!page_evictable(page))) { |
599d0c95 | 1662 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1663 | putback_lru_page(page); |
599d0c95 | 1664 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1665 | continue; |
1666 | } | |
fa9add64 | 1667 | |
599d0c95 | 1668 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1669 | |
7a608572 | 1670 | SetPageLRU(page); |
66635629 | 1671 | lru = page_lru(page); |
fa9add64 HD |
1672 | add_page_to_lru_list(page, lruvec, lru); |
1673 | ||
66635629 MG |
1674 | if (is_active_lru(lru)) { |
1675 | int file = is_file_lru(lru); | |
9992af10 RR |
1676 | int numpages = hpage_nr_pages(page); |
1677 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1678 | } |
2bcf8879 HD |
1679 | if (put_page_testzero(page)) { |
1680 | __ClearPageLRU(page); | |
1681 | __ClearPageActive(page); | |
fa9add64 | 1682 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1683 | |
1684 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1685 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1686 | mem_cgroup_uncharge(page); |
2bcf8879 | 1687 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1688 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1689 | } else |
1690 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1691 | } |
1692 | } | |
66635629 | 1693 | |
3f79768f HD |
1694 | /* |
1695 | * To save our caller's stack, now use input list for pages to free. | |
1696 | */ | |
1697 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1698 | } |
1699 | ||
399ba0b9 N |
1700 | /* |
1701 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1702 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1703 | * In that case we should only throttle if the backing device it is | |
1704 | * writing to is congested. In other cases it is safe to throttle. | |
1705 | */ | |
1706 | static int current_may_throttle(void) | |
1707 | { | |
1708 | return !(current->flags & PF_LESS_THROTTLE) || | |
1709 | current->backing_dev_info == NULL || | |
1710 | bdi_write_congested(current->backing_dev_info); | |
1711 | } | |
1712 | ||
1da177e4 | 1713 | /* |
b2e18757 | 1714 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1715 | * of reclaimed pages |
1da177e4 | 1716 | */ |
66635629 | 1717 | static noinline_for_stack unsigned long |
1a93be0e | 1718 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1719 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1720 | { |
1721 | LIST_HEAD(page_list); | |
e247dbce | 1722 | unsigned long nr_scanned; |
05ff5137 | 1723 | unsigned long nr_reclaimed = 0; |
e247dbce | 1724 | unsigned long nr_taken; |
3c710c1a | 1725 | struct reclaim_stat stat = {}; |
f3fd4a61 | 1726 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1727 | int file = is_file_lru(lru); |
599d0c95 | 1728 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1729 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
78dc583d | 1730 | |
599d0c95 | 1731 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
58355c78 | 1732 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1733 | |
1734 | /* We are about to die and free our memory. Return now. */ | |
1735 | if (fatal_signal_pending(current)) | |
1736 | return SWAP_CLUSTER_MAX; | |
1737 | } | |
1738 | ||
1da177e4 | 1739 | lru_add_drain(); |
f80c0673 MK |
1740 | |
1741 | if (!sc->may_unmap) | |
61317289 | 1742 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1743 | |
599d0c95 | 1744 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1745 | |
5dc35979 KK |
1746 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1747 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc | 1748 | |
599d0c95 | 1749 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1750 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1751 | |
89b5fae5 | 1752 | if (global_reclaim(sc)) { |
599d0c95 | 1753 | __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned); |
e247dbce | 1754 | if (current_is_kswapd()) |
599d0c95 | 1755 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
e247dbce | 1756 | else |
599d0c95 | 1757 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
e247dbce | 1758 | } |
599d0c95 | 1759 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1760 | |
d563c050 | 1761 | if (nr_taken == 0) |
66635629 | 1762 | return 0; |
5ad333eb | 1763 | |
599d0c95 | 1764 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP, |
3c710c1a | 1765 | &stat, false); |
c661b078 | 1766 | |
599d0c95 | 1767 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 1768 | |
904249aa YH |
1769 | if (global_reclaim(sc)) { |
1770 | if (current_is_kswapd()) | |
599d0c95 | 1771 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
904249aa | 1772 | else |
599d0c95 | 1773 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
904249aa | 1774 | } |
a74609fa | 1775 | |
27ac81d8 | 1776 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1777 | |
599d0c95 | 1778 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 1779 | |
599d0c95 | 1780 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 1781 | |
747db954 | 1782 | mem_cgroup_uncharge_list(&page_list); |
b745bc85 | 1783 | free_hot_cold_page_list(&page_list, true); |
e11da5b4 | 1784 | |
92df3a72 MG |
1785 | /* |
1786 | * If reclaim is isolating dirty pages under writeback, it implies | |
1787 | * that the long-lived page allocation rate is exceeding the page | |
1788 | * laundering rate. Either the global limits are not being effective | |
1789 | * at throttling processes due to the page distribution throughout | |
1790 | * zones or there is heavy usage of a slow backing device. The | |
1791 | * only option is to throttle from reclaim context which is not ideal | |
1792 | * as there is no guarantee the dirtying process is throttled in the | |
1793 | * same way balance_dirty_pages() manages. | |
1794 | * | |
8e950282 MG |
1795 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1796 | * of pages under pages flagged for immediate reclaim and stall if any | |
1797 | * are encountered in the nr_immediate check below. | |
92df3a72 | 1798 | */ |
3c710c1a | 1799 | if (stat.nr_writeback && stat.nr_writeback == nr_taken) |
599d0c95 | 1800 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); |
92df3a72 | 1801 | |
d43006d5 | 1802 | /* |
97c9341f TH |
1803 | * Legacy memcg will stall in page writeback so avoid forcibly |
1804 | * stalling here. | |
d43006d5 | 1805 | */ |
97c9341f | 1806 | if (sane_reclaim(sc)) { |
8e950282 MG |
1807 | /* |
1808 | * Tag a zone as congested if all the dirty pages scanned were | |
1809 | * backed by a congested BDI and wait_iff_congested will stall. | |
1810 | */ | |
3c710c1a | 1811 | if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested) |
599d0c95 | 1812 | set_bit(PGDAT_CONGESTED, &pgdat->flags); |
8e950282 | 1813 | |
b1a6f21e MG |
1814 | /* |
1815 | * If dirty pages are scanned that are not queued for IO, it | |
726d061f JW |
1816 | * implies that flushers are not doing their job. This can |
1817 | * happen when memory pressure pushes dirty pages to the end of | |
1818 | * the LRU before the dirty limits are breached and the dirty | |
1819 | * data has expired. It can also happen when the proportion of | |
1820 | * dirty pages grows not through writes but through memory | |
1821 | * pressure reclaiming all the clean cache. And in some cases, | |
1822 | * the flushers simply cannot keep up with the allocation | |
1823 | * rate. Nudge the flusher threads in case they are asleep, but | |
1824 | * also allow kswapd to start writing pages during reclaim. | |
b1a6f21e | 1825 | */ |
726d061f JW |
1826 | if (stat.nr_unqueued_dirty == nr_taken) { |
1827 | wakeup_flusher_threads(0, WB_REASON_VMSCAN); | |
599d0c95 | 1828 | set_bit(PGDAT_DIRTY, &pgdat->flags); |
726d061f | 1829 | } |
b1a6f21e MG |
1830 | |
1831 | /* | |
b738d764 LT |
1832 | * If kswapd scans pages marked marked for immediate |
1833 | * reclaim and under writeback (nr_immediate), it implies | |
1834 | * that pages are cycling through the LRU faster than | |
b1a6f21e MG |
1835 | * they are written so also forcibly stall. |
1836 | */ | |
3c710c1a | 1837 | if (stat.nr_immediate && current_may_throttle()) |
b1a6f21e | 1838 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
e2be15f6 | 1839 | } |
d43006d5 | 1840 | |
8e950282 MG |
1841 | /* |
1842 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1843 | * is congested. Allow kswapd to continue until it starts encountering | |
1844 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1845 | */ | |
399ba0b9 N |
1846 | if (!sc->hibernation_mode && !current_is_kswapd() && |
1847 | current_may_throttle()) | |
599d0c95 | 1848 | wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); |
8e950282 | 1849 | |
599d0c95 MG |
1850 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
1851 | nr_scanned, nr_reclaimed, | |
5bccd166 MH |
1852 | stat.nr_dirty, stat.nr_writeback, |
1853 | stat.nr_congested, stat.nr_immediate, | |
1854 | stat.nr_activate, stat.nr_ref_keep, | |
1855 | stat.nr_unmap_fail, | |
ba5e9579 | 1856 | sc->priority, file); |
05ff5137 | 1857 | return nr_reclaimed; |
1da177e4 LT |
1858 | } |
1859 | ||
1860 | /* | |
1861 | * This moves pages from the active list to the inactive list. | |
1862 | * | |
1863 | * We move them the other way if the page is referenced by one or more | |
1864 | * processes, from rmap. | |
1865 | * | |
1866 | * If the pages are mostly unmapped, the processing is fast and it is | |
a52633d8 | 1867 | * appropriate to hold zone_lru_lock across the whole operation. But if |
1da177e4 | 1868 | * the pages are mapped, the processing is slow (page_referenced()) so we |
a52633d8 | 1869 | * should drop zone_lru_lock around each page. It's impossible to balance |
1da177e4 LT |
1870 | * this, so instead we remove the pages from the LRU while processing them. |
1871 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1872 | * nobody will play with that bit on a non-LRU page. | |
1873 | * | |
0139aa7b | 1874 | * The downside is that we have to touch page->_refcount against each page. |
1da177e4 | 1875 | * But we had to alter page->flags anyway. |
9d998b4f MH |
1876 | * |
1877 | * Returns the number of pages moved to the given lru. | |
1da177e4 | 1878 | */ |
1cfb419b | 1879 | |
9d998b4f | 1880 | static unsigned move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1881 | struct list_head *list, |
2bcf8879 | 1882 | struct list_head *pages_to_free, |
3eb4140f WF |
1883 | enum lru_list lru) |
1884 | { | |
599d0c95 | 1885 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3eb4140f | 1886 | struct page *page; |
fa9add64 | 1887 | int nr_pages; |
9d998b4f | 1888 | int nr_moved = 0; |
3eb4140f | 1889 | |
3eb4140f WF |
1890 | while (!list_empty(list)) { |
1891 | page = lru_to_page(list); | |
599d0c95 | 1892 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
3eb4140f | 1893 | |
309381fe | 1894 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3eb4140f WF |
1895 | SetPageLRU(page); |
1896 | ||
fa9add64 | 1897 | nr_pages = hpage_nr_pages(page); |
599d0c95 | 1898 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
925b7673 | 1899 | list_move(&page->lru, &lruvec->lists[lru]); |
3eb4140f | 1900 | |
2bcf8879 HD |
1901 | if (put_page_testzero(page)) { |
1902 | __ClearPageLRU(page); | |
1903 | __ClearPageActive(page); | |
fa9add64 | 1904 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1905 | |
1906 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1907 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1908 | mem_cgroup_uncharge(page); |
2bcf8879 | 1909 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1910 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1911 | } else |
1912 | list_add(&page->lru, pages_to_free); | |
9d998b4f MH |
