]>
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> |
1da177e4 LT |
49 | |
50 | #include <asm/tlbflush.h> | |
51 | #include <asm/div64.h> | |
52 | ||
53 | #include <linux/swapops.h> | |
117aad1e | 54 | #include <linux/balloon_compaction.h> |
1da177e4 | 55 | |
0f8053a5 NP |
56 | #include "internal.h" |
57 | ||
33906bc5 MG |
58 | #define CREATE_TRACE_POINTS |
59 | #include <trace/events/vmscan.h> | |
60 | ||
1da177e4 | 61 | struct scan_control { |
22fba335 KM |
62 | /* How many pages shrink_list() should reclaim */ |
63 | unsigned long nr_to_reclaim; | |
64 | ||
1da177e4 | 65 | /* This context's GFP mask */ |
6daa0e28 | 66 | gfp_t gfp_mask; |
1da177e4 | 67 | |
ee814fe2 | 68 | /* Allocation order */ |
5ad333eb | 69 | int order; |
66e1707b | 70 | |
ee814fe2 JW |
71 | /* |
72 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
73 | * are scanned. | |
74 | */ | |
75 | nodemask_t *nodemask; | |
9e3b2f8c | 76 | |
f16015fb JW |
77 | /* |
78 | * The memory cgroup that hit its limit and as a result is the | |
79 | * primary target of this reclaim invocation. | |
80 | */ | |
81 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 82 | |
ee814fe2 JW |
83 | /* Scan (total_size >> priority) pages at once */ |
84 | int priority; | |
85 | ||
86 | unsigned int may_writepage:1; | |
87 | ||
88 | /* Can mapped pages be reclaimed? */ | |
89 | unsigned int may_unmap:1; | |
90 | ||
91 | /* Can pages be swapped as part of reclaim? */ | |
92 | unsigned int may_swap:1; | |
93 | ||
241994ed JW |
94 | /* Can cgroups be reclaimed below their normal consumption range? */ |
95 | unsigned int may_thrash:1; | |
96 | ||
ee814fe2 JW |
97 | unsigned int hibernation_mode:1; |
98 | ||
99 | /* One of the zones is ready for compaction */ | |
100 | unsigned int compaction_ready:1; | |
101 | ||
102 | /* Incremented by the number of inactive pages that were scanned */ | |
103 | unsigned long nr_scanned; | |
104 | ||
105 | /* Number of pages freed so far during a call to shrink_zones() */ | |
106 | unsigned long nr_reclaimed; | |
1da177e4 LT |
107 | }; |
108 | ||
1da177e4 LT |
109 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
110 | ||
111 | #ifdef ARCH_HAS_PREFETCH | |
112 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
113 | do { \ | |
114 | if ((_page)->lru.prev != _base) { \ | |
115 | struct page *prev; \ | |
116 | \ | |
117 | prev = lru_to_page(&(_page->lru)); \ | |
118 | prefetch(&prev->_field); \ | |
119 | } \ | |
120 | } while (0) | |
121 | #else | |
122 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
123 | #endif | |
124 | ||
125 | #ifdef ARCH_HAS_PREFETCHW | |
126 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
127 | do { \ | |
128 | if ((_page)->lru.prev != _base) { \ | |
129 | struct page *prev; \ | |
130 | \ | |
131 | prev = lru_to_page(&(_page->lru)); \ | |
132 | prefetchw(&prev->_field); \ | |
133 | } \ | |
134 | } while (0) | |
135 | #else | |
136 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
137 | #endif | |
138 | ||
139 | /* | |
140 | * From 0 .. 100. Higher means more swappy. | |
141 | */ | |
142 | int vm_swappiness = 60; | |
d0480be4 WSH |
143 | /* |
144 | * The total number of pages which are beyond the high watermark within all | |
145 | * zones. | |
146 | */ | |
147 | unsigned long vm_total_pages; | |
1da177e4 LT |
148 | |
149 | static LIST_HEAD(shrinker_list); | |
150 | static DECLARE_RWSEM(shrinker_rwsem); | |
151 | ||
c255a458 | 152 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
153 | static bool global_reclaim(struct scan_control *sc) |
154 | { | |
f16015fb | 155 | return !sc->target_mem_cgroup; |
89b5fae5 | 156 | } |
97c9341f TH |
157 | |
158 | /** | |
159 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
160 | * @sc: scan_control in question | |
161 | * | |
162 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
163 | * completely broken with the legacy memcg and direct stalling in | |
164 | * shrink_page_list() is used for throttling instead, which lacks all the | |
165 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
166 | * allocation and configurability. | |
167 | * | |
168 | * This function tests whether the vmscan currently in progress can assume | |
169 | * that the normal dirty throttling mechanism is operational. | |
170 | */ | |
171 | static bool sane_reclaim(struct scan_control *sc) | |
172 | { | |
173 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
174 | ||
175 | if (!memcg) | |
176 | return true; | |
177 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 178 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
179 | return true; |
180 | #endif | |
181 | return false; | |
182 | } | |
91a45470 | 183 | #else |
89b5fae5 JW |
184 | static bool global_reclaim(struct scan_control *sc) |
185 | { | |
186 | return true; | |
187 | } | |
97c9341f TH |
188 | |
189 | static bool sane_reclaim(struct scan_control *sc) | |
190 | { | |
191 | return true; | |
192 | } | |
91a45470 KH |
193 | #endif |
194 | ||
c630ec12 | 195 | unsigned long zone_reclaimable_pages(struct zone *zone) |
6e543d57 | 196 | { |
d031a157 | 197 | unsigned long nr; |
6e543d57 LD |
198 | |
199 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
200 | zone_page_state(zone, NR_INACTIVE_FILE); | |
201 | ||
202 | if (get_nr_swap_pages() > 0) | |
203 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
204 | zone_page_state(zone, NR_INACTIVE_ANON); | |
205 | ||
206 | return nr; | |
207 | } | |
208 | ||
209 | bool zone_reclaimable(struct zone *zone) | |
210 | { | |
0d5d823a MG |
211 | return zone_page_state(zone, NR_PAGES_SCANNED) < |
212 | zone_reclaimable_pages(zone) * 6; | |
6e543d57 LD |
213 | } |
214 | ||
4d7dcca2 | 215 | static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) |
c9f299d9 | 216 | { |
c3c787e8 | 217 | if (!mem_cgroup_disabled()) |
4d7dcca2 | 218 | return mem_cgroup_get_lru_size(lruvec, lru); |
a3d8e054 | 219 | |
074291fe | 220 | return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); |
c9f299d9 KM |
221 | } |
222 | ||
1da177e4 | 223 | /* |
1d3d4437 | 224 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 225 | */ |
1d3d4437 | 226 | int register_shrinker(struct shrinker *shrinker) |
1da177e4 | 227 | { |
1d3d4437 GC |
228 | size_t size = sizeof(*shrinker->nr_deferred); |
229 | ||
230 | /* | |
231 | * If we only have one possible node in the system anyway, save | |
232 | * ourselves the trouble and disable NUMA aware behavior. This way we | |
233 | * will save memory and some small loop time later. | |
234 | */ | |
235 | if (nr_node_ids == 1) | |
236 | shrinker->flags &= ~SHRINKER_NUMA_AWARE; | |
237 | ||
238 | if (shrinker->flags & SHRINKER_NUMA_AWARE) | |
239 | size *= nr_node_ids; | |
240 | ||
241 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
242 | if (!shrinker->nr_deferred) | |
243 | return -ENOMEM; | |
244 | ||
8e1f936b RR |
245 | down_write(&shrinker_rwsem); |
246 | list_add_tail(&shrinker->list, &shrinker_list); | |
247 | up_write(&shrinker_rwsem); | |
1d3d4437 | 248 | return 0; |
1da177e4 | 249 | } |
8e1f936b | 250 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
251 | |
252 | /* | |
253 | * Remove one | |
254 | */ | |
8e1f936b | 255 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
256 | { |
257 | down_write(&shrinker_rwsem); | |
258 | list_del(&shrinker->list); | |
259 | up_write(&shrinker_rwsem); | |
ae393321 | 260 | kfree(shrinker->nr_deferred); |
1da177e4 | 261 | } |
8e1f936b | 262 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
263 | |
264 | #define SHRINK_BATCH 128 | |
1d3d4437 | 265 | |
cb731d6c VD |
266 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
267 | struct shrinker *shrinker, | |
268 | unsigned long nr_scanned, | |
269 | unsigned long nr_eligible) | |
1d3d4437 GC |
270 | { |
271 | unsigned long freed = 0; | |
272 | unsigned long long delta; | |
273 | long total_scan; | |
d5bc5fd3 | 274 | long freeable; |
1d3d4437 GC |
275 | long nr; |
276 | long new_nr; | |
277 | int nid = shrinkctl->nid; | |
278 | long batch_size = shrinker->batch ? shrinker->batch | |
279 | : SHRINK_BATCH; | |
de51d37d | 280 | long scanned = 0, next_deferred; |
1d3d4437 | 281 | |
d5bc5fd3 VD |
282 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
283 | if (freeable == 0) | |
1d3d4437 GC |
284 | return 0; |
285 | ||
286 | /* | |
287 | * copy the current shrinker scan count into a local variable | |
288 | * and zero it so that other concurrent shrinker invocations | |
289 | * don't also do this scanning work. | |
290 | */ | |
291 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
292 | ||
293 | total_scan = nr; | |
6b4f7799 | 294 | delta = (4 * nr_scanned) / shrinker->seeks; |
d5bc5fd3 | 295 | delta *= freeable; |
6b4f7799 | 296 | do_div(delta, nr_eligible + 1); |
1d3d4437 GC |
297 | total_scan += delta; |
298 | if (total_scan < 0) { | |
8612c663 | 299 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
a0b02131 | 300 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 301 | total_scan = freeable; |
de51d37d SL |
302 | next_deferred = nr; |
303 | } else | |
304 | next_deferred = total_scan; | |
1d3d4437 GC |
305 | |
306 | /* | |
307 | * We need to avoid excessive windup on filesystem shrinkers | |
308 | * due to large numbers of GFP_NOFS allocations causing the | |
309 | * shrinkers to return -1 all the time. This results in a large | |
310 | * nr being built up so when a shrink that can do some work | |
311 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 312 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
313 | * memory. |
314 | * | |
315 | * Hence only allow the shrinker to scan the entire cache when | |
316 | * a large delta change is calculated directly. | |
317 | */ | |
d5bc5fd3 VD |
318 | if (delta < freeable / 4) |
319 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
320 | |
321 | /* | |
322 | * Avoid risking looping forever due to too large nr value: | |
323 | * never try to free more than twice the estimate number of | |
324 | * freeable entries. | |
325 | */ | |
d5bc5fd3 VD |
326 | if (total_scan > freeable * 2) |
327 | total_scan = freeable * 2; | |
1d3d4437 GC |
328 | |
329 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
6b4f7799 JW |
330 | nr_scanned, nr_eligible, |
331 | freeable, delta, total_scan); | |
1d3d4437 | 332 | |
0b1fb40a VD |
333 | /* |
334 | * Normally, we should not scan less than batch_size objects in one | |
335 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
336 | * than batch_size objects in total and we are really tight on memory, | |
337 | * we will try to reclaim all available objects, otherwise we can end | |
338 | * up failing allocations although there are plenty of reclaimable | |
339 | * objects spread over several slabs with usage less than the | |
340 | * batch_size. | |
341 | * | |
342 | * We detect the "tight on memory" situations by looking at the total | |
343 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 344 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
345 | * scanning at high prio and therefore should try to reclaim as much as |
346 | * possible. | |
347 | */ | |
348 | while (total_scan >= batch_size || | |
d5bc5fd3 | 349 | total_scan >= freeable) { |
a0b02131 | 350 | unsigned long ret; |
0b1fb40a | 351 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 352 | |
0b1fb40a | 353 | shrinkctl->nr_to_scan = nr_to_scan; |
a0b02131 DC |
354 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
355 | if (ret == SHRINK_STOP) | |
356 | break; | |
357 | freed += ret; | |
1d3d4437 | 358 | |
0b1fb40a VD |
359 | count_vm_events(SLABS_SCANNED, nr_to_scan); |
360 | total_scan -= nr_to_scan; | |
de51d37d | 361 | scanned += nr_to_scan; |
1d3d4437 GC |
362 | |
363 | cond_resched(); | |
364 | } | |
365 | ||
de51d37d SL |
366 | if (next_deferred >= scanned) |
367 | next_deferred -= scanned; | |
368 | else | |
369 | next_deferred = 0; | |
1d3d4437 GC |
370 | /* |
371 | * move the unused scan count back into the shrinker in a | |
372 | * manner that handles concurrent updates. If we exhausted the | |
373 | * scan, there is no need to do an update. | |
374 | */ | |
de51d37d SL |
375 | if (next_deferred > 0) |
376 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
377 | &shrinker->nr_deferred[nid]); |
378 | else | |
379 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
380 | ||
df9024a8 | 381 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 382 | return freed; |
1495f230 YH |
383 | } |
384 | ||
6b4f7799 | 385 | /** |
cb731d6c | 386 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
387 | * @gfp_mask: allocation context |
388 | * @nid: node whose slab caches to target | |
cb731d6c | 389 | * @memcg: memory cgroup whose slab caches to target |
