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 LT |
16 | #include <linux/mm.h> |
17 | #include <linux/module.h> | |
5a0e3ad6 | 18 | #include <linux/gfp.h> |
1da177e4 LT |
19 | #include <linux/kernel_stat.h> |
20 | #include <linux/swap.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/init.h> | |
23 | #include <linux/highmem.h> | |
70ddf637 | 24 | #include <linux/vmpressure.h> |
e129b5c2 | 25 | #include <linux/vmstat.h> |
1da177e4 LT |
26 | #include <linux/file.h> |
27 | #include <linux/writeback.h> | |
28 | #include <linux/blkdev.h> | |
29 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
30 | buffer_heads_over_limit */ | |
31 | #include <linux/mm_inline.h> | |
1da177e4 LT |
32 | #include <linux/backing-dev.h> |
33 | #include <linux/rmap.h> | |
34 | #include <linux/topology.h> | |
35 | #include <linux/cpu.h> | |
36 | #include <linux/cpuset.h> | |
3e7d3449 | 37 | #include <linux/compaction.h> |
1da177e4 LT |
38 | #include <linux/notifier.h> |
39 | #include <linux/rwsem.h> | |
248a0301 | 40 | #include <linux/delay.h> |
3218ae14 | 41 | #include <linux/kthread.h> |
7dfb7103 | 42 | #include <linux/freezer.h> |
66e1707b | 43 | #include <linux/memcontrol.h> |
873b4771 | 44 | #include <linux/delayacct.h> |
af936a16 | 45 | #include <linux/sysctl.h> |
929bea7c | 46 | #include <linux/oom.h> |
268bb0ce | 47 | #include <linux/prefetch.h> |
b1de0d13 | 48 | #include <linux/printk.h> |
f9fe48be | 49 | #include <linux/dax.h> |
1da177e4 LT |
50 | |
51 | #include <asm/tlbflush.h> | |
52 | #include <asm/div64.h> | |
53 | ||
54 | #include <linux/swapops.h> | |
117aad1e | 55 | #include <linux/balloon_compaction.h> |
1da177e4 | 56 | |
0f8053a5 NP |
57 | #include "internal.h" |
58 | ||
33906bc5 MG |
59 | #define CREATE_TRACE_POINTS |
60 | #include <trace/events/vmscan.h> | |
61 | ||
1da177e4 | 62 | struct scan_control { |
22fba335 KM |
63 | /* How many pages shrink_list() should reclaim */ |
64 | unsigned long nr_to_reclaim; | |
65 | ||
1da177e4 | 66 | /* This context's GFP mask */ |
6daa0e28 | 67 | gfp_t gfp_mask; |
1da177e4 | 68 | |
ee814fe2 | 69 | /* Allocation order */ |
5ad333eb | 70 | int order; |
66e1707b | 71 | |
ee814fe2 JW |
72 | /* |
73 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
74 | * are scanned. | |
75 | */ | |
76 | nodemask_t *nodemask; | |
9e3b2f8c | 77 | |
f16015fb JW |
78 | /* |
79 | * The memory cgroup that hit its limit and as a result is the | |
80 | * primary target of this reclaim invocation. | |
81 | */ | |
82 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 83 | |
ee814fe2 JW |
84 | /* Scan (total_size >> priority) pages at once */ |
85 | int priority; | |
86 | ||
b2e18757 MG |
87 | /* The highest zone to isolate pages for reclaim from */ |
88 | enum zone_type reclaim_idx; | |
89 | ||
ee814fe2 JW |
90 | unsigned int may_writepage:1; |
91 | ||
92 | /* Can mapped pages be reclaimed? */ | |
93 | unsigned int may_unmap:1; | |
94 | ||
95 | /* Can pages be swapped as part of reclaim? */ | |
96 | unsigned int may_swap:1; | |
97 | ||
241994ed JW |
98 | /* Can cgroups be reclaimed below their normal consumption range? */ |
99 | unsigned int may_thrash:1; | |
100 | ||
ee814fe2 JW |
101 | unsigned int hibernation_mode:1; |
102 | ||
103 | /* One of the zones is ready for compaction */ | |
104 | unsigned int compaction_ready:1; | |
105 | ||
106 | /* Incremented by the number of inactive pages that were scanned */ | |
107 | unsigned long nr_scanned; | |
108 | ||
109 | /* Number of pages freed so far during a call to shrink_zones() */ | |
110 | unsigned long nr_reclaimed; | |
1da177e4 LT |
111 | }; |
112 | ||
1da177e4 LT |
113 | #ifdef ARCH_HAS_PREFETCH |
114 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
115 | do { \ | |
116 | if ((_page)->lru.prev != _base) { \ | |
117 | struct page *prev; \ | |
118 | \ | |
119 | prev = lru_to_page(&(_page->lru)); \ | |
120 | prefetch(&prev->_field); \ | |
121 | } \ | |
122 | } while (0) | |
123 | #else | |
124 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
125 | #endif | |
126 | ||
127 | #ifdef ARCH_HAS_PREFETCHW | |
128 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
129 | do { \ | |
130 | if ((_page)->lru.prev != _base) { \ | |
131 | struct page *prev; \ | |
132 | \ | |
133 | prev = lru_to_page(&(_page->lru)); \ | |
134 | prefetchw(&prev->_field); \ | |
135 | } \ | |
136 | } while (0) | |
137 | #else | |
138 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
139 | #endif | |
140 | ||
141 | /* | |
142 | * From 0 .. 100. Higher means more swappy. | |
143 | */ | |
144 | int vm_swappiness = 60; | |
d0480be4 WSH |
145 | /* |
146 | * The total number of pages which are beyond the high watermark within all | |
147 | * zones. | |
148 | */ | |
149 | unsigned long vm_total_pages; | |
1da177e4 LT |
150 | |
151 | static LIST_HEAD(shrinker_list); | |
152 | static DECLARE_RWSEM(shrinker_rwsem); | |
153 | ||
c255a458 | 154 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
155 | static bool global_reclaim(struct scan_control *sc) |
156 | { | |
f16015fb | 157 | return !sc->target_mem_cgroup; |
89b5fae5 | 158 | } |
97c9341f TH |
159 | |
160 | /** | |
161 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
162 | * @sc: scan_control in question | |
163 | * | |
164 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
165 | * completely broken with the legacy memcg and direct stalling in | |
166 | * shrink_page_list() is used for throttling instead, which lacks all the | |
167 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
168 | * allocation and configurability. | |
169 | * | |
170 | * This function tests whether the vmscan currently in progress can assume | |
171 | * that the normal dirty throttling mechanism is operational. | |
172 | */ | |
173 | static bool sane_reclaim(struct scan_control *sc) | |
174 | { | |
175 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
176 | ||
177 | if (!memcg) | |
178 | return true; | |
179 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 180 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
181 | return true; |
182 | #endif | |
183 | return false; | |
184 | } | |
91a45470 | 185 | #else |
89b5fae5 JW |
186 | static bool global_reclaim(struct scan_control *sc) |
187 | { | |
188 | return true; | |
189 | } | |
97c9341f TH |
190 | |
191 | static bool sane_reclaim(struct scan_control *sc) | |
192 | { | |
193 | return true; | |
194 | } | |
91a45470 KH |
195 | #endif |
196 | ||
599d0c95 MG |
197 | /* |
198 | * This misses isolated pages which are not accounted for to save counters. | |
199 | * As the data only determines if reclaim or compaction continues, it is | |
200 | * not expected that isolated pages will be a dominating factor. | |
201 | */ | |
0a0337e0 | 202 | unsigned long zone_reclaimable_pages(struct zone *zone) |
6e543d57 | 203 | { |
d031a157 | 204 | unsigned long nr; |
6e543d57 | 205 | |
599d0c95 MG |
206 | nr = zone_page_state_snapshot(zone, NR_ZONE_LRU_FILE); |
207 | if (get_nr_swap_pages() > 0) | |
208 | nr += zone_page_state_snapshot(zone, NR_ZONE_LRU_ANON); | |
209 | ||
210 | return nr; | |
211 | } | |
212 | ||
213 | unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat) | |
214 | { | |
215 | unsigned long nr; | |
216 | ||
217 | nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) + | |
218 | node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) + | |
219 | node_page_state_snapshot(pgdat, NR_ISOLATED_FILE); | |
6e543d57 LD |
220 | |
221 | if (get_nr_swap_pages() > 0) | |
599d0c95 MG |
222 | nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) + |
223 | node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) + | |
224 | node_page_state_snapshot(pgdat, NR_ISOLATED_ANON); | |
6e543d57 LD |
225 | |
226 | return nr; | |
227 | } | |
228 | ||
599d0c95 | 229 | bool pgdat_reclaimable(struct pglist_data *pgdat) |
6e543d57 | 230 | { |
599d0c95 MG |
231 | return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) < |
232 | pgdat_reclaimable_pages(pgdat) * 6; | |
6e543d57 LD |
233 | } |
234 | ||
23047a96 | 235 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru) |
c9f299d9 | 236 | { |
c3c787e8 | 237 | if (!mem_cgroup_disabled()) |
4d7dcca2 | 238 | return mem_cgroup_get_lru_size(lruvec, lru); |
a3d8e054 | 239 | |
599d0c95 | 240 | return node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); |
c9f299d9 KM |
241 | } |
242 | ||
1da177e4 | 243 | /* |
1d3d4437 | 244 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 245 | */ |
1d3d4437 | 246 | int register_shrinker(struct shrinker *shrinker) |
1da177e4 | 247 | { |
1d3d4437 GC |
248 | size_t size = sizeof(*shrinker->nr_deferred); |
249 | ||
1d3d4437 GC |
250 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
251 | size *= nr_node_ids; | |
252 | ||
253 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
254 | if (!shrinker->nr_deferred) | |
255 | return -ENOMEM; | |
256 | ||
8e1f936b RR |
257 | down_write(&shrinker_rwsem); |
258 | list_add_tail(&shrinker->list, &shrinker_list); | |
259 | up_write(&shrinker_rwsem); | |
1d3d4437 | 260 | return 0; |
1da177e4 | 261 | } |
8e1f936b | 262 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
263 | |
264 | /* | |
265 | * Remove one | |
266 | */ | |
8e1f936b | 267 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
268 | { |
269 | down_write(&shrinker_rwsem); | |
270 | list_del(&shrinker->list); | |
271 | up_write(&shrinker_rwsem); | |
ae393321 | 272 | kfree(shrinker->nr_deferred); |
1da177e4 | 273 | } |
8e1f936b | 274 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
275 | |
276 | #define SHRINK_BATCH 128 | |
1d3d4437 | 277 | |
cb731d6c VD |
278 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
279 | struct shrinker *shrinker, | |
280 | unsigned long nr_scanned, | |
281 | unsigned long nr_eligible) | |
1d3d4437 GC |
282 | { |
283 | unsigned long freed = 0; | |
284 | unsigned long long delta; | |
285 | long total_scan; | |
d5bc5fd3 | 286 | long freeable; |
1d3d4437 GC |
287 | long nr; |
288 | long new_nr; | |
289 | int nid = shrinkctl->nid; | |
290 | long batch_size = shrinker->batch ? shrinker->batch | |
291 | : SHRINK_BATCH; | |
292 | ||
d5bc5fd3 VD |
293 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
294 | if (freeable == 0) | |
1d3d4437 GC |
295 | return 0; |
296 | ||
297 | /* | |
298 | * copy the current shrinker scan count into a local variable | |
299 | * and zero it so that other concurrent shrinker invocations | |
300 | * don't also do this scanning work. | |
301 | */ | |
302 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
303 | ||
304 | total_scan = nr; | |
6b4f7799 | 305 | delta = (4 * nr_scanned) / shrinker->seeks; |
d5bc5fd3 | 306 | delta *= freeable; |
6b4f7799 | 307 | do_div(delta, nr_eligible + 1); |
1d3d4437 GC |
308 | total_scan += delta; |
309 | if (total_scan < 0) { | |
8612c663 | 310 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
a0b02131 | 311 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 312 | total_scan = freeable; |
1d3d4437 GC |
313 | } |
314 | ||
315 | /* | |
316 | * We need to avoid excessive windup on filesystem shrinkers | |
317 | * due to large numbers of GFP_NOFS allocations causing the | |
318 | * shrinkers to return -1 all the time. This results in a large | |
319 | * nr being built up so when a shrink that can do some work | |
320 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 321 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
322 | * memory. |
323 | * | |
324 | * Hence only allow the shrinker to scan the entire cache when | |
325 | * a large delta change is calculated directly. | |
326 | */ | |
d5bc5fd3 VD |
327 | if (delta < freeable / 4) |
328 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
329 | |
330 | /* | |
331 | * Avoid risking looping forever due to too large nr value: | |
332 | * never try to free more than twice the estimate number of | |
333 | * freeable entries. | |
334 | */ | |
d5bc5fd3 VD |
335 | if (total_scan > freeable * 2) |
336 | total_scan = freeable * 2; | |
1d3d4437 GC |
337 | |
338 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
6b4f7799 JW |
339 | nr_scanned, nr_eligible, |
340 | freeable, delta, total_scan); | |
1d3d4437 | 341 | |
0b1fb40a VD |
342 | /* |
343 | * Normally, we should not scan less than batch_size objects in one | |
344 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
345 | * than batch_size objects in total and we are really tight on memory, | |
346 | * we will try to reclaim all available objects, otherwise we can end | |
347 | * up failing allocations although there are plenty of reclaimable | |
348 | * objects spread over several slabs with usage less than the | |
349 | * batch_size. | |
350 | * | |
351 | * We detect the "tight on memory" situations by looking at the total | |
352 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 353 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
354 | * scanning at high prio and therefore should try to reclaim as much as |
355 | * possible. | |
356 | */ | |
357 | while (total_scan >= batch_size || | |
d5bc5fd3 | 358 | total_scan >= freeable) { |
a0b02131 | 359 | unsigned long ret; |
0b1fb40a | 360 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 361 | |
0b1fb40a | 362 | shrinkctl->nr_to_scan = nr_to_scan; |
a0b02131 DC |
363 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
364 | if (ret == SHRINK_STOP) | |
365 | break; | |
366 | freed += ret; | |
1d3d4437 | 367 | |
0b1fb40a VD |
368 | count_vm_events(SLABS_SCANNED, nr_to_scan); |
369 | total_scan -= nr_to_scan; | |
1d3d4437 GC |
370 | |
371 | cond_resched(); | |
372 | } | |
373 | ||
374 | /* | |
375 | * move the unused scan count back into the shrinker in a | |
376 | * manner that handles concurrent updates. If we exhausted the | |
377 | * scan, there is no need to do an update. | |
378 | */ | |
379 | if (total_scan > 0) | |
380 | new_nr = atomic_long_add_return(total_scan, | |
381 | &shrinker->nr_deferred[nid]); | |
382 | else | |
383 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
384 | ||
df9024a8 | 385 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 386 | return freed; |
1495f230 YH |
387 | } |
388 | ||
6b4f7799 | 389 | /** |
cb731d6c | 390 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
391 | * @gfp_mask: allocation context |
392 | * @nid: node whose slab caches to target | |
cb731d6c | 393 | * @memcg: memory cgroup whose slab caches to target |
6b4f7799 JW |
394 | * @nr_scanned: pressure numerator |
395 | * @nr_eligible: pressure denominator | |
1da177e4 | 396 | * |
6b4f7799 | 397 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 398 | * |
6b4f7799 JW |
399 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
400 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 401 | * |
cb731d6c VD |
402 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
403 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan | |
0fc9f58a VD |
404 | * objects from the memory cgroup specified. Otherwise, only unaware |
405 | * shrinkers are called. | |
cb731d6c | 406 | * |
6b4f7799 JW |
407 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of |
408 | * the available objects should be scanned. Page reclaim for example | |
409 | * passes the number of pages scanned and the number of pages on the | |
410 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned | |
411 | * when it encountered mapped pages. The ratio is further biased by | |
412 | * the ->seeks setting of the shrink function, which indicates the | |
413 | * cost to recreate an object relative to that of an LRU page. | |
b15e0905 | 414 | * |
6b4f7799 | 415 | * Returns the number of reclaimed slab objects. |
1da177e4 | 416 | */ |
cb731d6c VD |
417 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
418 | struct mem_cgroup *memcg, | |
419 | unsigned long nr_scanned, | |
420 | unsigned long nr_eligible) | |
1da177e4 LT |
421 | { |
422 | struct shrinker *shrinker; | |
24f7c6b9 | 423 | unsigned long freed = 0; |
1da177e4 | 424 | |
0fc9f58a | 425 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) |
cb731d6c VD |
426 | return 0; |
427 | ||
6b4f7799 JW |
428 | if (nr_scanned == 0) |
429 | nr_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 430 | |
f06590bd | 431 | if (!down_read_trylock(&shrinker_rwsem)) { |
24f7c6b9 DC |
432 | /* |
433 | * If we would return 0, our callers would understand that we | |
434 | * have nothing else to shrink and give up trying. By returning | |
435 | * 1 we keep it going and assume we'll be able to shrink next | |
436 | * time. | |
437 | */ | |
438 | freed = 1; | |
f06590bd MK |
439 | goto out; |
440 | } | |
1da177e4 LT |
441 | |
442 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
443 | struct shrink_control sc = { |
444 | .gfp_mask = gfp_mask, | |
445 | .nid = nid, | |
cb731d6c | 446 | .memcg = memcg, |
6b4f7799 | 447 | }; |
ec97097b | 448 | |
0fc9f58a VD |
449 | /* |
450 | * If kernel memory accounting is disabled, we ignore | |
451 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers | |
452 | * passing NULL for memcg. | |
453 | */ | |
454 | if (memcg_kmem_enabled() && | |
455 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
cb731d6c VD |
456 | continue; |
457 | ||
6b4f7799 JW |
458 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
459 | sc.nid = 0; | |
1da177e4 | 460 | |
cb731d6c | 461 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); |
1da177e4 | 462 | } |
6b4f7799 | 463 | |
1da177e4 | 464 | up_read(&shrinker_rwsem); |
f06590bd MK |
465 | out: |
466 | cond_resched(); | |
24f7c6b9 | 467 | return freed; |
1da177e4 LT |
468 | } |
469 | ||
cb731d6c VD |
470 | void drop_slab_node(int nid) |
471 | { | |
472 | unsigned long freed; | |
473 | ||
474 | do { | |
475 | struct mem_cgroup *memcg = NULL; | |
476 | ||
477 | freed = 0; | |
478 | do { | |
479 | freed += shrink_slab(GFP_KERNEL, nid, memcg, | |
480 | 1000, 1000); | |
481 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
482 | } while (freed > 10); | |
483 | } | |
484 | ||
485 | void drop_slab(void) | |
486 | { | |
487 | int nid; | |
488 | ||
489 | for_each_online_node(nid) | |
490 | drop_slab_node(nid); | |
491 | } | |
492 | ||
1da177e4 LT |
493 | static inline int is_page_cache_freeable(struct page *page) |
494 | { | |
ceddc3a5 JW |
495 | /* |
496 | * A freeable page cache page is referenced only by the caller | |
497 | * that isolated the page, the page cache radix tree and | |
498 | * optional buffer heads at page->private. | |
499 | */ | |
edcf4748 | 500 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
501 | } |
502 | ||
703c2708 | 503 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
1da177e4 | 504 | { |
930d9152 | 505 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 506 | return 1; |
703c2708 | 507 | if (!inode_write_congested(inode)) |
1da177e4 | 508 | return 1; |
703c2708 | 509 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
510 | return 1; |
511 | return 0; | |
512 | } | |
513 | ||
514 | /* | |
515 | * We detected a synchronous write error writing a page out. Probably | |
516 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
517 | * fsync(), msync() or close(). | |
518 | * | |
519 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
520 | * prevents it from being freed up. But we have a ref on the page and once | |
521 | * that page is locked, the mapping is pinned. | |