1913 | } else { |
1914 | nr_moved += nr_pages; | |
3eb4140f WF |
1915 | } |
1916 | } | |
9d5e6a9f | 1917 | |
3eb4140f | 1918 | if (!is_active_lru(lru)) |
f0958906 | 1919 | __count_vm_events(PGDEACTIVATE, nr_moved); |
9d998b4f MH |
1920 | |
1921 | return nr_moved; | |
3eb4140f | 1922 | } |
1cfb419b | 1923 | |
f626012d | 1924 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1925 | struct lruvec *lruvec, |
f16015fb | 1926 | struct scan_control *sc, |
9e3b2f8c | 1927 | enum lru_list lru) |
1da177e4 | 1928 | { |
44c241f1 | 1929 | unsigned long nr_taken; |
f626012d | 1930 | unsigned long nr_scanned; |
6fe6b7e3 | 1931 | unsigned long vm_flags; |
1da177e4 | 1932 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1933 | LIST_HEAD(l_active); |
b69408e8 | 1934 | LIST_HEAD(l_inactive); |
1da177e4 | 1935 | struct page *page; |
1a93be0e | 1936 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
9d998b4f MH |
1937 | unsigned nr_deactivate, nr_activate; |
1938 | unsigned nr_rotated = 0; | |
f3fd4a61 | 1939 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1940 | int file = is_file_lru(lru); |
599d0c95 | 1941 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
1942 | |
1943 | lru_add_drain(); | |
f80c0673 MK |
1944 | |
1945 | if (!sc->may_unmap) | |
61317289 | 1946 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1947 | |
599d0c95 | 1948 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 1949 | |
5dc35979 KK |
1950 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1951 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1952 | |
599d0c95 | 1953 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 1954 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1955 | |
9d5e6a9f | 1956 | if (global_reclaim(sc)) |
599d0c95 MG |
1957 | __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned); |
1958 | __count_vm_events(PGREFILL, nr_scanned); | |
9d5e6a9f | 1959 | |
599d0c95 | 1960 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 1961 | |
1da177e4 LT |
1962 | while (!list_empty(&l_hold)) { |
1963 | cond_resched(); | |
1964 | page = lru_to_page(&l_hold); | |
1965 | list_del(&page->lru); | |
7e9cd484 | 1966 | |
39b5f29a | 1967 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
1968 | putback_lru_page(page); |
1969 | continue; | |
1970 | } | |
1971 | ||
cc715d99 MG |
1972 | if (unlikely(buffer_heads_over_limit)) { |
1973 | if (page_has_private(page) && trylock_page(page)) { | |
1974 | if (page_has_private(page)) | |
1975 | try_to_release_page(page, 0); | |
1976 | unlock_page(page); | |
1977 | } | |
1978 | } | |
1979 | ||
c3ac9a8a JW |
1980 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1981 | &vm_flags)) { | |
9992af10 | 1982 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1983 | /* |
1984 | * Identify referenced, file-backed active pages and | |
1985 | * give them one more trip around the active list. So | |
1986 | * that executable code get better chances to stay in | |
1987 | * memory under moderate memory pressure. Anon pages | |
1988 | * are not likely to be evicted by use-once streaming | |
1989 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1990 | * so we ignore them here. | |
1991 | */ | |
41e20983 | 1992 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1993 | list_add(&page->lru, &l_active); |
1994 | continue; | |
1995 | } | |
1996 | } | |
7e9cd484 | 1997 | |
5205e56e | 1998 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1999 | list_add(&page->lru, &l_inactive); |
2000 | } | |
2001 | ||
b555749a | 2002 | /* |
8cab4754 | 2003 | * Move pages back to the lru list. |
b555749a | 2004 | */ |
599d0c95 | 2005 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 2006 | /* |
8cab4754 WF |
2007 | * Count referenced pages from currently used mappings as rotated, |
2008 | * even though only some of them are actually re-activated. This | |
2009 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 2010 | * get_scan_count. |
7e9cd484 | 2011 | */ |
b7c46d15 | 2012 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 2013 | |
9d998b4f MH |
2014 | nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
2015 | nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
599d0c95 MG |
2016 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
2017 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 2018 | |
747db954 | 2019 | mem_cgroup_uncharge_list(&l_hold); |
b745bc85 | 2020 | free_hot_cold_page_list(&l_hold, true); |
9d998b4f MH |
2021 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2022 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2023 | } |
2024 | ||
59dc76b0 RR |
2025 | /* |
2026 | * The inactive anon list should be small enough that the VM never has | |
2027 | * to do too much work. | |
14797e23 | 2028 | * |
59dc76b0 RR |
2029 | * The inactive file list should be small enough to leave most memory |
2030 | * to the established workingset on the scan-resistant active list, | |
2031 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2032 | * |
59dc76b0 RR |
2033 | * Both inactive lists should also be large enough that each inactive |
2034 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2035 | * |
59dc76b0 RR |
2036 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
2037 | * on this LRU, maintained by the pageout code. A zone->inactive_ratio | |
2038 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. | |
56e49d21 | 2039 | * |
59dc76b0 RR |
2040 | * total target max |
2041 | * memory ratio inactive | |
2042 | * ------------------------------------- | |
2043 | * 10MB 1 5MB | |
2044 | * 100MB 1 50MB | |
2045 | * 1GB 3 250MB | |
2046 | * 10GB 10 0.9GB | |
2047 | * 100GB 31 3GB | |
2048 | * 1TB 101 10GB | |
2049 | * 10TB 320 32GB | |
56e49d21 | 2050 | */ |
f8d1a311 | 2051 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
dcec0b60 | 2052 | struct scan_control *sc, bool trace) |
56e49d21 | 2053 | { |
59dc76b0 | 2054 | unsigned long inactive_ratio; |
fd538803 MH |
2055 | unsigned long inactive, active; |
2056 | enum lru_list inactive_lru = file * LRU_FILE; | |
2057 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; | |
59dc76b0 | 2058 | unsigned long gb; |
e3790144 | 2059 | |
59dc76b0 RR |
2060 | /* |
2061 | * If we don't have swap space, anonymous page deactivation | |
2062 | * is pointless. | |
2063 | */ | |
2064 | if (!file && !total_swap_pages) | |
2065 | return false; | |
56e49d21 | 2066 | |
fd538803 MH |
2067 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
2068 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); | |
f8d1a311 | 2069 | |
59dc76b0 RR |
2070 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
2071 | if (gb) | |
2072 | inactive_ratio = int_sqrt(10 * gb); | |
b39415b2 | 2073 | else |
59dc76b0 RR |
2074 | inactive_ratio = 1; |
2075 | ||
dcec0b60 | 2076 | if (trace) |
fd538803 | 2077 | trace_mm_vmscan_inactive_list_is_low(lruvec_pgdat(lruvec)->node_id, |
dcec0b60 | 2078 | sc->reclaim_idx, |
fd538803 MH |
2079 | lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, |
2080 | lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, | |
2081 | inactive_ratio, file); | |
2082 | ||
59dc76b0 | 2083 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2084 | } |
2085 | ||
4f98a2fe | 2086 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
1a93be0e | 2087 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 2088 | { |
b39415b2 | 2089 | if (is_active_lru(lru)) { |
dcec0b60 | 2090 | if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true)) |
1a93be0e | 2091 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
2092 | return 0; |
2093 | } | |
2094 | ||
1a93be0e | 2095 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
2096 | } |
2097 | ||
9a265114 JW |
2098 | enum scan_balance { |
2099 | SCAN_EQUAL, | |
2100 | SCAN_FRACT, | |
2101 | SCAN_ANON, | |
2102 | SCAN_FILE, | |
2103 | }; | |
2104 | ||
4f98a2fe RR |
2105 | /* |
2106 | * Determine how aggressively the anon and file LRU lists should be | |
2107 | * scanned. The relative value of each set of LRU lists is determined | |
2108 | * by looking at the fraction of the pages scanned we did rotate back | |
2109 | * onto the active list instead of evict. | |
2110 | * | |
be7bd59d WL |
2111 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2112 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2113 | */ |
33377678 | 2114 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
6b4f7799 JW |
2115 | struct scan_control *sc, unsigned long *nr, |
2116 | unsigned long *lru_pages) | |
4f98a2fe | 2117 | { |
33377678 | 2118 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
2119 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2120 | u64 fraction[2]; | |
2121 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 2122 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 2123 | unsigned long anon_prio, file_prio; |
9a265114 | 2124 | enum scan_balance scan_balance; |
0bf1457f | 2125 | unsigned long anon, file; |
9a265114 | 2126 | bool force_scan = false; |
4f98a2fe | 2127 | unsigned long ap, fp; |
4111304d | 2128 | enum lru_list lru; |
6f04f48d SS |
2129 | bool some_scanned; |
2130 | int pass; | |
246e87a9 | 2131 | |
f11c0ca5 JW |
2132 | /* |
2133 | * If the zone or memcg is small, nr[l] can be 0. This | |
2134 | * results in no scanning on this priority and a potential | |
2135 | * priority drop. Global direct reclaim can go to the next | |
2136 | * zone and tends to have no problems. Global kswapd is for | |
2137 | * zone balancing and it needs to scan a minimum amount. When | |
2138 | * reclaiming for a memcg, a priority drop can cause high | |
2139 | * latencies, so it's better to scan a minimum amount there as | |
2140 | * well. | |
2141 | */ | |
90cbc250 | 2142 | if (current_is_kswapd()) { |
599d0c95 | 2143 | if (!pgdat_reclaimable(pgdat)) |
90cbc250 | 2144 | force_scan = true; |
eb01aaab | 2145 | if (!mem_cgroup_online(memcg)) |
90cbc250 VD |
2146 | force_scan = true; |
2147 | } | |
89b5fae5 | 2148 | if (!global_reclaim(sc)) |
a4d3e9e7 | 2149 | force_scan = true; |
76a33fc3 SL |
2150 | |
2151 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2152 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2153 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2154 | goto out; |
2155 | } | |
4f98a2fe | 2156 | |
10316b31 JW |
2157 | /* |
2158 | * Global reclaim will swap to prevent OOM even with no | |
2159 | * swappiness, but memcg users want to use this knob to | |
2160 | * disable swapping for individual groups completely when | |
2161 | * using the memory controller's swap limit feature would be | |
2162 | * too expensive. | |
2163 | */ | |
02695175 | 2164 | if (!global_reclaim(sc) && !swappiness) { |
9a265114 | 2165 | scan_balance = SCAN_FILE; |
10316b31 JW |
2166 | goto out; |
2167 | } | |
2168 | ||
2169 | /* | |
2170 | * Do not apply any pressure balancing cleverness when the | |
2171 | * system is close to OOM, scan both anon and file equally | |
2172 | * (unless the swappiness setting disagrees with swapping). | |
2173 | */ | |
02695175 | 2174 | if (!sc->priority && swappiness) { |
9a265114 | 2175 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2176 | goto out; |
2177 | } | |
2178 | ||
62376251 JW |
2179 | /* |
2180 | * Prevent the reclaimer from falling into the cache trap: as | |
2181 | * cache pages start out inactive, every cache fault will tip | |
2182 | * the scan balance towards the file LRU. And as the file LRU | |
2183 | * shrinks, so does the window for rotation from references. | |
2184 | * This means we have a runaway feedback loop where a tiny | |
2185 | * thrashing file LRU becomes infinitely more attractive than | |
2186 | * anon pages. Try to detect this based on file LRU size. | |
2187 | */ | |