6b4f7799 JW |
390 | * @nr_scanned: pressure numerator |
391 | * @nr_eligible: pressure denominator | |
1da177e4 | 392 | * |
6b4f7799 | 393 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 394 | * |
6b4f7799 JW |
395 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
396 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 397 | * |
cb731d6c VD |
398 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
399 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan | |
400 | * objects from the memory cgroup specified. Otherwise all shrinkers | |
401 | * are called, and memcg aware shrinkers are supposed to scan the | |
402 | * global list then. | |
403 | * | |
6b4f7799 JW |
404 | * @nr_scanned and @nr_eligible form a ratio that indicate how much of |
405 | * the available objects should be scanned. Page reclaim for example | |
406 | * passes the number of pages scanned and the number of pages on the | |
407 | * LRU lists that it considered on @nid, plus a bias in @nr_scanned | |
408 | * when it encountered mapped pages. The ratio is further biased by | |
409 | * the ->seeks setting of the shrink function, which indicates the | |
410 | * cost to recreate an object relative to that of an LRU page. | |
b15e0905 | 411 | * |
6b4f7799 | 412 | * Returns the number of reclaimed slab objects. |
1da177e4 | 413 | */ |
cb731d6c VD |
414 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
415 | struct mem_cgroup *memcg, | |
416 | unsigned long nr_scanned, | |
417 | unsigned long nr_eligible) | |
1da177e4 LT |
418 | { |
419 | struct shrinker *shrinker; | |
24f7c6b9 | 420 | unsigned long freed = 0; |
1da177e4 | 421 | |
cb731d6c VD |
422 | if (memcg && !memcg_kmem_is_active(memcg)) |
423 | return 0; | |
424 | ||
6b4f7799 JW |
425 | if (nr_scanned == 0) |
426 | nr_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 427 | |
f06590bd | 428 | if (!down_read_trylock(&shrinker_rwsem)) { |
24f7c6b9 DC |
429 | /* |
430 | * If we would return 0, our callers would understand that we | |
431 | * have nothing else to shrink and give up trying. By returning | |
432 | * 1 we keep it going and assume we'll be able to shrink next | |
433 | * time. | |
434 | */ | |
435 | freed = 1; | |
f06590bd MK |
436 | goto out; |
437 | } | |
1da177e4 LT |
438 | |
439 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
440 | struct shrink_control sc = { |
441 | .gfp_mask = gfp_mask, | |
442 | .nid = nid, | |
cb731d6c | 443 | .memcg = memcg, |
6b4f7799 | 444 | }; |
ec97097b | 445 | |
cb731d6c VD |
446 | if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE)) |
447 | continue; | |
448 | ||
6b4f7799 JW |
449 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
450 | sc.nid = 0; | |
1da177e4 | 451 | |
cb731d6c | 452 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); |
1da177e4 | 453 | } |
6b4f7799 | 454 | |
1da177e4 | 455 | up_read(&shrinker_rwsem); |
f06590bd MK |
456 | out: |
457 | cond_resched(); | |
24f7c6b9 | 458 | return freed; |
1da177e4 LT |
459 | } |
460 | ||
cb731d6c VD |
461 | void drop_slab_node(int nid) |
462 | { | |
463 | unsigned long freed; | |
464 | ||
465 | do { | |
466 | struct mem_cgroup *memcg = NULL; | |
467 | ||
468 | freed = 0; | |
469 | do { | |
470 | freed += shrink_slab(GFP_KERNEL, nid, memcg, | |
471 | 1000, 1000); | |
472 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
473 | } while (freed > 10); | |
474 | } | |
475 | ||
476 | void drop_slab(void) | |
477 | { | |
478 | int nid; | |
479 | ||
480 | for_each_online_node(nid) | |
481 | drop_slab_node(nid); | |
482 | } | |
483 | ||
1da177e4 LT |
484 | static inline int is_page_cache_freeable(struct page *page) |
485 | { | |
ceddc3a5 JW |
486 | /* |
487 | * A freeable page cache page is referenced only by the caller | |
488 | * that isolated the page, the page cache radix tree and | |
489 | * optional buffer heads at page->private. | |
490 | */ | |
edcf4748 | 491 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
492 | } |
493 | ||
703c2708 | 494 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
1da177e4 | 495 | { |
930d9152 | 496 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 497 | return 1; |
703c2708 | 498 | if (!inode_write_congested(inode)) |
1da177e4 | 499 | return 1; |
703c2708 | 500 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
501 | return 1; |
502 | return 0; | |
503 | } | |
504 | ||
505 | /* | |
506 | * We detected a synchronous write error writing a page out. Probably | |
507 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
508 | * fsync(), msync() or close(). | |
509 | * | |
510 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
511 | * prevents it from being freed up. But we have a ref on the page and once | |
512 | * that page is locked, the mapping is pinned. | |
513 | * | |
514 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
515 | * __GFP_FS. | |
516 | */ | |
517 | static void handle_write_error(struct address_space *mapping, | |
518 | struct page *page, int error) | |
519 | { | |
7eaceacc | 520 | lock_page(page); |
3e9f45bd GC |
521 | if (page_mapping(page) == mapping) |
522 | mapping_set_error(mapping, error); | |
1da177e4 LT |
523 | unlock_page(page); |
524 | } | |
525 | ||
04e62a29 CL |
526 | /* possible outcome of pageout() */ |
527 | typedef enum { | |
528 | /* failed to write page out, page is locked */ | |
529 | PAGE_KEEP, | |
530 | /* move page to the active list, page is locked */ | |
531 | PAGE_ACTIVATE, | |
532 | /* page has been sent to the disk successfully, page is unlocked */ | |
533 | PAGE_SUCCESS, | |
534 | /* page is clean and locked */ | |
535 | PAGE_CLEAN, | |
536 | } pageout_t; | |
537 | ||
1da177e4 | 538 | /* |
1742f19f AM |
539 | * pageout is called by shrink_page_list() for each dirty page. |
540 | * Calls ->writepage(). | |
1da177e4 | 541 | */ |
c661b078 | 542 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 543 | struct scan_control *sc) |
1da177e4 LT |
544 | { |
545 | /* | |
546 | * If the page is dirty, only perform writeback if that write | |
547 | * will be non-blocking. To prevent this allocation from being | |
548 | * stalled by pagecache activity. But note that there may be | |
549 | * stalls if we need to run get_block(). We could test | |
550 | * PagePrivate for that. | |
551 | * | |
8174202b | 552 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
553 | * this page's queue, we can perform writeback even if that |
554 | * will block. | |
555 | * | |
556 | * If the page is swapcache, write it back even if that would | |
557 | * block, for some throttling. This happens by accident, because | |
558 | * swap_backing_dev_info is bust: it doesn't reflect the | |
559 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
560 | */ |
561 | if (!is_page_cache_freeable(page)) | |
562 | return PAGE_KEEP; | |
563 | if (!mapping) { | |
564 | /* | |
565 | * Some data journaling orphaned pages can have | |
566 | * page->mapping == NULL while being dirty with clean buffers. | |
567 | */ | |
266cf658 | 568 | if (page_has_private(page)) { |
1da177e4 LT |
569 | if (try_to_free_buffers(page)) { |
570 | ClearPageDirty(page); | |
b1de0d13 | 571 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
572 | return PAGE_CLEAN; |
573 | } | |
574 | } | |
575 | return PAGE_KEEP; | |
576 | } | |
577 | if (mapping->a_ops->writepage == NULL) | |
578 | return PAGE_ACTIVATE; | |
703c2708 | 579 | if (!may_write_to_inode(mapping->host, sc)) |
1da177e4 LT |
580 | return PAGE_KEEP; |
581 | ||
582 | if (clear_page_dirty_for_io(page)) { | |
583 | int res; | |
584 | struct writeback_control wbc = { | |
585 | .sync_mode = WB_SYNC_NONE, | |
586 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
587 | .range_start = 0, |
588 | .range_end = LLONG_MAX, | |
1da177e4 LT |
589 | .for_reclaim = 1, |
590 | }; | |
591 | ||
592 | SetPageReclaim(page); | |
593 | res = mapping->a_ops->writepage(page, &wbc); | |
594 | if (res < 0) | |
595 | handle_write_error(mapping, page, res); | |
994fc28c | 596 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
597 | ClearPageReclaim(page); |
598 | return PAGE_ACTIVATE; | |
599 | } | |
c661b078 | 600 | |
1da177e4 LT |
601 | if (!PageWriteback(page)) { |
602 | /* synchronous write or broken a_ops? */ | |
603 | ClearPageReclaim(page); | |
604 | } | |
23b9da55 | 605 | trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); |
e129b5c2 | 606 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
607 | return PAGE_SUCCESS; |
608 | } | |
609 | ||
610 | return PAGE_CLEAN; | |
611 | } | |
612 | ||
a649fd92 | 613 | /* |
e286781d NP |
614 | * Same as remove_mapping, but if the page is removed from the mapping, it |
615 | * gets returned with a refcount of 0. | |
a649fd92 | 616 | */ |
a528910e JW |
617 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
618 | bool reclaimed) | |
49d2e9cc | 619 | { |
c4843a75 GT |
620 | unsigned long flags; |
621 | struct mem_cgroup *memcg; | |
622 | ||
28e4d965 NP |
623 | BUG_ON(!PageLocked(page)); |
624 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 625 | |
c4843a75 GT |
626 | memcg = mem_cgroup_begin_page_stat(page); |
627 | spin_lock_irqsave(&mapping->tree_lock, flags); | |
49d2e9cc | 628 | /* |
0fd0e6b0 NP |
629 | * The non racy check for a busy page. |
630 | * | |
631 | * Must be careful with the order of the tests. When someone has | |
632 | * a ref to the page, it may be possible that they dirty it then | |
633 | * drop the reference. So if PageDirty is tested before page_count | |
634 | * here, then the following race may occur: | |
635 | * | |
636 | * get_user_pages(&page); | |
637 | * [user mapping goes away] | |
638 | * write_to(page); | |
639 | * !PageDirty(page) [good] | |
640 | * SetPageDirty(page); | |
641 | * put_page(page); | |
642 | * !page_count(page) [good, discard it] | |
643 | * | |
644 | * [oops, our write_to data is lost] | |
645 | * | |
646 | * Reversing the order of the tests ensures such a situation cannot | |
647 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
648 | * load is not satisfied before that of page->_count. | |
649 | * | |
650 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
651 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 652 | */ |
e286781d | 653 | if (!page_freeze_refs(page, 2)) |
49d2e9cc | 654 | goto cannot_free; |
e286781d NP |
655 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
656 | if (unlikely(PageDirty(page))) { | |
657 | page_unfreeze_refs(page, 2); | |
49d2e9cc | 658 | goto cannot_free; |
e286781d | 659 | } |
49d2e9cc CL |
660 | |
661 | if (PageSwapCache(page)) { | |
662 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 663 | mem_cgroup_swapout(page, swap); |
49d2e9cc | 664 | __delete_from_swap_cache(page); |
c4843a75 GT |
665 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
666 | mem_cgroup_end_page_stat(memcg); | |
0a31bc97 | 667 | swapcache_free(swap); |
e286781d | 668 | } else { |
6072d13c | 669 | void (*freepage)(struct page *); |
a528910e | 670 | void *shadow = NULL; |
6072d13c LT |
671 | |
672 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
673 | /* |
674 | * Remember a shadow entry for reclaimed file cache in | |
675 | * order to detect refaults, thus thrashing, later on. | |
676 | * | |
677 | * But don't store shadows in an address space that is | |
678 | * already exiting. This is not just an optizimation, | |
679 | * inode reclaim needs to empty out the radix tree or | |
680 | * the nodes are lost. Don't plant shadows behind its | |
681 | * back. | |
682 | */ | |
683 | if (reclaimed && page_is_file_cache(page) && | |
684 | !mapping_exiting(mapping)) | |
685 | shadow = workingset_eviction(mapping, page); | |
c4843a75 GT |
686 | __delete_from_page_cache(page, shadow, memcg); |
687 | spin_unlock_irqrestore(&mapping->tree_lock, flags); | |
688 | mem_cgroup_end_page_stat(memcg); | |
6072d13c LT |
689 | |
690 | if (freepage != NULL) | |
691 | freepage(page); | |
49d2e9cc CL |
692 | } |
693 | ||
49d2e9cc CL |
694 | return 1; |
695 | ||
696 | cannot_free: | |
c4843a75 GT |
697 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
698 | mem_cgroup_end_page_stat(memcg); | |
49d2e9cc CL |
699 | return 0; |
700 | } | |
701 | ||
e286781d NP |
702 | /* |
703 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
704 | * someone else has a ref on the page, abort and return 0. If it was | |
705 | * successfully detached, return 1. Assumes the caller has a single ref on | |
706 | * this page. | |
707 | */ | |
708 | int remove_mapping(struct address_space *mapping, struct page *page) | |
709 | { | |
a528910e | 710 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
711 | /* |
712 | * Unfreezing the refcount with 1 rather than 2 effectively | |