522 | * | |
523 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
524 | * __GFP_FS. | |
525 | */ | |
526 | static void handle_write_error(struct address_space *mapping, | |
527 | struct page *page, int error) | |
528 | { | |
7eaceacc | 529 | lock_page(page); |
3e9f45bd GC |
530 | if (page_mapping(page) == mapping) |
531 | mapping_set_error(mapping, error); | |
1da177e4 LT |
532 | unlock_page(page); |
533 | } | |
534 | ||
04e62a29 CL |
535 | /* possible outcome of pageout() */ |
536 | typedef enum { | |
537 | /* failed to write page out, page is locked */ | |
538 | PAGE_KEEP, | |
539 | /* move page to the active list, page is locked */ | |
540 | PAGE_ACTIVATE, | |
541 | /* page has been sent to the disk successfully, page is unlocked */ | |
542 | PAGE_SUCCESS, | |
543 | /* page is clean and locked */ | |
544 | PAGE_CLEAN, | |
545 | } pageout_t; | |
546 | ||
1da177e4 | 547 | /* |
1742f19f AM |
548 | * pageout is called by shrink_page_list() for each dirty page. |
549 | * Calls ->writepage(). | |
1da177e4 | 550 | */ |
c661b078 | 551 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 552 | struct scan_control *sc) |
1da177e4 LT |
553 | { |
554 | /* | |
555 | * If the page is dirty, only perform writeback if that write | |
556 | * will be non-blocking. To prevent this allocation from being | |
557 | * stalled by pagecache activity. But note that there may be | |
558 | * stalls if we need to run get_block(). We could test | |
559 | * PagePrivate for that. | |
560 | * | |
8174202b | 561 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
562 | * this page's queue, we can perform writeback even if that |
563 | * will block. | |
564 | * | |
565 | * If the page is swapcache, write it back even if that would | |
566 | * block, for some throttling. This happens by accident, because | |
567 | * swap_backing_dev_info is bust: it doesn't reflect the | |
568 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
569 | */ |
570 | if (!is_page_cache_freeable(page)) | |
571 | return PAGE_KEEP; | |
572 | if (!mapping) { | |
573 | /* | |
574 | * Some data journaling orphaned pages can have | |
575 | * page->mapping == NULL while being dirty with clean buffers. | |
576 | */ | |
266cf658 | 577 | if (page_has_private(page)) { |
1da177e4 LT |
578 | if (try_to_free_buffers(page)) { |
579 | ClearPageDirty(page); | |
b1de0d13 | 580 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
581 | return PAGE_CLEAN; |
582 | } | |
583 | } | |
584 | return PAGE_KEEP; | |
585 | } | |
586 | if (mapping->a_ops->writepage == NULL) | |
587 | return PAGE_ACTIVATE; | |
703c2708 | 588 | if (!may_write_to_inode(mapping->host, sc)) |
1da177e4 LT |
589 | return PAGE_KEEP; |
590 | ||
591 | if (clear_page_dirty_for_io(page)) { | |
592 | int res; | |
593 | struct writeback_control wbc = { | |
594 | .sync_mode = WB_SYNC_NONE, | |
595 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
596 | .range_start = 0, |
597 | .range_end = LLONG_MAX, | |
1da177e4 LT |
598 | .for_reclaim = 1, |
599 | }; | |
600 | ||
601 | SetPageReclaim(page); | |
602 | res = mapping->a_ops->writepage(page, &wbc); | |
603 | if (res < 0) | |
604 | handle_write_error(mapping, page, res); | |
994fc28c | 605 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
606 | ClearPageReclaim(page); |
607 | return PAGE_ACTIVATE; | |
608 | } | |
c661b078 | 609 | |
1da177e4 LT |
610 | if (!PageWriteback(page)) { |
611 | /* synchronous write or broken a_ops? */ | |
612 | ClearPageReclaim(page); | |
613 | } | |
3aa23851 | 614 | trace_mm_vmscan_writepage(page); |
c4a25635 | 615 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
616 | return PAGE_SUCCESS; |
617 | } | |
618 | ||
619 | return PAGE_CLEAN; | |
620 | } | |
621 | ||
a649fd92 | 622 | /* |
e286781d NP |
623 | * Same as remove_mapping, but if the page is removed from the mapping, it |
624 | * gets returned with a refcount of 0. | |
a649fd92 | 625 | */ |
a528910e JW |
626 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
627 | bool reclaimed) | |
49d2e9cc | 628 | { |
c4843a75 | 629 | unsigned long flags; |
c4843a75 | 630 | |
28e4d965 NP |
631 | BUG_ON(!PageLocked(page)); |
632 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 633 | |
c4843a75 | 634 | spin_lock_irqsave(&mapping->tree_lock, flags); |
49d2e9cc | 635 | /* |
0fd0e6b0 NP |
636 | * The non racy check for a busy page. |
637 | * | |
638 | * Must be careful with the order of the tests. When someone has | |
639 | * a ref to the page, it may be possible that they dirty it then | |
640 | * drop the reference. So if PageDirty is tested before page_count | |
641 | * here, then the following race may occur: | |
642 | * | |
643 | * get_user_pages(&page); | |
644 | * [user mapping goes away] | |
645 | * write_to(page); | |
646 | * !PageDirty(page) [good] | |
647 | * SetPageDirty(page); | |
648 | * put_page(page); | |
649 | * !page_count(page) [good, discard it] | |
650 | * | |
651 | * [oops, our write_to data is lost] | |
652 | * | |
653 | * Reversing the order of the tests ensures such a situation cannot | |
654 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 655 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
656 | * |
657 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
658 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 659 | */ |
fe896d18 | 660 | if (!page_ref_freeze(page, 2)) |
49d2e9cc | 661 | goto cannot_free; |
e286781d NP |
662 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
663 | if (unlikely(PageDirty(page))) { | |
fe896d18 | 664 | page_ref_unfreeze(page, 2); |
49d2e9cc | 665 | goto cannot_free; |
e286781d | 666 | } |
49d2e9cc CL |
667 | |
668 | if (PageSwapCache(page)) { | |
669 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 670 | mem_cgroup_swapout(page, swap); |
49d2e9cc | 671 | __delete_from_swap_cache(page); |
c4843a75 | 672 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
0a31bc97 | 673 | swapcache_free(swap); |
e286781d | 674 | } else { |
6072d13c | 675 | void (*freepage)(struct page *); |
a528910e | 676 | void *shadow = NULL; |
6072d13c LT |
677 | |
678 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
679 | /* |
680 | * Remember a shadow entry for reclaimed file cache in | |
681 | * order to detect refaults, thus thrashing, later on. | |
682 | * | |
683 | * But don't store shadows in an address space that is | |
684 | * already exiting. This is not just an optizimation, | |
685 | * inode reclaim needs to empty out the radix tree or | |
686 | * the nodes are lost. Don't plant shadows behind its | |
687 | * back. | |
f9fe48be RZ |
688 | * |
689 | * We also don't store shadows for DAX mappings because the | |
690 | * only page cache pages found in these are zero pages | |
691 | * covering holes, and because we don't want to mix DAX | |
692 | * exceptional entries and shadow exceptional entries in the | |
693 | * same page_tree. | |
a528910e JW |
694 | */ |
695 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 696 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
a528910e | 697 | shadow = workingset_eviction(mapping, page); |
62cccb8c | 698 | __delete_from_page_cache(page, shadow); |
c4843a75 | 699 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
6072d13c LT |
700 | |
701 | if (freepage != NULL) | |
702 | freepage(page); | |
49d2e9cc CL |
703 | } |
704 | ||
49d2e9cc CL |
705 | return 1; |
706 | ||
707 | cannot_free: | |
c4843a75 | 708 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
49d2e9cc CL |
709 | return 0; |
710 | } | |
711 | ||
e286781d NP |
712 | /* |
713 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
714 | * someone else has a ref on the page, abort and return 0. If it was | |
715 | * successfully detached, return 1. Assumes the caller has a single ref on | |
716 | * this page. | |
717 | */ | |
718 | int remove_mapping(struct address_space *mapping, struct page *page) | |
719 | { | |
a528910e | 720 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
721 | /* |
722 | * Unfreezing the refcount with 1 rather than 2 effectively | |
723 | * drops the pagecache ref for us without requiring another | |
724 | * atomic operation. | |
725 | */ | |
fe896d18 | 726 | page_ref_unfreeze(page, 1); |
e286781d NP |
727 | return 1; |
728 | } | |
729 | return 0; | |
730 | } | |
731 | ||
894bc310 LS |
732 | /** |
733 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
734 | * @page: page to be put back to appropriate lru list | |
735 | * | |
736 | * Add previously isolated @page to appropriate LRU list. | |
737 | * Page may still be unevictable for other reasons. | |
738 | * | |
739 | * lru_lock must not be held, interrupts must be enabled. | |
740 | */ | |
894bc310 LS |
741 | void putback_lru_page(struct page *page) |
742 | { | |
0ec3b74c | 743 | bool is_unevictable; |
bbfd28ee | 744 | int was_unevictable = PageUnevictable(page); |
894bc310 | 745 | |
309381fe | 746 | VM_BUG_ON_PAGE(PageLRU(page), page); |
894bc310 LS |
747 | |
748 | redo: | |
749 | ClearPageUnevictable(page); | |
750 | ||
39b5f29a | 751 | if (page_evictable(page)) { |
894bc310 LS |
752 | /* |
753 | * For evictable pages, we can use the cache. | |
754 | * In event of a race, worst case is we end up with an | |
755 | * unevictable page on [in]active list. | |
756 | * We know how to handle that. | |
757 | */ | |
0ec3b74c | 758 | is_unevictable = false; |
c53954a0 | 759 | lru_cache_add(page); |
894bc310 LS |
760 | } else { |
761 | /* | |
762 | * Put unevictable pages directly on zone's unevictable | |
763 | * list. | |
764 | */ | |
0ec3b74c | 765 | is_unevictable = true; |
894bc310 | 766 | add_page_to_unevictable_list(page); |
6a7b9548 | 767 | /* |
21ee9f39 MK |
768 | * When racing with an mlock or AS_UNEVICTABLE clearing |
769 | * (page is unlocked) make sure that if the other thread | |
770 | * does not observe our setting of PG_lru and fails | |
24513264 | 771 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 772 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
773 | * the page back to the evictable list. |
774 | * | |
21ee9f39 | 775 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
776 | */ |
777 | smp_mb(); | |
894bc310 | 778 | } |
894bc310 LS |
779 | |
780 | /* | |
781 | * page's status can change while we move it among lru. If an evictable | |
782 | * page is on unevictable list, it never be freed. To avoid that, | |
783 | * check after we added it to the list, again. | |
784 | */ | |
0ec3b74c | 785 | if (is_unevictable && page_evictable(page)) { |
894bc310 LS |
786 | if (!isolate_lru_page(page)) { |
787 | put_page(page); | |
788 | goto redo; | |
789 | } | |
790 | /* This means someone else dropped this page from LRU | |
791 | * So, it will be freed or putback to LRU again. There is | |
792 | * nothing to do here. | |
793 | */ | |
794 | } | |
795 | ||
0ec3b74c | 796 | if (was_unevictable && !is_unevictable) |
bbfd28ee | 797 | count_vm_event(UNEVICTABLE_PGRESCUED); |
0ec3b74c | 798 | else if (!was_unevictable && is_unevictable) |
bbfd28ee LS |
799 | count_vm_event(UNEVICTABLE_PGCULLED); |
800 | ||
894bc310 LS |
801 | put_page(page); /* drop ref from isolate */ |
802 | } | |
803 | ||
dfc8d636 JW |
804 | enum page_references { |
805 | PAGEREF_RECLAIM, | |
806 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 807 | PAGEREF_KEEP, |
dfc8d636 JW |
808 | PAGEREF_ACTIVATE, |
809 | }; | |
810 | ||
811 | static enum page_references page_check_references(struct page *page, | |
812 | struct scan_control *sc) | |
813 | { | |
64574746 | 814 | int referenced_ptes, referenced_page; |
dfc8d636 | 815 | unsigned long vm_flags; |
dfc8d636 | 816 | |
c3ac9a8a JW |
817 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
818 | &vm_flags); | |
64574746 | 819 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 820 | |
dfc8d636 JW |
821 | /* |
822 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
823 | * move the page to the unevictable list. | |
824 | */ | |
825 | if (vm_flags & VM_LOCKED) | |
826 | return PAGEREF_RECLAIM; | |
827 | ||
64574746 | 828 | if (referenced_ptes) { |
e4898273 | 829 | if (PageSwapBacked(page)) |
64574746 JW |
830 | return PAGEREF_ACTIVATE; |
831 | /* | |
832 | * All mapped pages start out with page table | |
833 | * references from the instantiating fault, so we need | |
834 | * to look twice if a mapped file page is used more | |
835 | * than once. | |
836 | * | |
837 | * Mark it and spare it for another trip around the | |
838 | * inactive list. Another page table reference will | |
839 | * lead to its activation. | |
840 | * | |
841 | * Note: the mark is set for activated pages as well | |
842 | * so that recently deactivated but used pages are | |
843 | * quickly recovered. | |
844 | */ | |
845 | SetPageReferenced(page); | |
846 | ||
34dbc67a | 847 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
848 | return PAGEREF_ACTIVATE; |
849 | ||
c909e993 KK |
850 | /* |
851 | * Activate file-backed executable pages after first usage. | |
852 | */ | |
853 | if (vm_flags & VM_EXEC) | |
854 | return PAGEREF_ACTIVATE; | |
855 | ||
64574746 JW |
856 | return PAGEREF_KEEP; |
857 | } | |
dfc8d636 JW |
858 | |
859 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 860 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
861 | return PAGEREF_RECLAIM_CLEAN; |
862 | ||
863 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
864 | } |
865 | ||
e2be15f6 MG |
866 | /* Check if a page is dirty or under writeback */ |
867 | static void page_check_dirty_writeback(struct page *page, | |
868 | bool *dirty, bool *writeback) | |
869 | { | |
b4597226 MG |
870 | struct address_space *mapping; |
871 | ||
e2be15f6 MG |
872 | /* |
873 | * Anonymous pages are not handled by flushers and must be written | |
874 | * from reclaim context. Do not stall reclaim based on them | |
875 | */ | |
876 | if (!page_is_file_cache(page)) { | |
877 | *dirty = false; | |
878 | *writeback = false; | |
879 | return; | |
880 | } | |
881 | ||
882 | /* By default assume that the page flags are accurate */ | |
883 | *dirty = PageDirty(page); | |
884 | *writeback = PageWriteback(page); | |
b4597226 MG |
885 | |
886 | /* Verify dirty/writeback state if the filesystem supports it */ | |
887 | if (!page_has_private(page)) | |
888 | return; | |
889 | ||
890 | mapping = page_mapping(page); | |
891 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
892 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
893 | } |
894 | ||
1da177e4 | 895 | /* |
1742f19f | 896 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 897 | */ |
1742f19f | 898 | static unsigned long shrink_page_list(struct list_head *page_list, |
599d0c95 | 899 | struct pglist_data *pgdat, |
f84f6e2b | 900 | struct scan_control *sc, |
02c6de8d | 901 | enum ttu_flags ttu_flags, |
8e950282 | 902 | unsigned long *ret_nr_dirty, |
d43006d5 | 903 | unsigned long *ret_nr_unqueued_dirty, |
8e950282 | 904 | unsigned long *ret_nr_congested, |
02c6de8d | 905 | unsigned long *ret_nr_writeback, |
b1a6f21e | 906 | unsigned long *ret_nr_immediate, |
02c6de8d | 907 | bool force_reclaim) |
1da177e4 LT |
908 | { |
909 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 910 | LIST_HEAD(free_pages); |
1da177e4 | 911 | int pgactivate = 0; |
d43006d5 | 912 | unsigned long nr_unqueued_dirty = 0; |
0e093d99 MG |
913 | unsigned long nr_dirty = 0; |
914 | unsigned long nr_congested = 0; | |
05ff5137 | 915 | unsigned long nr_reclaimed = 0; |
92df3a72 | 916 | unsigned long nr_writeback = 0; |
b1a6f21e | 917 | unsigned long nr_immediate = 0; |
1da177e4 LT |
918 | |
919 | cond_resched(); | |
920 | ||
1da177e4 LT |
921 | while (!list_empty(page_list)) { |
922 | struct address_space *mapping; | |
923 | struct page *page; | |
924 | int may_enter_fs; | |
02c6de8d | 925 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
e2be15f6 | 926 | bool dirty, writeback; |
854e9ed0 MK |
927 | bool lazyfree = false; |
928 | int ret = SWAP_SUCCESS; | |
1da177e4 LT |
929 | |
930 | cond_resched(); | |
931 | ||
932 | page = lru_to_page(page_list); | |
933 | list_del(&page->lru); | |
934 | ||
529ae9aa | 935 | if (!trylock_page(page)) |
1da177e4 LT |
936 | goto keep; |
937 | ||
309381fe | 938 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
939 | |
940 | sc->nr_scanned++; | |
80e43426 | 941 | |
39b5f29a | 942 | if (unlikely(!page_evictable(page))) |
b291f000 | 943 | goto cull_mlocked; |
894bc310 | 944 | |
a6dc60f8 | 945 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
946 | goto keep_locked; |
947 | ||
1da177e4 LT |
948 | /* Double the slab pressure for mapped and swapcache pages */ |
949 | if (page_mapped(page) || PageSwapCache(page)) | |
950 | sc->nr_scanned++; | |
951 | ||
c661b078 AW |
952 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
953 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
954 | ||
e2be15f6 MG |
955 | /* |
956 | * The number of dirty pages determines if a zone is marked | |
957 | * reclaim_congested which affects wait_iff_congested. kswapd | |
958 | * will stall and start writing pages if the tail of the LRU | |
959 | * is all dirty unqueued pages. | |
960 | */ | |
961 | page_check_dirty_writeback(page, &dirty, &writeback); | |
962 | if (dirty || writeback) | |
963 | nr_dirty++; | |
964 | ||
965 | if (dirty && !writeback) | |
966 | nr_unqueued_dirty++; | |
967 | ||
d04e8acd MG |
968 | /* |
969 | * Treat this page as congested if the underlying BDI is or if | |
970 | * pages are cycling through the LRU so quickly that the | |
971 | * pages marked for immediate reclaim are making it to the | |
972 | * end of the LRU a second time. | |
973 | */ | |
e2be15f6 | 974 | mapping = page_mapping(page); |
1da58ee2 | 975 | if (((dirty || writeback) && mapping && |
703c2708 | 976 | inode_write_congested(mapping->host)) || |
d04e8acd | 977 | (writeback && PageReclaim(page))) |
e2be15f6 MG |
978 | nr_congested++; |
979 | ||
283aba9f MG |
980 | /* |
981 | * If a page at the tail of the LRU is under writeback, there | |
982 | * are three cases to consider. | |
983 | * | |
984 | * 1) If reclaim is encountering an excessive number of pages | |
985 | * under writeback and this page is both under writeback and | |
986 | * PageReclaim then it indicates that pages are being queued | |
987 | * for IO but are being recycled through the LRU before the | |
988 | * IO can complete. Waiting on the page itself risks an | |
989 | * indefinite stall if it is impossible to writeback the | |
990 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
991 | * note that the LRU is being scanned too quickly and the |
992 | * caller can stall after page list has been processed. | |
283aba9f | 993 | * |
97c9341f | 994 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
995 | * not marked for immediate reclaim, or the caller does not |
996 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
997 | * not to fs). In this case mark the page for immediate | |
97c9341f | 998 | * reclaim and continue scanning. |
283aba9f | 999 | * |
ecf5fc6e MH |
1000 | * Require may_enter_fs because we would wait on fs, which |
1001 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1002 | * enter reclaim, and deadlock if it waits on a page for |
1003 | * which it is needed to do the write (loop masks off | |
1004 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1005 | * would probably show more reasons. | |