2188 | if (global_reclaim(sc)) { | |
599d0c95 MG |
2189 | unsigned long pgdatfile; |
2190 | unsigned long pgdatfree; | |
2191 | int z; | |
2192 | unsigned long total_high_wmark = 0; | |
2ab051e1 | 2193 | |
599d0c95 MG |
2194 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2195 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2196 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2197 | ||
2198 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2199 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2200 | if (!managed_zone(zone)) |
599d0c95 MG |
2201 | continue; |
2202 | ||
2203 | total_high_wmark += high_wmark_pages(zone); | |
2204 | } | |
62376251 | 2205 | |
599d0c95 | 2206 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
62376251 JW |
2207 | scan_balance = SCAN_ANON; |
2208 | goto out; | |
2209 | } | |
2210 | } | |
2211 | ||
7c5bd705 | 2212 | /* |
316bda0e VD |
2213 | * If there is enough inactive page cache, i.e. if the size of the |
2214 | * inactive list is greater than that of the active list *and* the | |
2215 | * inactive list actually has some pages to scan on this priority, we | |
2216 | * do not reclaim anything from the anonymous working set right now. | |
2217 | * Without the second condition we could end up never scanning an | |
2218 | * lruvec even if it has plenty of old anonymous pages unless the | |
2219 | * system is under heavy pressure. | |
7c5bd705 | 2220 | */ |
dcec0b60 | 2221 | if (!inactive_list_is_low(lruvec, true, sc, false) && |
71ab6cfe | 2222 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
9a265114 | 2223 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2224 | goto out; |
2225 | } | |
2226 | ||
9a265114 JW |
2227 | scan_balance = SCAN_FRACT; |
2228 | ||
58c37f6e KM |
2229 | /* |
2230 | * With swappiness at 100, anonymous and file have the same priority. | |
2231 | * This scanning priority is essentially the inverse of IO cost. | |
2232 | */ | |
02695175 | 2233 | anon_prio = swappiness; |
75b00af7 | 2234 | file_prio = 200 - anon_prio; |
58c37f6e | 2235 | |
4f98a2fe RR |
2236 | /* |
2237 | * OK, so we have swap space and a fair amount of page cache | |
2238 | * pages. We use the recently rotated / recently scanned | |
2239 | * ratios to determine how valuable each cache is. | |
2240 | * | |
2241 | * Because workloads change over time (and to avoid overflow) | |
2242 | * we keep these statistics as a floating average, which ends | |
2243 | * up weighing recent references more than old ones. | |
2244 | * | |
2245 | * anon in [0], file in [1] | |
2246 | */ | |
2ab051e1 | 2247 | |
fd538803 MH |
2248 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
2249 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); | |
2250 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + | |
2251 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); | |
2ab051e1 | 2252 | |
599d0c95 | 2253 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2254 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2255 | reclaim_stat->recent_scanned[0] /= 2; |
2256 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2257 | } |
2258 | ||
6e901571 | 2259 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2260 | reclaim_stat->recent_scanned[1] /= 2; |
2261 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2262 | } |
2263 | ||
4f98a2fe | 2264 | /* |
00d8089c RR |
2265 | * The amount of pressure on anon vs file pages is inversely |
2266 | * proportional to the fraction of recently scanned pages on | |
2267 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2268 | */ |
fe35004f | 2269 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2270 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2271 | |
fe35004f | 2272 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2273 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2274 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2275 | |
76a33fc3 SL |
2276 | fraction[0] = ap; |
2277 | fraction[1] = fp; | |
2278 | denominator = ap + fp + 1; | |
2279 | out: | |
6f04f48d SS |
2280 | some_scanned = false; |
2281 | /* Only use force_scan on second pass. */ | |
2282 | for (pass = 0; !some_scanned && pass < 2; pass++) { | |
6b4f7799 | 2283 | *lru_pages = 0; |
6f04f48d SS |
2284 | for_each_evictable_lru(lru) { |
2285 | int file = is_file_lru(lru); | |
2286 | unsigned long size; | |
2287 | unsigned long scan; | |
6e08a369 | 2288 | |
71ab6cfe | 2289 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); |
6f04f48d | 2290 | scan = size >> sc->priority; |
9a265114 | 2291 | |
6f04f48d SS |
2292 | if (!scan && pass && force_scan) |
2293 | scan = min(size, SWAP_CLUSTER_MAX); | |
9a265114 | 2294 | |
6f04f48d SS |
2295 | switch (scan_balance) { |
2296 | case SCAN_EQUAL: | |
2297 | /* Scan lists relative to size */ | |
2298 | break; | |
2299 | case SCAN_FRACT: | |
2300 | /* | |
2301 | * Scan types proportional to swappiness and | |
2302 | * their relative recent reclaim efficiency. | |
2303 | */ | |
2304 | scan = div64_u64(scan * fraction[file], | |
2305 | denominator); | |
2306 | break; | |
2307 | case SCAN_FILE: | |
2308 | case SCAN_ANON: | |
2309 | /* Scan one type exclusively */ | |
6b4f7799 JW |
2310 | if ((scan_balance == SCAN_FILE) != file) { |
2311 | size = 0; | |
6f04f48d | 2312 | scan = 0; |
6b4f7799 | 2313 | } |
6f04f48d SS |
2314 | break; |
2315 | default: | |
2316 | /* Look ma, no brain */ | |
2317 | BUG(); | |
2318 | } | |
6b4f7799 JW |
2319 | |
2320 | *lru_pages += size; | |
6f04f48d | 2321 | nr[lru] = scan; |
6b4f7799 | 2322 | |
9a265114 | 2323 | /* |
6f04f48d SS |
2324 | * Skip the second pass and don't force_scan, |
2325 | * if we found something to scan. | |
9a265114 | 2326 | */ |
6f04f48d | 2327 | some_scanned |= !!scan; |
9a265114 | 2328 | } |
76a33fc3 | 2329 | } |
6e08a369 | 2330 | } |
4f98a2fe | 2331 | |
9b4f98cd | 2332 | /* |
a9dd0a83 | 2333 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. |
9b4f98cd | 2334 | */ |
a9dd0a83 | 2335 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, |
33377678 | 2336 | struct scan_control *sc, unsigned long *lru_pages) |
9b4f98cd | 2337 | { |
ef8f2327 | 2338 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
9b4f98cd | 2339 | unsigned long nr[NR_LRU_LISTS]; |
e82e0561 | 2340 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2341 | unsigned long nr_to_scan; |
2342 | enum lru_list lru; | |
2343 | unsigned long nr_reclaimed = 0; | |
2344 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2345 | struct blk_plug plug; | |
1a501907 | 2346 | bool scan_adjusted; |
9b4f98cd | 2347 | |
33377678 | 2348 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
9b4f98cd | 2349 | |
e82e0561 MG |
2350 | /* Record the original scan target for proportional adjustments later */ |
2351 | memcpy(targets, nr, sizeof(nr)); | |
2352 | ||
1a501907 MG |
2353 | /* |
2354 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2355 | * event that can occur when there is little memory pressure e.g. | |
2356 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2357 | * when the requested number of pages are reclaimed when scanning at | |
2358 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2359 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2360 | * do a batch of work at once. For memcg reclaim one check is made to | |
2361 | * abort proportional reclaim if either the file or anon lru has already | |
2362 | * dropped to zero at the first pass. | |
2363 | */ | |
2364 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2365 | sc->priority == DEF_PRIORITY); | |
2366 | ||
9b4f98cd JW |
2367 | blk_start_plug(&plug); |
2368 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2369 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2370 | unsigned long nr_anon, nr_file, percentage; |
2371 | unsigned long nr_scanned; | |
2372 | ||
9b4f98cd JW |
2373 | for_each_evictable_lru(lru) { |
2374 | if (nr[lru]) { | |
2375 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2376 | nr[lru] -= nr_to_scan; | |
2377 | ||
2378 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2379 | lruvec, sc); | |
2380 | } | |
2381 | } | |
e82e0561 | 2382 | |
bd041733 MH |
2383 | cond_resched(); |
2384 | ||
e82e0561 MG |
2385 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2386 | continue; | |
2387 | ||
e82e0561 MG |
2388 | /* |
2389 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2390 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2391 | * proportionally what was requested by get_scan_count(). We |
2392 | * stop reclaiming one LRU and reduce the amount scanning | |
2393 | * proportional to the original scan target. | |
2394 | */ | |
2395 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2396 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2397 | ||
1a501907 MG |
2398 | /* |
2399 | * It's just vindictive to attack the larger once the smaller | |
2400 | * has gone to zero. And given the way we stop scanning the | |
2401 | * smaller below, this makes sure that we only make one nudge | |
2402 | * towards proportionality once we've got nr_to_reclaim. | |
2403 | */ | |
2404 | if (!nr_file || !nr_anon) | |
2405 | break; | |
2406 | ||
e82e0561 MG |
2407 | if (nr_file > nr_anon) { |
2408 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2409 | targets[LRU_ACTIVE_ANON] + 1; | |
2410 | lru = LRU_BASE; | |
2411 | percentage = nr_anon * 100 / scan_target; | |
2412 | } else { | |
2413 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2414 | targets[LRU_ACTIVE_FILE] + 1; | |
2415 | lru = LRU_FILE; | |
2416 | percentage = nr_file * 100 / scan_target; | |
2417 | } | |
2418 | ||
2419 | /* Stop scanning the smaller of the LRU */ | |
2420 | nr[lru] = 0; | |
2421 | nr[lru + LRU_ACTIVE] = 0; | |
2422 | ||
2423 | /* | |
2424 | * Recalculate the other LRU scan count based on its original | |
2425 | * scan target and the percentage scanning already complete | |
2426 | */ | |
2427 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2428 | nr_scanned = targets[lru] - nr[lru]; | |
2429 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2430 | nr[lru] -= min(nr[lru], nr_scanned); | |
2431 | ||
2432 | lru += LRU_ACTIVE; | |
2433 | nr_scanned = targets[lru] - nr[lru]; | |
2434 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2435 | nr[lru] -= min(nr[lru], nr_scanned); | |
2436 | ||
2437 | scan_adjusted = true; | |
9b4f98cd JW |
2438 | } |
2439 | blk_finish_plug(&plug); | |
2440 | sc->nr_reclaimed += nr_reclaimed; | |
2441 | ||
2442 | /* | |
2443 | * Even if we did not try to evict anon pages at all, we want to | |
2444 | * rebalance the anon lru active/inactive ratio. | |
2445 | */ | |
dcec0b60 | 2446 | if (inactive_list_is_low(lruvec, false, sc, true)) |
9b4f98cd JW |
2447 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2448 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
2449 | } |
2450 | ||
23b9da55 | 2451 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2452 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2453 | { |
d84da3f9 | 2454 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2455 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2456 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2457 | return true; |
2458 | ||
2459 | return false; | |
2460 | } | |
2461 | ||
3e7d3449 | 2462 | /* |
23b9da55 MG |
2463 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2464 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2465 | * true if more pages should be reclaimed such that when the page allocator | |
2466 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2467 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2468 | */ |
a9dd0a83 | 2469 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 MG |
2470 | unsigned long nr_reclaimed, |