713 | * drops the pagecache ref for us without requiring another | |
714 | * atomic operation. | |
715 | */ | |
716 | page_unfreeze_refs(page, 1); | |
717 | return 1; | |
718 | } | |
719 | return 0; | |
720 | } | |
721 | ||
894bc310 LS |
722 | /** |
723 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
724 | * @page: page to be put back to appropriate lru list | |
725 | * | |
726 | * Add previously isolated @page to appropriate LRU list. | |
727 | * Page may still be unevictable for other reasons. | |
728 | * | |
729 | * lru_lock must not be held, interrupts must be enabled. | |
730 | */ | |
894bc310 LS |
731 | void putback_lru_page(struct page *page) |
732 | { | |
0ec3b74c | 733 | bool is_unevictable; |
bbfd28ee | 734 | int was_unevictable = PageUnevictable(page); |
894bc310 | 735 | |
309381fe | 736 | VM_BUG_ON_PAGE(PageLRU(page), page); |
894bc310 LS |
737 | |
738 | redo: | |
739 | ClearPageUnevictable(page); | |
740 | ||
39b5f29a | 741 | if (page_evictable(page)) { |
894bc310 LS |
742 | /* |
743 | * For evictable pages, we can use the cache. | |
744 | * In event of a race, worst case is we end up with an | |
745 | * unevictable page on [in]active list. | |
746 | * We know how to handle that. | |
747 | */ | |
0ec3b74c | 748 | is_unevictable = false; |
c53954a0 | 749 | lru_cache_add(page); |
894bc310 LS |
750 | } else { |
751 | /* | |
752 | * Put unevictable pages directly on zone's unevictable | |
753 | * list. | |
754 | */ | |
0ec3b74c | 755 | is_unevictable = true; |
894bc310 | 756 | add_page_to_unevictable_list(page); |
6a7b9548 | 757 | /* |
21ee9f39 MK |
758 | * When racing with an mlock or AS_UNEVICTABLE clearing |
759 | * (page is unlocked) make sure that if the other thread | |
760 | * does not observe our setting of PG_lru and fails | |
24513264 | 761 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 762 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
763 | * the page back to the evictable list. |
764 | * | |
21ee9f39 | 765 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
766 | */ |
767 | smp_mb(); | |
894bc310 | 768 | } |
894bc310 LS |
769 | |
770 | /* | |
771 | * page's status can change while we move it among lru. If an evictable | |
772 | * page is on unevictable list, it never be freed. To avoid that, | |
773 | * check after we added it to the list, again. | |
774 | */ | |
0ec3b74c | 775 | if (is_unevictable && page_evictable(page)) { |
894bc310 LS |
776 | if (!isolate_lru_page(page)) { |
777 | put_page(page); | |
778 | goto redo; | |
779 | } | |
780 | /* This means someone else dropped this page from LRU | |
781 | * So, it will be freed or putback to LRU again. There is | |
782 | * nothing to do here. | |
783 | */ | |
784 | } | |
785 | ||
0ec3b74c | 786 | if (was_unevictable && !is_unevictable) |
bbfd28ee | 787 | count_vm_event(UNEVICTABLE_PGRESCUED); |
0ec3b74c | 788 | else if (!was_unevictable && is_unevictable) |
bbfd28ee LS |
789 | count_vm_event(UNEVICTABLE_PGCULLED); |
790 | ||
894bc310 LS |
791 | put_page(page); /* drop ref from isolate */ |
792 | } | |
793 | ||
dfc8d636 JW |
794 | enum page_references { |
795 | PAGEREF_RECLAIM, | |
796 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 797 | PAGEREF_KEEP, |
dfc8d636 JW |
798 | PAGEREF_ACTIVATE, |
799 | }; | |
800 | ||
801 | static enum page_references page_check_references(struct page *page, | |
802 | struct scan_control *sc) | |
803 | { | |
64574746 | 804 | int referenced_ptes, referenced_page; |
dfc8d636 | 805 | unsigned long vm_flags; |
dfc8d636 | 806 | |
c3ac9a8a JW |
807 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
808 | &vm_flags); | |
64574746 | 809 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 810 | |
dfc8d636 JW |
811 | /* |
812 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
813 | * move the page to the unevictable list. | |
814 | */ | |
815 | if (vm_flags & VM_LOCKED) | |
816 | return PAGEREF_RECLAIM; | |
817 | ||
64574746 | 818 | if (referenced_ptes) { |
e4898273 | 819 | if (PageSwapBacked(page)) |
64574746 JW |
820 | return PAGEREF_ACTIVATE; |
821 | /* | |
822 | * All mapped pages start out with page table | |
823 | * references from the instantiating fault, so we need | |
824 | * to look twice if a mapped file page is used more | |
825 | * than once. | |
826 | * | |
827 | * Mark it and spare it for another trip around the | |
828 | * inactive list. Another page table reference will | |
829 | * lead to its activation. | |
830 | * | |
831 | * Note: the mark is set for activated pages as well | |
832 | * so that recently deactivated but used pages are | |
833 | * quickly recovered. | |
834 | */ | |
835 | SetPageReferenced(page); | |
836 | ||
34dbc67a | 837 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
838 | return PAGEREF_ACTIVATE; |
839 | ||
c909e993 KK |
840 | /* |
841 | * Activate file-backed executable pages after first usage. | |
842 | */ | |
843 | if (vm_flags & VM_EXEC) | |
844 | return PAGEREF_ACTIVATE; | |
845 | ||
64574746 JW |
846 | return PAGEREF_KEEP; |
847 | } | |
dfc8d636 JW |
848 | |
849 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 850 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
851 | return PAGEREF_RECLAIM_CLEAN; |
852 | ||
853 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
854 | } |
855 | ||
e2be15f6 MG |
856 | /* Check if a page is dirty or under writeback */ |
857 | static void page_check_dirty_writeback(struct page *page, | |
858 | bool *dirty, bool *writeback) | |
859 | { | |
b4597226 MG |
860 | struct address_space *mapping; |
861 | ||
e2be15f6 MG |
862 | /* |
863 | * Anonymous pages are not handled by flushers and must be written | |
864 | * from reclaim context. Do not stall reclaim based on them | |
865 | */ | |
866 | if (!page_is_file_cache(page)) { | |
867 | *dirty = false; | |
868 | *writeback = false; | |
869 | return; | |
870 | } | |
871 | ||
872 | /* By default assume that the page flags are accurate */ | |
873 | *dirty = PageDirty(page); | |
874 | *writeback = PageWriteback(page); | |
b4597226 MG |
875 | |
876 | /* Verify dirty/writeback state if the filesystem supports it */ | |
877 | if (!page_has_private(page)) | |
878 | return; | |
879 | ||
880 | mapping = page_mapping(page); | |
881 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
882 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
883 | } |
884 | ||
1da177e4 | 885 | /* |
1742f19f | 886 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 887 | */ |
1742f19f | 888 | static unsigned long shrink_page_list(struct list_head *page_list, |
6a18adb3 | 889 | struct zone *zone, |
f84f6e2b | 890 | struct scan_control *sc, |
02c6de8d | 891 | enum ttu_flags ttu_flags, |
8e950282 | 892 | unsigned long *ret_nr_dirty, |
d43006d5 | 893 | unsigned long *ret_nr_unqueued_dirty, |
8e950282 | 894 | unsigned long *ret_nr_congested, |
02c6de8d | 895 | unsigned long *ret_nr_writeback, |
b1a6f21e | 896 | unsigned long *ret_nr_immediate, |
02c6de8d | 897 | bool force_reclaim) |
1da177e4 LT |
898 | { |
899 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 900 | LIST_HEAD(free_pages); |
1da177e4 | 901 | int pgactivate = 0; |
d43006d5 | 902 | unsigned long nr_unqueued_dirty = 0; |
0e093d99 MG |
903 | unsigned long nr_dirty = 0; |
904 | unsigned long nr_congested = 0; | |
05ff5137 | 905 | unsigned long nr_reclaimed = 0; |
92df3a72 | 906 | unsigned long nr_writeback = 0; |
b1a6f21e | 907 | unsigned long nr_immediate = 0; |
1da177e4 LT |
908 | |
909 | cond_resched(); | |
910 | ||
1da177e4 LT |
911 | while (!list_empty(page_list)) { |
912 | struct address_space *mapping; | |
913 | struct page *page; | |
914 | int may_enter_fs; | |
02c6de8d | 915 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
e2be15f6 | 916 | bool dirty, writeback; |
1da177e4 LT |
917 | |
918 | cond_resched(); | |
919 | ||
920 | page = lru_to_page(page_list); | |
921 | list_del(&page->lru); | |
922 | ||
529ae9aa | 923 | if (!trylock_page(page)) |
1da177e4 LT |
924 | goto keep; |
925 | ||
309381fe SL |
926 | VM_BUG_ON_PAGE(PageActive(page), page); |
927 | VM_BUG_ON_PAGE(page_zone(page) != zone, page); | |
1da177e4 LT |
928 | |
929 | sc->nr_scanned++; | |
80e43426 | 930 | |
39b5f29a | 931 | if (unlikely(!page_evictable(page))) |
b291f000 | 932 | goto cull_mlocked; |
894bc310 | 933 | |
a6dc60f8 | 934 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
935 | goto keep_locked; |
936 | ||
1da177e4 LT |
937 | /* Double the slab pressure for mapped and swapcache pages */ |
938 | if (page_mapped(page) || PageSwapCache(page)) | |
939 | sc->nr_scanned++; | |
940 | ||
c661b078 AW |
941 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
942 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
943 | ||
e2be15f6 MG |
944 | /* |
945 | * The number of dirty pages determines if a zone is marked | |
946 | * reclaim_congested which affects wait_iff_congested. kswapd | |
947 | * will stall and start writing pages if the tail of the LRU | |
948 | * is all dirty unqueued pages. | |
949 | */ | |
950 | page_check_dirty_writeback(page, &dirty, &writeback); | |
951 | if (dirty || writeback) | |
952 | nr_dirty++; | |
953 | ||
954 | if (dirty && !writeback) | |
955 | nr_unqueued_dirty++; | |
956 | ||
d04e8acd MG |
957 | /* |
958 | * Treat this page as congested if the underlying BDI is or if | |
959 | * pages are cycling through the LRU so quickly that the | |
960 | * pages marked for immediate reclaim are making it to the | |
961 | * end of the LRU a second time. | |
962 | */ | |
e2be15f6 | 963 | mapping = page_mapping(page); |
1da58ee2 | 964 | if (((dirty || writeback) && mapping && |
703c2708 | 965 | inode_write_congested(mapping->host)) || |
d04e8acd | 966 | (writeback && PageReclaim(page))) |
e2be15f6 MG |
967 | nr_congested++; |
968 | ||
283aba9f MG |
969 | /* |
970 | * If a page at the tail of the LRU is under writeback, there | |
971 | * are three cases to consider. | |
972 | * | |
973 | * 1) If reclaim is encountering an excessive number of pages | |
974 | * under writeback and this page is both under writeback and | |
975 | * PageReclaim then it indicates that pages are being queued | |
976 | * for IO but are being recycled through the LRU before the | |
977 | * IO can complete. Waiting on the page itself risks an | |
978 | * indefinite stall if it is impossible to writeback the | |
979 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
980 | * note that the LRU is being scanned too quickly and the |
981 | * caller can stall after page list has been processed. | |
283aba9f | 982 | * |
97c9341f | 983 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
984 | * not marked for immediate reclaim, or the caller does not |
985 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
986 | * not to fs). In this case mark the page for immediate | |
97c9341f | 987 | * reclaim and continue scanning. |
283aba9f | 988 | * |
ecf5fc6e MH |
989 | * Require may_enter_fs because we would wait on fs, which |
990 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
991 | * enter reclaim, and deadlock if it waits on a page for |
992 | * which it is needed to do the write (loop masks off | |
993 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
994 | * would probably show more reasons. | |
995 | * | |
7fadc820 | 996 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
997 | * PageReclaim. memcg does not have any dirty pages |
998 | * throttling so we could easily OOM just because too many | |
999 | * pages are in writeback and there is nothing else to | |
1000 | * reclaim. Wait for the writeback to complete. | |
1001 | */ | |
c661b078 | 1002 | if (PageWriteback(page)) { |
283aba9f MG |
1003 | /* Case 1 above */ |
1004 | if (current_is_kswapd() && | |
1005 | PageReclaim(page) && | |
57054651 | 1006 | test_bit(ZONE_WRITEBACK, &zone->flags)) { |
b1a6f21e MG |
1007 | nr_immediate++; |
1008 | goto keep_locked; | |
283aba9f MG |
1009 | |
1010 | /* Case 2 above */ | |
97c9341f | 1011 | } else if (sane_reclaim(sc) || |
ecf5fc6e | 1012 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1013 | /* |
1014 | * This is slightly racy - end_page_writeback() | |
1015 | * might have just cleared PageReclaim, then | |
1016 | * setting PageReclaim here end up interpreted | |
1017 | * as PageReadahead - but that does not matter | |
1018 | * enough to care. What we do want is for this | |
1019 | * page to have PageReclaim set next time memcg | |
1020 | * reclaim reaches the tests above, so it will | |
1021 | * then wait_on_page_writeback() to avoid OOM; | |
1022 | * and it's also appropriate in global reclaim. | |
1023 | */ | |
1024 | SetPageReclaim(page); | |
e62e384e | 1025 | nr_writeback++; |
c3b94f44 | 1026 | goto keep_locked; |