1006 | * | |
7fadc820 | 1007 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1008 | * PageReclaim. memcg does not have any dirty pages |
1009 | * throttling so we could easily OOM just because too many | |
1010 | * pages are in writeback and there is nothing else to | |
1011 | * reclaim. Wait for the writeback to complete. | |
1012 | */ | |
c661b078 | 1013 | if (PageWriteback(page)) { |
283aba9f MG |
1014 | /* Case 1 above */ |
1015 | if (current_is_kswapd() && | |
1016 | PageReclaim(page) && | |
599d0c95 | 1017 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
b1a6f21e MG |
1018 | nr_immediate++; |
1019 | goto keep_locked; | |
283aba9f MG |
1020 | |
1021 | /* Case 2 above */ | |
97c9341f | 1022 | } else if (sane_reclaim(sc) || |
ecf5fc6e | 1023 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1024 | /* |
1025 | * This is slightly racy - end_page_writeback() | |
1026 | * might have just cleared PageReclaim, then | |
1027 | * setting PageReclaim here end up interpreted | |
1028 | * as PageReadahead - but that does not matter | |
1029 | * enough to care. What we do want is for this | |
1030 | * page to have PageReclaim set next time memcg | |
1031 | * reclaim reaches the tests above, so it will | |
1032 | * then wait_on_page_writeback() to avoid OOM; | |
1033 | * and it's also appropriate in global reclaim. | |
1034 | */ | |
1035 | SetPageReclaim(page); | |
e62e384e | 1036 | nr_writeback++; |
c3b94f44 | 1037 | goto keep_locked; |
283aba9f MG |
1038 | |
1039 | /* Case 3 above */ | |
1040 | } else { | |
7fadc820 | 1041 | unlock_page(page); |
283aba9f | 1042 | wait_on_page_writeback(page); |
7fadc820 HD |
1043 | /* then go back and try same page again */ |
1044 | list_add_tail(&page->lru, page_list); | |
1045 | continue; | |
e62e384e | 1046 | } |
c661b078 | 1047 | } |
1da177e4 | 1048 | |
02c6de8d MK |
1049 | if (!force_reclaim) |
1050 | references = page_check_references(page, sc); | |
1051 | ||
dfc8d636 JW |
1052 | switch (references) { |
1053 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1054 | goto activate_locked; |
64574746 JW |
1055 | case PAGEREF_KEEP: |
1056 | goto keep_locked; | |
dfc8d636 JW |
1057 | case PAGEREF_RECLAIM: |
1058 | case PAGEREF_RECLAIM_CLEAN: | |
1059 | ; /* try to reclaim the page below */ | |
1060 | } | |
1da177e4 | 1061 | |
1da177e4 LT |
1062 | /* |
1063 | * Anonymous process memory has backing store? | |
1064 | * Try to allocate it some swap space here. | |
1065 | */ | |
b291f000 | 1066 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
1067 | if (!(sc->gfp_mask & __GFP_IO)) |
1068 | goto keep_locked; | |
5bc7b8ac | 1069 | if (!add_to_swap(page, page_list)) |
1da177e4 | 1070 | goto activate_locked; |
854e9ed0 | 1071 | lazyfree = true; |
63eb6b93 | 1072 | may_enter_fs = 1; |
1da177e4 | 1073 | |
e2be15f6 MG |
1074 | /* Adding to swap updated mapping */ |
1075 | mapping = page_mapping(page); | |
7751b2da KS |
1076 | } else if (unlikely(PageTransHuge(page))) { |
1077 | /* Split file THP */ | |
1078 | if (split_huge_page_to_list(page, page_list)) | |
1079 | goto keep_locked; | |
e2be15f6 | 1080 | } |
1da177e4 | 1081 | |
7751b2da KS |
1082 | VM_BUG_ON_PAGE(PageTransHuge(page), page); |
1083 | ||
1da177e4 LT |
1084 | /* |
1085 | * The page is mapped into the page tables of one or more | |
1086 | * processes. Try to unmap it here. | |
1087 | */ | |
1088 | if (page_mapped(page) && mapping) { | |
854e9ed0 MK |
1089 | switch (ret = try_to_unmap(page, lazyfree ? |
1090 | (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) : | |
1091 | (ttu_flags | TTU_BATCH_FLUSH))) { | |
1da177e4 LT |
1092 | case SWAP_FAIL: |
1093 | goto activate_locked; | |
1094 | case SWAP_AGAIN: | |
1095 | goto keep_locked; | |
b291f000 NP |
1096 | case SWAP_MLOCK: |
1097 | goto cull_mlocked; | |
854e9ed0 MK |
1098 | case SWAP_LZFREE: |
1099 | goto lazyfree; | |
1da177e4 LT |
1100 | case SWAP_SUCCESS: |
1101 | ; /* try to free the page below */ | |
1102 | } | |
1103 | } | |
1104 | ||
1105 | if (PageDirty(page)) { | |
ee72886d MG |
1106 | /* |
1107 | * Only kswapd can writeback filesystem pages to | |
d43006d5 MG |
1108 | * avoid risk of stack overflow but only writeback |
1109 | * if many dirty pages have been encountered. | |
ee72886d | 1110 | */ |
f84f6e2b | 1111 | if (page_is_file_cache(page) && |
9e3b2f8c | 1112 | (!current_is_kswapd() || |
599d0c95 | 1113 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { |
49ea7eb6 MG |
1114 | /* |
1115 | * Immediately reclaim when written back. | |
1116 | * Similar in principal to deactivate_page() | |
1117 | * except we already have the page isolated | |
1118 | * and know it's dirty | |
1119 | */ | |
c4a25635 | 1120 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1121 | SetPageReclaim(page); |
1122 | ||
ee72886d MG |
1123 | goto keep_locked; |
1124 | } | |
1125 | ||
dfc8d636 | 1126 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1127 | goto keep_locked; |
4dd4b920 | 1128 | if (!may_enter_fs) |
1da177e4 | 1129 | goto keep_locked; |
52a8363e | 1130 | if (!sc->may_writepage) |
1da177e4 LT |
1131 | goto keep_locked; |
1132 | ||
d950c947 MG |
1133 | /* |
1134 | * Page is dirty. Flush the TLB if a writable entry | |
1135 | * potentially exists to avoid CPU writes after IO | |
1136 | * starts and then write it out here. | |
1137 | */ | |
1138 | try_to_unmap_flush_dirty(); | |
7d3579e8 | 1139 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1140 | case PAGE_KEEP: |
1141 | goto keep_locked; | |
1142 | case PAGE_ACTIVATE: | |
1143 | goto activate_locked; | |
1144 | case PAGE_SUCCESS: | |
7d3579e8 | 1145 | if (PageWriteback(page)) |
41ac1999 | 1146 | goto keep; |
7d3579e8 | 1147 | if (PageDirty(page)) |
1da177e4 | 1148 | goto keep; |
7d3579e8 | 1149 | |
1da177e4 LT |
1150 | /* |
1151 | * A synchronous write - probably a ramdisk. Go | |
1152 | * ahead and try to reclaim the page. | |
1153 | */ | |
529ae9aa | 1154 | if (!trylock_page(page)) |
1da177e4 LT |
1155 | goto keep; |
1156 | if (PageDirty(page) || PageWriteback(page)) | |
1157 | goto keep_locked; | |
1158 | mapping = page_mapping(page); | |
1159 | case PAGE_CLEAN: | |
1160 | ; /* try to free the page below */ | |
1161 | } | |
1162 | } | |
1163 | ||
1164 | /* | |
1165 | * If the page has buffers, try to free the buffer mappings | |
1166 | * associated with this page. If we succeed we try to free | |
1167 | * the page as well. | |
1168 | * | |
1169 | * We do this even if the page is PageDirty(). | |
1170 | * try_to_release_page() does not perform I/O, but it is | |
1171 | * possible for a page to have PageDirty set, but it is actually | |
1172 | * clean (all its buffers are clean). This happens if the | |
1173 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1174 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1175 | * try_to_release_page() will discover that cleanness and will |
1176 | * drop the buffers and mark the page clean - it can be freed. | |
1177 | * | |
1178 | * Rarely, pages can have buffers and no ->mapping. These are | |
1179 | * the pages which were not successfully invalidated in | |
1180 | * truncate_complete_page(). We try to drop those buffers here | |
1181 | * and if that worked, and the page is no longer mapped into | |
1182 | * process address space (page_count == 1) it can be freed. | |
1183 | * Otherwise, leave the page on the LRU so it is swappable. | |
1184 | */ | |
266cf658 | 1185 | if (page_has_private(page)) { |
1da177e4 LT |
1186 | if (!try_to_release_page(page, sc->gfp_mask)) |
1187 | goto activate_locked; | |
e286781d NP |
1188 | if (!mapping && page_count(page) == 1) { |
1189 | unlock_page(page); | |
1190 | if (put_page_testzero(page)) | |
1191 | goto free_it; | |
1192 | else { | |
1193 | /* | |
1194 | * rare race with speculative reference. | |
1195 | * the speculative reference will free | |
1196 | * this page shortly, so we may | |
1197 | * increment nr_reclaimed here (and | |
1198 | * leave it off the LRU). | |
1199 | */ | |
1200 | nr_reclaimed++; | |
1201 | continue; | |
1202 | } | |
1203 | } | |
1da177e4 LT |
1204 | } |
1205 | ||
854e9ed0 | 1206 | lazyfree: |
a528910e | 1207 | if (!mapping || !__remove_mapping(mapping, page, true)) |
49d2e9cc | 1208 | goto keep_locked; |
1da177e4 | 1209 | |
a978d6f5 NP |
1210 | /* |
1211 | * At this point, we have no other references and there is | |
1212 | * no way to pick any more up (removed from LRU, removed | |
1213 | * from pagecache). Can use non-atomic bitops now (and | |
1214 | * we obviously don't have to worry about waking up a process | |
1215 | * waiting on the page lock, because there are no references. | |
1216 | */ | |
48c935ad | 1217 | __ClearPageLocked(page); |
e286781d | 1218 | free_it: |
854e9ed0 MK |
1219 | if (ret == SWAP_LZFREE) |
1220 | count_vm_event(PGLAZYFREED); | |
1221 | ||
05ff5137 | 1222 | nr_reclaimed++; |
abe4c3b5 MG |
1223 | |
1224 | /* | |
1225 | * Is there need to periodically free_page_list? It would | |
1226 | * appear not as the counts should be low | |
1227 | */ | |
1228 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1229 | continue; |
1230 | ||
b291f000 | 1231 | cull_mlocked: |
63d6c5ad HD |
1232 | if (PageSwapCache(page)) |
1233 | try_to_free_swap(page); | |
b291f000 | 1234 | unlock_page(page); |
c54839a7 | 1235 | list_add(&page->lru, &ret_pages); |
b291f000 NP |
1236 | continue; |
1237 | ||
1da177e4 | 1238 | activate_locked: |
68a22394 | 1239 | /* Not a candidate for swapping, so reclaim swap space. */ |
5ccc5aba | 1240 | if (PageSwapCache(page) && mem_cgroup_swap_full(page)) |
a2c43eed | 1241 | try_to_free_swap(page); |
309381fe | 1242 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
1243 | SetPageActive(page); |
1244 | pgactivate++; | |
1245 | keep_locked: | |
1246 | unlock_page(page); | |
1247 | keep: | |
1248 | list_add(&page->lru, &ret_pages); | |
309381fe | 1249 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1250 | } |
abe4c3b5 | 1251 | |
747db954 | 1252 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1253 | try_to_unmap_flush(); |
b745bc85 | 1254 | free_hot_cold_page_list(&free_pages, true); |
abe4c3b5 | 1255 | |
1da177e4 | 1256 | list_splice(&ret_pages, page_list); |
f8891e5e | 1257 | count_vm_events(PGACTIVATE, pgactivate); |
0a31bc97 | 1258 | |
8e950282 MG |
1259 | *ret_nr_dirty += nr_dirty; |
1260 | *ret_nr_congested += nr_congested; | |
d43006d5 | 1261 | *ret_nr_unqueued_dirty += nr_unqueued_dirty; |
92df3a72 | 1262 | *ret_nr_writeback += nr_writeback; |
b1a6f21e | 1263 | *ret_nr_immediate += nr_immediate; |
05ff5137 | 1264 | return nr_reclaimed; |
1da177e4 LT |
1265 | } |
1266 | ||
02c6de8d MK |
1267 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1268 | struct list_head *page_list) | |
1269 | { | |
1270 | struct scan_control sc = { | |
1271 | .gfp_mask = GFP_KERNEL, | |
1272 | .priority = DEF_PRIORITY, | |
1273 | .may_unmap = 1, | |
1274 | }; | |
8e950282 | 1275 | unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; |
02c6de8d MK |
1276 | struct page *page, *next; |
1277 | LIST_HEAD(clean_pages); | |
1278 | ||
1279 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1280 | if (page_is_file_cache(page) && !PageDirty(page) && |
b1123ea6 | 1281 | !__PageMovable(page)) { |
02c6de8d MK |
1282 | ClearPageActive(page); |
1283 | list_move(&page->lru, &clean_pages); | |
1284 | } | |
1285 | } | |
1286 | ||
599d0c95 | 1287 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
8e950282 MG |
1288 | TTU_UNMAP|TTU_IGNORE_ACCESS, |
1289 | &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); | |
02c6de8d | 1290 | list_splice(&clean_pages, page_list); |
599d0c95 | 1291 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1292 | return ret; |
1293 | } | |
1294 | ||
5ad333eb AW |
1295 | /* |
1296 | * Attempt to remove the specified page from its LRU. Only take this page | |
1297 | * if it is of the appropriate PageActive status. Pages which are being | |
1298 | * freed elsewhere are also ignored. | |
1299 | * | |
1300 | * page: page to consider | |
1301 | * mode: one of the LRU isolation modes defined above | |
1302 | * | |
1303 | * returns 0 on success, -ve errno on failure. | |
1304 | */ | |
f3fd4a61 | 1305 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1306 | { |
1307 | int ret = -EINVAL; | |
1308 | ||
1309 | /* Only take pages on the LRU. */ | |
1310 | if (!PageLRU(page)) | |
1311 | return ret; | |
1312 | ||
e46a2879 MK |
1313 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1314 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1315 | return ret; |
1316 | ||
5ad333eb | 1317 | ret = -EBUSY; |
08e552c6 | 1318 | |
c8244935 MG |
1319 | /* |
1320 | * To minimise LRU disruption, the caller can indicate that it only | |
1321 | * wants to isolate pages it will be able to operate on without | |
1322 | * blocking - clean pages for the most part. | |
1323 | * | |
1324 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1325 | * is used by reclaim when it is cannot write to backing storage | |
1326 | * | |
1327 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1328 | * that it is possible to migrate without blocking | |
1329 | */ | |
1330 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1331 | /* All the caller can do on PageWriteback is block */ | |
1332 | if (PageWriteback(page)) | |
1333 | return ret; | |
1334 | ||
1335 | if (PageDirty(page)) { | |
1336 | struct address_space *mapping; | |
1337 | ||
1338 | /* ISOLATE_CLEAN means only clean pages */ | |
1339 | if (mode & ISOLATE_CLEAN) | |
1340 | return ret; | |
1341 | ||
1342 | /* | |
1343 | * Only pages without mappings or that have a | |
1344 | * ->migratepage callback are possible to migrate | |
1345 | * without blocking | |
1346 | */ | |
1347 | mapping = page_mapping(page); | |
1348 | if (mapping && !mapping->a_ops->migratepage) | |
1349 | return ret; | |
1350 | } | |
1351 | } | |
39deaf85 | 1352 | |
f80c0673 MK |
1353 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1354 | return ret; | |
1355 | ||
5ad333eb AW |
1356 | if (likely(get_page_unless_zero(page))) { |
1357 | /* | |
1358 | * Be careful not to clear PageLRU until after we're | |
1359 | * sure the page is not being freed elsewhere -- the | |
1360 | * page release code relies on it. | |
1361 | */ | |
1362 | ClearPageLRU(page); | |
1363 | ret = 0; | |
1364 | } | |
1365 | ||
1366 | return ret; | |
1367 | } | |
1368 | ||
1da177e4 | 1369 | /* |
a52633d8 | 1370 | * zone_lru_lock is heavily contended. Some of the functions that |
1da177e4 LT |
1371 | * shrink the lists perform better by taking out a batch of pages |
1372 | * and working on them outside the LRU lock. | |
1373 | * | |
1374 | * For pagecache intensive workloads, this function is the hottest | |
1375 | * spot in the kernel (apart from copy_*_user functions). | |
1376 | * | |
1377 | * Appropriate locks must be held before calling this function. | |
1378 | * | |
1379 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1380 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1381 | * @dst: The temp list to put pages on to. |
f626012d | 1382 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1383 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1384 | * @mode: One of the LRU isolation modes |
3cb99451 | 1385 | * @lru: LRU list id for isolating |
1da177e4 LT |
1386 | * |
1387 | * returns how many pages were moved onto *@dst. | |
1388 | */ | |
69e05944 | 1389 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1390 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1391 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1392 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1393 | { |
75b00af7 | 1394 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1395 | unsigned long nr_taken = 0; |
599d0c95 MG |
1396 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
1397 | unsigned long scan, nr_pages; | |
b2e18757 | 1398 | LIST_HEAD(pages_skipped); |
1da177e4 | 1399 | |
0b802f10 VD |
1400 | for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan && |
1401 | !list_empty(src); scan++) { | |
5ad333eb | 1402 | struct page *page; |
5ad333eb | 1403 | |
1da177e4 LT |
1404 | page = lru_to_page(src); |
1405 | prefetchw_prev_lru_page(page, src, flags); | |
1406 | ||
309381fe | 1407 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1408 | |
b2e18757 MG |
1409 | if (page_zonenum(page) > sc->reclaim_idx) { |
1410 | list_move(&page->lru, &pages_skipped); | |
1411 | continue; | |
1412 | } | |
1413 | ||
f3fd4a61 | 1414 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1415 | case 0: |
599d0c95 MG |
1416 | nr_pages = hpage_nr_pages(page); |
1417 | nr_taken += nr_pages; | |
1418 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1419 | list_move(&page->lru, dst); |
5ad333eb AW |
1420 | break; |
1421 | ||
1422 | case -EBUSY: | |
1423 | /* else it is being freed elsewhere */ | |
1424 | list_move(&page->lru, src); | |
1425 | continue; | |
46453a6e | 1426 | |
5ad333eb AW |
1427 | default: |
1428 | BUG(); | |
1429 | } | |
1da177e4 LT |
1430 | } |
1431 | ||
b2e18757 MG |
1432 | /* |
1433 | * Splice any skipped pages to the start of the LRU list. Note that | |
1434 | * this disrupts the LRU order when reclaiming for lower zones but | |
1435 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1436 | * scanning would soon rescan the same pages to skip and put the | |
1437 | * system at risk of premature OOM. | |
1438 | */ | |
1439 | if (!list_empty(&pages_skipped)) | |
1440 | list_splice(&pages_skipped, src); | |
f626012d | 1441 | *nr_scanned = scan; |
75b00af7 HD |
1442 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, |
1443 | nr_taken, mode, is_file_lru(lru)); | |
599d0c95 MG |
1444 | for (scan = 0; scan < MAX_NR_ZONES; scan++) { |
1445 | nr_pages = nr_zone_taken[scan]; | |
1446 | if (!nr_pages) | |
1447 | continue; | |
1448 | ||
1449 | update_lru_size(lruvec, lru, scan, -nr_pages); | |
1450 | } | |
1da177e4 LT |
1451 | return nr_taken; |
1452 | } | |
1453 | ||
62695a84 NP |
1454 | /** |
1455 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1456 | * @page: page to isolate from its LRU list | |
1457 | * | |
1458 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1459 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1460 | * | |
1461 | * Returns 0 if the page was removed from an LRU list. | |
1462 | * Returns -EBUSY if the page was not on an LRU list. | |
1463 | * | |
1464 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1465 | * the active list, it will have PageActive set. If it was found on |
1466 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1467 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1468 | * |
1469 | * The vmstat statistic corresponding to the list on which the page was | |
1470 | * found will be decremented. | |
1471 | * | |
1472 | * Restrictions: | |
1473 | * (1) Must be called with an elevated refcount on the page. This is a | |
1474 | * fundamentnal difference from isolate_lru_pages (which is called | |
1475 | * without a stable reference). | |
1476 | * (2) the lru_lock must not be held. | |
1477 | * (3) interrupts must be enabled. | |
1478 | */ | |
1479 | int isolate_lru_page(struct page *page) | |
1480 | { | |
1481 | int ret = -EBUSY; | |
1482 | ||
309381fe | 1483 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1484 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1485 | |