2471 | unsigned long nr_scanned, | |
2472 | struct scan_control *sc) | |
2473 | { | |
2474 | unsigned long pages_for_compaction; | |
2475 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2476 | int z; |
3e7d3449 MG |
2477 | |
2478 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2479 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2480 | return false; |
2481 | ||
2876592f MG |
2482 | /* Consider stopping depending on scan and reclaim activity */ |
2483 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2484 | /* | |
2485 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2486 | * full LRU list has been scanned and we are still failing | |
2487 | * to reclaim pages. This full LRU scan is potentially | |
2488 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2489 | */ | |
2490 | if (!nr_reclaimed && !nr_scanned) | |
2491 | return false; | |
2492 | } else { | |
2493 | /* | |
2494 | * For non-__GFP_REPEAT allocations which can presumably | |
2495 | * fail without consequence, stop if we failed to reclaim | |
2496 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2497 | * pages that were scanned. This will return to the | |
2498 | * caller faster at the risk reclaim/compaction and | |
2499 | * the resulting allocation attempt fails | |
2500 | */ | |
2501 | if (!nr_reclaimed) | |
2502 | return false; | |
2503 | } | |
3e7d3449 MG |
2504 | |
2505 | /* | |
2506 | * If we have not reclaimed enough pages for compaction and the | |
2507 | * inactive lists are large enough, continue reclaiming | |
2508 | */ | |
9861a62c | 2509 | pages_for_compaction = compact_gap(sc->order); |
a9dd0a83 | 2510 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
ec8acf20 | 2511 | if (get_nr_swap_pages() > 0) |
a9dd0a83 | 2512 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
3e7d3449 MG |
2513 | if (sc->nr_reclaimed < pages_for_compaction && |
2514 | inactive_lru_pages > pages_for_compaction) | |
2515 | return true; | |
2516 | ||
2517 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
a9dd0a83 MG |
2518 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2519 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2520 | if (!managed_zone(zone)) |
a9dd0a83 MG |
2521 | continue; |
2522 | ||
2523 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 2524 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
2525 | case COMPACT_CONTINUE: |
2526 | return false; | |
2527 | default: | |
2528 | /* check next zone */ | |
2529 | ; | |
2530 | } | |
3e7d3449 | 2531 | } |
a9dd0a83 | 2532 | return true; |
3e7d3449 MG |
2533 | } |
2534 | ||
970a39a3 | 2535 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2536 | { |
cb731d6c | 2537 | struct reclaim_state *reclaim_state = current->reclaim_state; |
f0fdc5e8 | 2538 | unsigned long nr_reclaimed, nr_scanned; |
2344d7e4 | 2539 | bool reclaimable = false; |
1da177e4 | 2540 | |
9b4f98cd JW |
2541 | do { |
2542 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2543 | struct mem_cgroup_reclaim_cookie reclaim = { | |
ef8f2327 | 2544 | .pgdat = pgdat, |
9b4f98cd JW |
2545 | .priority = sc->priority, |
2546 | }; | |
a9dd0a83 | 2547 | unsigned long node_lru_pages = 0; |
694fbc0f | 2548 | struct mem_cgroup *memcg; |
3e7d3449 | 2549 | |
9b4f98cd JW |
2550 | nr_reclaimed = sc->nr_reclaimed; |
2551 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2552 | |
694fbc0f AM |
2553 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2554 | do { | |
6b4f7799 | 2555 | unsigned long lru_pages; |
8e8ae645 | 2556 | unsigned long reclaimed; |
cb731d6c | 2557 | unsigned long scanned; |
5660048c | 2558 | |
241994ed JW |
2559 | if (mem_cgroup_low(root, memcg)) { |
2560 | if (!sc->may_thrash) | |
2561 | continue; | |
2562 | mem_cgroup_events(memcg, MEMCG_LOW, 1); | |
2563 | } | |
2564 | ||
8e8ae645 | 2565 | reclaimed = sc->nr_reclaimed; |
cb731d6c | 2566 | scanned = sc->nr_scanned; |
f9be23d6 | 2567 | |
a9dd0a83 MG |
2568 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
2569 | node_lru_pages += lru_pages; | |
f16015fb | 2570 | |
b5afba29 | 2571 | if (memcg) |
a9dd0a83 | 2572 | shrink_slab(sc->gfp_mask, pgdat->node_id, |
cb731d6c VD |
2573 | memcg, sc->nr_scanned - scanned, |
2574 | lru_pages); | |
2575 | ||
8e8ae645 JW |
2576 | /* Record the group's reclaim efficiency */ |
2577 | vmpressure(sc->gfp_mask, memcg, false, | |
2578 | sc->nr_scanned - scanned, | |
2579 | sc->nr_reclaimed - reclaimed); | |
2580 | ||
9b4f98cd | 2581 | /* |
a394cb8e MH |
2582 | * Direct reclaim and kswapd have to scan all memory |
2583 | * cgroups to fulfill the overall scan target for the | |
a9dd0a83 | 2584 | * node. |
a394cb8e MH |
2585 | * |
2586 | * Limit reclaim, on the other hand, only cares about | |
2587 | * nr_to_reclaim pages to be reclaimed and it will | |
2588 | * retry with decreasing priority if one round over the | |
2589 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2590 | */ |
a394cb8e MH |
2591 | if (!global_reclaim(sc) && |
2592 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2593 | mem_cgroup_iter_break(root, memcg); |
2594 | break; | |
2595 | } | |
241994ed | 2596 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
70ddf637 | 2597 | |
6b4f7799 JW |
2598 | /* |
2599 | * Shrink the slab caches in the same proportion that | |
2600 | * the eligible LRU pages were scanned. | |
2601 | */ | |
b2e18757 | 2602 | if (global_reclaim(sc)) |
a9dd0a83 | 2603 | shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, |
cb731d6c | 2604 | sc->nr_scanned - nr_scanned, |
a9dd0a83 | 2605 | node_lru_pages); |
cb731d6c VD |
2606 | |
2607 | if (reclaim_state) { | |
2608 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2609 | reclaim_state->reclaimed_slab = 0; | |
6b4f7799 JW |
2610 | } |
2611 | ||
8e8ae645 JW |
2612 | /* Record the subtree's reclaim efficiency */ |
2613 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
70ddf637 AV |
2614 | sc->nr_scanned - nr_scanned, |
2615 | sc->nr_reclaimed - nr_reclaimed); | |
2616 | ||
2344d7e4 JW |
2617 | if (sc->nr_reclaimed - nr_reclaimed) |
2618 | reclaimable = true; | |
2619 | ||
a9dd0a83 | 2620 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
9b4f98cd | 2621 | sc->nr_scanned - nr_scanned, sc)); |
2344d7e4 JW |
2622 | |
2623 | return reclaimable; | |
f16015fb JW |
2624 | } |
2625 | ||
53853e2d | 2626 | /* |
fdd4c614 VB |
2627 | * Returns true if compaction should go ahead for a costly-order request, or |
2628 | * the allocation would already succeed without compaction. Return false if we | |
2629 | * should reclaim first. | |
53853e2d | 2630 | */ |
4f588331 | 2631 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2632 | { |
31483b6a | 2633 | unsigned long watermark; |
fdd4c614 | 2634 | enum compact_result suitable; |
fe4b1b24 | 2635 | |
fdd4c614 VB |
2636 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2637 | if (suitable == COMPACT_SUCCESS) | |
2638 | /* Allocation should succeed already. Don't reclaim. */ | |
2639 | return true; | |
2640 | if (suitable == COMPACT_SKIPPED) | |
2641 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2642 | return false; | |
fe4b1b24 | 2643 | |
53853e2d | 2644 | /* |
fdd4c614 VB |
2645 | * Compaction is already possible, but it takes time to run and there |
2646 | * are potentially other callers using the pages just freed. So proceed | |
2647 | * with reclaim to make a buffer of free pages available to give | |
2648 | * compaction a reasonable chance of completing and allocating the page. | |
2649 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2650 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2651 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 2652 | */ |
fdd4c614 | 2653 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 2654 | |
fdd4c614 | 2655 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
2656 | } |
2657 | ||
1da177e4 LT |
2658 | /* |
2659 | * This is the direct reclaim path, for page-allocating processes. We only | |
2660 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2661 | * request. | |
2662 | * | |
1da177e4 LT |
2663 | * If a zone is deemed to be full of pinned pages then just give it a light |
2664 | * scan then give up on it. | |
2665 | */ | |
0a0337e0 | 2666 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2667 | { |
dd1a239f | 2668 | struct zoneref *z; |
54a6eb5c | 2669 | struct zone *zone; |
0608f43d AM |
2670 | unsigned long nr_soft_reclaimed; |
2671 | unsigned long nr_soft_scanned; | |
619d0d76 | 2672 | gfp_t orig_mask; |
79dafcdc | 2673 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2674 | |
cc715d99 MG |
2675 | /* |
2676 | * If the number of buffer_heads in the machine exceeds the maximum | |
2677 | * allowed level, force direct reclaim to scan the highmem zone as | |
2678 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2679 | */ | |
619d0d76 | 2680 | orig_mask = sc->gfp_mask; |
b2e18757 | 2681 | if (buffer_heads_over_limit) { |
cc715d99 | 2682 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2683 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2684 | } |
cc715d99 | 2685 | |
d4debc66 | 2686 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2687 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
2688 | /* |
2689 | * Take care memory controller reclaiming has small influence | |
2690 | * to global LRU. | |
2691 | */ | |
89b5fae5 | 2692 | if (global_reclaim(sc)) { |
344736f2 VD |
2693 | if (!cpuset_zone_allowed(zone, |
2694 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2695 | continue; |
65ec02cb | 2696 | |
6e543d57 | 2697 | if (sc->priority != DEF_PRIORITY && |
599d0c95 | 2698 | !pgdat_reclaimable(zone->zone_pgdat)) |
1cfb419b | 2699 | continue; /* Let kswapd poll it */ |
0b06496a JW |
2700 | |
2701 | /* | |
2702 | * If we already have plenty of memory free for | |
2703 | * compaction in this zone, don't free any more. | |
2704 | * Even though compaction is invoked for any | |
2705 | * non-zero order, only frequent costly order | |
2706 | * reclamation is disruptive enough to become a | |
2707 | * noticeable problem, like transparent huge | |
2708 | * page allocations. | |
2709 | */ | |
2710 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2711 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2712 | compaction_ready(zone, sc)) { |
0b06496a JW |
2713 | sc->compaction_ready = true; |
2714 | continue; | |
e0887c19 | 2715 | } |
0b06496a | 2716 | |
79dafcdc MG |
2717 | /* |
2718 | * Shrink each node in the zonelist once. If the | |
2719 | * zonelist is ordered by zone (not the default) then a | |
2720 | * node may be shrunk multiple times but in that case | |
2721 | * the user prefers lower zones being preserved. | |
2722 | */ | |
2723 | if (zone->zone_pgdat == last_pgdat) | |
2724 | continue; | |
2725 | ||
0608f43d AM |
2726 | /* |
2727 | * This steals pages from memory cgroups over softlimit | |
2728 | * and returns the number of reclaimed pages and | |
2729 | * scanned pages. This works for global memory pressure | |
2730 | * and balancing, not for a memcg's limit. | |
2731 | */ | |
2732 | nr_soft_scanned = 0; | |
ef8f2327 | 2733 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2734 | sc->order, sc->gfp_mask, |
2735 | &nr_soft_scanned); | |
2736 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2737 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2738 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2739 | } |
408d8544 | 2740 | |
79dafcdc MG |
2741 | /* See comment about same check for global reclaim above */ |
2742 | if (zone->zone_pgdat == last_pgdat) | |
2743 | continue; | |
2744 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2745 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2746 | } |
e0c23279 | 2747 | |
619d0d76 WY |