283aba9f MG |
1027 | |
1028 | /* Case 3 above */ | |
1029 | } else { | |
7fadc820 | 1030 | unlock_page(page); |
283aba9f | 1031 | wait_on_page_writeback(page); |
7fadc820 HD |
1032 | /* then go back and try same page again */ |
1033 | list_add_tail(&page->lru, page_list); | |
1034 | continue; | |
e62e384e | 1035 | } |
c661b078 | 1036 | } |
1da177e4 | 1037 | |
02c6de8d MK |
1038 | if (!force_reclaim) |
1039 | references = page_check_references(page, sc); | |
1040 | ||
dfc8d636 JW |
1041 | switch (references) { |
1042 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1043 | goto activate_locked; |
64574746 JW |
1044 | case PAGEREF_KEEP: |
1045 | goto keep_locked; | |
dfc8d636 JW |
1046 | case PAGEREF_RECLAIM: |
1047 | case PAGEREF_RECLAIM_CLEAN: | |
1048 | ; /* try to reclaim the page below */ | |
1049 | } | |
1da177e4 | 1050 | |
1da177e4 LT |
1051 | /* |
1052 | * Anonymous process memory has backing store? | |
1053 | * Try to allocate it some swap space here. | |
1054 | */ | |
b291f000 | 1055 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
1056 | if (!(sc->gfp_mask & __GFP_IO)) |
1057 | goto keep_locked; | |
5bc7b8ac | 1058 | if (!add_to_swap(page, page_list)) |
1da177e4 | 1059 | goto activate_locked; |
63eb6b93 | 1060 | may_enter_fs = 1; |
1da177e4 | 1061 | |
e2be15f6 MG |
1062 | /* Adding to swap updated mapping */ |
1063 | mapping = page_mapping(page); | |
1064 | } | |
1da177e4 LT |
1065 | |
1066 | /* | |
1067 | * The page is mapped into the page tables of one or more | |
1068 | * processes. Try to unmap it here. | |
1069 | */ | |
1070 | if (page_mapped(page) && mapping) { | |
72b252ae MG |
1071 | switch (try_to_unmap(page, |
1072 | ttu_flags|TTU_BATCH_FLUSH)) { | |
1da177e4 LT |
1073 | case SWAP_FAIL: |
1074 | goto activate_locked; | |
1075 | case SWAP_AGAIN: | |
1076 | goto keep_locked; | |
b291f000 NP |
1077 | case SWAP_MLOCK: |
1078 | goto cull_mlocked; | |
1da177e4 LT |
1079 | case SWAP_SUCCESS: |
1080 | ; /* try to free the page below */ | |
1081 | } | |
1082 | } | |
1083 | ||
1084 | if (PageDirty(page)) { | |
ee72886d MG |
1085 | /* |
1086 | * Only kswapd can writeback filesystem pages to | |
d43006d5 MG |
1087 | * avoid risk of stack overflow but only writeback |
1088 | * if many dirty pages have been encountered. | |
ee72886d | 1089 | */ |
f84f6e2b | 1090 | if (page_is_file_cache(page) && |
9e3b2f8c | 1091 | (!current_is_kswapd() || |
57054651 | 1092 | !test_bit(ZONE_DIRTY, &zone->flags))) { |
49ea7eb6 MG |
1093 | /* |
1094 | * Immediately reclaim when written back. | |
1095 | * Similar in principal to deactivate_page() | |
1096 | * except we already have the page isolated | |
1097 | * and know it's dirty | |
1098 | */ | |
1099 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
1100 | SetPageReclaim(page); | |
1101 | ||
ee72886d MG |
1102 | goto keep_locked; |
1103 | } | |
1104 | ||
dfc8d636 | 1105 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1106 | goto keep_locked; |
4dd4b920 | 1107 | if (!may_enter_fs) |
1da177e4 | 1108 | goto keep_locked; |
52a8363e | 1109 | if (!sc->may_writepage) |
1da177e4 LT |
1110 | goto keep_locked; |
1111 | ||
d950c947 MG |
1112 | /* |
1113 | * Page is dirty. Flush the TLB if a writable entry | |
1114 | * potentially exists to avoid CPU writes after IO | |
1115 | * starts and then write it out here. | |
1116 | */ | |
1117 | try_to_unmap_flush_dirty(); | |
7d3579e8 | 1118 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1119 | case PAGE_KEEP: |
1120 | goto keep_locked; | |
1121 | case PAGE_ACTIVATE: | |
1122 | goto activate_locked; | |
1123 | case PAGE_SUCCESS: | |
7d3579e8 | 1124 | if (PageWriteback(page)) |
41ac1999 | 1125 | goto keep; |
7d3579e8 | 1126 | if (PageDirty(page)) |
1da177e4 | 1127 | goto keep; |
7d3579e8 | 1128 | |
1da177e4 LT |
1129 | /* |
1130 | * A synchronous write - probably a ramdisk. Go | |
1131 | * ahead and try to reclaim the page. | |
1132 | */ | |
529ae9aa | 1133 | if (!trylock_page(page)) |
1da177e4 LT |
1134 | goto keep; |
1135 | if (PageDirty(page) || PageWriteback(page)) | |
1136 | goto keep_locked; | |
1137 | mapping = page_mapping(page); | |
1138 | case PAGE_CLEAN: | |
1139 | ; /* try to free the page below */ | |
1140 | } | |
1141 | } | |
1142 | ||
1143 | /* | |
1144 | * If the page has buffers, try to free the buffer mappings | |
1145 | * associated with this page. If we succeed we try to free | |
1146 | * the page as well. | |
1147 | * | |
1148 | * We do this even if the page is PageDirty(). | |
1149 | * try_to_release_page() does not perform I/O, but it is | |
1150 | * possible for a page to have PageDirty set, but it is actually | |
1151 | * clean (all its buffers are clean). This happens if the | |
1152 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1153 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1154 | * try_to_release_page() will discover that cleanness and will |
1155 | * drop the buffers and mark the page clean - it can be freed. | |
1156 | * | |
1157 | * Rarely, pages can have buffers and no ->mapping. These are | |
1158 | * the pages which were not successfully invalidated in | |
1159 | * truncate_complete_page(). We try to drop those buffers here | |
1160 | * and if that worked, and the page is no longer mapped into | |
1161 | * process address space (page_count == 1) it can be freed. | |
1162 | * Otherwise, leave the page on the LRU so it is swappable. | |
1163 | */ | |
266cf658 | 1164 | if (page_has_private(page)) { |
1da177e4 LT |
1165 | if (!try_to_release_page(page, sc->gfp_mask)) |
1166 | goto activate_locked; | |
e286781d NP |
1167 | if (!mapping && page_count(page) == 1) { |
1168 | unlock_page(page); | |
1169 | if (put_page_testzero(page)) | |
1170 | goto free_it; | |
1171 | else { | |
1172 | /* | |
1173 | * rare race with speculative reference. | |
1174 | * the speculative reference will free | |
1175 | * this page shortly, so we may | |
1176 | * increment nr_reclaimed here (and | |
1177 | * leave it off the LRU). | |
1178 | */ | |
1179 | nr_reclaimed++; | |
1180 | continue; | |
1181 | } | |
1182 | } | |
1da177e4 LT |
1183 | } |
1184 | ||
a528910e | 1185 | if (!mapping || !__remove_mapping(mapping, page, true)) |
49d2e9cc | 1186 | goto keep_locked; |
1da177e4 | 1187 | |
a978d6f5 NP |
1188 | /* |
1189 | * At this point, we have no other references and there is | |
1190 | * no way to pick any more up (removed from LRU, removed | |
1191 | * from pagecache). Can use non-atomic bitops now (and | |
1192 | * we obviously don't have to worry about waking up a process | |
1193 | * waiting on the page lock, because there are no references. | |
1194 | */ | |
1195 | __clear_page_locked(page); | |
e286781d | 1196 | free_it: |
05ff5137 | 1197 | nr_reclaimed++; |
abe4c3b5 MG |
1198 | |
1199 | /* | |
1200 | * Is there need to periodically free_page_list? It would | |
1201 | * appear not as the counts should be low | |
1202 | */ | |
1203 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1204 | continue; |
1205 | ||
b291f000 | 1206 | cull_mlocked: |
63d6c5ad HD |
1207 | if (PageSwapCache(page)) |
1208 | try_to_free_swap(page); | |
b291f000 | 1209 | unlock_page(page); |
c54839a7 | 1210 | list_add(&page->lru, &ret_pages); |
b291f000 NP |
1211 | continue; |
1212 | ||
1da177e4 | 1213 | activate_locked: |
68a22394 RR |
1214 | /* Not a candidate for swapping, so reclaim swap space. */ |
1215 | if (PageSwapCache(page) && vm_swap_full()) | |
a2c43eed | 1216 | try_to_free_swap(page); |
309381fe | 1217 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
1218 | SetPageActive(page); |
1219 | pgactivate++; | |
1220 | keep_locked: | |
1221 | unlock_page(page); | |
1222 | keep: | |
1223 | list_add(&page->lru, &ret_pages); | |
309381fe | 1224 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1225 | } |
abe4c3b5 | 1226 | |
747db954 | 1227 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1228 | try_to_unmap_flush(); |
b745bc85 | 1229 | free_hot_cold_page_list(&free_pages, true); |
abe4c3b5 | 1230 | |
1da177e4 | 1231 | list_splice(&ret_pages, page_list); |
f8891e5e | 1232 | count_vm_events(PGACTIVATE, pgactivate); |
0a31bc97 | 1233 | |
8e950282 MG |
1234 | *ret_nr_dirty += nr_dirty; |
1235 | *ret_nr_congested += nr_congested; | |
d43006d5 | 1236 | *ret_nr_unqueued_dirty += nr_unqueued_dirty; |
92df3a72 | 1237 | *ret_nr_writeback += nr_writeback; |
b1a6f21e | 1238 | *ret_nr_immediate += nr_immediate; |
05ff5137 | 1239 | return nr_reclaimed; |
1da177e4 LT |
1240 | } |
1241 | ||
02c6de8d MK |
1242 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1243 | struct list_head *page_list) | |
1244 | { | |
1245 | struct scan_control sc = { | |
1246 | .gfp_mask = GFP_KERNEL, | |
1247 | .priority = DEF_PRIORITY, | |
1248 | .may_unmap = 1, | |
1249 | }; | |
8e950282 | 1250 | unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; |
02c6de8d MK |
1251 | struct page *page, *next; |
1252 | LIST_HEAD(clean_pages); | |
1253 | ||
1254 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e RA |
1255 | if (page_is_file_cache(page) && !PageDirty(page) && |
1256 | !isolated_balloon_page(page)) { | |
02c6de8d MK |
1257 | ClearPageActive(page); |
1258 | list_move(&page->lru, &clean_pages); | |
1259 | } | |
1260 | } | |
1261 | ||
1262 | ret = shrink_page_list(&clean_pages, zone, &sc, | |
8e950282 MG |
1263 | TTU_UNMAP|TTU_IGNORE_ACCESS, |
1264 | &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); | |
02c6de8d | 1265 | list_splice(&clean_pages, page_list); |
83da7510 | 1266 | mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1267 | return ret; |
1268 | } | |
1269 | ||
5ad333eb AW |
1270 | /* |
1271 | * Attempt to remove the specified page from its LRU. Only take this page | |
1272 | * if it is of the appropriate PageActive status. Pages which are being | |
1273 | * freed elsewhere are also ignored. | |
1274 | * | |
1275 | * page: page to consider | |
1276 | * mode: one of the LRU isolation modes defined above | |
1277 | * | |
1278 | * returns 0 on success, -ve errno on failure. | |
1279 | */ | |
f3fd4a61 | 1280 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1281 | { |
1282 | int ret = -EINVAL; | |
1283 | ||
1284 | /* Only take pages on the LRU. */ | |
1285 | if (!PageLRU(page)) | |
1286 | return ret; | |
1287 | ||
e46a2879 MK |
1288 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1289 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1290 | return ret; |
1291 | ||
5ad333eb | 1292 | ret = -EBUSY; |
08e552c6 | 1293 | |
c8244935 MG |
1294 | /* |
1295 | * To minimise LRU disruption, the caller can indicate that it only | |
1296 | * wants to isolate pages it will be able to operate on without | |
1297 | * blocking - clean pages for the most part. | |
1298 | * | |
1299 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1300 | * is used by reclaim when it is cannot write to backing storage | |
1301 | * | |
1302 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1303 | * that it is possible to migrate without blocking | |
1304 | */ | |
1305 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1306 | /* All the caller can do on PageWriteback is block */ | |
1307 | if (PageWriteback(page)) | |
1308 | return ret; | |
1309 | ||
1310 | if (PageDirty(page)) { | |
1311 | struct address_space *mapping; | |
1312 | ||
1313 | /* ISOLATE_CLEAN means only clean pages */ | |
1314 | if (mode & ISOLATE_CLEAN) | |
1315 | return ret; | |
1316 | ||
1317 | /* | |
1318 | * Only pages without mappings or that have a | |
1319 | * ->migratepage callback are possible to migrate | |
1320 | * without blocking | |
1321 | */ | |
1322 | mapping = page_mapping(page); | |
1323 | if (mapping && !mapping->a_ops->migratepage) | |
1324 | return ret; | |
1325 | } | |
1326 | } | |
39deaf85 | 1327 | |
f80c0673 MK |
1328 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1329 | return ret; | |
1330 | ||
5ad333eb AW |
1331 | if (likely(get_page_unless_zero(page))) { |
1332 | /* | |
1333 | * Be careful not to clear PageLRU until after we're | |
1334 | * sure the page is not being freed elsewhere -- the | |
1335 | * page release code relies on it. | |
1336 | */ | |
1337 | ClearPageLRU(page); | |
1338 | ret = 0; | |
1339 | } | |
1340 | ||
1341 | return ret; | |
1342 | } | |
1343 | ||
1da177e4 LT |
1344 | /* |
1345 | * zone->lru_lock is heavily contended. Some of the functions that | |
1346 | * shrink the lists perform better by taking out a batch of pages | |
1347 | * and working on them outside the LRU lock. | |
1348 | * | |
1349 | * For pagecache intensive workloads, this function is the hottest | |
1350 | * spot in the kernel (apart from copy_*_user functions). | |
1351 | * | |