62695a84 NP |
1486 | if (PageLRU(page)) { |
1487 | struct zone *zone = page_zone(page); | |
fa9add64 | 1488 | struct lruvec *lruvec; |
62695a84 | 1489 | |
a52633d8 | 1490 | spin_lock_irq(zone_lru_lock(zone)); |
599d0c95 | 1491 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
0c917313 | 1492 | if (PageLRU(page)) { |
894bc310 | 1493 | int lru = page_lru(page); |
0c917313 | 1494 | get_page(page); |
62695a84 | 1495 | ClearPageLRU(page); |
fa9add64 HD |
1496 | del_page_from_lru_list(page, lruvec, lru); |
1497 | ret = 0; | |
62695a84 | 1498 | } |
a52633d8 | 1499 | spin_unlock_irq(zone_lru_lock(zone)); |
62695a84 NP |
1500 | } |
1501 | return ret; | |
1502 | } | |
1503 | ||
35cd7815 | 1504 | /* |
d37dd5dc FW |
1505 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1506 | * then get resheduled. When there are massive number of tasks doing page | |
1507 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1508 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1509 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1510 | */ |
599d0c95 | 1511 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1512 | struct scan_control *sc) |
1513 | { | |
1514 | unsigned long inactive, isolated; | |
1515 | ||
1516 | if (current_is_kswapd()) | |
1517 | return 0; | |
1518 | ||
97c9341f | 1519 | if (!sane_reclaim(sc)) |
35cd7815 RR |
1520 | return 0; |
1521 | ||
1522 | if (file) { | |
599d0c95 MG |
1523 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1524 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1525 | } else { |
599d0c95 MG |
1526 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1527 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1528 | } |
1529 | ||
3cf23841 FW |
1530 | /* |
1531 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1532 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1533 | * deadlock. | |
1534 | */ | |
d0164adc | 1535 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1536 | inactive >>= 3; |
1537 | ||
35cd7815 RR |
1538 | return isolated > inactive; |
1539 | } | |
1540 | ||
66635629 | 1541 | static noinline_for_stack void |
75b00af7 | 1542 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1543 | { |
27ac81d8 | 1544 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
599d0c95 | 1545 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3f79768f | 1546 | LIST_HEAD(pages_to_free); |
66635629 | 1547 | |
66635629 MG |
1548 | /* |
1549 | * Put back any unfreeable pages. | |
1550 | */ | |
66635629 | 1551 | while (!list_empty(page_list)) { |
3f79768f | 1552 | struct page *page = lru_to_page(page_list); |
66635629 | 1553 | int lru; |
3f79768f | 1554 | |
309381fe | 1555 | VM_BUG_ON_PAGE(PageLRU(page), page); |
66635629 | 1556 | list_del(&page->lru); |
39b5f29a | 1557 | if (unlikely(!page_evictable(page))) { |
599d0c95 | 1558 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1559 | putback_lru_page(page); |
599d0c95 | 1560 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1561 | continue; |
1562 | } | |
fa9add64 | 1563 | |
599d0c95 | 1564 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1565 | |
7a608572 | 1566 | SetPageLRU(page); |
66635629 | 1567 | lru = page_lru(page); |
fa9add64 HD |
1568 | add_page_to_lru_list(page, lruvec, lru); |
1569 | ||
66635629 MG |
1570 | if (is_active_lru(lru)) { |
1571 | int file = is_file_lru(lru); | |
9992af10 RR |
1572 | int numpages = hpage_nr_pages(page); |
1573 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1574 | } |
2bcf8879 HD |
1575 | if (put_page_testzero(page)) { |
1576 | __ClearPageLRU(page); | |
1577 | __ClearPageActive(page); | |
fa9add64 | 1578 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1579 | |
1580 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1581 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1582 | mem_cgroup_uncharge(page); |
2bcf8879 | 1583 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1584 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1585 | } else |
1586 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1587 | } |
1588 | } | |
66635629 | 1589 | |
3f79768f HD |
1590 | /* |
1591 | * To save our caller's stack, now use input list for pages to free. | |
1592 | */ | |
1593 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1594 | } |
1595 | ||
399ba0b9 N |
1596 | /* |
1597 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1598 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1599 | * In that case we should only throttle if the backing device it is | |
1600 | * writing to is congested. In other cases it is safe to throttle. | |
1601 | */ | |
1602 | static int current_may_throttle(void) | |
1603 | { | |
1604 | return !(current->flags & PF_LESS_THROTTLE) || | |
1605 | current->backing_dev_info == NULL || | |
1606 | bdi_write_congested(current->backing_dev_info); | |
1607 | } | |
1608 | ||
1da177e4 | 1609 | /* |
b2e18757 | 1610 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1611 | * of reclaimed pages |
1da177e4 | 1612 | */ |
66635629 | 1613 | static noinline_for_stack unsigned long |
1a93be0e | 1614 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1615 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1616 | { |
1617 | LIST_HEAD(page_list); | |
e247dbce | 1618 | unsigned long nr_scanned; |
05ff5137 | 1619 | unsigned long nr_reclaimed = 0; |
e247dbce | 1620 | unsigned long nr_taken; |
8e950282 MG |
1621 | unsigned long nr_dirty = 0; |
1622 | unsigned long nr_congested = 0; | |
e2be15f6 | 1623 | unsigned long nr_unqueued_dirty = 0; |
92df3a72 | 1624 | unsigned long nr_writeback = 0; |
b1a6f21e | 1625 | unsigned long nr_immediate = 0; |
f3fd4a61 | 1626 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1627 | int file = is_file_lru(lru); |
599d0c95 | 1628 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1629 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
78dc583d | 1630 | |
599d0c95 | 1631 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
58355c78 | 1632 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1633 | |
1634 | /* We are about to die and free our memory. Return now. */ | |
1635 | if (fatal_signal_pending(current)) | |
1636 | return SWAP_CLUSTER_MAX; | |
1637 | } | |
1638 | ||
1da177e4 | 1639 | lru_add_drain(); |
f80c0673 MK |
1640 | |
1641 | if (!sc->may_unmap) | |
61317289 | 1642 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1643 | if (!sc->may_writepage) |
61317289 | 1644 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1645 | |
599d0c95 | 1646 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1647 | |
5dc35979 KK |
1648 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1649 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc | 1650 | |
599d0c95 | 1651 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1652 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1653 | |
89b5fae5 | 1654 | if (global_reclaim(sc)) { |
599d0c95 | 1655 | __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned); |
e247dbce | 1656 | if (current_is_kswapd()) |
599d0c95 | 1657 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
e247dbce | 1658 | else |
599d0c95 | 1659 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
e247dbce | 1660 | } |
599d0c95 | 1661 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1662 | |
d563c050 | 1663 | if (nr_taken == 0) |
66635629 | 1664 | return 0; |
5ad333eb | 1665 | |
599d0c95 | 1666 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP, |
8e950282 MG |
1667 | &nr_dirty, &nr_unqueued_dirty, &nr_congested, |
1668 | &nr_writeback, &nr_immediate, | |
1669 | false); | |
c661b078 | 1670 | |
599d0c95 | 1671 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 1672 | |
904249aa YH |
1673 | if (global_reclaim(sc)) { |
1674 | if (current_is_kswapd()) | |
599d0c95 | 1675 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
904249aa | 1676 | else |
599d0c95 | 1677 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
904249aa | 1678 | } |
a74609fa | 1679 | |
27ac81d8 | 1680 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1681 | |
599d0c95 | 1682 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 1683 | |
599d0c95 | 1684 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 1685 | |
747db954 | 1686 | mem_cgroup_uncharge_list(&page_list); |
b745bc85 | 1687 | free_hot_cold_page_list(&page_list, true); |
e11da5b4 | 1688 | |
92df3a72 MG |
1689 | /* |
1690 | * If reclaim is isolating dirty pages under writeback, it implies | |
1691 | * that the long-lived page allocation rate is exceeding the page | |
1692 | * laundering rate. Either the global limits are not being effective | |
1693 | * at throttling processes due to the page distribution throughout | |
1694 | * zones or there is heavy usage of a slow backing device. The | |
1695 | * only option is to throttle from reclaim context which is not ideal | |
1696 | * as there is no guarantee the dirtying process is throttled in the | |
1697 | * same way balance_dirty_pages() manages. | |
1698 | * | |
8e950282 MG |
1699 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1700 | * of pages under pages flagged for immediate reclaim and stall if any | |
1701 | * are encountered in the nr_immediate check below. | |
92df3a72 | 1702 | */ |
918fc718 | 1703 | if (nr_writeback && nr_writeback == nr_taken) |
599d0c95 | 1704 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); |
92df3a72 | 1705 | |
d43006d5 | 1706 | /* |
97c9341f TH |
1707 | * Legacy memcg will stall in page writeback so avoid forcibly |
1708 | * stalling here. | |
d43006d5 | 1709 | */ |
97c9341f | 1710 | if (sane_reclaim(sc)) { |
8e950282 MG |
1711 | /* |
1712 | * Tag a zone as congested if all the dirty pages scanned were | |
1713 | * backed by a congested BDI and wait_iff_congested will stall. | |
1714 | */ | |
1715 | if (nr_dirty && nr_dirty == nr_congested) | |
599d0c95 | 1716 | set_bit(PGDAT_CONGESTED, &pgdat->flags); |
8e950282 | 1717 | |
b1a6f21e MG |
1718 | /* |
1719 | * If dirty pages are scanned that are not queued for IO, it | |
1720 | * implies that flushers are not keeping up. In this case, flag | |
599d0c95 | 1721 | * the pgdat PGDAT_DIRTY and kswapd will start writing pages from |
57054651 | 1722 | * reclaim context. |
b1a6f21e MG |
1723 | */ |
1724 | if (nr_unqueued_dirty == nr_taken) | |
599d0c95 | 1725 | set_bit(PGDAT_DIRTY, &pgdat->flags); |
b1a6f21e MG |
1726 | |
1727 | /* | |
b738d764 LT |
1728 | * If kswapd scans pages marked marked for immediate |
1729 | * reclaim and under writeback (nr_immediate), it implies | |
1730 | * that pages are cycling through the LRU faster than | |
b1a6f21e MG |
1731 | * they are written so also forcibly stall. |
1732 | */ | |
b738d764 | 1733 | if (nr_immediate && current_may_throttle()) |
b1a6f21e | 1734 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
e2be15f6 | 1735 | } |
d43006d5 | 1736 | |
8e950282 MG |
1737 | /* |
1738 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1739 | * is congested. Allow kswapd to continue until it starts encountering | |
1740 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1741 | */ | |
399ba0b9 N |
1742 | if (!sc->hibernation_mode && !current_is_kswapd() && |
1743 | current_may_throttle()) | |
599d0c95 | 1744 | wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); |
8e950282 | 1745 | |
599d0c95 MG |
1746 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
1747 | nr_scanned, nr_reclaimed, | |
ba5e9579 | 1748 | sc->priority, file); |
05ff5137 | 1749 | return nr_reclaimed; |
1da177e4 LT |
1750 | } |
1751 | ||
1752 | /* | |
1753 | * This moves pages from the active list to the inactive list. | |
1754 | * | |
1755 | * We move them the other way if the page is referenced by one or more | |
1756 | * processes, from rmap. | |
1757 | * | |
1758 | * If the pages are mostly unmapped, the processing is fast and it is | |
a52633d8 | 1759 | * appropriate to hold zone_lru_lock across the whole operation. But if |
1da177e4 | 1760 | * the pages are mapped, the processing is slow (page_referenced()) so we |
a52633d8 | 1761 | * should drop zone_lru_lock around each page. It's impossible to balance |
1da177e4 LT |
1762 | * this, so instead we remove the pages from the LRU while processing them. |
1763 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1764 | * nobody will play with that bit on a non-LRU page. | |
1765 | * | |
0139aa7b | 1766 | * The downside is that we have to touch page->_refcount against each page. |
1da177e4 LT |
1767 | * But we had to alter page->flags anyway. |
1768 | */ | |
1cfb419b | 1769 | |
fa9add64 | 1770 | static void move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1771 | struct list_head *list, |
2bcf8879 | 1772 | struct list_head *pages_to_free, |
3eb4140f WF |
1773 | enum lru_list lru) |
1774 | { | |
599d0c95 | 1775 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3eb4140f | 1776 | unsigned long pgmoved = 0; |
3eb4140f | 1777 | struct page *page; |
fa9add64 | 1778 | int nr_pages; |
3eb4140f | 1779 | |
3eb4140f WF |
1780 | while (!list_empty(list)) { |
1781 | page = lru_to_page(list); | |
599d0c95 | 1782 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
3eb4140f | 1783 | |
309381fe | 1784 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3eb4140f WF |
1785 | SetPageLRU(page); |
1786 | ||
fa9add64 | 1787 | nr_pages = hpage_nr_pages(page); |
599d0c95 | 1788 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
925b7673 | 1789 | list_move(&page->lru, &lruvec->lists[lru]); |
fa9add64 | 1790 | pgmoved += nr_pages; |
3eb4140f | 1791 | |
2bcf8879 HD |
1792 | if (put_page_testzero(page)) { |
1793 | __ClearPageLRU(page); | |
1794 | __ClearPageActive(page); | |
fa9add64 | 1795 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1796 | |
1797 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1798 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1799 | mem_cgroup_uncharge(page); |
2bcf8879 | 1800 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1801 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1802 | } else |
1803 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1804 | } |
1805 | } | |
9d5e6a9f | 1806 | |
3eb4140f WF |
1807 | if (!is_active_lru(lru)) |
1808 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1809 | } | |
1cfb419b | 1810 | |
f626012d | 1811 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1812 | struct lruvec *lruvec, |
f16015fb | 1813 | struct scan_control *sc, |
9e3b2f8c | 1814 | enum lru_list lru) |
1da177e4 | 1815 | { |
44c241f1 | 1816 | unsigned long nr_taken; |
f626012d | 1817 | unsigned long nr_scanned; |
6fe6b7e3 | 1818 | unsigned long vm_flags; |
1da177e4 | 1819 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1820 | LIST_HEAD(l_active); |
b69408e8 | 1821 | LIST_HEAD(l_inactive); |
1da177e4 | 1822 | struct page *page; |
1a93be0e | 1823 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
44c241f1 | 1824 | unsigned long nr_rotated = 0; |
f3fd4a61 | 1825 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1826 | int file = is_file_lru(lru); |
599d0c95 | 1827 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
1828 | |
1829 | lru_add_drain(); | |
f80c0673 MK |
1830 | |
1831 | if (!sc->may_unmap) | |
61317289 | 1832 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1833 | if (!sc->may_writepage) |
61317289 | 1834 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1835 | |
599d0c95 | 1836 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 1837 | |
5dc35979 KK |
1838 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1839 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1840 | |
599d0c95 | 1841 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 1842 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1843 | |
9d5e6a9f | 1844 | if (global_reclaim(sc)) |
599d0c95 MG |
1845 | __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned); |
1846 | __count_vm_events(PGREFILL, nr_scanned); | |
9d5e6a9f | 1847 | |
599d0c95 | 1848 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 1849 | |
1da177e4 LT |
1850 | while (!list_empty(&l_hold)) { |
1851 | cond_resched(); | |
1852 | page = lru_to_page(&l_hold); | |
1853 | list_del(&page->lru); | |
7e9cd484 | 1854 | |
39b5f29a | 1855 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
1856 | putback_lru_page(page); |
1857 | continue; | |
1858 | } | |
1859 | ||
cc715d99 MG |
1860 | if (unlikely(buffer_heads_over_limit)) { |
1861 | if (page_has_private(page) && trylock_page(page)) { | |
1862 | if (page_has_private(page)) | |
1863 | try_to_release_page(page, 0); | |
1864 | unlock_page(page); | |
1865 | } | |
1866 | } | |
1867 | ||
c3ac9a8a JW |
1868 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1869 | &vm_flags)) { | |
9992af10 | 1870 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1871 | /* |
1872 | * Identify referenced, file-backed active pages and | |
1873 | * give them one more trip around the active list. So | |
1874 | * that executable code get better chances to stay in | |
1875 | * memory under moderate memory pressure. Anon pages | |
1876 | * are not likely to be evicted by use-once streaming | |
1877 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1878 | * so we ignore them here. | |
1879 | */ | |
41e20983 | 1880 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1881 | list_add(&page->lru, &l_active); |
1882 | continue; | |
1883 | } | |
1884 | } | |
7e9cd484 | 1885 | |
5205e56e | 1886 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1887 | list_add(&page->lru, &l_inactive); |
1888 | } | |
1889 | ||
b555749a | 1890 | /* |
8cab4754 | 1891 | * Move pages back to the lru list. |
b555749a | 1892 | */ |
599d0c95 | 1893 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 1894 | /* |
8cab4754 WF |
1895 | * Count referenced pages from currently used mappings as rotated, |
1896 | * even though only some of them are actually re-activated. This | |
1897 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 1898 | * get_scan_count. |
7e9cd484 | 1899 | */ |
b7c46d15 | 1900 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1901 | |
fa9add64 HD |
1902 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
1903 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
599d0c95 MG |
1904 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
1905 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 1906 | |
747db954 | 1907 | mem_cgroup_uncharge_list(&l_hold); |
b745bc85 | 1908 | free_hot_cold_page_list(&l_hold, true); |
1da177e4 LT |
1909 | } |
1910 | ||
59dc76b0 RR |
1911 | /* |
1912 | * The inactive anon list should be small enough that the VM never has | |
1913 | * to do too much work. | |
14797e23 | 1914 | * |
59dc76b0 RR |
1915 | * The inactive file list should be small enough to leave most memory |
1916 | * to the established workingset on the scan-resistant active list, | |
1917 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 1918 | * |
59dc76b0 RR |
1919 | * Both inactive lists should also be large enough that each inactive |
1920 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 1921 | * |
59dc76b0 RR |
1922 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
1923 | * on this LRU, maintained by the pageout code. A zone->inactive_ratio | |
1924 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. | |
56e49d21 | 1925 | * |