2748 | /* |
2749 | * Restore to original mask to avoid the impact on the caller if we | |
2750 | * promoted it to __GFP_HIGHMEM. | |
2751 | */ | |
2752 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2753 | } |
4f98a2fe | 2754 | |
1da177e4 LT |
2755 | /* |
2756 | * This is the main entry point to direct page reclaim. | |
2757 | * | |
2758 | * If a full scan of the inactive list fails to free enough memory then we | |
2759 | * are "out of memory" and something needs to be killed. | |
2760 | * | |
2761 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2762 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2763 | * caller can't do much about. We kick the writeback threads and take explicit |
2764 | * naps in the hope that some of these pages can be written. But if the | |
2765 | * allocating task holds filesystem locks which prevent writeout this might not | |
2766 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2767 | * |
2768 | * returns: 0, if no pages reclaimed | |
2769 | * else, the number of pages reclaimed | |
1da177e4 | 2770 | */ |
dac1d27b | 2771 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2772 | struct scan_control *sc) |
1da177e4 | 2773 | { |
241994ed | 2774 | int initial_priority = sc->priority; |
241994ed | 2775 | retry: |
873b4771 KK |
2776 | delayacct_freepages_start(); |
2777 | ||
89b5fae5 | 2778 | if (global_reclaim(sc)) |
7cc30fcf | 2779 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 2780 | |
9e3b2f8c | 2781 | do { |
70ddf637 AV |
2782 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2783 | sc->priority); | |
66e1707b | 2784 | sc->nr_scanned = 0; |
0a0337e0 | 2785 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 2786 | |
bb21c7ce | 2787 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
2788 | break; |
2789 | ||
2790 | if (sc->compaction_ready) | |
2791 | break; | |
1da177e4 | 2792 | |
0e50ce3b MK |
2793 | /* |
2794 | * If we're getting trouble reclaiming, start doing | |
2795 | * writepage even in laptop mode. | |
2796 | */ | |
2797 | if (sc->priority < DEF_PRIORITY - 2) | |
2798 | sc->may_writepage = 1; | |
0b06496a | 2799 | } while (--sc->priority >= 0); |
bb21c7ce | 2800 | |
873b4771 KK |
2801 | delayacct_freepages_end(); |
2802 | ||
bb21c7ce KM |
2803 | if (sc->nr_reclaimed) |
2804 | return sc->nr_reclaimed; | |
2805 | ||
0cee34fd | 2806 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 2807 | if (sc->compaction_ready) |
7335084d MG |
2808 | return 1; |
2809 | ||
241994ed JW |
2810 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
2811 | if (!sc->may_thrash) { | |
2812 | sc->priority = initial_priority; | |
2813 | sc->may_thrash = 1; | |
2814 | goto retry; | |
2815 | } | |
2816 | ||
bb21c7ce | 2817 | return 0; |
1da177e4 LT |
2818 | } |
2819 | ||
5515061d MG |
2820 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) |
2821 | { | |
2822 | struct zone *zone; | |
2823 | unsigned long pfmemalloc_reserve = 0; | |
2824 | unsigned long free_pages = 0; | |
2825 | int i; | |
2826 | bool wmark_ok; | |
2827 | ||
2828 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2829 | zone = &pgdat->node_zones[i]; | |
6aa303de | 2830 | if (!managed_zone(zone) || |
599d0c95 | 2831 | pgdat_reclaimable_pages(pgdat) == 0) |
675becce MG |
2832 | continue; |
2833 | ||
5515061d MG |
2834 | pfmemalloc_reserve += min_wmark_pages(zone); |
2835 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2836 | } | |
2837 | ||
675becce MG |
2838 | /* If there are no reserves (unexpected config) then do not throttle */ |
2839 | if (!pfmemalloc_reserve) | |
2840 | return true; | |
2841 | ||
5515061d MG |
2842 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
2843 | ||
2844 | /* kswapd must be awake if processes are being throttled */ | |
2845 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
38087d9b | 2846 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
5515061d MG |
2847 | (enum zone_type)ZONE_NORMAL); |
2848 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2849 | } | |
2850 | ||
2851 | return wmark_ok; | |
2852 | } | |
2853 | ||
2854 | /* | |
2855 | * Throttle direct reclaimers if backing storage is backed by the network | |
2856 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2857 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2858 | * when the low watermark is reached. |
2859 | * | |
2860 | * Returns true if a fatal signal was delivered during throttling. If this | |
2861 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2862 | */ |
50694c28 | 2863 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2864 | nodemask_t *nodemask) |
2865 | { | |
675becce | 2866 | struct zoneref *z; |
5515061d | 2867 | struct zone *zone; |
675becce | 2868 | pg_data_t *pgdat = NULL; |
5515061d MG |
2869 | |
2870 | /* | |
2871 | * Kernel threads should not be throttled as they may be indirectly | |
2872 | * responsible for cleaning pages necessary for reclaim to make forward | |
2873 | * progress. kjournald for example may enter direct reclaim while | |
2874 | * committing a transaction where throttling it could forcing other | |
2875 | * processes to block on log_wait_commit(). | |
2876 | */ | |
2877 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2878 | goto out; |
2879 | ||
2880 | /* | |
2881 | * If a fatal signal is pending, this process should not throttle. | |
2882 | * It should return quickly so it can exit and free its memory | |
2883 | */ | |
2884 | if (fatal_signal_pending(current)) | |
2885 | goto out; | |
5515061d | 2886 | |
675becce MG |
2887 | /* |
2888 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2889 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2890 | * GFP_KERNEL will be required for allocating network buffers when | |
2891 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2892 | * | |
2893 | * Throttling is based on the first usable node and throttled processes | |
2894 | * wait on a queue until kswapd makes progress and wakes them. There | |
2895 | * is an affinity then between processes waking up and where reclaim | |
2896 | * progress has been made assuming the process wakes on the same node. | |
2897 | * More importantly, processes running on remote nodes will not compete | |
2898 | * for remote pfmemalloc reserves and processes on different nodes | |
2899 | * should make reasonable progress. | |
2900 | */ | |
2901 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 2902 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
2903 | if (zone_idx(zone) > ZONE_NORMAL) |
2904 | continue; | |
2905 | ||
2906 | /* Throttle based on the first usable node */ | |
2907 | pgdat = zone->zone_pgdat; | |
2908 | if (pfmemalloc_watermark_ok(pgdat)) | |
2909 | goto out; | |
2910 | break; | |
2911 | } | |
2912 | ||
2913 | /* If no zone was usable by the allocation flags then do not throttle */ | |
2914 | if (!pgdat) | |
50694c28 | 2915 | goto out; |
5515061d | 2916 | |
68243e76 MG |
2917 | /* Account for the throttling */ |
2918 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2919 | ||
5515061d MG |
2920 | /* |
2921 | * If the caller cannot enter the filesystem, it's possible that it | |
2922 | * is due to the caller holding an FS lock or performing a journal | |
2923 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2924 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2925 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2926 | * second before continuing. | |
2927 | */ | |
2928 | if (!(gfp_mask & __GFP_FS)) { | |
2929 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2930 | pfmemalloc_watermark_ok(pgdat), HZ); | |
50694c28 MG |
2931 | |
2932 | goto check_pending; | |
5515061d MG |
2933 | } |
2934 | ||
2935 | /* Throttle until kswapd wakes the process */ | |
2936 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2937 | pfmemalloc_watermark_ok(pgdat)); | |
50694c28 MG |
2938 | |
2939 | check_pending: | |
2940 | if (fatal_signal_pending(current)) | |
2941 | return true; | |
2942 | ||
2943 | out: | |
2944 | return false; | |
5515061d MG |
2945 | } |
2946 | ||
dac1d27b | 2947 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2948 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2949 | { |
33906bc5 | 2950 | unsigned long nr_reclaimed; |
66e1707b | 2951 | struct scan_control sc = { |
ee814fe2 | 2952 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
21caf2fc | 2953 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
b2e18757 | 2954 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
2955 | .order = order, |
2956 | .nodemask = nodemask, | |
2957 | .priority = DEF_PRIORITY, | |
66e1707b | 2958 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2959 | .may_unmap = 1, |
2e2e4259 | 2960 | .may_swap = 1, |
66e1707b BS |
2961 | }; |
2962 | ||
5515061d | 2963 | /* |
50694c28 MG |
2964 | * Do not enter reclaim if fatal signal was delivered while throttled. |
2965 | * 1 is returned so that the page allocator does not OOM kill at this | |
2966 | * point. | |
5515061d | 2967 | */ |
50694c28 | 2968 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) |
5515061d MG |
2969 | return 1; |
2970 | ||
33906bc5 MG |
2971 | trace_mm_vmscan_direct_reclaim_begin(order, |
2972 | sc.may_writepage, | |
e5146b12 MG |
2973 | gfp_mask, |
2974 | sc.reclaim_idx); | |
33906bc5 | 2975 | |
3115cd91 | 2976 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
2977 | |
2978 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2979 | ||
2980 | return nr_reclaimed; | |
66e1707b BS |
2981 | } |
2982 | ||
c255a458 | 2983 | #ifdef CONFIG_MEMCG |
66e1707b | 2984 | |
a9dd0a83 | 2985 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 2986 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 2987 | pg_data_t *pgdat, |
0ae5e89c | 2988 | unsigned long *nr_scanned) |
4e416953 BS |
2989 | { |
2990 | struct scan_control sc = { | |
b8f5c566 | 2991 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 2992 | .target_mem_cgroup = memcg, |
4e416953 BS |
2993 | .may_writepage = !laptop_mode, |
2994 | .may_unmap = 1, | |
b2e18757 | 2995 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 2996 | .may_swap = !noswap, |
4e416953 | 2997 | }; |
6b4f7799 | 2998 | unsigned long lru_pages; |
0ae5e89c | 2999 | |
4e416953 BS |
3000 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3001 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3002 | |
9e3b2f8c | 3003 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e | 3004 | sc.may_writepage, |
e5146b12 MG |
3005 | sc.gfp_mask, |
3006 | sc.reclaim_idx); | |
bdce6d9e | 3007 | |
4e416953 BS |
3008 | /* |
3009 | * NOTE: Although we can get the priority field, using it | |
3010 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3011 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3012 | * will pick up pages from other mem cgroup's as well. We hack |
3013 | * the priority and make it zero. | |
3014 | */ | |
ef8f2327 | 3015 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
bdce6d9e KM |
3016 | |
3017 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3018 | ||
0ae5e89c | 3019 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
3020 | return sc.nr_reclaimed; |
3021 | } | |
3022 | ||
72835c86 | 3023 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3024 | unsigned long nr_pages, |
a7885eb8 | 3025 | gfp_t gfp_mask, |
b70a2a21 | 3026 | bool may_swap) |
66e1707b | 3027 | { |
4e416953 | 3028 | struct zonelist *zonelist; |
bdce6d9e | 3029 | unsigned long nr_reclaimed; |
889976db | 3030 | int nid; |
66e1707b | 3031 | struct scan_control sc = { |
b70a2a21 | 3032 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
a09ed5e0 YH |
3033 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3034 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