1352 | * Appropriate locks must be held before calling this function. | |
1353 | * | |
1354 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1355 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1356 | * @dst: The temp list to put pages on to. |
f626012d | 1357 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1358 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1359 | * @mode: One of the LRU isolation modes |
3cb99451 | 1360 | * @lru: LRU list id for isolating |
1da177e4 LT |
1361 | * |
1362 | * returns how many pages were moved onto *@dst. | |
1363 | */ | |
69e05944 | 1364 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1365 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1366 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1367 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1368 | { |
75b00af7 | 1369 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1370 | unsigned long nr_taken = 0; |
c9b02d97 | 1371 | unsigned long scan; |
1da177e4 | 1372 | |
0b802f10 VD |
1373 | for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan && |
1374 | !list_empty(src); scan++) { | |
5ad333eb | 1375 | struct page *page; |
fa9add64 | 1376 | int nr_pages; |
5ad333eb | 1377 | |
1da177e4 LT |
1378 | page = lru_to_page(src); |
1379 | prefetchw_prev_lru_page(page, src, flags); | |
1380 | ||
309381fe | 1381 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1382 | |
f3fd4a61 | 1383 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1384 | case 0: |
fa9add64 HD |
1385 | nr_pages = hpage_nr_pages(page); |
1386 | mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); | |
5ad333eb | 1387 | list_move(&page->lru, dst); |
fa9add64 | 1388 | nr_taken += nr_pages; |
5ad333eb AW |
1389 | break; |
1390 | ||
1391 | case -EBUSY: | |
1392 | /* else it is being freed elsewhere */ | |
1393 | list_move(&page->lru, src); | |
1394 | continue; | |
46453a6e | 1395 | |
5ad333eb AW |
1396 | default: |
1397 | BUG(); | |
1398 | } | |
1da177e4 LT |
1399 | } |
1400 | ||
f626012d | 1401 | *nr_scanned = scan; |
75b00af7 HD |
1402 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, |
1403 | nr_taken, mode, is_file_lru(lru)); | |
1da177e4 LT |
1404 | return nr_taken; |
1405 | } | |
1406 | ||
62695a84 NP |
1407 | /** |
1408 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1409 | * @page: page to isolate from its LRU list | |
1410 | * | |
1411 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1412 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1413 | * | |
1414 | * Returns 0 if the page was removed from an LRU list. | |
1415 | * Returns -EBUSY if the page was not on an LRU list. | |
1416 | * | |
1417 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1418 | * the active list, it will have PageActive set. If it was found on |
1419 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1420 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1421 | * |
1422 | * The vmstat statistic corresponding to the list on which the page was | |
1423 | * found will be decremented. | |
1424 | * | |
1425 | * Restrictions: | |
1426 | * (1) Must be called with an elevated refcount on the page. This is a | |
1427 | * fundamentnal difference from isolate_lru_pages (which is called | |
1428 | * without a stable reference). | |
1429 | * (2) the lru_lock must not be held. | |
1430 | * (3) interrupts must be enabled. | |
1431 | */ | |
1432 | int isolate_lru_page(struct page *page) | |
1433 | { | |
1434 | int ret = -EBUSY; | |
1435 | ||
309381fe | 1436 | VM_BUG_ON_PAGE(!page_count(page), page); |
0c917313 | 1437 | |
62695a84 NP |
1438 | if (PageLRU(page)) { |
1439 | struct zone *zone = page_zone(page); | |
fa9add64 | 1440 | struct lruvec *lruvec; |
62695a84 NP |
1441 | |
1442 | spin_lock_irq(&zone->lru_lock); | |
fa9add64 | 1443 | lruvec = mem_cgroup_page_lruvec(page, zone); |
0c917313 | 1444 | if (PageLRU(page)) { |
894bc310 | 1445 | int lru = page_lru(page); |
0c917313 | 1446 | get_page(page); |
62695a84 | 1447 | ClearPageLRU(page); |
fa9add64 HD |
1448 | del_page_from_lru_list(page, lruvec, lru); |
1449 | ret = 0; | |
62695a84 NP |
1450 | } |
1451 | spin_unlock_irq(&zone->lru_lock); | |
1452 | } | |
1453 | return ret; | |
1454 | } | |
1455 | ||
35cd7815 | 1456 | /* |
d37dd5dc FW |
1457 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1458 | * then get resheduled. When there are massive number of tasks doing page | |
1459 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1460 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1461 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 RR |
1462 | */ |
1463 | static int too_many_isolated(struct zone *zone, int file, | |
1464 | struct scan_control *sc) | |
1465 | { | |
1466 | unsigned long inactive, isolated; | |
1467 | ||
1468 | if (current_is_kswapd()) | |
1469 | return 0; | |
1470 | ||
97c9341f | 1471 | if (!sane_reclaim(sc)) |
35cd7815 RR |
1472 | return 0; |
1473 | ||
1474 | if (file) { | |
1475 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1476 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1477 | } else { | |
1478 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1479 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1480 | } | |
1481 | ||
3cf23841 FW |
1482 | /* |
1483 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1484 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1485 | * deadlock. | |
1486 | */ | |
d0164adc | 1487 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1488 | inactive >>= 3; |
1489 | ||
35cd7815 RR |
1490 | return isolated > inactive; |
1491 | } | |
1492 | ||
66635629 | 1493 | static noinline_for_stack void |
75b00af7 | 1494 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1495 | { |
27ac81d8 KK |
1496 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1497 | struct zone *zone = lruvec_zone(lruvec); | |
3f79768f | 1498 | LIST_HEAD(pages_to_free); |
66635629 | 1499 | |
66635629 MG |
1500 | /* |
1501 | * Put back any unfreeable pages. | |
1502 | */ | |
66635629 | 1503 | while (!list_empty(page_list)) { |
3f79768f | 1504 | struct page *page = lru_to_page(page_list); |
66635629 | 1505 | int lru; |
3f79768f | 1506 | |
309381fe | 1507 | VM_BUG_ON_PAGE(PageLRU(page), page); |
66635629 | 1508 | list_del(&page->lru); |
39b5f29a | 1509 | if (unlikely(!page_evictable(page))) { |
66635629 MG |
1510 | spin_unlock_irq(&zone->lru_lock); |
1511 | putback_lru_page(page); | |
1512 | spin_lock_irq(&zone->lru_lock); | |
1513 | continue; | |
1514 | } | |
fa9add64 HD |
1515 | |
1516 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1517 | ||
7a608572 | 1518 | SetPageLRU(page); |
66635629 | 1519 | lru = page_lru(page); |
fa9add64 HD |
1520 | add_page_to_lru_list(page, lruvec, lru); |
1521 | ||
66635629 MG |
1522 | if (is_active_lru(lru)) { |
1523 | int file = is_file_lru(lru); | |
9992af10 RR |
1524 | int numpages = hpage_nr_pages(page); |
1525 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1526 | } |
2bcf8879 HD |
1527 | if (put_page_testzero(page)) { |
1528 | __ClearPageLRU(page); | |
1529 | __ClearPageActive(page); | |
fa9add64 | 1530 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1531 | |
1532 | if (unlikely(PageCompound(page))) { | |
1533 | spin_unlock_irq(&zone->lru_lock); | |
747db954 | 1534 | mem_cgroup_uncharge(page); |
2bcf8879 HD |
1535 | (*get_compound_page_dtor(page))(page); |
1536 | spin_lock_irq(&zone->lru_lock); | |
1537 | } else | |
1538 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1539 | } |
1540 | } | |
66635629 | 1541 | |
3f79768f HD |
1542 | /* |
1543 | * To save our caller's stack, now use input list for pages to free. | |
1544 | */ | |
1545 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1546 | } |
1547 | ||
399ba0b9 N |
1548 | /* |
1549 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1550 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1551 | * In that case we should only throttle if the backing device it is | |
1552 | * writing to is congested. In other cases it is safe to throttle. | |
1553 | */ | |
1554 | static int current_may_throttle(void) | |
1555 | { | |
1556 | return !(current->flags & PF_LESS_THROTTLE) || | |
1557 | current->backing_dev_info == NULL || | |
1558 | bdi_write_congested(current->backing_dev_info); | |
1559 | } | |
1560 | ||
1da177e4 | 1561 | /* |
1742f19f AM |
1562 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1563 | * of reclaimed pages | |
1da177e4 | 1564 | */ |
66635629 | 1565 | static noinline_for_stack unsigned long |
1a93be0e | 1566 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1567 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1568 | { |
1569 | LIST_HEAD(page_list); | |
e247dbce | 1570 | unsigned long nr_scanned; |
05ff5137 | 1571 | unsigned long nr_reclaimed = 0; |
e247dbce | 1572 | unsigned long nr_taken; |
8e950282 MG |
1573 | unsigned long nr_dirty = 0; |
1574 | unsigned long nr_congested = 0; | |
e2be15f6 | 1575 | unsigned long nr_unqueued_dirty = 0; |
92df3a72 | 1576 | unsigned long nr_writeback = 0; |
b1a6f21e | 1577 | unsigned long nr_immediate = 0; |
f3fd4a61 | 1578 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1579 | int file = is_file_lru(lru); |
1a93be0e KK |
1580 | struct zone *zone = lruvec_zone(lruvec); |
1581 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
78dc583d | 1582 | |
35cd7815 | 1583 | while (unlikely(too_many_isolated(zone, file, sc))) { |
58355c78 | 1584 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1585 | |
1586 | /* We are about to die and free our memory. Return now. */ | |
1587 | if (fatal_signal_pending(current)) | |
1588 | return SWAP_CLUSTER_MAX; | |
1589 | } | |
1590 | ||
1da177e4 | 1591 | lru_add_drain(); |
f80c0673 MK |
1592 | |
1593 | if (!sc->may_unmap) | |
61317289 | 1594 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1595 | if (!sc->may_writepage) |
61317289 | 1596 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1597 | |
1da177e4 | 1598 | spin_lock_irq(&zone->lru_lock); |
b35ea17b | 1599 | |
5dc35979 KK |
1600 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1601 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc KK |
1602 | |
1603 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1604 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1605 | ||
89b5fae5 | 1606 | if (global_reclaim(sc)) { |
0d5d823a | 1607 | __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); |
e247dbce | 1608 | if (current_is_kswapd()) |
75b00af7 | 1609 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); |
e247dbce | 1610 | else |
75b00af7 | 1611 | __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); |
e247dbce | 1612 | } |
d563c050 | 1613 | spin_unlock_irq(&zone->lru_lock); |
b35ea17b | 1614 | |
d563c050 | 1615 | if (nr_taken == 0) |
66635629 | 1616 | return 0; |
5ad333eb | 1617 | |
02c6de8d | 1618 | nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, |
8e950282 MG |
1619 | &nr_dirty, &nr_unqueued_dirty, &nr_congested, |
1620 | &nr_writeback, &nr_immediate, | |
1621 | false); | |
c661b078 | 1622 | |
3f79768f HD |
1623 | spin_lock_irq(&zone->lru_lock); |
1624 | ||
95d918fc | 1625 | reclaim_stat->recent_scanned[file] += nr_taken; |
d563c050 | 1626 | |
904249aa YH |
1627 | if (global_reclaim(sc)) { |
1628 | if (current_is_kswapd()) | |
1629 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1630 | nr_reclaimed); | |
1631 | else | |
1632 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1633 | nr_reclaimed); | |
1634 | } | |
a74609fa | 1635 | |
27ac81d8 | 1636 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1637 | |
95d918fc | 1638 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f HD |
1639 | |
1640 | spin_unlock_irq(&zone->lru_lock); | |
1641 | ||
747db954 | 1642 | mem_cgroup_uncharge_list(&page_list); |
b745bc85 | 1643 | free_hot_cold_page_list(&page_list, true); |
e11da5b4 | 1644 | |
92df3a72 MG |
1645 | /* |
1646 | * If reclaim is isolating dirty pages under writeback, it implies | |
1647 | * that the long-lived page allocation rate is exceeding the page | |
1648 | * laundering rate. Either the global limits are not being effective | |
1649 | * at throttling processes due to the page distribution throughout | |
1650 | * zones or there is heavy usage of a slow backing device. The | |
1651 | * only option is to throttle from reclaim context which is not ideal | |
1652 | * as there is no guarantee the dirtying process is throttled in the | |
1653 | * same way balance_dirty_pages() manages. | |
1654 | * | |
8e950282 MG |
1655 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1656 | * of pages under pages flagged for immediate reclaim and stall if any | |
1657 | * are encountered in the nr_immediate check below. | |
92df3a72 | 1658 | */ |