59dc76b0 RR |
1926 | * total target max |
1927 | * memory ratio inactive | |
1928 | * ------------------------------------- | |
1929 | * 10MB 1 5MB | |
1930 | * 100MB 1 50MB | |
1931 | * 1GB 3 250MB | |
1932 | * 10GB 10 0.9GB | |
1933 | * 100GB 31 3GB | |
1934 | * 1TB 101 10GB | |
1935 | * 10TB 320 32GB | |
56e49d21 | 1936 | */ |
59dc76b0 | 1937 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file) |
56e49d21 | 1938 | { |
59dc76b0 | 1939 | unsigned long inactive_ratio; |
e3790144 JW |
1940 | unsigned long inactive; |
1941 | unsigned long active; | |
59dc76b0 | 1942 | unsigned long gb; |
e3790144 | 1943 | |
59dc76b0 RR |
1944 | /* |
1945 | * If we don't have swap space, anonymous page deactivation | |
1946 | * is pointless. | |
1947 | */ | |
1948 | if (!file && !total_swap_pages) | |
1949 | return false; | |
56e49d21 | 1950 | |
59dc76b0 RR |
1951 | inactive = lruvec_lru_size(lruvec, file * LRU_FILE); |
1952 | active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE); | |
56e49d21 | 1953 | |
59dc76b0 RR |
1954 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
1955 | if (gb) | |
1956 | inactive_ratio = int_sqrt(10 * gb); | |
b39415b2 | 1957 | else |
59dc76b0 RR |
1958 | inactive_ratio = 1; |
1959 | ||
1960 | return inactive * inactive_ratio < active; | |
b39415b2 RR |
1961 | } |
1962 | ||
4f98a2fe | 1963 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
1a93be0e | 1964 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 1965 | { |
b39415b2 | 1966 | if (is_active_lru(lru)) { |
59dc76b0 | 1967 | if (inactive_list_is_low(lruvec, is_file_lru(lru))) |
1a93be0e | 1968 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
1969 | return 0; |
1970 | } | |
1971 | ||
1a93be0e | 1972 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
1973 | } |
1974 | ||
9a265114 JW |
1975 | enum scan_balance { |
1976 | SCAN_EQUAL, | |
1977 | SCAN_FRACT, | |
1978 | SCAN_ANON, | |
1979 | SCAN_FILE, | |
1980 | }; | |
1981 | ||
4f98a2fe RR |
1982 | /* |
1983 | * Determine how aggressively the anon and file LRU lists should be | |
1984 | * scanned. The relative value of each set of LRU lists is determined | |
1985 | * by looking at the fraction of the pages scanned we did rotate back | |
1986 | * onto the active list instead of evict. | |
1987 | * | |
be7bd59d WL |
1988 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
1989 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 1990 | */ |
33377678 | 1991 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
6b4f7799 JW |
1992 | struct scan_control *sc, unsigned long *nr, |
1993 | unsigned long *lru_pages) | |
4f98a2fe | 1994 | { |
33377678 | 1995 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
1996 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1997 | u64 fraction[2]; | |
1998 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 1999 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 2000 | unsigned long anon_prio, file_prio; |
9a265114 | 2001 | enum scan_balance scan_balance; |
0bf1457f | 2002 | unsigned long anon, file; |
9a265114 | 2003 | bool force_scan = false; |
4f98a2fe | 2004 | unsigned long ap, fp; |
4111304d | 2005 | enum lru_list lru; |
6f04f48d SS |
2006 | bool some_scanned; |
2007 | int pass; | |
246e87a9 | 2008 | |
f11c0ca5 JW |
2009 | /* |
2010 | * If the zone or memcg is small, nr[l] can be 0. This | |
2011 | * results in no scanning on this priority and a potential | |
2012 | * priority drop. Global direct reclaim can go to the next | |
2013 | * zone and tends to have no problems. Global kswapd is for | |
2014 | * zone balancing and it needs to scan a minimum amount. When | |
2015 | * reclaiming for a memcg, a priority drop can cause high | |
2016 | * latencies, so it's better to scan a minimum amount there as | |
2017 | * well. | |
2018 | */ | |
90cbc250 | 2019 | if (current_is_kswapd()) { |
599d0c95 | 2020 | if (!pgdat_reclaimable(pgdat)) |
90cbc250 | 2021 | force_scan = true; |
eb01aaab | 2022 | if (!mem_cgroup_online(memcg)) |
90cbc250 VD |
2023 | force_scan = true; |
2024 | } | |
89b5fae5 | 2025 | if (!global_reclaim(sc)) |
a4d3e9e7 | 2026 | force_scan = true; |
76a33fc3 SL |
2027 | |
2028 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2029 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2030 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2031 | goto out; |
2032 | } | |
4f98a2fe | 2033 | |
10316b31 JW |
2034 | /* |
2035 | * Global reclaim will swap to prevent OOM even with no | |
2036 | * swappiness, but memcg users want to use this knob to | |
2037 | * disable swapping for individual groups completely when | |
2038 | * using the memory controller's swap limit feature would be | |
2039 | * too expensive. | |
2040 | */ | |
02695175 | 2041 | if (!global_reclaim(sc) && !swappiness) { |
9a265114 | 2042 | scan_balance = SCAN_FILE; |
10316b31 JW |
2043 | goto out; |
2044 | } | |
2045 | ||
2046 | /* | |
2047 | * Do not apply any pressure balancing cleverness when the | |
2048 | * system is close to OOM, scan both anon and file equally | |
2049 | * (unless the swappiness setting disagrees with swapping). | |
2050 | */ | |
02695175 | 2051 | if (!sc->priority && swappiness) { |
9a265114 | 2052 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2053 | goto out; |
2054 | } | |
2055 | ||
62376251 JW |
2056 | /* |
2057 | * Prevent the reclaimer from falling into the cache trap: as | |
2058 | * cache pages start out inactive, every cache fault will tip | |
2059 | * the scan balance towards the file LRU. And as the file LRU | |
2060 | * shrinks, so does the window for rotation from references. | |
2061 | * This means we have a runaway feedback loop where a tiny | |
2062 | * thrashing file LRU becomes infinitely more attractive than | |
2063 | * anon pages. Try to detect this based on file LRU size. | |
2064 | */ | |
2065 | if (global_reclaim(sc)) { | |
599d0c95 MG |
2066 | unsigned long pgdatfile; |
2067 | unsigned long pgdatfree; | |
2068 | int z; | |
2069 | unsigned long total_high_wmark = 0; | |
2ab051e1 | 2070 | |
599d0c95 MG |
2071 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2072 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2073 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2074 | ||
2075 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2076 | struct zone *zone = &pgdat->node_zones[z]; | |
2077 | if (!populated_zone(zone)) | |
2078 | continue; | |
2079 | ||
2080 | total_high_wmark += high_wmark_pages(zone); | |
2081 | } | |
62376251 | 2082 | |
599d0c95 | 2083 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
62376251 JW |
2084 | scan_balance = SCAN_ANON; |
2085 | goto out; | |
2086 | } | |
2087 | } | |
2088 | ||
7c5bd705 | 2089 | /* |
316bda0e VD |
2090 | * If there is enough inactive page cache, i.e. if the size of the |
2091 | * inactive list is greater than that of the active list *and* the | |
2092 | * inactive list actually has some pages to scan on this priority, we | |
2093 | * do not reclaim anything from the anonymous working set right now. | |
2094 | * Without the second condition we could end up never scanning an | |
2095 | * lruvec even if it has plenty of old anonymous pages unless the | |
2096 | * system is under heavy pressure. | |
7c5bd705 | 2097 | */ |
59dc76b0 | 2098 | if (!inactive_list_is_low(lruvec, true) && |
23047a96 | 2099 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) { |
9a265114 | 2100 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2101 | goto out; |
2102 | } | |
2103 | ||
9a265114 JW |
2104 | scan_balance = SCAN_FRACT; |
2105 | ||
58c37f6e KM |
2106 | /* |
2107 | * With swappiness at 100, anonymous and file have the same priority. | |
2108 | * This scanning priority is essentially the inverse of IO cost. | |
2109 | */ | |
02695175 | 2110 | anon_prio = swappiness; |
75b00af7 | 2111 | file_prio = 200 - anon_prio; |
58c37f6e | 2112 | |
4f98a2fe RR |
2113 | /* |
2114 | * OK, so we have swap space and a fair amount of page cache | |
2115 | * pages. We use the recently rotated / recently scanned | |
2116 | * ratios to determine how valuable each cache is. | |
2117 | * | |
2118 | * Because workloads change over time (and to avoid overflow) | |
2119 | * we keep these statistics as a floating average, which ends | |
2120 | * up weighing recent references more than old ones. | |
2121 | * | |
2122 | * anon in [0], file in [1] | |
2123 | */ | |
2ab051e1 | 2124 | |
23047a96 JW |
2125 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) + |
2126 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON); | |
2127 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
2128 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE); | |
2ab051e1 | 2129 | |
599d0c95 | 2130 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2131 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2132 | reclaim_stat->recent_scanned[0] /= 2; |
2133 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2134 | } |
2135 | ||
6e901571 | 2136 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2137 | reclaim_stat->recent_scanned[1] /= 2; |
2138 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2139 | } |
2140 | ||
4f98a2fe | 2141 | /* |
00d8089c RR |
2142 | * The amount of pressure on anon vs file pages is inversely |
2143 | * proportional to the fraction of recently scanned pages on | |
2144 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2145 | */ |
fe35004f | 2146 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2147 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2148 | |
fe35004f | 2149 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2150 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2151 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2152 | |
76a33fc3 SL |
2153 | fraction[0] = ap; |
2154 | fraction[1] = fp; | |
2155 | denominator = ap + fp + 1; | |
2156 | out: | |
6f04f48d SS |
2157 | some_scanned = false; |
2158 | /* Only use force_scan on second pass. */ | |
2159 | for (pass = 0; !some_scanned && pass < 2; pass++) { | |
6b4f7799 | 2160 | *lru_pages = 0; |
6f04f48d SS |
2161 | for_each_evictable_lru(lru) { |
2162 | int file = is_file_lru(lru); | |
2163 | unsigned long size; | |
2164 | unsigned long scan; | |
6e08a369 | 2165 | |
23047a96 | 2166 | size = lruvec_lru_size(lruvec, lru); |
6f04f48d | 2167 | scan = size >> sc->priority; |
9a265114 | 2168 | |
6f04f48d SS |
2169 | if (!scan && pass && force_scan) |
2170 | scan = min(size, SWAP_CLUSTER_MAX); | |
9a265114 | 2171 | |
6f04f48d SS |
2172 | switch (scan_balance) { |
2173 | case SCAN_EQUAL: | |
2174 | /* Scan lists relative to size */ | |
2175 | break; | |
2176 | case SCAN_FRACT: | |
2177 | /* | |
2178 | * Scan types proportional to swappiness and | |
2179 | * their relative recent reclaim efficiency. | |
2180 | */ | |
2181 | scan = div64_u64(scan * fraction[file], | |
2182 | denominator); | |
2183 | break; | |
2184 | case SCAN_FILE: | |
2185 | case SCAN_ANON: | |
2186 | /* Scan one type exclusively */ | |
6b4f7799 JW |
2187 | if ((scan_balance == SCAN_FILE) != file) { |
2188 | size = 0; | |
6f04f48d | 2189 | scan = 0; |
6b4f7799 | 2190 | } |
6f04f48d SS |
2191 | break; |
2192 | default: | |
2193 | /* Look ma, no brain */ | |
2194 | BUG(); | |
2195 | } | |
6b4f7799 JW |
2196 | |
2197 | *lru_pages += size; | |
6f04f48d | 2198 | nr[lru] = scan; |
6b4f7799 | 2199 | |
9a265114 | 2200 | /* |
6f04f48d SS |
2201 | * Skip the second pass and don't force_scan, |
2202 | * if we found something to scan. | |
9a265114 | 2203 | */ |
6f04f48d | 2204 | some_scanned |= !!scan; |
9a265114 | 2205 | } |
76a33fc3 | 2206 | } |
6e08a369 | 2207 | } |
4f98a2fe | 2208 | |
72b252ae MG |
2209 | #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH |
2210 | static void init_tlb_ubc(void) | |
2211 | { | |
2212 | /* | |
2213 | * This deliberately does not clear the cpumask as it's expensive | |
2214 | * and unnecessary. If there happens to be data in there then the | |
2215 | * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and | |
2216 | * then will be cleared. | |
2217 | */ | |
2218 | current->tlb_ubc.flush_required = false; | |
2219 | } | |
2220 | #else | |
2221 | static inline void init_tlb_ubc(void) | |
2222 | { | |
2223 | } | |
2224 | #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ | |
2225 | ||
9b4f98cd | 2226 | /* |
a9dd0a83 | 2227 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. |
9b4f98cd | 2228 | */ |
a9dd0a83 | 2229 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, |
33377678 | 2230 | struct scan_control *sc, unsigned long *lru_pages) |
9b4f98cd | 2231 | { |
ef8f2327 | 2232 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
9b4f98cd | 2233 | unsigned long nr[NR_LRU_LISTS]; |
e82e0561 | 2234 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2235 | unsigned long nr_to_scan; |
2236 | enum lru_list lru; | |
2237 | unsigned long nr_reclaimed = 0; | |
2238 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2239 | struct blk_plug plug; | |
1a501907 | 2240 | bool scan_adjusted; |
9b4f98cd | 2241 | |
33377678 | 2242 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
9b4f98cd | 2243 | |
e82e0561 MG |
2244 | /* Record the original scan target for proportional adjustments later */ |
2245 | memcpy(targets, nr, sizeof(nr)); | |
2246 | ||
1a501907 MG |
2247 | /* |
2248 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2249 | * event that can occur when there is little memory pressure e.g. | |
2250 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2251 | * when the requested number of pages are reclaimed when scanning at | |
2252 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2253 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2254 | * do a batch of work at once. For memcg reclaim one check is made to | |
2255 | * abort proportional reclaim if either the file or anon lru has already | |
2256 | * dropped to zero at the first pass. | |
2257 | */ | |
2258 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2259 | sc->priority == DEF_PRIORITY); | |
2260 | ||
72b252ae MG |
2261 | init_tlb_ubc(); |
2262 | ||
9b4f98cd JW |
2263 | blk_start_plug(&plug); |
2264 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2265 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2266 | unsigned long nr_anon, nr_file, percentage; |
2267 | unsigned long nr_scanned; | |
2268 | ||
9b4f98cd JW |
2269 | for_each_evictable_lru(lru) { |
2270 | if (nr[lru]) { | |
2271 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2272 | nr[lru] -= nr_to_scan; | |
2273 | ||
2274 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2275 | lruvec, sc); | |
2276 | } | |
2277 | } | |
e82e0561 MG |
2278 | |
2279 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) | |
2280 | continue; | |
2281 | ||
e82e0561 MG |
2282 | /* |
2283 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2284 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2285 | * proportionally what was requested by get_scan_count(). We |
2286 | * stop reclaiming one LRU and reduce the amount scanning | |
2287 | * proportional to the original scan target. | |
2288 | */ | |
2289 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2290 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2291 | ||
1a501907 MG |
2292 | /* |
2293 | * It's just vindictive to attack the larger once the smaller | |
2294 | * has gone to zero. And given the way we stop scanning the | |
2295 | * smaller below, this makes sure that we only make one nudge | |
2296 | * towards proportionality once we've got nr_to_reclaim. | |
2297 | */ | |
2298 | if (!nr_file || !nr_anon) | |
2299 | break; | |
2300 | ||
e82e0561 MG |
2301 | if (nr_file > nr_anon) { |
2302 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2303 | targets[LRU_ACTIVE_ANON] + 1; | |
2304 | lru = LRU_BASE; | |
2305 | percentage = nr_anon * 100 / scan_target; | |
2306 | } else { | |
2307 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2308 | targets[LRU_ACTIVE_FILE] + 1; | |
2309 | lru = LRU_FILE; | |
2310 | percentage = nr_file * 100 / scan_target; | |
2311 | } | |
2312 | ||
2313 | /* Stop scanning the smaller of the LRU */ | |
2314 | nr[lru] = 0; | |
2315 | nr[lru + LRU_ACTIVE] = 0; | |
2316 | ||
2317 | /* | |
2318 | * Recalculate the other LRU scan count based on its original | |
2319 | * scan target and the percentage scanning already complete | |
2320 | */ | |
2321 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2322 | nr_scanned = targets[lru] - nr[lru]; | |
2323 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2324 | nr[lru] -= min(nr[lru], nr_scanned); | |
2325 | ||
2326 | lru += LRU_ACTIVE; | |
2327 | nr_scanned = targets[lru] - nr[lru]; | |
2328 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2329 | nr[lru] -= min(nr[lru], nr_scanned); | |
2330 | ||
2331 | scan_adjusted = true; | |
9b4f98cd JW |
2332 | } |
2333 | blk_finish_plug(&plug); | |
2334 | sc->nr_reclaimed += nr_reclaimed; | |
2335 | ||
2336 | /* | |
2337 | * Even if we did not try to evict anon pages at all, we want to | |
2338 | * rebalance the anon lru active/inactive ratio. | |
2339 | */ | |
59dc76b0 | 2340 | if (inactive_list_is_low(lruvec, false)) |
9b4f98cd JW |
2341 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2342 | sc, LRU_ACTIVE_ANON); | |
2343 | ||
2344 | throttle_vm_writeout(sc->gfp_mask); | |
2345 | } | |
2346 | ||
23b9da55 | 2347 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2348 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2349 | { |
d84da3f9 | 2350 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2351 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2352 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2353 | return true; |
2354 | ||
2355 | return false; | |
2356 | } | |
2357 | ||
3e7d3449 | 2358 | /* |
23b9da55 MG |
2359 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2360 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2361 | * true if more pages should be reclaimed such that when the page allocator | |
2362 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2363 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2364 | */ |
a9dd0a83 | 2365 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 MG |
2366 | unsigned long nr_reclaimed, |
2367 | unsigned long nr_scanned, | |
2368 | struct scan_control *sc) | |
2369 | { | |
2370 | unsigned long pages_for_compaction; | |
2371 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2372 | int z; |
3e7d3449 MG |
2373 | |
2374 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2375 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2376 | return false; |
2377 | ||
2876592f MG |
2378 | /* Consider stopping depending on scan and reclaim activity */ |
2379 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2380 | /* | |
2381 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2382 | * full LRU list has been scanned and we are still failing | |
2383 | * to reclaim pages. This full LRU scan is potentially | |
2384 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2385 | */ | |
2386 | if (!nr_reclaimed && !nr_scanned) | |
2387 | return false; | |
2388 | } else { | |
2389 | /* | |
2390 | * For non-__GFP_REPEAT allocations which can presumably | |
2391 | * fail without consequence, stop if we failed to reclaim | |