b2e18757 | 3035 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3036 | .target_mem_cgroup = memcg, |
3037 | .priority = DEF_PRIORITY, | |
3038 | .may_writepage = !laptop_mode, | |
3039 | .may_unmap = 1, | |
b70a2a21 | 3040 | .may_swap = may_swap, |
a09ed5e0 | 3041 | }; |
66e1707b | 3042 | |
889976db YH |
3043 | /* |
3044 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
3045 | * take care of from where we get pages. So the node where we start the | |
3046 | * scan does not need to be the current node. | |
3047 | */ | |
72835c86 | 3048 | nid = mem_cgroup_select_victim_node(memcg); |
889976db | 3049 | |
c9634cf0 | 3050 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; |
bdce6d9e KM |
3051 | |
3052 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
3053 | sc.may_writepage, | |
e5146b12 MG |
3054 | sc.gfp_mask, |
3055 | sc.reclaim_idx); | |
bdce6d9e | 3056 | |
89a28483 | 3057 | current->flags |= PF_MEMALLOC; |
3115cd91 | 3058 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
89a28483 | 3059 | current->flags &= ~PF_MEMALLOC; |
bdce6d9e KM |
3060 | |
3061 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
3062 | ||
3063 | return nr_reclaimed; | |
66e1707b BS |
3064 | } |
3065 | #endif | |
3066 | ||
1d82de61 | 3067 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3068 | struct scan_control *sc) |
f16015fb | 3069 | { |
b95a2f2d | 3070 | struct mem_cgroup *memcg; |
f16015fb | 3071 | |
b95a2f2d JW |
3072 | if (!total_swap_pages) |
3073 | return; | |
3074 | ||
3075 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
3076 | do { | |
ef8f2327 | 3077 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
b95a2f2d | 3078 | |
dcec0b60 | 3079 | if (inactive_list_is_low(lruvec, false, sc, true)) |
1a93be0e | 3080 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 3081 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
3082 | |
3083 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
3084 | } while (memcg); | |
f16015fb JW |
3085 | } |
3086 | ||
31483b6a | 3087 | static bool zone_balanced(struct zone *zone, int order, int classzone_idx) |
60cefed4 | 3088 | { |
31483b6a | 3089 | unsigned long mark = high_wmark_pages(zone); |
60cefed4 | 3090 | |
6256c6b4 MG |
3091 | if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx)) |
3092 | return false; | |
3093 | ||
3094 | /* | |
3095 | * If any eligible zone is balanced then the node is not considered | |
3096 | * to be congested or dirty | |
3097 | */ | |
3098 | clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags); | |
3099 | clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags); | |
c2f83143 | 3100 | clear_bit(PGDAT_WRITEBACK, &zone->zone_pgdat->flags); |
6256c6b4 MG |
3101 | |
3102 | return true; | |
60cefed4 JW |
3103 | } |
3104 | ||
5515061d MG |
3105 | /* |
3106 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3107 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3108 | * | |
3109 | * Returns true if kswapd is ready to sleep | |
3110 | */ | |
d9f21d42 | 3111 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f50de2d3 | 3112 | { |
1d82de61 MG |
3113 | int i; |
3114 | ||
5515061d | 3115 | /* |
9e5e3661 VB |
3116 | * The throttled processes are normally woken up in balance_pgdat() as |
3117 | * soon as pfmemalloc_watermark_ok() is true. But there is a potential | |
3118 | * race between when kswapd checks the watermarks and a process gets | |
3119 | * throttled. There is also a potential race if processes get | |
3120 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3121 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3122 | * the wake up checks. If kswapd is going to sleep, no process should | |
3123 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3124 | * the wake up is premature, processes will wake kswapd and get | |
3125 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3126 | * that here we are under prepare_to_wait(). | |
5515061d | 3127 | */ |
9e5e3661 VB |
3128 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3129 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3130 | |
1d82de61 MG |
3131 | for (i = 0; i <= classzone_idx; i++) { |
3132 | struct zone *zone = pgdat->node_zones + i; | |
3133 | ||
6aa303de | 3134 | if (!managed_zone(zone)) |
1d82de61 MG |
3135 | continue; |
3136 | ||
38087d9b MG |
3137 | if (!zone_balanced(zone, order, classzone_idx)) |
3138 | return false; | |
1d82de61 MG |
3139 | } |
3140 | ||
38087d9b | 3141 | return true; |
f50de2d3 MG |
3142 | } |
3143 | ||
75485363 | 3144 | /* |
1d82de61 MG |
3145 | * kswapd shrinks a node of pages that are at or below the highest usable |
3146 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3147 | * |
3148 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3149 | * reclaim or if the lack of progress was due to pages under writeback. |
3150 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3151 | */ |
1d82de61 | 3152 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3153 | struct scan_control *sc) |
75485363 | 3154 | { |
1d82de61 MG |
3155 | struct zone *zone; |
3156 | int z; | |
75485363 | 3157 | |
1d82de61 MG |
3158 | /* Reclaim a number of pages proportional to the number of zones */ |
3159 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3160 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 3161 | zone = pgdat->node_zones + z; |
6aa303de | 3162 | if (!managed_zone(zone)) |
1d82de61 | 3163 | continue; |
7c954f6d | 3164 | |
1d82de61 MG |
3165 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3166 | } | |
7c954f6d MG |
3167 | |
3168 | /* | |
1d82de61 MG |
3169 | * Historically care was taken to put equal pressure on all zones but |
3170 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3171 | */ |
970a39a3 | 3172 | shrink_node(pgdat, sc); |
283aba9f | 3173 | |
7c954f6d | 3174 | /* |
1d82de61 MG |
3175 | * Fragmentation may mean that the system cannot be rebalanced for |
3176 | * high-order allocations. If twice the allocation size has been | |
3177 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3178 | * excessive reclaim. Assume that a process requested a high-order | |
3179 | * can direct reclaim/compact. | |
7c954f6d | 3180 | */ |
9861a62c | 3181 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 3182 | sc->order = 0; |
7c954f6d | 3183 | |
b8e83b94 | 3184 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3185 | } |
3186 | ||
1da177e4 | 3187 | /* |
1d82de61 MG |
3188 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3189 | * that are eligible for use by the caller until at least one zone is | |
3190 | * balanced. | |
1da177e4 | 3191 | * |
1d82de61 | 3192 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3193 | * |
3194 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3195 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
1d82de61 MG |
3196 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone |
3197 | * or lower is eligible for reclaim until at least one usable zone is | |
3198 | * balanced. | |
1da177e4 | 3199 | */ |
accf6242 | 3200 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3201 | { |
1da177e4 | 3202 | int i; |
0608f43d AM |
3203 | unsigned long nr_soft_reclaimed; |
3204 | unsigned long nr_soft_scanned; | |
1d82de61 | 3205 | struct zone *zone; |
179e9639 AM |
3206 | struct scan_control sc = { |
3207 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3208 | .order = order, |
b8e83b94 | 3209 | .priority = DEF_PRIORITY, |
ee814fe2 | 3210 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3211 | .may_unmap = 1, |
2e2e4259 | 3212 | .may_swap = 1, |
179e9639 | 3213 | }; |
f8891e5e | 3214 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3215 | |
9e3b2f8c | 3216 | do { |
b8e83b94 MG |
3217 | bool raise_priority = true; |
3218 | ||
3219 | sc.nr_reclaimed = 0; | |
84c7a777 | 3220 | sc.reclaim_idx = classzone_idx; |
1da177e4 | 3221 | |
86c79f6b | 3222 | /* |
84c7a777 MG |
3223 | * If the number of buffer_heads exceeds the maximum allowed |
3224 | * then consider reclaiming from all zones. This has a dual | |
3225 | * purpose -- on 64-bit systems it is expected that | |
3226 | * buffer_heads are stripped during active rotation. On 32-bit | |
3227 | * systems, highmem pages can pin lowmem memory and shrinking | |
3228 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3229 | * go ahead if all eligible zones for the original allocation | |
3230 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3231 | */ |
3232 | if (buffer_heads_over_limit) { | |
3233 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3234 | zone = pgdat->node_zones + i; | |
6aa303de | 3235 | if (!managed_zone(zone)) |
86c79f6b | 3236 | continue; |
cc715d99 | 3237 | |
970a39a3 | 3238 | sc.reclaim_idx = i; |
e1dbeda6 | 3239 | break; |
1da177e4 | 3240 | } |
1da177e4 | 3241 | } |
dafcb73e | 3242 | |
86c79f6b MG |
3243 | /* |
3244 | * Only reclaim if there are no eligible zones. Check from | |
3245 | * high to low zone as allocations prefer higher zones. | |
3246 | * Scanning from low to high zone would allow congestion to be | |
3247 | * cleared during a very small window when a small low | |
3248 | * zone was balanced even under extreme pressure when the | |
84c7a777 MG |
3249 | * overall node may be congested. Note that sc.reclaim_idx |
3250 | * is not used as buffer_heads_over_limit may have adjusted | |
3251 | * it. | |
86c79f6b | 3252 | */ |
84c7a777 | 3253 | for (i = classzone_idx; i >= 0; i--) { |
86c79f6b | 3254 | zone = pgdat->node_zones + i; |
6aa303de | 3255 | if (!managed_zone(zone)) |
86c79f6b MG |
3256 | continue; |
3257 | ||
84c7a777 | 3258 | if (zone_balanced(zone, sc.order, classzone_idx)) |
86c79f6b MG |
3259 | goto out; |
3260 | } | |
e1dbeda6 | 3261 | |
1d82de61 MG |
3262 | /* |
3263 | * Do some background aging of the anon list, to give | |
3264 | * pages a chance to be referenced before reclaiming. All | |
3265 | * pages are rotated regardless of classzone as this is | |
3266 | * about consistent aging. | |
3267 | */ | |
ef8f2327 | 3268 | age_active_anon(pgdat, &sc); |
1d82de61 | 3269 | |
b7ea3c41 MG |
3270 | /* |
3271 | * If we're getting trouble reclaiming, start doing writepage | |
3272 | * even in laptop mode. | |
3273 | */ | |
1d82de61 | 3274 | if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat)) |
b7ea3c41 MG |
3275 | sc.may_writepage = 1; |
3276 | ||
1d82de61 MG |
3277 | /* Call soft limit reclaim before calling shrink_node. */ |
3278 | sc.nr_scanned = 0; | |
3279 | nr_soft_scanned = 0; | |
ef8f2327 | 3280 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3281 | sc.gfp_mask, &nr_soft_scanned); |
3282 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3283 | ||
1da177e4 | 3284 | /* |
1d82de61 MG |
3285 | * There should be no need to raise the scanning priority if |
3286 | * enough pages are already being scanned that that high | |
3287 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3288 | */ |
970a39a3 | 3289 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3290 | raise_priority = false; |
5515061d MG |
3291 | |
3292 | /* | |
3293 | * If the low watermark is met there is no need for processes | |
3294 | * to be throttled on pfmemalloc_wait as they should not be | |
3295 | * able to safely make forward progress. Wake them | |
3296 | */ | |
3297 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
3298 | pfmemalloc_watermark_ok(pgdat)) | |
cfc51155 | 3299 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3300 | |
b8e83b94 MG |
3301 | /* Check if kswapd should be suspending */ |
3302 | if (try_to_freeze() || kthread_should_stop()) | |
3303 | break; | |
8357376d | 3304 | |
73ce02e9 | 3305 | /* |
b8e83b94 MG |
3306 | * Raise priority if scanning rate is too low or there was no |