918fc718 | 1659 | if (nr_writeback && nr_writeback == nr_taken) |
57054651 | 1660 | set_bit(ZONE_WRITEBACK, &zone->flags); |
92df3a72 | 1661 | |
d43006d5 | 1662 | /* |
97c9341f TH |
1663 | * Legacy memcg will stall in page writeback so avoid forcibly |
1664 | * stalling here. | |
d43006d5 | 1665 | */ |
97c9341f | 1666 | if (sane_reclaim(sc)) { |
8e950282 MG |
1667 | /* |
1668 | * Tag a zone as congested if all the dirty pages scanned were | |
1669 | * backed by a congested BDI and wait_iff_congested will stall. | |
1670 | */ | |
1671 | if (nr_dirty && nr_dirty == nr_congested) | |
57054651 | 1672 | set_bit(ZONE_CONGESTED, &zone->flags); |
8e950282 | 1673 | |
b1a6f21e MG |
1674 | /* |
1675 | * If dirty pages are scanned that are not queued for IO, it | |
1676 | * implies that flushers are not keeping up. In this case, flag | |
57054651 JW |
1677 | * the zone ZONE_DIRTY and kswapd will start writing pages from |
1678 | * reclaim context. | |
b1a6f21e MG |
1679 | */ |
1680 | if (nr_unqueued_dirty == nr_taken) | |
57054651 | 1681 | set_bit(ZONE_DIRTY, &zone->flags); |
b1a6f21e MG |
1682 | |
1683 | /* | |
b738d764 LT |
1684 | * If kswapd scans pages marked marked for immediate |
1685 | * reclaim and under writeback (nr_immediate), it implies | |
1686 | * that pages are cycling through the LRU faster than | |
b1a6f21e MG |
1687 | * they are written so also forcibly stall. |
1688 | */ | |
b738d764 | 1689 | if (nr_immediate && current_may_throttle()) |
b1a6f21e | 1690 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
e2be15f6 | 1691 | } |
d43006d5 | 1692 | |
8e950282 MG |
1693 | /* |
1694 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1695 | * is congested. Allow kswapd to continue until it starts encountering | |
1696 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1697 | */ | |
399ba0b9 N |
1698 | if (!sc->hibernation_mode && !current_is_kswapd() && |
1699 | current_may_throttle()) | |
8e950282 MG |
1700 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); |
1701 | ||
e11da5b4 MG |
1702 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, |
1703 | zone_idx(zone), | |
1704 | nr_scanned, nr_reclaimed, | |
9e3b2f8c | 1705 | sc->priority, |
23b9da55 | 1706 | trace_shrink_flags(file)); |
05ff5137 | 1707 | return nr_reclaimed; |
1da177e4 LT |
1708 | } |
1709 | ||
1710 | /* | |
1711 | * This moves pages from the active list to the inactive list. | |
1712 | * | |
1713 | * We move them the other way if the page is referenced by one or more | |
1714 | * processes, from rmap. | |
1715 | * | |
1716 | * If the pages are mostly unmapped, the processing is fast and it is | |
1717 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1718 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1719 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1720 | * this, so instead we remove the pages from the LRU while processing them. | |
1721 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1722 | * nobody will play with that bit on a non-LRU page. | |
1723 | * | |
1724 | * The downside is that we have to touch page->_count against each page. | |
1725 | * But we had to alter page->flags anyway. | |
1726 | */ | |
1cfb419b | 1727 | |
fa9add64 | 1728 | static void move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1729 | struct list_head *list, |
2bcf8879 | 1730 | struct list_head *pages_to_free, |
3eb4140f WF |
1731 | enum lru_list lru) |
1732 | { | |
fa9add64 | 1733 | struct zone *zone = lruvec_zone(lruvec); |
3eb4140f | 1734 | unsigned long pgmoved = 0; |
3eb4140f | 1735 | struct page *page; |
fa9add64 | 1736 | int nr_pages; |
3eb4140f | 1737 | |
3eb4140f WF |
1738 | while (!list_empty(list)) { |
1739 | page = lru_to_page(list); | |
fa9add64 | 1740 | lruvec = mem_cgroup_page_lruvec(page, zone); |
3eb4140f | 1741 | |
309381fe | 1742 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3eb4140f WF |
1743 | SetPageLRU(page); |
1744 | ||
fa9add64 HD |
1745 | nr_pages = hpage_nr_pages(page); |
1746 | mem_cgroup_update_lru_size(lruvec, lru, nr_pages); | |
925b7673 | 1747 | list_move(&page->lru, &lruvec->lists[lru]); |
fa9add64 | 1748 | pgmoved += nr_pages; |
3eb4140f | 1749 | |
2bcf8879 HD |
1750 | if (put_page_testzero(page)) { |
1751 | __ClearPageLRU(page); | |
1752 | __ClearPageActive(page); | |
fa9add64 | 1753 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1754 | |
1755 | if (unlikely(PageCompound(page))) { | |
1756 | spin_unlock_irq(&zone->lru_lock); | |
747db954 | 1757 | mem_cgroup_uncharge(page); |
2bcf8879 HD |
1758 | (*get_compound_page_dtor(page))(page); |
1759 | spin_lock_irq(&zone->lru_lock); | |
1760 | } else | |
1761 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1762 | } |
1763 | } | |
1764 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1765 | if (!is_active_lru(lru)) | |
1766 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1767 | } | |
1cfb419b | 1768 | |
f626012d | 1769 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1770 | struct lruvec *lruvec, |
f16015fb | 1771 | struct scan_control *sc, |
9e3b2f8c | 1772 | enum lru_list lru) |
1da177e4 | 1773 | { |
44c241f1 | 1774 | unsigned long nr_taken; |
f626012d | 1775 | unsigned long nr_scanned; |
6fe6b7e3 | 1776 | unsigned long vm_flags; |
1da177e4 | 1777 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1778 | LIST_HEAD(l_active); |
b69408e8 | 1779 | LIST_HEAD(l_inactive); |
1da177e4 | 1780 | struct page *page; |
1a93be0e | 1781 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
44c241f1 | 1782 | unsigned long nr_rotated = 0; |
f3fd4a61 | 1783 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1784 | int file = is_file_lru(lru); |
1a93be0e | 1785 | struct zone *zone = lruvec_zone(lruvec); |
1da177e4 LT |
1786 | |
1787 | lru_add_drain(); | |
f80c0673 MK |
1788 | |
1789 | if (!sc->may_unmap) | |
61317289 | 1790 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1791 | if (!sc->may_writepage) |
61317289 | 1792 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1793 | |
1da177e4 | 1794 | spin_lock_irq(&zone->lru_lock); |
925b7673 | 1795 | |
5dc35979 KK |
1796 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1797 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1798 | if (global_reclaim(sc)) |
0d5d823a | 1799 | __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); |
89b5fae5 | 1800 | |
b7c46d15 | 1801 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1802 | |
f626012d | 1803 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); |
3cb99451 | 1804 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); |
a731286d | 1805 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
1da177e4 LT |
1806 | spin_unlock_irq(&zone->lru_lock); |
1807 | ||
1da177e4 LT |
1808 | while (!list_empty(&l_hold)) { |
1809 | cond_resched(); | |
1810 | page = lru_to_page(&l_hold); | |
1811 | list_del(&page->lru); | |
7e9cd484 | 1812 | |
39b5f29a | 1813 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
1814 | putback_lru_page(page); |
1815 | continue; | |
1816 | } | |
1817 | ||
cc715d99 MG |
1818 | if (unlikely(buffer_heads_over_limit)) { |
1819 | if (page_has_private(page) && trylock_page(page)) { | |
1820 | if (page_has_private(page)) | |
1821 | try_to_release_page(page, 0); | |
1822 | unlock_page(page); | |
1823 | } | |
1824 | } | |
1825 | ||
c3ac9a8a JW |
1826 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1827 | &vm_flags)) { | |
9992af10 | 1828 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1829 | /* |
1830 | * Identify referenced, file-backed active pages and | |
1831 | * give them one more trip around the active list. So | |
1832 | * that executable code get better chances to stay in | |
1833 | * memory under moderate memory pressure. Anon pages | |
1834 | * are not likely to be evicted by use-once streaming | |
1835 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1836 | * so we ignore them here. | |
1837 | */ | |
41e20983 | 1838 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1839 | list_add(&page->lru, &l_active); |
1840 | continue; | |
1841 | } | |
1842 | } | |
7e9cd484 | 1843 | |
5205e56e | 1844 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1845 | list_add(&page->lru, &l_inactive); |
1846 | } | |
1847 | ||
b555749a | 1848 | /* |
8cab4754 | 1849 | * Move pages back to the lru list. |
b555749a | 1850 | */ |
2a1dc509 | 1851 | spin_lock_irq(&zone->lru_lock); |
556adecb | 1852 | /* |
8cab4754 WF |
1853 | * Count referenced pages from currently used mappings as rotated, |
1854 | * even though only some of them are actually re-activated. This | |
1855 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 1856 | * get_scan_count. |
7e9cd484 | 1857 | */ |
b7c46d15 | 1858 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1859 | |
fa9add64 HD |
1860 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
1861 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
a731286d | 1862 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
f8891e5e | 1863 | spin_unlock_irq(&zone->lru_lock); |
2bcf8879 | 1864 | |
747db954 | 1865 | mem_cgroup_uncharge_list(&l_hold); |
b745bc85 | 1866 | free_hot_cold_page_list(&l_hold, true); |
1da177e4 LT |
1867 | } |
1868 | ||
74e3f3c3 | 1869 | #ifdef CONFIG_SWAP |
42e2e457 | 1870 | static bool inactive_anon_is_low_global(struct zone *zone) |
f89eb90e KM |
1871 | { |
1872 | unsigned long active, inactive; | |
1873 | ||
1874 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1875 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1876 | ||
29d06bbb | 1877 | return inactive * zone->inactive_ratio < active; |
f89eb90e KM |
1878 | } |
1879 | ||
14797e23 KM |
1880 | /** |
1881 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
c56d5c7d | 1882 | * @lruvec: LRU vector to check |
14797e23 KM |
1883 | * |
1884 | * Returns true if the zone does not have enough inactive anon pages, | |
1885 | * meaning some active anon pages need to be deactivated. | |
1886 | */ | |
42e2e457 | 1887 | static bool inactive_anon_is_low(struct lruvec *lruvec) |
14797e23 | 1888 | { |
74e3f3c3 MK |
1889 | /* |
1890 | * If we don't have swap space, anonymous page deactivation | |
1891 | * is pointless. | |
1892 | */ | |
1893 | if (!total_swap_pages) | |
42e2e457 | 1894 | return false; |
74e3f3c3 | 1895 | |
c3c787e8 | 1896 | if (!mem_cgroup_disabled()) |
c56d5c7d | 1897 | return mem_cgroup_inactive_anon_is_low(lruvec); |
f16015fb | 1898 | |
c56d5c7d | 1899 | return inactive_anon_is_low_global(lruvec_zone(lruvec)); |
14797e23 | 1900 | } |
74e3f3c3 | 1901 | #else |
42e2e457 | 1902 | static inline bool inactive_anon_is_low(struct lruvec *lruvec) |
74e3f3c3 | 1903 | { |
42e2e457 | 1904 | return false; |
74e3f3c3 MK |
1905 | } |
1906 | #endif | |
14797e23 | 1907 | |
56e49d21 RR |
1908 | /** |
1909 | * inactive_file_is_low - check if file pages need to be deactivated | |
c56d5c7d | 1910 | * @lruvec: LRU vector to check |
56e49d21 RR |
1911 | * |
1912 | * When the system is doing streaming IO, memory pressure here | |
1913 | * ensures that active file pages get deactivated, until more | |
1914 | * than half of the file pages are on the inactive list. | |
1915 | * | |
1916 | * Once we get to that situation, protect the system's working | |
1917 | * set from being evicted by disabling active file page aging. | |
1918 | * | |
1919 | * This uses a different ratio than the anonymous pages, because | |
1920 | * the page cache uses a use-once replacement algorithm. | |
1921 | */ | |
42e2e457 | 1922 | static bool inactive_file_is_low(struct lruvec *lruvec) |
56e49d21 | 1923 | { |
e3790144 JW |
1924 | unsigned long inactive; |
1925 | unsigned long active; | |
1926 | ||
1927 | inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
1928 | active = get_lru_size(lruvec, LRU_ACTIVE_FILE); | |
56e49d21 | 1929 | |
e3790144 | 1930 | return active > inactive; |
56e49d21 RR |
1931 | } |
1932 | ||
42e2e457 | 1933 | static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) |
b39415b2 | 1934 | { |
75b00af7 | 1935 | if (is_file_lru(lru)) |
c56d5c7d | 1936 | return inactive_file_is_low(lruvec); |
b39415b2 | 1937 | else |
c56d5c7d | 1938 | return inactive_anon_is_low(lruvec); |
b39415b2 RR |
1939 | } |
1940 | ||
4f98a2fe | 1941 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
1a93be0e | 1942 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 1943 | { |
b39415b2 | 1944 | if (is_active_lru(lru)) { |
75b00af7 | 1945 | if (inactive_list_is_low(lruvec, lru)) |
1a93be0e | 1946 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
1947 | return 0; |
1948 | } | |
1949 | ||
1a93be0e | 1950 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
1951 | } |
1952 | ||
9a265114 JW |
1953 | enum scan_balance { |
1954 | SCAN_EQUAL, | |