2392 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2393 | * pages that were scanned. This will return to the | |
2394 | * caller faster at the risk reclaim/compaction and | |
2395 | * the resulting allocation attempt fails | |
2396 | */ | |
2397 | if (!nr_reclaimed) | |
2398 | return false; | |
2399 | } | |
3e7d3449 MG |
2400 | |
2401 | /* | |
2402 | * If we have not reclaimed enough pages for compaction and the | |
2403 | * inactive lists are large enough, continue reclaiming | |
2404 | */ | |
2405 | pages_for_compaction = (2UL << sc->order); | |
a9dd0a83 | 2406 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
ec8acf20 | 2407 | if (get_nr_swap_pages() > 0) |
a9dd0a83 | 2408 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
3e7d3449 MG |
2409 | if (sc->nr_reclaimed < pages_for_compaction && |
2410 | inactive_lru_pages > pages_for_compaction) | |
2411 | return true; | |
2412 | ||
2413 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
a9dd0a83 MG |
2414 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2415 | struct zone *zone = &pgdat->node_zones[z]; | |
2416 | if (!populated_zone(zone)) | |
2417 | continue; | |
2418 | ||
2419 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
2420 | case COMPACT_PARTIAL: | |
2421 | case COMPACT_CONTINUE: | |
2422 | return false; | |
2423 | default: | |
2424 | /* check next zone */ | |
2425 | ; | |
2426 | } | |
3e7d3449 | 2427 | } |
a9dd0a83 | 2428 | return true; |
3e7d3449 MG |
2429 | } |
2430 | ||
970a39a3 | 2431 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2432 | { |
cb731d6c | 2433 | struct reclaim_state *reclaim_state = current->reclaim_state; |
f0fdc5e8 | 2434 | unsigned long nr_reclaimed, nr_scanned; |
2344d7e4 | 2435 | bool reclaimable = false; |
1da177e4 | 2436 | |
9b4f98cd JW |
2437 | do { |
2438 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2439 | struct mem_cgroup_reclaim_cookie reclaim = { | |
ef8f2327 | 2440 | .pgdat = pgdat, |
9b4f98cd JW |
2441 | .priority = sc->priority, |
2442 | }; | |
a9dd0a83 | 2443 | unsigned long node_lru_pages = 0; |
694fbc0f | 2444 | struct mem_cgroup *memcg; |
3e7d3449 | 2445 | |
9b4f98cd JW |
2446 | nr_reclaimed = sc->nr_reclaimed; |
2447 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2448 | |
694fbc0f AM |
2449 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2450 | do { | |
6b4f7799 | 2451 | unsigned long lru_pages; |
8e8ae645 | 2452 | unsigned long reclaimed; |
cb731d6c | 2453 | unsigned long scanned; |
5660048c | 2454 | |
241994ed JW |
2455 | if (mem_cgroup_low(root, memcg)) { |
2456 | if (!sc->may_thrash) | |
2457 | continue; | |
2458 | mem_cgroup_events(memcg, MEMCG_LOW, 1); | |
2459 | } | |
2460 | ||
8e8ae645 | 2461 | reclaimed = sc->nr_reclaimed; |
cb731d6c | 2462 | scanned = sc->nr_scanned; |
f9be23d6 | 2463 | |
a9dd0a83 MG |
2464 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
2465 | node_lru_pages += lru_pages; | |
f16015fb | 2466 | |
b2e18757 | 2467 | if (!global_reclaim(sc)) |
a9dd0a83 | 2468 | shrink_slab(sc->gfp_mask, pgdat->node_id, |
cb731d6c VD |
2469 | memcg, sc->nr_scanned - scanned, |
2470 | lru_pages); | |
2471 | ||
8e8ae645 JW |
2472 | /* Record the group's reclaim efficiency */ |
2473 | vmpressure(sc->gfp_mask, memcg, false, | |
2474 | sc->nr_scanned - scanned, | |
2475 | sc->nr_reclaimed - reclaimed); | |
2476 | ||
9b4f98cd | 2477 | /* |
a394cb8e MH |
2478 | * Direct reclaim and kswapd have to scan all memory |
2479 | * cgroups to fulfill the overall scan target for the | |
a9dd0a83 | 2480 | * node. |
a394cb8e MH |
2481 | * |
2482 | * Limit reclaim, on the other hand, only cares about | |
2483 | * nr_to_reclaim pages to be reclaimed and it will | |
2484 | * retry with decreasing priority if one round over the | |
2485 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2486 | */ |
a394cb8e MH |
2487 | if (!global_reclaim(sc) && |
2488 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2489 | mem_cgroup_iter_break(root, memcg); |
2490 | break; | |
2491 | } | |
241994ed | 2492 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
70ddf637 | 2493 | |
6b4f7799 JW |
2494 | /* |
2495 | * Shrink the slab caches in the same proportion that | |
2496 | * the eligible LRU pages were scanned. | |
2497 | */ | |
b2e18757 | 2498 | if (global_reclaim(sc)) |
a9dd0a83 | 2499 | shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, |
cb731d6c | 2500 | sc->nr_scanned - nr_scanned, |
a9dd0a83 | 2501 | node_lru_pages); |
cb731d6c VD |
2502 | |
2503 | if (reclaim_state) { | |
2504 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2505 | reclaim_state->reclaimed_slab = 0; | |
6b4f7799 JW |
2506 | } |
2507 | ||
8e8ae645 JW |
2508 | /* Record the subtree's reclaim efficiency */ |
2509 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
70ddf637 AV |
2510 | sc->nr_scanned - nr_scanned, |
2511 | sc->nr_reclaimed - nr_reclaimed); | |
2512 | ||
2344d7e4 JW |
2513 | if (sc->nr_reclaimed - nr_reclaimed) |
2514 | reclaimable = true; | |
2515 | ||
a9dd0a83 | 2516 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
9b4f98cd | 2517 | sc->nr_scanned - nr_scanned, sc)); |
2344d7e4 JW |
2518 | |
2519 | return reclaimable; | |
f16015fb JW |
2520 | } |
2521 | ||
53853e2d VB |
2522 | /* |
2523 | * Returns true if compaction should go ahead for a high-order request, or | |
2524 | * the high-order allocation would succeed without compaction. | |
2525 | */ | |
4f588331 | 2526 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2527 | { |
31483b6a | 2528 | unsigned long watermark; |
fe4b1b24 MG |
2529 | bool watermark_ok; |
2530 | ||
fe4b1b24 MG |
2531 | /* |
2532 | * Compaction takes time to run and there are potentially other | |
2533 | * callers using the pages just freed. Continue reclaiming until | |
2534 | * there is a buffer of free pages available to give compaction | |
2535 | * a reasonable chance of completing and allocating the page | |
2536 | */ | |
4f588331 MG |
2537 | watermark = high_wmark_pages(zone) + (2UL << sc->order); |
2538 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); | |
fe4b1b24 MG |
2539 | |
2540 | /* | |
2541 | * If compaction is deferred, reclaim up to a point where | |
2542 | * compaction will have a chance of success when re-enabled | |
2543 | */ | |
4f588331 | 2544 | if (compaction_deferred(zone, sc->order)) |
fe4b1b24 MG |
2545 | return watermark_ok; |
2546 | ||
53853e2d VB |
2547 | /* |
2548 | * If compaction is not ready to start and allocation is not likely | |
2549 | * to succeed without it, then keep reclaiming. | |
2550 | */ | |
4f588331 | 2551 | if (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx) == COMPACT_SKIPPED) |
fe4b1b24 MG |
2552 | return false; |
2553 | ||
2554 | return watermark_ok; | |
2555 | } | |
2556 | ||
1da177e4 LT |
2557 | /* |
2558 | * This is the direct reclaim path, for page-allocating processes. We only | |
2559 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2560 | * request. | |
2561 | * | |
1da177e4 LT |
2562 | * If a zone is deemed to be full of pinned pages then just give it a light |
2563 | * scan then give up on it. | |
2564 | */ | |
0a0337e0 | 2565 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2566 | { |
dd1a239f | 2567 | struct zoneref *z; |
54a6eb5c | 2568 | struct zone *zone; |
0608f43d AM |
2569 | unsigned long nr_soft_reclaimed; |
2570 | unsigned long nr_soft_scanned; | |
619d0d76 | 2571 | gfp_t orig_mask; |
79dafcdc | 2572 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2573 | |
cc715d99 MG |
2574 | /* |
2575 | * If the number of buffer_heads in the machine exceeds the maximum | |
2576 | * allowed level, force direct reclaim to scan the highmem zone as | |
2577 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2578 | */ | |
619d0d76 | 2579 | orig_mask = sc->gfp_mask; |
b2e18757 | 2580 | if (buffer_heads_over_limit) { |
cc715d99 | 2581 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2582 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2583 | } |
cc715d99 | 2584 | |
d4debc66 | 2585 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2586 | sc->reclaim_idx, sc->nodemask) { |
f3fe6512 | 2587 | if (!populated_zone(zone)) |
1da177e4 | 2588 | continue; |
6b4f7799 | 2589 | |
1cfb419b KH |
2590 | /* |
2591 | * Take care memory controller reclaiming has small influence | |
2592 | * to global LRU. | |
2593 | */ | |
89b5fae5 | 2594 | if (global_reclaim(sc)) { |
344736f2 VD |
2595 | if (!cpuset_zone_allowed(zone, |
2596 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2597 | continue; |
65ec02cb | 2598 | |
6e543d57 | 2599 | if (sc->priority != DEF_PRIORITY && |
599d0c95 | 2600 | !pgdat_reclaimable(zone->zone_pgdat)) |
1cfb419b | 2601 | continue; /* Let kswapd poll it */ |
0b06496a JW |
2602 | |
2603 | /* | |
2604 | * If we already have plenty of memory free for | |
2605 | * compaction in this zone, don't free any more. | |
2606 | * Even though compaction is invoked for any | |
2607 | * non-zero order, only frequent costly order | |
2608 | * reclamation is disruptive enough to become a | |
2609 | * noticeable problem, like transparent huge | |
2610 | * page allocations. | |
2611 | */ | |
2612 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2613 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2614 | compaction_ready(zone, sc)) { |
0b06496a JW |
2615 | sc->compaction_ready = true; |
2616 | continue; | |
e0887c19 | 2617 | } |
0b06496a | 2618 | |
79dafcdc MG |
2619 | /* |
2620 | * Shrink each node in the zonelist once. If the | |
2621 | * zonelist is ordered by zone (not the default) then a | |
2622 | * node may be shrunk multiple times but in that case | |
2623 | * the user prefers lower zones being preserved. | |
2624 | */ | |
2625 | if (zone->zone_pgdat == last_pgdat) | |
2626 | continue; | |
2627 | ||
0608f43d AM |
2628 | /* |
2629 | * This steals pages from memory cgroups over softlimit | |
2630 | * and returns the number of reclaimed pages and | |
2631 | * scanned pages. This works for global memory pressure | |
2632 | * and balancing, not for a memcg's limit. | |
2633 | */ | |
2634 | nr_soft_scanned = 0; | |
ef8f2327 | 2635 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2636 | sc->order, sc->gfp_mask, |
2637 | &nr_soft_scanned); | |
2638 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2639 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2640 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2641 | } |
408d8544 | 2642 | |
79dafcdc MG |
2643 | /* See comment about same check for global reclaim above */ |
2644 | if (zone->zone_pgdat == last_pgdat) | |
2645 | continue; | |
2646 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2647 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2648 | } |
e0c23279 | 2649 | |
619d0d76 WY |
2650 | /* |
2651 | * Restore to original mask to avoid the impact on the caller if we | |
2652 | * promoted it to __GFP_HIGHMEM. | |
2653 | */ | |
2654 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2655 | } |
4f98a2fe | 2656 | |
1da177e4 LT |
2657 | /* |
2658 | * This is the main entry point to direct page reclaim. | |
2659 | * | |
2660 | * If a full scan of the inactive list fails to free enough memory then we | |
2661 | * are "out of memory" and something needs to be killed. | |
2662 | * | |
2663 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2664 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2665 | * caller can't do much about. We kick the writeback threads and take explicit |
2666 | * naps in the hope that some of these pages can be written. But if the | |
2667 | * allocating task holds filesystem locks which prevent writeout this might not | |
2668 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2669 | * |
2670 | * returns: 0, if no pages reclaimed | |
2671 | * else, the number of pages reclaimed | |
1da177e4 | 2672 | */ |
dac1d27b | 2673 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2674 | struct scan_control *sc) |
1da177e4 | 2675 | { |
241994ed | 2676 | int initial_priority = sc->priority; |
69e05944 | 2677 | unsigned long total_scanned = 0; |
22fba335 | 2678 | unsigned long writeback_threshold; |
241994ed | 2679 | retry: |
873b4771 KK |
2680 | delayacct_freepages_start(); |
2681 | ||
89b5fae5 | 2682 | if (global_reclaim(sc)) |
1cfb419b | 2683 | count_vm_event(ALLOCSTALL); |
1da177e4 | 2684 | |
9e3b2f8c | 2685 | do { |
70ddf637 AV |
2686 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2687 | sc->priority); | |
66e1707b | 2688 | sc->nr_scanned = 0; |
0a0337e0 | 2689 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 2690 | |
66e1707b | 2691 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2692 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
2693 | break; |
2694 | ||
2695 | if (sc->compaction_ready) | |
2696 | break; | |
1da177e4 | 2697 | |
0e50ce3b MK |
2698 | /* |
2699 | * If we're getting trouble reclaiming, start doing | |
2700 | * writepage even in laptop mode. | |
2701 | */ | |
2702 | if (sc->priority < DEF_PRIORITY - 2) | |
2703 | sc->may_writepage = 1; | |
2704 | ||
1da177e4 LT |
2705 | /* |
2706 | * Try to write back as many pages as we just scanned. This | |
2707 | * tends to cause slow streaming writers to write data to the | |
2708 | * disk smoothly, at the dirtying rate, which is nice. But | |
2709 | * that's undesirable in laptop mode, where we *want* lumpy | |
2710 | * writeout. So in laptop mode, write out the whole world. | |
2711 | */ | |
22fba335 KM |
2712 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2713 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2714 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2715 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2716 | sc->may_writepage = 1; |
1da177e4 | 2717 | } |
0b06496a | 2718 | } while (--sc->priority >= 0); |
bb21c7ce | 2719 | |
873b4771 KK |
2720 | delayacct_freepages_end(); |
2721 | ||
bb21c7ce KM |
2722 | if (sc->nr_reclaimed) |
2723 | return sc->nr_reclaimed; | |
2724 | ||
0cee34fd | 2725 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 2726 | if (sc->compaction_ready) |
7335084d MG |
2727 | return 1; |
2728 | ||
241994ed JW |
2729 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
2730 | if (!sc->may_thrash) { | |
2731 | sc->priority = initial_priority; | |
2732 | sc->may_thrash = 1; | |
2733 | goto retry; | |
2734 | } | |
2735 | ||
bb21c7ce | 2736 | return 0; |
1da177e4 LT |
2737 | } |
2738 | ||
5515061d MG |
2739 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) |
2740 | { | |
2741 | struct zone *zone; | |
2742 | unsigned long pfmemalloc_reserve = 0; | |
2743 | unsigned long free_pages = 0; | |
2744 | int i; | |
2745 | bool wmark_ok; | |
2746 | ||
2747 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2748 | zone = &pgdat->node_zones[i]; | |
f012a84a | 2749 | if (!populated_zone(zone) || |
599d0c95 | 2750 | pgdat_reclaimable_pages(pgdat) == 0) |
675becce MG |
2751 | continue; |
2752 | ||
5515061d MG |
2753 | pfmemalloc_reserve += min_wmark_pages(zone); |
2754 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2755 | } | |
2756 | ||
675becce MG |
2757 | /* If there are no reserves (unexpected config) then do not throttle */ |
2758 | if (!pfmemalloc_reserve) | |
2759 | return true; | |
2760 | ||
5515061d MG |
2761 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
2762 | ||
2763 | /* kswapd must be awake if processes are being throttled */ | |
2764 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
38087d9b | 2765 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
5515061d MG |
2766 | (enum zone_type)ZONE_NORMAL); |
2767 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2768 | } | |
2769 | ||
2770 | return wmark_ok; | |
2771 | } | |
2772 | ||
2773 | /* | |
2774 | * Throttle direct reclaimers if backing storage is backed by the network | |
2775 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2776 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2777 | * when the low watermark is reached. |
2778 | * | |
2779 | * Returns true if a fatal signal was delivered during throttling. If this | |
2780 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2781 | */ |
50694c28 | 2782 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2783 | nodemask_t *nodemask) |
2784 | { | |
675becce | 2785 | struct zoneref *z; |
5515061d | 2786 | struct zone *zone; |
675becce | 2787 | pg_data_t *pgdat = NULL; |
5515061d MG |
2788 | |
2789 | /* | |
2790 | * Kernel threads should not be throttled as they may be indirectly | |
2791 | * responsible for cleaning pages necessary for reclaim to make forward | |
2792 | * progress. kjournald for example may enter direct reclaim while | |
2793 | * committing a transaction where throttling it could forcing other | |
2794 | * processes to block on log_wait_commit(). | |
2795 | */ | |
2796 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2797 | goto out; |
2798 | ||
2799 | /* | |
2800 | * If a fatal signal is pending, this process should not throttle. | |
2801 | * It should return quickly so it can exit and free its memory | |
2802 | */ | |
2803 | if (fatal_signal_pending(current)) | |
2804 | goto out; | |
5515061d | 2805 | |
675becce MG |
2806 | /* |
2807 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2808 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2809 | * GFP_KERNEL will be required for allocating network buffers when | |
2810 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2811 | * | |
2812 | * Throttling is based on the first usable node and throttled processes | |
2813 | * wait on a queue until kswapd makes progress and wakes them. There | |
2814 | * is an affinity then between processes waking up and where reclaim | |
2815 | * progress has been made assuming the process wakes on the same node. | |
2816 | * More importantly, processes running on remote nodes will not compete | |
2817 | * for remote pfmemalloc reserves and processes on different nodes | |
2818 | * should make reasonable progress. | |
2819 | */ | |
2820 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 2821 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
2822 | if (zone_idx(zone) > ZONE_NORMAL) |
2823 | continue; | |
2824 | ||
2825 | /* Throttle based on the first usable node */ | |
2826 | pgdat = zone->zone_pgdat; | |
2827 | if (pfmemalloc_watermark_ok(pgdat)) | |
2828 | goto out; | |
2829 | break; | |
2830 | } | |
2831 | ||
2832 | /* If no zone was usable by the allocation flags then do not throttle */ | |
2833 | if (!pgdat) | |
50694c28 | 2834 | goto out; |
5515061d | 2835 | |
68243e76 MG |
2836 | /* Account for the throttling */ |
2837 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2838 | ||
5515061d MG |
2839 | /* |
2840 | * If the caller cannot enter the filesystem, it's possible that it | |
2841 | * is due to the caller holding an FS lock or performing a journal | |
2842 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2843 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2844 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2845 | * second before continuing. | |
2846 | */ | |
2847 | if (!(gfp_mask & __GFP_FS)) { | |
2848 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2849 | pfmemalloc_watermark_ok(pgdat), HZ); | |
50694c28 MG |
2850 | |
2851 | goto check_pending; | |
5515061d MG |
2852 | } |
2853 | ||
2854 | /* Throttle until kswapd wakes the process */ | |