3307 | * progress in reclaiming pages | |
73ce02e9 | 3308 | */ |
b8e83b94 MG |
3309 | if (raise_priority || !sc.nr_reclaimed) |
3310 | sc.priority--; | |
1d82de61 | 3311 | } while (sc.priority >= 1); |
1da177e4 | 3312 | |
b8e83b94 | 3313 | out: |
0abdee2b | 3314 | /* |
1d82de61 MG |
3315 | * Return the order kswapd stopped reclaiming at as |
3316 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3317 | * entered the allocator slow path while kswapd was awake, order will | |
3318 | * remain at the higher level. | |
0abdee2b | 3319 | */ |
1d82de61 | 3320 | return sc.order; |
1da177e4 LT |
3321 | } |
3322 | ||
38087d9b MG |
3323 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3324 | unsigned int classzone_idx) | |
f0bc0a60 KM |
3325 | { |
3326 | long remaining = 0; | |
3327 | DEFINE_WAIT(wait); | |
3328 | ||
3329 | if (freezing(current) || kthread_should_stop()) | |
3330 | return; | |
3331 | ||
3332 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3333 | ||
3334 | /* Try to sleep for a short interval */ | |
d9f21d42 | 3335 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
fd901c95 VB |
3336 | /* |
3337 | * Compaction records what page blocks it recently failed to | |
3338 | * isolate pages from and skips them in the future scanning. | |
3339 | * When kswapd is going to sleep, it is reasonable to assume | |
3340 | * that pages and compaction may succeed so reset the cache. | |
3341 | */ | |
3342 | reset_isolation_suitable(pgdat); | |
3343 | ||
3344 | /* | |
3345 | * We have freed the memory, now we should compact it to make | |
3346 | * allocation of the requested order possible. | |
3347 | */ | |
38087d9b | 3348 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
fd901c95 | 3349 | |
f0bc0a60 | 3350 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3351 | |
3352 | /* | |
3353 | * If woken prematurely then reset kswapd_classzone_idx and | |
3354 | * order. The values will either be from a wakeup request or | |
3355 | * the previous request that slept prematurely. | |
3356 | */ | |
3357 | if (remaining) { | |
3358 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx); | |
3359 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); | |
3360 | } | |
3361 | ||
f0bc0a60 KM |
3362 | finish_wait(&pgdat->kswapd_wait, &wait); |
3363 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3364 | } | |
3365 | ||
3366 | /* | |
3367 | * After a short sleep, check if it was a premature sleep. If not, then | |
3368 | * go fully to sleep until explicitly woken up. | |
3369 | */ | |
d9f21d42 MG |
3370 | if (!remaining && |
3371 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
f0bc0a60 KM |
3372 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3373 | ||
3374 | /* | |
3375 | * vmstat counters are not perfectly accurate and the estimated | |
3376 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3377 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3378 | * watermarks being breached while under pressure, we reduce the | |
3379 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3380 | * them before going back to sleep. | |
3381 | */ | |
3382 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3383 | |
3384 | if (!kthread_should_stop()) | |
3385 | schedule(); | |
3386 | ||
f0bc0a60 KM |
3387 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3388 | } else { | |
3389 | if (remaining) | |
3390 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3391 | else | |
3392 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3393 | } | |
3394 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3395 | } | |
3396 | ||
1da177e4 LT |
3397 | /* |
3398 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3399 | * from the init process. |
1da177e4 LT |
3400 | * |
3401 | * This basically trickles out pages so that we have _some_ | |
3402 | * free memory available even if there is no other activity | |
3403 | * that frees anything up. This is needed for things like routing | |
3404 | * etc, where we otherwise might have all activity going on in | |
3405 | * asynchronous contexts that cannot page things out. | |
3406 | * | |
3407 | * If there are applications that are active memory-allocators | |
3408 | * (most normal use), this basically shouldn't matter. | |
3409 | */ | |
3410 | static int kswapd(void *p) | |
3411 | { | |
38087d9b | 3412 | unsigned int alloc_order, reclaim_order, classzone_idx; |
1da177e4 LT |
3413 | pg_data_t *pgdat = (pg_data_t*)p; |
3414 | struct task_struct *tsk = current; | |
f0bc0a60 | 3415 | |
1da177e4 LT |
3416 | struct reclaim_state reclaim_state = { |
3417 | .reclaimed_slab = 0, | |
3418 | }; | |
a70f7302 | 3419 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3420 | |
cf40bd16 NP |
3421 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3422 | ||
174596a0 | 3423 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3424 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3425 | current->reclaim_state = &reclaim_state; |
3426 | ||
3427 | /* | |
3428 | * Tell the memory management that we're a "memory allocator", | |
3429 | * and that if we need more memory we should get access to it | |
3430 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3431 | * never get caught in the normal page freeing logic. | |
3432 | * | |
3433 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3434 | * you need a small amount of memory in order to be able to | |
3435 | * page out something else, and this flag essentially protects | |
3436 | * us from recursively trying to free more memory as we're | |
3437 | * trying to free the first piece of memory in the first place). | |
3438 | */ | |
930d9152 | 3439 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3440 | set_freezable(); |
1da177e4 | 3441 | |
38087d9b MG |
3442 | pgdat->kswapd_order = alloc_order = reclaim_order = 0; |
3443 | pgdat->kswapd_classzone_idx = classzone_idx = 0; | |
1da177e4 | 3444 | for ( ; ; ) { |
6f6313d4 | 3445 | bool ret; |
3e1d1d28 | 3446 | |
38087d9b MG |
3447 | kswapd_try_sleep: |
3448 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3449 | classzone_idx); | |
215ddd66 | 3450 | |
38087d9b MG |
3451 | /* Read the new order and classzone_idx */ |
3452 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
3453 | classzone_idx = pgdat->kswapd_classzone_idx; | |
3454 | pgdat->kswapd_order = 0; | |
3455 | pgdat->kswapd_classzone_idx = 0; | |
1da177e4 | 3456 | |
8fe23e05 DR |
3457 | ret = try_to_freeze(); |
3458 | if (kthread_should_stop()) | |
3459 | break; | |
3460 | ||
3461 | /* | |
3462 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3463 | * after returning from the refrigerator | |
3464 | */ | |
38087d9b MG |
3465 | if (ret) |
3466 | continue; | |
3467 | ||
3468 | /* | |
3469 | * Reclaim begins at the requested order but if a high-order | |
3470 | * reclaim fails then kswapd falls back to reclaiming for | |
3471 | * order-0. If that happens, kswapd will consider sleeping | |
3472 | * for the order it finished reclaiming at (reclaim_order) | |
3473 | * but kcompactd is woken to compact for the original | |
3474 | * request (alloc_order). | |
3475 | */ | |
e5146b12 MG |
3476 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
3477 | alloc_order); | |
38087d9b MG |
3478 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
3479 | if (reclaim_order < alloc_order) | |
3480 | goto kswapd_try_sleep; | |
1d82de61 | 3481 | |
38087d9b MG |
3482 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3483 | classzone_idx = pgdat->kswapd_classzone_idx; | |
1da177e4 | 3484 | } |
b0a8cc58 | 3485 | |
71abdc15 | 3486 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
b0a8cc58 | 3487 | current->reclaim_state = NULL; |
71abdc15 JW |
3488 | lockdep_clear_current_reclaim_state(); |
3489 | ||
1da177e4 LT |
3490 | return 0; |
3491 | } | |
3492 | ||
3493 | /* | |
3494 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3495 | */ | |
99504748 | 3496 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3497 | { |
3498 | pg_data_t *pgdat; | |
e1a55637 | 3499 | int z; |
1da177e4 | 3500 | |
6aa303de | 3501 | if (!managed_zone(zone)) |
1da177e4 LT |
3502 | return; |
3503 | ||
344736f2 | 3504 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) |
1da177e4 | 3505 | return; |
88f5acf8 | 3506 | pgdat = zone->zone_pgdat; |
38087d9b MG |
3507 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx); |
3508 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); | |
8d0986e2 | 3509 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3510 | return; |
e1a55637 MG |
3511 | |
3512 | /* Only wake kswapd if all zones are unbalanced */ | |
3513 | for (z = 0; z <= classzone_idx; z++) { | |
3514 | zone = pgdat->node_zones + z; | |
6aa303de | 3515 | if (!managed_zone(zone)) |
e1a55637 MG |
3516 | continue; |
3517 | ||
3518 | if (zone_balanced(zone, order, classzone_idx)) | |
3519 | return; | |
3520 | } | |
88f5acf8 MG |
3521 | |
3522 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3523 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3524 | } |
3525 | ||
c6f37f12 | 3526 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3527 | /* |
7b51755c | 3528 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3529 | * freed pages. |
3530 | * | |
3531 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3532 | * LRU order by reclaiming preferentially | |
3533 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3534 | */ |
7b51755c | 3535 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3536 | { |
d6277db4 | 3537 | struct reclaim_state reclaim_state; |
d6277db4 | 3538 | struct scan_control sc = { |
ee814fe2 | 3539 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3540 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 3541 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 3542 | .priority = DEF_PRIORITY, |
d6277db4 | 3543 | .may_writepage = 1, |
ee814fe2 JW |
3544 | .may_unmap = 1, |
3545 | .may_swap = 1, | |
7b51755c | 3546 | .hibernation_mode = 1, |
1da177e4 | 3547 | }; |
a09ed5e0 | 3548 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c KM |
3549 | struct task_struct *p = current; |
3550 | unsigned long nr_reclaimed; | |
1da177e4 | 3551 | |
7b51755c KM |
3552 | p->flags |= PF_MEMALLOC; |
3553 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3554 | reclaim_state.reclaimed_slab = 0; | |
3555 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3556 | |
3115cd91 | 3557 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3558 | |
7b51755c KM |
3559 | p->reclaim_state = NULL; |
3560 | lockdep_clear_current_reclaim_state(); | |
3561 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3562 | |
7b51755c | 3563 | return nr_reclaimed; |
1da177e4 | 3564 | } |
c6f37f12 | 3565 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3566 | |
1da177e4 LT |
3567 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3568 | not required for correctness. So if the last cpu in a node goes | |
3569 | away, we get changed to run anywhere: as the first one comes back, | |
3570 | restore their cpu bindings. */ | |
517bbed9 | 3571 | static int kswapd_cpu_online(unsigned int cpu) |
1da177e4 | 3572 | { |
58c0a4a7 | 3573 | int nid; |
1da177e4 | 3574 | |
517bbed9 SAS |
3575 | for_each_node_state(nid, N_MEMORY) { |
3576 | pg_data_t *pgdat = NODE_DATA(nid); | |
3577 | const struct cpumask *mask; | |
a70f7302 | 3578 | |
517bbed9 | 3579 | mask = cpumask_of_node(pgdat->node_id); |
c5f59f08 | 3580 | |
517bbed9 SAS |
3581 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3582 | /* One of our CPUs online: restore mask */ | |
3583 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
1da177e4 | 3584 | } |
517bbed9 | 3585 | return 0; |
1da177e4 | 3586 | } |
1da177e4 | 3587 | |
3218ae14 YG |
3588 | /* |