1955 | SCAN_FRACT, | |
1956 | SCAN_ANON, | |
1957 | SCAN_FILE, | |
1958 | }; | |
1959 | ||
4f98a2fe RR |
1960 | /* |
1961 | * Determine how aggressively the anon and file LRU lists should be | |
1962 | * scanned. The relative value of each set of LRU lists is determined | |
1963 | * by looking at the fraction of the pages scanned we did rotate back | |
1964 | * onto the active list instead of evict. | |
1965 | * | |
be7bd59d WL |
1966 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
1967 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 1968 | */ |
02695175 | 1969 | static void get_scan_count(struct lruvec *lruvec, int swappiness, |
6b4f7799 JW |
1970 | struct scan_control *sc, unsigned long *nr, |
1971 | unsigned long *lru_pages) | |
4f98a2fe | 1972 | { |
9a265114 JW |
1973 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1974 | u64 fraction[2]; | |
1975 | u64 denominator = 0; /* gcc */ | |
1976 | struct zone *zone = lruvec_zone(lruvec); | |
4f98a2fe | 1977 | unsigned long anon_prio, file_prio; |
9a265114 | 1978 | enum scan_balance scan_balance; |
0bf1457f | 1979 | unsigned long anon, file; |
9a265114 | 1980 | bool force_scan = false; |
4f98a2fe | 1981 | unsigned long ap, fp; |
4111304d | 1982 | enum lru_list lru; |
6f04f48d SS |
1983 | bool some_scanned; |
1984 | int pass; | |
246e87a9 | 1985 | |
f11c0ca5 JW |
1986 | /* |
1987 | * If the zone or memcg is small, nr[l] can be 0. This | |
1988 | * results in no scanning on this priority and a potential | |
1989 | * priority drop. Global direct reclaim can go to the next | |
1990 | * zone and tends to have no problems. Global kswapd is for | |
1991 | * zone balancing and it needs to scan a minimum amount. When | |
1992 | * reclaiming for a memcg, a priority drop can cause high | |
1993 | * latencies, so it's better to scan a minimum amount there as | |
1994 | * well. | |
1995 | */ | |
90cbc250 VD |
1996 | if (current_is_kswapd()) { |
1997 | if (!zone_reclaimable(zone)) | |
1998 | force_scan = true; | |
1999 | if (!mem_cgroup_lruvec_online(lruvec)) | |
2000 | force_scan = true; | |
2001 | } | |
89b5fae5 | 2002 | if (!global_reclaim(sc)) |
a4d3e9e7 | 2003 | force_scan = true; |
76a33fc3 SL |
2004 | |
2005 | /* If we have no swap space, do not bother scanning anon pages. */ | |
ec8acf20 | 2006 | if (!sc->may_swap || (get_nr_swap_pages() <= 0)) { |
9a265114 | 2007 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2008 | goto out; |
2009 | } | |
4f98a2fe | 2010 | |
10316b31 JW |
2011 | /* |
2012 | * Global reclaim will swap to prevent OOM even with no | |
2013 | * swappiness, but memcg users want to use this knob to | |
2014 | * disable swapping for individual groups completely when | |
2015 | * using the memory controller's swap limit feature would be | |
2016 | * too expensive. | |
2017 | */ | |
02695175 | 2018 | if (!global_reclaim(sc) && !swappiness) { |
9a265114 | 2019 | scan_balance = SCAN_FILE; |
10316b31 JW |
2020 | goto out; |
2021 | } | |
2022 | ||
2023 | /* | |
2024 | * Do not apply any pressure balancing cleverness when the | |
2025 | * system is close to OOM, scan both anon and file equally | |
2026 | * (unless the swappiness setting disagrees with swapping). | |
2027 | */ | |
02695175 | 2028 | if (!sc->priority && swappiness) { |
9a265114 | 2029 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2030 | goto out; |
2031 | } | |
2032 | ||
62376251 JW |
2033 | /* |
2034 | * Prevent the reclaimer from falling into the cache trap: as | |
2035 | * cache pages start out inactive, every cache fault will tip | |
2036 | * the scan balance towards the file LRU. And as the file LRU | |
2037 | * shrinks, so does the window for rotation from references. | |
2038 | * This means we have a runaway feedback loop where a tiny | |
2039 | * thrashing file LRU becomes infinitely more attractive than | |
2040 | * anon pages. Try to detect this based on file LRU size. | |
2041 | */ | |
2042 | if (global_reclaim(sc)) { | |
2ab051e1 JM |
2043 | unsigned long zonefile; |
2044 | unsigned long zonefree; | |
2045 | ||
2046 | zonefree = zone_page_state(zone, NR_FREE_PAGES); | |
2047 | zonefile = zone_page_state(zone, NR_ACTIVE_FILE) + | |
2048 | zone_page_state(zone, NR_INACTIVE_FILE); | |
62376251 | 2049 | |
2ab051e1 | 2050 | if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) { |
62376251 JW |
2051 | scan_balance = SCAN_ANON; |
2052 | goto out; | |
2053 | } | |
2054 | } | |
2055 | ||
7c5bd705 JW |
2056 | /* |
2057 | * There is enough inactive page cache, do not reclaim | |
2058 | * anything from the anonymous working set right now. | |
2059 | */ | |
2060 | if (!inactive_file_is_low(lruvec)) { | |
9a265114 | 2061 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2062 | goto out; |
2063 | } | |
2064 | ||
9a265114 JW |
2065 | scan_balance = SCAN_FRACT; |
2066 | ||
58c37f6e KM |
2067 | /* |
2068 | * With swappiness at 100, anonymous and file have the same priority. | |
2069 | * This scanning priority is essentially the inverse of IO cost. | |
2070 | */ | |
02695175 | 2071 | anon_prio = swappiness; |
75b00af7 | 2072 | file_prio = 200 - anon_prio; |
58c37f6e | 2073 | |
4f98a2fe RR |
2074 | /* |
2075 | * OK, so we have swap space and a fair amount of page cache | |
2076 | * pages. We use the recently rotated / recently scanned | |
2077 | * ratios to determine how valuable each cache is. | |
2078 | * | |
2079 | * Because workloads change over time (and to avoid overflow) | |
2080 | * we keep these statistics as a floating average, which ends | |
2081 | * up weighing recent references more than old ones. | |
2082 | * | |
2083 | * anon in [0], file in [1] | |
2084 | */ | |
2ab051e1 JM |
2085 | |
2086 | anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + | |
2087 | get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
2088 | file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
2089 | get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
2090 | ||
90126375 | 2091 | spin_lock_irq(&zone->lru_lock); |
6e901571 | 2092 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2093 | reclaim_stat->recent_scanned[0] /= 2; |
2094 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2095 | } |
2096 | ||
6e901571 | 2097 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2098 | reclaim_stat->recent_scanned[1] /= 2; |
2099 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2100 | } |
2101 | ||
4f98a2fe | 2102 | /* |
00d8089c RR |
2103 | * The amount of pressure on anon vs file pages is inversely |
2104 | * proportional to the fraction of recently scanned pages on | |
2105 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2106 | */ |
fe35004f | 2107 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2108 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2109 | |
fe35004f | 2110 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2111 | fp /= reclaim_stat->recent_rotated[1] + 1; |
90126375 | 2112 | spin_unlock_irq(&zone->lru_lock); |
4f98a2fe | 2113 | |
76a33fc3 SL |
2114 | fraction[0] = ap; |
2115 | fraction[1] = fp; | |
2116 | denominator = ap + fp + 1; | |
2117 | out: | |
6f04f48d SS |
2118 | some_scanned = false; |
2119 | /* Only use force_scan on second pass. */ | |
2120 | for (pass = 0; !some_scanned && pass < 2; pass++) { | |
6b4f7799 | 2121 | *lru_pages = 0; |
6f04f48d SS |
2122 | for_each_evictable_lru(lru) { |
2123 | int file = is_file_lru(lru); | |
2124 | unsigned long size; | |
2125 | unsigned long scan; | |
6e08a369 | 2126 | |
6f04f48d SS |
2127 | size = get_lru_size(lruvec, lru); |
2128 | scan = size >> sc->priority; | |
9a265114 | 2129 | |
6f04f48d SS |
2130 | if (!scan && pass && force_scan) |
2131 | scan = min(size, SWAP_CLUSTER_MAX); | |
9a265114 | 2132 | |
6f04f48d SS |
2133 | switch (scan_balance) { |
2134 | case SCAN_EQUAL: | |
2135 | /* Scan lists relative to size */ | |
2136 | break; | |
2137 | case SCAN_FRACT: | |
2138 | /* | |
2139 | * Scan types proportional to swappiness and | |
2140 | * their relative recent reclaim efficiency. | |
2141 | */ | |
2142 | scan = div64_u64(scan * fraction[file], | |
2143 | denominator); | |
2144 | break; | |
2145 | case SCAN_FILE: | |
2146 | case SCAN_ANON: | |
2147 | /* Scan one type exclusively */ | |
6b4f7799 JW |
2148 | if ((scan_balance == SCAN_FILE) != file) { |
2149 | size = 0; | |
6f04f48d | 2150 | scan = 0; |
6b4f7799 | 2151 | } |
6f04f48d SS |
2152 | break; |
2153 | default: | |
2154 | /* Look ma, no brain */ | |
2155 | BUG(); | |
2156 | } | |
6b4f7799 JW |
2157 | |
2158 | *lru_pages += size; | |
6f04f48d | 2159 | nr[lru] = scan; |
6b4f7799 | 2160 | |
9a265114 | 2161 | /* |
6f04f48d SS |
2162 | * Skip the second pass and don't force_scan, |
2163 | * if we found something to scan. | |
9a265114 | 2164 | */ |
6f04f48d | 2165 | some_scanned |= !!scan; |
9a265114 | 2166 | } |
76a33fc3 | 2167 | } |
6e08a369 | 2168 | } |
4f98a2fe | 2169 | |
9b4f98cd JW |
2170 | /* |
2171 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
2172 | */ | |
02695175 | 2173 | static void shrink_lruvec(struct lruvec *lruvec, int swappiness, |
6b4f7799 | 2174 | struct scan_control *sc, unsigned long *lru_pages) |
9b4f98cd JW |
2175 | { |
2176 | unsigned long nr[NR_LRU_LISTS]; | |
e82e0561 | 2177 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2178 | unsigned long nr_to_scan; |
2179 | enum lru_list lru; | |
2180 | unsigned long nr_reclaimed = 0; | |
2181 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2182 | struct blk_plug plug; | |
1a501907 | 2183 | bool scan_adjusted; |
9b4f98cd | 2184 | |
6b4f7799 | 2185 | get_scan_count(lruvec, swappiness, sc, nr, lru_pages); |
9b4f98cd | 2186 | |
e82e0561 MG |
2187 | /* Record the original scan target for proportional adjustments later */ |
2188 | memcpy(targets, nr, sizeof(nr)); | |
2189 | ||
1a501907 MG |
2190 | /* |
2191 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2192 | * event that can occur when there is little memory pressure e.g. | |
2193 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2194 | * when the requested number of pages are reclaimed when scanning at | |
2195 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2196 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2197 | * do a batch of work at once. For memcg reclaim one check is made to | |
2198 | * abort proportional reclaim if either the file or anon lru has already | |
2199 | * dropped to zero at the first pass. | |
2200 | */ | |
2201 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2202 | sc->priority == DEF_PRIORITY); | |
2203 | ||
9b4f98cd JW |
2204 | blk_start_plug(&plug); |
2205 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2206 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2207 | unsigned long nr_anon, nr_file, percentage; |
2208 | unsigned long nr_scanned; | |
2209 | ||
9b4f98cd JW |
2210 | for_each_evictable_lru(lru) { |
2211 | if (nr[lru]) { | |
2212 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2213 | nr[lru] -= nr_to_scan; | |
2214 | ||
2215 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2216 | lruvec, sc); | |
2217 | } | |
2218 | } | |
e82e0561 MG |
2219 | |
2220 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) | |
2221 | continue; | |
2222 | ||
e82e0561 MG |
2223 | /* |
2224 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2225 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2226 | * proportionally what was requested by get_scan_count(). We |
2227 | * stop reclaiming one LRU and reduce the amount scanning | |
2228 | * proportional to the original scan target. | |
2229 | */ | |
2230 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2231 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2232 | ||
1a501907 MG |
2233 | /* |
2234 | * It's just vindictive to attack the larger once the smaller | |
2235 | * has gone to zero. And given the way we stop scanning the | |
2236 | * smaller below, this makes sure that we only make one nudge | |
2237 | * towards proportionality once we've got nr_to_reclaim. | |
2238 | */ | |
2239 | if (!nr_file || !nr_anon) | |
2240 | break; | |
2241 | ||
e82e0561 MG |
2242 | if (nr_file > nr_anon) { |
2243 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2244 | targets[LRU_ACTIVE_ANON] + 1; | |
2245 | lru = LRU_BASE; | |
2246 | percentage = nr_anon * 100 / scan_target; | |
2247 | } else { | |
2248 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2249 | targets[LRU_ACTIVE_FILE] + 1; | |
2250 | lru = LRU_FILE; | |
2251 | percentage = nr_file * 100 / scan_target; | |
2252 | } | |
2253 | ||
2254 | /* Stop scanning the smaller of the LRU */ | |
2255 | nr[lru] = 0; | |
2256 | nr[lru + LRU_ACTIVE] = 0; | |
2257 | ||
2258 | /* | |
2259 | * Recalculate the other LRU scan count based on its original | |