2855 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2856 | pfmemalloc_watermark_ok(pgdat)); | |
50694c28 MG |
2857 | |
2858 | check_pending: | |
2859 | if (fatal_signal_pending(current)) | |
2860 | return true; | |
2861 | ||
2862 | out: | |
2863 | return false; | |
5515061d MG |
2864 | } |
2865 | ||
dac1d27b | 2866 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2867 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2868 | { |
33906bc5 | 2869 | unsigned long nr_reclaimed; |
66e1707b | 2870 | struct scan_control sc = { |
ee814fe2 | 2871 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
21caf2fc | 2872 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
b2e18757 | 2873 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
2874 | .order = order, |
2875 | .nodemask = nodemask, | |
2876 | .priority = DEF_PRIORITY, | |
66e1707b | 2877 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2878 | .may_unmap = 1, |
2e2e4259 | 2879 | .may_swap = 1, |
66e1707b BS |
2880 | }; |
2881 | ||
5515061d | 2882 | /* |
50694c28 MG |
2883 | * Do not enter reclaim if fatal signal was delivered while throttled. |
2884 | * 1 is returned so that the page allocator does not OOM kill at this | |
2885 | * point. | |
5515061d | 2886 | */ |
50694c28 | 2887 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) |
5515061d MG |
2888 | return 1; |
2889 | ||
33906bc5 MG |
2890 | trace_mm_vmscan_direct_reclaim_begin(order, |
2891 | sc.may_writepage, | |
2892 | gfp_mask); | |
2893 | ||
3115cd91 | 2894 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
2895 | |
2896 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2897 | ||
2898 | return nr_reclaimed; | |
66e1707b BS |
2899 | } |
2900 | ||
c255a458 | 2901 | #ifdef CONFIG_MEMCG |
66e1707b | 2902 | |
a9dd0a83 | 2903 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 2904 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 2905 | pg_data_t *pgdat, |
0ae5e89c | 2906 | unsigned long *nr_scanned) |
4e416953 BS |
2907 | { |
2908 | struct scan_control sc = { | |
b8f5c566 | 2909 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 2910 | .target_mem_cgroup = memcg, |
4e416953 BS |
2911 | .may_writepage = !laptop_mode, |
2912 | .may_unmap = 1, | |
b2e18757 | 2913 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 2914 | .may_swap = !noswap, |
4e416953 | 2915 | }; |
6b4f7799 | 2916 | unsigned long lru_pages; |
0ae5e89c | 2917 | |
4e416953 BS |
2918 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2919 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 2920 | |
9e3b2f8c | 2921 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e KM |
2922 | sc.may_writepage, |
2923 | sc.gfp_mask); | |
2924 | ||
4e416953 BS |
2925 | /* |
2926 | * NOTE: Although we can get the priority field, using it | |
2927 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 2928 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
2929 | * will pick up pages from other mem cgroup's as well. We hack |
2930 | * the priority and make it zero. | |
2931 | */ | |
ef8f2327 | 2932 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
bdce6d9e KM |
2933 | |
2934 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2935 | ||
0ae5e89c | 2936 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2937 | return sc.nr_reclaimed; |
2938 | } | |
2939 | ||
72835c86 | 2940 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 2941 | unsigned long nr_pages, |
a7885eb8 | 2942 | gfp_t gfp_mask, |
b70a2a21 | 2943 | bool may_swap) |
66e1707b | 2944 | { |
4e416953 | 2945 | struct zonelist *zonelist; |
bdce6d9e | 2946 | unsigned long nr_reclaimed; |
889976db | 2947 | int nid; |
66e1707b | 2948 | struct scan_control sc = { |
b70a2a21 | 2949 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
a09ed5e0 YH |
2950 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2951 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
b2e18757 | 2952 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
2953 | .target_mem_cgroup = memcg, |
2954 | .priority = DEF_PRIORITY, | |
2955 | .may_writepage = !laptop_mode, | |
2956 | .may_unmap = 1, | |
b70a2a21 | 2957 | .may_swap = may_swap, |
a09ed5e0 | 2958 | }; |
66e1707b | 2959 | |
889976db YH |
2960 | /* |
2961 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2962 | * take care of from where we get pages. So the node where we start the | |
2963 | * scan does not need to be the current node. | |
2964 | */ | |
72835c86 | 2965 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2966 | |
2967 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2968 | |
2969 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2970 | sc.may_writepage, | |
2971 | sc.gfp_mask); | |
2972 | ||
3115cd91 | 2973 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
bdce6d9e KM |
2974 | |
2975 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2976 | ||
2977 | return nr_reclaimed; | |
66e1707b BS |
2978 | } |
2979 | #endif | |
2980 | ||
1d82de61 | 2981 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 2982 | struct scan_control *sc) |
f16015fb | 2983 | { |
b95a2f2d | 2984 | struct mem_cgroup *memcg; |
f16015fb | 2985 | |
b95a2f2d JW |
2986 | if (!total_swap_pages) |
2987 | return; | |
2988 | ||
2989 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2990 | do { | |
ef8f2327 | 2991 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
b95a2f2d | 2992 | |
59dc76b0 | 2993 | if (inactive_list_is_low(lruvec, false)) |
1a93be0e | 2994 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 2995 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
2996 | |
2997 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2998 | } while (memcg); | |
f16015fb JW |
2999 | } |
3000 | ||
31483b6a | 3001 | static bool zone_balanced(struct zone *zone, int order, int classzone_idx) |
60cefed4 | 3002 | { |
31483b6a | 3003 | unsigned long mark = high_wmark_pages(zone); |
60cefed4 | 3004 | |
6256c6b4 MG |
3005 | if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx)) |
3006 | return false; | |
3007 | ||
3008 | /* | |
3009 | * If any eligible zone is balanced then the node is not considered | |
3010 | * to be congested or dirty | |
3011 | */ | |
3012 | clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags); | |
3013 | clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags); | |
3014 | ||
3015 | return true; | |
60cefed4 JW |
3016 | } |
3017 | ||
5515061d MG |
3018 | /* |
3019 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3020 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3021 | * | |
3022 | * Returns true if kswapd is ready to sleep | |
3023 | */ | |
d9f21d42 | 3024 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f50de2d3 | 3025 | { |
1d82de61 MG |
3026 | int i; |
3027 | ||
5515061d | 3028 | /* |
9e5e3661 VB |
3029 | * The throttled processes are normally woken up in balance_pgdat() as |
3030 | * soon as pfmemalloc_watermark_ok() is true. But there is a potential | |
3031 | * race between when kswapd checks the watermarks and a process gets | |
3032 | * throttled. There is also a potential race if processes get | |
3033 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3034 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3035 | * the wake up checks. If kswapd is going to sleep, no process should | |
3036 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3037 | * the wake up is premature, processes will wake kswapd and get | |
3038 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3039 | * that here we are under prepare_to_wait(). | |
5515061d | 3040 | */ |
9e5e3661 VB |
3041 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3042 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3043 | |
1d82de61 MG |
3044 | for (i = 0; i <= classzone_idx; i++) { |
3045 | struct zone *zone = pgdat->node_zones + i; | |
3046 | ||
3047 | if (!populated_zone(zone)) | |
3048 | continue; | |
3049 | ||
38087d9b MG |
3050 | if (!zone_balanced(zone, order, classzone_idx)) |
3051 | return false; | |
1d82de61 MG |
3052 | } |
3053 | ||
38087d9b | 3054 | return true; |
f50de2d3 MG |
3055 | } |
3056 | ||
75485363 | 3057 | /* |
1d82de61 MG |
3058 | * kswapd shrinks a node of pages that are at or below the highest usable |
3059 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3060 | * |
3061 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3062 | * reclaim or if the lack of progress was due to pages under writeback. |
3063 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3064 | */ |
1d82de61 | 3065 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3066 | struct scan_control *sc) |
75485363 | 3067 | { |
1d82de61 MG |
3068 | struct zone *zone; |
3069 | int z; | |
75485363 | 3070 | |
1d82de61 MG |
3071 | /* Reclaim a number of pages proportional to the number of zones */ |
3072 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3073 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 MG |
3074 | zone = pgdat->node_zones + z; |
3075 | if (!populated_zone(zone)) | |
3076 | continue; | |
7c954f6d | 3077 | |
1d82de61 MG |
3078 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3079 | } | |
7c954f6d MG |
3080 | |
3081 | /* | |
1d82de61 MG |
3082 | * Historically care was taken to put equal pressure on all zones but |
3083 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3084 | */ |
970a39a3 | 3085 | shrink_node(pgdat, sc); |
283aba9f | 3086 | |
7c954f6d | 3087 | /* |
1d82de61 MG |
3088 | * Fragmentation may mean that the system cannot be rebalanced for |
3089 | * high-order allocations. If twice the allocation size has been | |
3090 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3091 | * excessive reclaim. Assume that a process requested a high-order | |
3092 | * can direct reclaim/compact. | |
7c954f6d | 3093 | */ |
1d82de61 MG |
3094 | if (sc->order && sc->nr_reclaimed >= 2UL << sc->order) |
3095 | sc->order = 0; | |
7c954f6d | 3096 | |
b8e83b94 | 3097 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3098 | } |
3099 | ||
1da177e4 | 3100 | /* |
1d82de61 MG |
3101 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3102 | * that are eligible for use by the caller until at least one zone is | |
3103 | * balanced. | |
1da177e4 | 3104 | * |
1d82de61 | 3105 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3106 | * |
3107 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3108 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
1d82de61 MG |
3109 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone |
3110 | * or lower is eligible for reclaim until at least one usable zone is | |
3111 | * balanced. | |
1da177e4 | 3112 | */ |
accf6242 | 3113 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3114 | { |
1da177e4 | 3115 | int i; |
0608f43d AM |
3116 | unsigned long nr_soft_reclaimed; |
3117 | unsigned long nr_soft_scanned; | |
1d82de61 | 3118 | struct zone *zone; |
179e9639 AM |
3119 | struct scan_control sc = { |
3120 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3121 | .order = order, |
b8e83b94 | 3122 | .priority = DEF_PRIORITY, |
ee814fe2 | 3123 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3124 | .may_unmap = 1, |
2e2e4259 | 3125 | .may_swap = 1, |
179e9639 | 3126 | }; |
f8891e5e | 3127 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3128 | |
9e3b2f8c | 3129 | do { |
b8e83b94 MG |
3130 | bool raise_priority = true; |
3131 | ||
3132 | sc.nr_reclaimed = 0; | |
84c7a777 | 3133 | sc.reclaim_idx = classzone_idx; |
1da177e4 | 3134 | |
86c79f6b | 3135 | /* |
84c7a777 MG |
3136 | * If the number of buffer_heads exceeds the maximum allowed |
3137 | * then consider reclaiming from all zones. This has a dual | |
3138 | * purpose -- on 64-bit systems it is expected that | |
3139 | * buffer_heads are stripped during active rotation. On 32-bit | |
3140 | * systems, highmem pages can pin lowmem memory and shrinking | |
3141 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3142 | * go ahead if all eligible zones for the original allocation | |
3143 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3144 | */ |
3145 | if (buffer_heads_over_limit) { | |
3146 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3147 | zone = pgdat->node_zones + i; | |
3148 | if (!populated_zone(zone)) | |
3149 | continue; | |
cc715d99 | 3150 | |
970a39a3 | 3151 | sc.reclaim_idx = i; |
e1dbeda6 | 3152 | break; |
1da177e4 | 3153 | } |
1da177e4 | 3154 | } |
dafcb73e | 3155 | |
86c79f6b MG |
3156 | /* |
3157 | * Only reclaim if there are no eligible zones. Check from | |
3158 | * high to low zone as allocations prefer higher zones. | |
3159 | * Scanning from low to high zone would allow congestion to be | |
3160 | * cleared during a very small window when a small low | |
3161 | * zone was balanced even under extreme pressure when the | |
84c7a777 MG |
3162 | * overall node may be congested. Note that sc.reclaim_idx |
3163 | * is not used as buffer_heads_over_limit may have adjusted | |
3164 | * it. | |
86c79f6b | 3165 | */ |
84c7a777 | 3166 | for (i = classzone_idx; i >= 0; i--) { |
86c79f6b MG |
3167 | zone = pgdat->node_zones + i; |
3168 | if (!populated_zone(zone)) | |
3169 | continue; | |
3170 | ||
84c7a777 | 3171 | if (zone_balanced(zone, sc.order, classzone_idx)) |
86c79f6b MG |
3172 | goto out; |
3173 | } | |
e1dbeda6 | 3174 | |
1d82de61 MG |
3175 | /* |
3176 | * Do some background aging of the anon list, to give | |
3177 | * pages a chance to be referenced before reclaiming. All | |
3178 | * pages are rotated regardless of classzone as this is | |
3179 | * about consistent aging. | |
3180 | */ | |
ef8f2327 | 3181 | age_active_anon(pgdat, &sc); |
1d82de61 | 3182 | |
b7ea3c41 MG |
3183 | /* |
3184 | * If we're getting trouble reclaiming, start doing writepage | |
3185 | * even in laptop mode. | |
3186 | */ | |
1d82de61 | 3187 | if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat)) |
b7ea3c41 MG |
3188 | sc.may_writepage = 1; |
3189 | ||
1d82de61 MG |
3190 | /* Call soft limit reclaim before calling shrink_node. */ |
3191 | sc.nr_scanned = 0; | |
3192 | nr_soft_scanned = 0; | |
ef8f2327 | 3193 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3194 | sc.gfp_mask, &nr_soft_scanned); |
3195 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3196 | ||
1da177e4 | 3197 | /* |
1d82de61 MG |
3198 | * There should be no need to raise the scanning priority if |
3199 | * enough pages are already being scanned that that high | |
3200 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3201 | */ |
970a39a3 | 3202 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3203 | raise_priority = false; |
5515061d MG |
3204 | |
3205 | /* | |
3206 | * If the low watermark is met there is no need for processes | |
3207 | * to be throttled on pfmemalloc_wait as they should not be | |
3208 | * able to safely make forward progress. Wake them | |
3209 | */ | |
3210 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
3211 | pfmemalloc_watermark_ok(pgdat)) | |
cfc51155 | 3212 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3213 | |
b8e83b94 MG |
3214 | /* Check if kswapd should be suspending */ |
3215 | if (try_to_freeze() || kthread_should_stop()) | |
3216 | break; | |
8357376d | 3217 | |
73ce02e9 | 3218 | /* |
b8e83b94 MG |
3219 | * Raise priority if scanning rate is too low or there was no |
3220 | * progress in reclaiming pages | |
73ce02e9 | 3221 | */ |
b8e83b94 MG |
3222 | if (raise_priority || !sc.nr_reclaimed) |
3223 | sc.priority--; | |
1d82de61 | 3224 | } while (sc.priority >= 1); |
1da177e4 | 3225 | |
b8e83b94 | 3226 | out: |
0abdee2b | 3227 | /* |
1d82de61 MG |
3228 | * Return the order kswapd stopped reclaiming at as |
3229 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3230 | * entered the allocator slow path while kswapd was awake, order will | |
3231 | * remain at the higher level. | |
0abdee2b | 3232 | */ |
1d82de61 | 3233 | return sc.order; |
1da177e4 LT |
3234 | } |
3235 | ||
38087d9b MG |
3236 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3237 | unsigned int classzone_idx) | |
f0bc0a60 KM |
3238 | { |
3239 | long remaining = 0; | |
3240 | DEFINE_WAIT(wait); | |
3241 | ||
3242 | if (freezing(current) || kthread_should_stop()) | |
3243 | return; | |
3244 | ||
3245 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3246 | ||
3247 | /* Try to sleep for a short interval */ | |
d9f21d42 | 3248 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
fd901c95 VB |
3249 | /* |
3250 | * Compaction records what page blocks it recently failed to | |
3251 | * isolate pages from and skips them in the future scanning. | |
3252 | * When kswapd is going to sleep, it is reasonable to assume | |
3253 | * that pages and compaction may succeed so reset the cache. | |
3254 | */ | |
3255 | reset_isolation_suitable(pgdat); | |
3256 | ||
3257 | /* | |
3258 | * We have freed the memory, now we should compact it to make | |
3259 | * allocation of the requested order possible. | |
3260 | */ | |
38087d9b | 3261 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
fd901c95 | 3262 | |
f0bc0a60 | 3263 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3264 | |
3265 | /* | |
3266 | * If woken prematurely then reset kswapd_classzone_idx and | |
3267 | * order. The values will either be from a wakeup request or | |
3268 | * the previous request that slept prematurely. | |
3269 | */ | |
3270 | if (remaining) { | |
3271 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx); | |
3272 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); | |
3273 | } | |
3274 | ||
f0bc0a60 KM |
3275 | finish_wait(&pgdat->kswapd_wait, &wait); |
3276 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3277 | } | |
3278 | ||
3279 | /* | |
3280 | * After a short sleep, check if it was a premature sleep. If not, then | |
3281 | * go fully to sleep until explicitly woken up. | |
3282 | */ | |
d9f21d42 MG |
3283 | if (!remaining && |
3284 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
f0bc0a60 KM |
3285 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3286 | ||
3287 | /* | |
3288 | * vmstat counters are not perfectly accurate and the estimated | |
3289 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3290 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3291 | * watermarks being breached while under pressure, we reduce the | |
3292 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3293 | * them before going back to sleep. | |
3294 | */ | |
3295 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3296 | |
3297 | if (!kthread_should_stop()) | |
3298 | schedule(); | |
3299 | ||
f0bc0a60 KM |
3300 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3301 | } else { | |
3302 | if (remaining) | |
3303 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3304 | else | |
3305 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3306 | } | |
3307 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3308 | } | |
3309 | ||
1da177e4 LT |
3310 | /* |
3311 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3312 | * from the init process. |
1da177e4 LT |
3313 | * |
3314 | * This basically trickles out pages so that we have _some_ | |
3315 | * free memory available even if there is no other activity | |
3316 | * that frees anything up. This is needed for things like routing | |
3317 | * etc, where we otherwise might have all activity going on in | |