3589 | * This kswapd start function will be called by init and node-hot-add. | |
3590 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3591 | */ | |
3592 | int kswapd_run(int nid) | |
3593 | { | |
3594 | pg_data_t *pgdat = NODE_DATA(nid); | |
3595 | int ret = 0; | |
3596 | ||
3597 | if (pgdat->kswapd) | |
3598 | return 0; | |
3599 | ||
3600 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3601 | if (IS_ERR(pgdat->kswapd)) { | |
3602 | /* failure at boot is fatal */ | |
3603 | BUG_ON(system_state == SYSTEM_BOOTING); | |
d5dc0ad9 GS |
3604 | pr_err("Failed to start kswapd on node %d\n", nid); |
3605 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3606 | pgdat->kswapd = NULL; |
3218ae14 YG |
3607 | } |
3608 | return ret; | |
3609 | } | |
3610 | ||
8fe23e05 | 3611 | /* |
d8adde17 | 3612 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 3613 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
3614 | */ |
3615 | void kswapd_stop(int nid) | |
3616 | { | |
3617 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3618 | ||
d8adde17 | 3619 | if (kswapd) { |
8fe23e05 | 3620 | kthread_stop(kswapd); |
d8adde17 JL |
3621 | NODE_DATA(nid)->kswapd = NULL; |
3622 | } | |
8fe23e05 DR |
3623 | } |
3624 | ||
1da177e4 LT |
3625 | static int __init kswapd_init(void) |
3626 | { | |
517bbed9 | 3627 | int nid, ret; |
69e05944 | 3628 | |
1da177e4 | 3629 | swap_setup(); |
48fb2e24 | 3630 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3631 | kswapd_run(nid); |
517bbed9 SAS |
3632 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
3633 | "mm/vmscan:online", kswapd_cpu_online, | |
3634 | NULL); | |
3635 | WARN_ON(ret < 0); | |
1da177e4 LT |
3636 | return 0; |
3637 | } | |
3638 | ||
3639 | module_init(kswapd_init) | |
9eeff239 CL |
3640 | |
3641 | #ifdef CONFIG_NUMA | |
3642 | /* | |
a5f5f91d | 3643 | * Node reclaim mode |
9eeff239 | 3644 | * |
a5f5f91d | 3645 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 3646 | * the watermarks. |
9eeff239 | 3647 | */ |
a5f5f91d | 3648 | int node_reclaim_mode __read_mostly; |
9eeff239 | 3649 | |
1b2ffb78 | 3650 | #define RECLAIM_OFF 0 |
7d03431c | 3651 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 3652 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 3653 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 3654 | |
a92f7126 | 3655 | /* |
a5f5f91d | 3656 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
3657 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
3658 | * a zone. | |
3659 | */ | |
a5f5f91d | 3660 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 3661 | |
9614634f | 3662 | /* |
a5f5f91d | 3663 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
3664 | * occur. |
3665 | */ | |
3666 | int sysctl_min_unmapped_ratio = 1; | |
3667 | ||
0ff38490 CL |
3668 | /* |
3669 | * If the number of slab pages in a zone grows beyond this percentage then | |
3670 | * slab reclaim needs to occur. | |
3671 | */ | |
3672 | int sysctl_min_slab_ratio = 5; | |
3673 | ||
11fb9989 | 3674 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 3675 | { |
11fb9989 MG |
3676 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
3677 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
3678 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
3679 | |
3680 | /* | |
3681 | * It's possible for there to be more file mapped pages than | |
3682 | * accounted for by the pages on the file LRU lists because | |
3683 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3684 | */ | |
3685 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3686 | } | |
3687 | ||
3688 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 3689 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 3690 | { |
d031a157 AM |
3691 | unsigned long nr_pagecache_reclaimable; |
3692 | unsigned long delta = 0; | |
90afa5de MG |
3693 | |
3694 | /* | |
95bbc0c7 | 3695 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 3696 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 3697 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
3698 | * a better estimate |
3699 | */ | |
a5f5f91d MG |
3700 | if (node_reclaim_mode & RECLAIM_UNMAP) |
3701 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 3702 | else |
a5f5f91d | 3703 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
3704 | |
3705 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
3706 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
3707 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
3708 | |
3709 | /* Watch for any possible underflows due to delta */ | |
3710 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3711 | delta = nr_pagecache_reclaimable; | |
3712 | ||
3713 | return nr_pagecache_reclaimable - delta; | |
3714 | } | |
3715 | ||
9eeff239 | 3716 | /* |
a5f5f91d | 3717 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 3718 | */ |
a5f5f91d | 3719 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3720 | { |
7fb2d46d | 3721 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3722 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3723 | struct task_struct *p = current; |
3724 | struct reclaim_state reclaim_state; | |
a5f5f91d | 3725 | int classzone_idx = gfp_zone(gfp_mask); |
179e9639 | 3726 | struct scan_control sc = { |
62b726c1 | 3727 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
21caf2fc | 3728 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
bd2f6199 | 3729 | .order = order, |
a5f5f91d MG |
3730 | .priority = NODE_RECLAIM_PRIORITY, |
3731 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
3732 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 3733 | .may_swap = 1, |
a5f5f91d | 3734 | .reclaim_idx = classzone_idx, |
179e9639 | 3735 | }; |
9eeff239 | 3736 | |
9eeff239 | 3737 | cond_resched(); |
d4f7796e | 3738 | /* |
95bbc0c7 | 3739 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 3740 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 3741 | * and RECLAIM_UNMAP. |
d4f7796e CL |
3742 | */ |
3743 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3744 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3745 | reclaim_state.reclaimed_slab = 0; |
3746 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3747 | |
a5f5f91d | 3748 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 CL |
3749 | /* |
3750 | * Free memory by calling shrink zone with increasing | |
3751 | * priorities until we have enough memory freed. | |
3752 | */ | |
0ff38490 | 3753 | do { |
970a39a3 | 3754 | shrink_node(pgdat, &sc); |
9e3b2f8c | 3755 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 3756 | } |
c84db23c | 3757 | |
9eeff239 | 3758 | p->reclaim_state = NULL; |
d4f7796e | 3759 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3760 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3761 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3762 | } |
179e9639 | 3763 | |
a5f5f91d | 3764 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 3765 | { |
d773ed6b | 3766 | int ret; |
179e9639 AM |
3767 | |
3768 | /* | |
a5f5f91d | 3769 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 3770 | * slab pages if we are over the defined limits. |
34aa1330 | 3771 | * |
9614634f CL |
3772 | * A small portion of unmapped file backed pages is needed for |
3773 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
3774 | * thrown out if the node is overallocated. So we do not reclaim |
3775 | * if less than a specified percentage of the node is used by | |
9614634f | 3776 | * unmapped file backed pages. |
179e9639 | 3777 | */ |
a5f5f91d MG |
3778 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
3779 | sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) | |
3780 | return NODE_RECLAIM_FULL; | |
179e9639 | 3781 | |
a5f5f91d MG |
3782 | if (!pgdat_reclaimable(pgdat)) |
3783 | return NODE_RECLAIM_FULL; | |
d773ed6b | 3784 | |
179e9639 | 3785 | /* |
d773ed6b | 3786 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3787 | */ |
d0164adc | 3788 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 3789 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
3790 | |
3791 | /* | |
a5f5f91d | 3792 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
3793 | * have associated processors. This will favor the local processor |
3794 | * over remote processors and spread off node memory allocations | |
3795 | * as wide as possible. | |
3796 | */ | |
a5f5f91d MG |
3797 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
3798 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 3799 | |
a5f5f91d MG |
3800 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
3801 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 3802 | |
a5f5f91d MG |
3803 | ret = __node_reclaim(pgdat, gfp_mask, order); |
3804 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 3805 | |
24cf7251 MG |
3806 | if (!ret) |
3807 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3808 | ||
d773ed6b | 3809 | return ret; |
179e9639 | 3810 | } |
9eeff239 | 3811 | #endif |
894bc310 | 3812 | |
894bc310 LS |
3813 | /* |
3814 | * page_evictable - test whether a page is evictable | |
3815 | * @page: the page to test | |
894bc310 LS |
3816 | * |
3817 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3818 | * lists vs unevictable list. |
894bc310 LS |
3819 | * |
3820 | * Reasons page might not be evictable: | |
ba9ddf49 | 3821 | * (1) page's mapping marked unevictable |
b291f000 | 3822 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3823 | * |
894bc310 | 3824 | */ |
39b5f29a | 3825 | int page_evictable(struct page *page) |
894bc310 | 3826 | { |
39b5f29a | 3827 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
894bc310 | 3828 | } |
89e004ea | 3829 | |
85046579 | 3830 | #ifdef CONFIG_SHMEM |
89e004ea | 3831 | /** |
24513264 HD |
3832 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3833 | * @pages: array of pages to check | |
3834 | * @nr_pages: number of pages to check | |
89e004ea | 3835 | * |
24513264 | 3836 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3837 | * |
3838 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3839 | */ |
24513264 | 3840 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3841 | { |
925b7673 | 3842 | struct lruvec *lruvec; |
785b99fe | 3843 | struct pglist_data *pgdat = NULL; |
24513264 HD |
3844 | int pgscanned = 0; |
3845 | int pgrescued = 0; | |
3846 | int i; | |
89e004ea | 3847 | |
24513264 HD |
3848 | for (i = 0; i < nr_pages; i++) { |
3849 | struct page *page = pages[i]; | |
785b99fe | 3850 | struct pglist_data *pagepgdat = page_pgdat(page); |
89e004ea | 3851 | |
24513264 | 3852 | pgscanned++; |
785b99fe MG |
3853 | if (pagepgdat != pgdat) { |
3854 | if (pgdat) | |
3855 | spin_unlock_irq(&pgdat->lru_lock); | |
3856 | pgdat = pagepgdat; | |
3857 | spin_lock_irq(&pgdat->lru_lock); | |
24513264 | 3858 | } |
785b99fe | 3859 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
89e004ea | 3860 | |
24513264 HD |
3861 | if (!PageLRU(page) || !PageUnevictable(page)) |
3862 | continue; | |
89e004ea | 3863 | |
39b5f29a | 3864 | if (page_evictable(page)) { |
24513264 HD |
3865 | enum lru_list lru = page_lru_base_type(page); |
3866 | ||
309381fe | 3867 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 3868 | ClearPageUnevictable(page); |
fa9add64 HD |
3869 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
3870 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 3871 | pgrescued++; |
89e004ea | 3872 | } |
24513264 | 3873 | } |
89e004ea | 3874 | |
785b99fe | 3875 | if (pgdat) { |
24513264 HD |
3876 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
3877 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
785b99fe | 3878 | spin_unlock_irq(&pgdat->lru_lock); |
89e004ea | 3879 | } |
89e004ea | 3880 | } |
85046579 | 3881 | #endif /* CONFIG_SHMEM */ |