3318 | * asynchronous contexts that cannot page things out. | |
3319 | * | |
3320 | * If there are applications that are active memory-allocators | |
3321 | * (most normal use), this basically shouldn't matter. | |
3322 | */ | |
3323 | static int kswapd(void *p) | |
3324 | { | |
38087d9b | 3325 | unsigned int alloc_order, reclaim_order, classzone_idx; |
1da177e4 LT |
3326 | pg_data_t *pgdat = (pg_data_t*)p; |
3327 | struct task_struct *tsk = current; | |
f0bc0a60 | 3328 | |
1da177e4 LT |
3329 | struct reclaim_state reclaim_state = { |
3330 | .reclaimed_slab = 0, | |
3331 | }; | |
a70f7302 | 3332 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3333 | |
cf40bd16 NP |
3334 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3335 | ||
174596a0 | 3336 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3337 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3338 | current->reclaim_state = &reclaim_state; |
3339 | ||
3340 | /* | |
3341 | * Tell the memory management that we're a "memory allocator", | |
3342 | * and that if we need more memory we should get access to it | |
3343 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3344 | * never get caught in the normal page freeing logic. | |
3345 | * | |
3346 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3347 | * you need a small amount of memory in order to be able to | |
3348 | * page out something else, and this flag essentially protects | |
3349 | * us from recursively trying to free more memory as we're | |
3350 | * trying to free the first piece of memory in the first place). | |
3351 | */ | |
930d9152 | 3352 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3353 | set_freezable(); |
1da177e4 | 3354 | |
38087d9b MG |
3355 | pgdat->kswapd_order = alloc_order = reclaim_order = 0; |
3356 | pgdat->kswapd_classzone_idx = classzone_idx = 0; | |
1da177e4 | 3357 | for ( ; ; ) { |
6f6313d4 | 3358 | bool ret; |
3e1d1d28 | 3359 | |
38087d9b MG |
3360 | kswapd_try_sleep: |
3361 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3362 | classzone_idx); | |
215ddd66 | 3363 | |
38087d9b MG |
3364 | /* Read the new order and classzone_idx */ |
3365 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
3366 | classzone_idx = pgdat->kswapd_classzone_idx; | |
3367 | pgdat->kswapd_order = 0; | |
3368 | pgdat->kswapd_classzone_idx = 0; | |
1da177e4 | 3369 | |
8fe23e05 DR |
3370 | ret = try_to_freeze(); |
3371 | if (kthread_should_stop()) | |
3372 | break; | |
3373 | ||
3374 | /* | |
3375 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3376 | * after returning from the refrigerator | |
3377 | */ | |
38087d9b MG |
3378 | if (ret) |
3379 | continue; | |
3380 | ||
3381 | /* | |
3382 | * Reclaim begins at the requested order but if a high-order | |
3383 | * reclaim fails then kswapd falls back to reclaiming for | |
3384 | * order-0. If that happens, kswapd will consider sleeping | |
3385 | * for the order it finished reclaiming at (reclaim_order) | |
3386 | * but kcompactd is woken to compact for the original | |
3387 | * request (alloc_order). | |
3388 | */ | |
3389 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, alloc_order); | |
3390 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); | |
3391 | if (reclaim_order < alloc_order) | |
3392 | goto kswapd_try_sleep; | |
1d82de61 | 3393 | |
38087d9b MG |
3394 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3395 | classzone_idx = pgdat->kswapd_classzone_idx; | |
1da177e4 | 3396 | } |
b0a8cc58 | 3397 | |
71abdc15 | 3398 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
b0a8cc58 | 3399 | current->reclaim_state = NULL; |
71abdc15 JW |
3400 | lockdep_clear_current_reclaim_state(); |
3401 | ||
1da177e4 LT |
3402 | return 0; |
3403 | } | |
3404 | ||
3405 | /* | |
3406 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3407 | */ | |
99504748 | 3408 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3409 | { |
3410 | pg_data_t *pgdat; | |
e1a55637 | 3411 | int z; |
1da177e4 | 3412 | |
f3fe6512 | 3413 | if (!populated_zone(zone)) |
1da177e4 LT |
3414 | return; |
3415 | ||
344736f2 | 3416 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) |
1da177e4 | 3417 | return; |
88f5acf8 | 3418 | pgdat = zone->zone_pgdat; |
38087d9b MG |
3419 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx); |
3420 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); | |
8d0986e2 | 3421 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3422 | return; |
e1a55637 MG |
3423 | |
3424 | /* Only wake kswapd if all zones are unbalanced */ | |
3425 | for (z = 0; z <= classzone_idx; z++) { | |
3426 | zone = pgdat->node_zones + z; | |
3427 | if (!populated_zone(zone)) | |
3428 | continue; | |
3429 | ||
3430 | if (zone_balanced(zone, order, classzone_idx)) | |
3431 | return; | |
3432 | } | |
88f5acf8 MG |
3433 | |
3434 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3435 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3436 | } |
3437 | ||
c6f37f12 | 3438 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3439 | /* |
7b51755c | 3440 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3441 | * freed pages. |
3442 | * | |
3443 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3444 | * LRU order by reclaiming preferentially | |
3445 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3446 | */ |
7b51755c | 3447 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3448 | { |
d6277db4 | 3449 | struct reclaim_state reclaim_state; |
d6277db4 | 3450 | struct scan_control sc = { |
ee814fe2 | 3451 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3452 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 3453 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 3454 | .priority = DEF_PRIORITY, |
d6277db4 | 3455 | .may_writepage = 1, |
ee814fe2 JW |
3456 | .may_unmap = 1, |
3457 | .may_swap = 1, | |
7b51755c | 3458 | .hibernation_mode = 1, |
1da177e4 | 3459 | }; |
a09ed5e0 | 3460 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c KM |
3461 | struct task_struct *p = current; |
3462 | unsigned long nr_reclaimed; | |
1da177e4 | 3463 | |
7b51755c KM |
3464 | p->flags |= PF_MEMALLOC; |
3465 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3466 | reclaim_state.reclaimed_slab = 0; | |
3467 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3468 | |
3115cd91 | 3469 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3470 | |
7b51755c KM |
3471 | p->reclaim_state = NULL; |
3472 | lockdep_clear_current_reclaim_state(); | |
3473 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3474 | |
7b51755c | 3475 | return nr_reclaimed; |
1da177e4 | 3476 | } |
c6f37f12 | 3477 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3478 | |
1da177e4 LT |
3479 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3480 | not required for correctness. So if the last cpu in a node goes | |
3481 | away, we get changed to run anywhere: as the first one comes back, | |
3482 | restore their cpu bindings. */ | |
fcb35a9b GKH |
3483 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
3484 | void *hcpu) | |
1da177e4 | 3485 | { |
58c0a4a7 | 3486 | int nid; |
1da177e4 | 3487 | |
8bb78442 | 3488 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
48fb2e24 | 3489 | for_each_node_state(nid, N_MEMORY) { |
c5f59f08 | 3490 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3491 | const struct cpumask *mask; |
3492 | ||
3493 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3494 | |
3e597945 | 3495 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3496 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3497 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3498 | } |
3499 | } | |
3500 | return NOTIFY_OK; | |
3501 | } | |
1da177e4 | 3502 | |
3218ae14 YG |
3503 | /* |
3504 | * This kswapd start function will be called by init and node-hot-add. | |
3505 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3506 | */ | |
3507 | int kswapd_run(int nid) | |
3508 | { | |
3509 | pg_data_t *pgdat = NODE_DATA(nid); | |
3510 | int ret = 0; | |
3511 | ||
3512 | if (pgdat->kswapd) | |
3513 | return 0; | |
3514 | ||
3515 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3516 | if (IS_ERR(pgdat->kswapd)) { | |
3517 | /* failure at boot is fatal */ | |
3518 | BUG_ON(system_state == SYSTEM_BOOTING); | |
d5dc0ad9 GS |
3519 | pr_err("Failed to start kswapd on node %d\n", nid); |
3520 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3521 | pgdat->kswapd = NULL; |
3218ae14 YG |
3522 | } |
3523 | return ret; | |
3524 | } | |
3525 | ||
8fe23e05 | 3526 | /* |
d8adde17 | 3527 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 3528 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
3529 | */ |
3530 | void kswapd_stop(int nid) | |
3531 | { | |
3532 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3533 | ||
d8adde17 | 3534 | if (kswapd) { |
8fe23e05 | 3535 | kthread_stop(kswapd); |
d8adde17 JL |
3536 | NODE_DATA(nid)->kswapd = NULL; |
3537 | } | |
8fe23e05 DR |
3538 | } |
3539 | ||
1da177e4 LT |
3540 | static int __init kswapd_init(void) |
3541 | { | |
3218ae14 | 3542 | int nid; |
69e05944 | 3543 | |
1da177e4 | 3544 | swap_setup(); |
48fb2e24 | 3545 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3546 | kswapd_run(nid); |
1da177e4 LT |
3547 | hotcpu_notifier(cpu_callback, 0); |
3548 | return 0; | |
3549 | } | |
3550 | ||
3551 | module_init(kswapd_init) | |
9eeff239 CL |
3552 | |
3553 | #ifdef CONFIG_NUMA | |
3554 | /* | |
a5f5f91d | 3555 | * Node reclaim mode |
9eeff239 | 3556 | * |
a5f5f91d | 3557 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 3558 | * the watermarks. |
9eeff239 | 3559 | */ |
a5f5f91d | 3560 | int node_reclaim_mode __read_mostly; |
9eeff239 | 3561 | |
1b2ffb78 | 3562 | #define RECLAIM_OFF 0 |
7d03431c | 3563 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 3564 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 3565 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 3566 | |
a92f7126 | 3567 | /* |
a5f5f91d | 3568 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
3569 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
3570 | * a zone. | |
3571 | */ | |
a5f5f91d | 3572 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 3573 | |
9614634f | 3574 | /* |
a5f5f91d | 3575 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
3576 | * occur. |
3577 | */ | |
3578 | int sysctl_min_unmapped_ratio = 1; | |
3579 | ||
0ff38490 CL |
3580 | /* |
3581 | * If the number of slab pages in a zone grows beyond this percentage then | |
3582 | * slab reclaim needs to occur. | |
3583 | */ | |
3584 | int sysctl_min_slab_ratio = 5; | |
3585 | ||
11fb9989 | 3586 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 3587 | { |
11fb9989 MG |
3588 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
3589 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
3590 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
3591 | |
3592 | /* | |
3593 | * It's possible for there to be more file mapped pages than | |
3594 | * accounted for by the pages on the file LRU lists because | |
3595 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3596 | */ | |
3597 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3598 | } | |
3599 | ||
3600 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 3601 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 3602 | { |
d031a157 AM |
3603 | unsigned long nr_pagecache_reclaimable; |
3604 | unsigned long delta = 0; | |
90afa5de MG |
3605 | |
3606 | /* | |
95bbc0c7 | 3607 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 3608 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 3609 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
3610 | * a better estimate |
3611 | */ | |
a5f5f91d MG |
3612 | if (node_reclaim_mode & RECLAIM_UNMAP) |
3613 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 3614 | else |
a5f5f91d | 3615 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
3616 | |
3617 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
3618 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
3619 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
3620 | |
3621 | /* Watch for any possible underflows due to delta */ | |
3622 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3623 | delta = nr_pagecache_reclaimable; | |
3624 | ||
3625 | return nr_pagecache_reclaimable - delta; | |
3626 | } | |
3627 | ||
9eeff239 | 3628 | /* |
a5f5f91d | 3629 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 3630 | */ |
a5f5f91d | 3631 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3632 | { |
7fb2d46d | 3633 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3634 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3635 | struct task_struct *p = current; |
3636 | struct reclaim_state reclaim_state; | |
a5f5f91d | 3637 | int classzone_idx = gfp_zone(gfp_mask); |
179e9639 | 3638 | struct scan_control sc = { |
62b726c1 | 3639 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
21caf2fc | 3640 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
bd2f6199 | 3641 | .order = order, |
a5f5f91d MG |
3642 | .priority = NODE_RECLAIM_PRIORITY, |
3643 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
3644 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 3645 | .may_swap = 1, |
a5f5f91d | 3646 | .reclaim_idx = classzone_idx, |
179e9639 | 3647 | }; |
9eeff239 | 3648 | |
9eeff239 | 3649 | cond_resched(); |
d4f7796e | 3650 | /* |
95bbc0c7 | 3651 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 3652 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 3653 | * and RECLAIM_UNMAP. |
d4f7796e CL |
3654 | */ |
3655 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3656 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3657 | reclaim_state.reclaimed_slab = 0; |
3658 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3659 | |
a5f5f91d | 3660 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 CL |
3661 | /* |
3662 | * Free memory by calling shrink zone with increasing | |
3663 | * priorities until we have enough memory freed. | |
3664 | */ | |
0ff38490 | 3665 | do { |
970a39a3 | 3666 | shrink_node(pgdat, &sc); |
9e3b2f8c | 3667 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 3668 | } |
c84db23c | 3669 | |
9eeff239 | 3670 | p->reclaim_state = NULL; |
d4f7796e | 3671 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3672 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3673 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3674 | } |
179e9639 | 3675 | |
a5f5f91d | 3676 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 3677 | { |
d773ed6b | 3678 | int ret; |
179e9639 AM |
3679 | |
3680 | /* | |
a5f5f91d | 3681 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 3682 | * slab pages if we are over the defined limits. |
34aa1330 | 3683 | * |
9614634f CL |
3684 | * A small portion of unmapped file backed pages is needed for |
3685 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
3686 | * thrown out if the node is overallocated. So we do not reclaim |
3687 | * if less than a specified percentage of the node is used by | |
9614634f | 3688 | * unmapped file backed pages. |
179e9639 | 3689 | */ |
a5f5f91d MG |
3690 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
3691 | sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) | |
3692 | return NODE_RECLAIM_FULL; | |
179e9639 | 3693 | |
a5f5f91d MG |
3694 | if (!pgdat_reclaimable(pgdat)) |
3695 | return NODE_RECLAIM_FULL; | |
d773ed6b | 3696 | |
179e9639 | 3697 | /* |
d773ed6b | 3698 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3699 | */ |
d0164adc | 3700 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 3701 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
3702 | |
3703 | /* | |
a5f5f91d | 3704 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
3705 | * have associated processors. This will favor the local processor |
3706 | * over remote processors and spread off node memory allocations | |
3707 | * as wide as possible. | |
3708 | */ | |
a5f5f91d MG |
3709 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
3710 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 3711 | |
a5f5f91d MG |
3712 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
3713 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 3714 | |
a5f5f91d MG |
3715 | ret = __node_reclaim(pgdat, gfp_mask, order); |
3716 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 3717 | |
24cf7251 MG |
3718 | if (!ret) |
3719 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3720 | ||
d773ed6b | 3721 | return ret; |
179e9639 | 3722 | } |
9eeff239 | 3723 | #endif |
894bc310 | 3724 | |
894bc310 LS |
3725 | /* |
3726 | * page_evictable - test whether a page is evictable | |
3727 | * @page: the page to test | |
894bc310 LS |
3728 | * |
3729 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3730 | * lists vs unevictable list. |
894bc310 LS |
3731 | * |
3732 | * Reasons page might not be evictable: | |
ba9ddf49 | 3733 | * (1) page's mapping marked unevictable |
b291f000 | 3734 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3735 | * |
894bc310 | 3736 | */ |
39b5f29a | 3737 | int page_evictable(struct page *page) |
894bc310 | 3738 | { |
39b5f29a | 3739 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
894bc310 | 3740 | } |
89e004ea | 3741 | |
85046579 | 3742 | #ifdef CONFIG_SHMEM |
89e004ea | 3743 | /** |
24513264 HD |
3744 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3745 | * @pages: array of pages to check | |
3746 | * @nr_pages: number of pages to check | |
89e004ea | 3747 | * |
24513264 | 3748 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3749 | * |
3750 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3751 | */ |
24513264 | 3752 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3753 | { |
925b7673 | 3754 | struct lruvec *lruvec; |
24513264 HD |
3755 | struct zone *zone = NULL; |
3756 | int pgscanned = 0; | |
3757 | int pgrescued = 0; | |
3758 | int i; | |
89e004ea | 3759 | |
24513264 HD |
3760 | for (i = 0; i < nr_pages; i++) { |
3761 | struct page *page = pages[i]; | |
3762 | struct zone *pagezone; | |
89e004ea | 3763 | |
24513264 HD |
3764 | pgscanned++; |
3765 | pagezone = page_zone(page); | |
3766 | if (pagezone != zone) { | |
3767 | if (zone) | |
a52633d8 | 3768 | spin_unlock_irq(zone_lru_lock(zone)); |
24513264 | 3769 | zone = pagezone; |
a52633d8 | 3770 | spin_lock_irq(zone_lru_lock(zone)); |
24513264 | 3771 | } |
599d0c95 | 3772 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
89e004ea | 3773 | |
24513264 HD |
3774 | if (!PageLRU(page) || !PageUnevictable(page)) |
3775 | continue; | |
89e004ea | 3776 | |
39b5f29a | 3777 | if (page_evictable(page)) { |
24513264 HD |
3778 | enum lru_list lru = page_lru_base_type(page); |
3779 | ||
309381fe | 3780 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 3781 | ClearPageUnevictable(page); |
fa9add64 HD |
3782 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
3783 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 3784 | pgrescued++; |
89e004ea | 3785 | } |
24513264 | 3786 | } |
89e004ea | 3787 | |
24513264 HD |
3788 | if (zone) { |
3789 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3790 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
a52633d8 | 3791 | spin_unlock_irq(zone_lru_lock(zone)); |
89e004ea | 3792 | } |
89e004ea | 3793 | } |
85046579 | 3794 | #endif /* CONFIG_SHMEM */ |