2260 | * scan target and the percentage scanning already complete | |
2261 | */ | |
2262 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2263 | nr_scanned = targets[lru] - nr[lru]; | |
2264 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2265 | nr[lru] -= min(nr[lru], nr_scanned); | |
2266 | ||
2267 | lru += LRU_ACTIVE; | |
2268 | nr_scanned = targets[lru] - nr[lru]; | |
2269 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2270 | nr[lru] -= min(nr[lru], nr_scanned); | |
2271 | ||
2272 | scan_adjusted = true; | |
9b4f98cd JW |
2273 | } |
2274 | blk_finish_plug(&plug); | |
2275 | sc->nr_reclaimed += nr_reclaimed; | |
2276 | ||
2277 | /* | |
2278 | * Even if we did not try to evict anon pages at all, we want to | |
2279 | * rebalance the anon lru active/inactive ratio. | |
2280 | */ | |
2281 | if (inactive_anon_is_low(lruvec)) | |
2282 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
2283 | sc, LRU_ACTIVE_ANON); | |
2284 | ||
2285 | throttle_vm_writeout(sc->gfp_mask); | |
2286 | } | |
2287 | ||
23b9da55 | 2288 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2289 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2290 | { |
d84da3f9 | 2291 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2292 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2293 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2294 | return true; |
2295 | ||
2296 | return false; | |
2297 | } | |
2298 | ||
3e7d3449 | 2299 | /* |
23b9da55 MG |
2300 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2301 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2302 | * true if more pages should be reclaimed such that when the page allocator | |
2303 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2304 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2305 | */ |
9b4f98cd | 2306 | static inline bool should_continue_reclaim(struct zone *zone, |
3e7d3449 MG |
2307 | unsigned long nr_reclaimed, |
2308 | unsigned long nr_scanned, | |
2309 | struct scan_control *sc) | |
2310 | { | |
2311 | unsigned long pages_for_compaction; | |
2312 | unsigned long inactive_lru_pages; | |
2313 | ||
2314 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2315 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2316 | return false; |
2317 | ||
2876592f MG |
2318 | /* Consider stopping depending on scan and reclaim activity */ |
2319 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2320 | /* | |
2321 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2322 | * full LRU list has been scanned and we are still failing | |
2323 | * to reclaim pages. This full LRU scan is potentially | |
2324 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2325 | */ | |
2326 | if (!nr_reclaimed && !nr_scanned) | |
2327 | return false; | |
2328 | } else { | |
2329 | /* | |
2330 | * For non-__GFP_REPEAT allocations which can presumably | |
2331 | * fail without consequence, stop if we failed to reclaim | |
2332 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2333 | * pages that were scanned. This will return to the | |
2334 | * caller faster at the risk reclaim/compaction and | |
2335 | * the resulting allocation attempt fails | |
2336 | */ | |
2337 | if (!nr_reclaimed) | |
2338 | return false; | |
2339 | } | |
3e7d3449 MG |
2340 | |
2341 | /* | |
2342 | * If we have not reclaimed enough pages for compaction and the | |
2343 | * inactive lists are large enough, continue reclaiming | |
2344 | */ | |
2345 | pages_for_compaction = (2UL << sc->order); | |
9b4f98cd | 2346 | inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); |
ec8acf20 | 2347 | if (get_nr_swap_pages() > 0) |
9b4f98cd | 2348 | inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); |
3e7d3449 MG |
2349 | if (sc->nr_reclaimed < pages_for_compaction && |
2350 | inactive_lru_pages > pages_for_compaction) | |
2351 | return true; | |
2352 | ||
2353 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
ebff3980 | 2354 | switch (compaction_suitable(zone, sc->order, 0, 0)) { |
3e7d3449 MG |
2355 | case COMPACT_PARTIAL: |
2356 | case COMPACT_CONTINUE: | |
2357 | return false; | |
2358 | default: | |
2359 | return true; | |
2360 | } | |
2361 | } | |
2362 | ||
6b4f7799 JW |
2363 | static bool shrink_zone(struct zone *zone, struct scan_control *sc, |
2364 | bool is_classzone) | |
1da177e4 | 2365 | { |
cb731d6c | 2366 | struct reclaim_state *reclaim_state = current->reclaim_state; |
f0fdc5e8 | 2367 | unsigned long nr_reclaimed, nr_scanned; |
2344d7e4 | 2368 | bool reclaimable = false; |
1da177e4 | 2369 | |
9b4f98cd JW |
2370 | do { |
2371 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2372 | struct mem_cgroup_reclaim_cookie reclaim = { | |
2373 | .zone = zone, | |
2374 | .priority = sc->priority, | |
2375 | }; | |
6b4f7799 | 2376 | unsigned long zone_lru_pages = 0; |
694fbc0f | 2377 | struct mem_cgroup *memcg; |
3e7d3449 | 2378 | |
9b4f98cd JW |
2379 | nr_reclaimed = sc->nr_reclaimed; |
2380 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2381 | |
694fbc0f AM |
2382 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2383 | do { | |
6b4f7799 | 2384 | unsigned long lru_pages; |
cb731d6c | 2385 | unsigned long scanned; |
9b4f98cd | 2386 | struct lruvec *lruvec; |
02695175 | 2387 | int swappiness; |
5660048c | 2388 | |
241994ed JW |
2389 | if (mem_cgroup_low(root, memcg)) { |
2390 | if (!sc->may_thrash) | |
2391 | continue; | |
2392 | mem_cgroup_events(memcg, MEMCG_LOW, 1); | |
2393 | } | |
2394 | ||
9b4f98cd | 2395 | lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
02695175 | 2396 | swappiness = mem_cgroup_swappiness(memcg); |
cb731d6c | 2397 | scanned = sc->nr_scanned; |
f9be23d6 | 2398 | |
6b4f7799 JW |
2399 | shrink_lruvec(lruvec, swappiness, sc, &lru_pages); |
2400 | zone_lru_pages += lru_pages; | |
f16015fb | 2401 | |
cb731d6c VD |
2402 | if (memcg && is_classzone) |
2403 | shrink_slab(sc->gfp_mask, zone_to_nid(zone), | |
2404 | memcg, sc->nr_scanned - scanned, | |
2405 | lru_pages); | |
2406 | ||
9b4f98cd | 2407 | /* |
a394cb8e MH |
2408 | * Direct reclaim and kswapd have to scan all memory |
2409 | * cgroups to fulfill the overall scan target for the | |
9b4f98cd | 2410 | * zone. |
a394cb8e MH |
2411 | * |
2412 | * Limit reclaim, on the other hand, only cares about | |
2413 | * nr_to_reclaim pages to be reclaimed and it will | |
2414 | * retry with decreasing priority if one round over the | |
2415 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2416 | */ |
a394cb8e MH |
2417 | if (!global_reclaim(sc) && |
2418 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2419 | mem_cgroup_iter_break(root, memcg); |
2420 | break; | |
2421 | } | |
241994ed | 2422 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
70ddf637 | 2423 | |
6b4f7799 JW |
2424 | /* |
2425 | * Shrink the slab caches in the same proportion that | |
2426 | * the eligible LRU pages were scanned. | |
2427 | */ | |
cb731d6c VD |
2428 | if (global_reclaim(sc) && is_classzone) |
2429 | shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL, | |
2430 | sc->nr_scanned - nr_scanned, | |
2431 | zone_lru_pages); | |
2432 | ||
2433 | if (reclaim_state) { | |
2434 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2435 | reclaim_state->reclaimed_slab = 0; | |
6b4f7799 JW |
2436 | } |
2437 | ||
70ddf637 AV |
2438 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, |
2439 | sc->nr_scanned - nr_scanned, | |
2440 | sc->nr_reclaimed - nr_reclaimed); | |
2441 | ||
2344d7e4 JW |
2442 | if (sc->nr_reclaimed - nr_reclaimed) |
2443 | reclaimable = true; | |
2444 | ||
9b4f98cd JW |
2445 | } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, |
2446 | sc->nr_scanned - nr_scanned, sc)); | |
2344d7e4 JW |
2447 | |
2448 | return reclaimable; | |
f16015fb JW |
2449 | } |
2450 | ||
53853e2d VB |
2451 | /* |
2452 | * Returns true if compaction should go ahead for a high-order request, or | |
2453 | * the high-order allocation would succeed without compaction. | |
2454 | */ | |
0b06496a | 2455 | static inline bool compaction_ready(struct zone *zone, int order) |
fe4b1b24 MG |
2456 | { |
2457 | unsigned long balance_gap, watermark; | |
2458 | bool watermark_ok; | |
2459 | ||
fe4b1b24 MG |
2460 | /* |
2461 | * Compaction takes time to run and there are potentially other | |
2462 | * callers using the pages just freed. Continue reclaiming until | |
2463 | * there is a buffer of free pages available to give compaction | |
2464 | * a reasonable chance of completing and allocating the page | |
2465 | */ | |
4be89a34 JZ |
2466 | balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( |
2467 | zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); | |
0b06496a | 2468 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << order); |
e2b19197 | 2469 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0); |
fe4b1b24 MG |
2470 | |
2471 | /* | |
2472 | * If compaction is deferred, reclaim up to a point where | |
2473 | * compaction will have a chance of success when re-enabled | |
2474 | */ | |
0b06496a | 2475 | if (compaction_deferred(zone, order)) |
fe4b1b24 MG |
2476 | return watermark_ok; |
2477 | ||
53853e2d VB |
2478 | /* |
2479 | * If compaction is not ready to start and allocation is not likely | |
2480 | * to succeed without it, then keep reclaiming. | |
2481 | */ | |
ebff3980 | 2482 | if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED) |
fe4b1b24 MG |
2483 | return false; |
2484 | ||
2485 | return watermark_ok; | |
2486 | } | |
2487 | ||
1da177e4 LT |
2488 | /* |
2489 | * This is the direct reclaim path, for page-allocating processes. We only | |
2490 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2491 | * request. | |
2492 | * | |
41858966 MG |
2493 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
2494 | * Because: | |
1da177e4 LT |
2495 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
2496 | * allocation or | |
41858966 MG |
2497 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
2498 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2499 | * zone defense algorithm. | |
1da177e4 | 2500 | * |
1da177e4 LT |
2501 | * If a zone is deemed to be full of pinned pages then just give it a light |
2502 | * scan then give up on it. | |
2503 | */ | |
c630ec12 | 2504 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2505 | { |
dd1a239f | 2506 | struct zoneref *z; |
54a6eb5c | 2507 | struct zone *zone; |
0608f43d AM |
2508 | unsigned long nr_soft_reclaimed; |
2509 | unsigned long nr_soft_scanned; | |
619d0d76 | 2510 | gfp_t orig_mask; |
9bbc04ee | 2511 | enum zone_type requested_highidx = gfp_zone(sc->gfp_mask); |
1cfb419b | 2512 | |
cc715d99 MG |
2513 | /* |
2514 | * If the number of buffer_heads in the machine exceeds the maximum | |
2515 | * allowed level, force direct reclaim to scan the highmem zone as | |
2516 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2517 | */ | |
619d0d76 | 2518 | orig_mask = sc->gfp_mask; |
cc715d99 MG |
2519 | if (buffer_heads_over_limit) |
2520 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2521 | ||
d4debc66 | 2522 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
5fa47439 | 2523 | gfp_zone(sc->gfp_mask), sc->nodemask) { |
6b4f7799 JW |
2524 | enum zone_type classzone_idx; |
2525 | ||
f3fe6512 | 2526 | if (!populated_zone(zone)) |
1da177e4 | 2527 | continue; |
6b4f7799 | 2528 | |
1c2cb085 | 2529 | classzone_idx = gfp_zone(sc->gfp_mask); |
6b4f7799 JW |
2530 | while (!populated_zone(zone->zone_pgdat->node_zones + |
2531 | classzone_idx)) | |
2532 | classzone_idx--; | |
2533 | ||
1cfb419b KH |
2534 | /* |
2535 | * Take care memory controller reclaiming has small influence | |
2536 | * to global LRU. | |
2537 | */ | |
89b5fae5 | 2538 | if (global_reclaim(sc)) { |
344736f2 VD |
2539 | if (!cpuset_zone_allowed(zone, |
2540 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2541 | continue; |
65ec02cb | 2542 | |
6e543d57 LD |
2543 | if (sc->priority != DEF_PRIORITY && |
2544 | !zone_reclaimable(zone)) | |
1cfb419b | 2545 | continue; /* Let kswapd poll it */ |
0b06496a JW |
2546 | |
2547 | /* | |
2548 | * If we already have plenty of memory free for | |
2549 | * compaction in this zone, don't free any more. | |
2550 | * Even though compaction is invoked for any | |
2551 | * non-zero order, only frequent costly order | |
2552 | * reclamation is disruptive enough to become a | |
2553 | * noticeable problem, like transparent huge | |
2554 | * page allocations. | |
2555 | */ | |
2556 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2557 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
2558 | zonelist_zone_idx(z) <= requested_highidx && | |
2559 | compaction_ready(zone, sc->order)) { | |
2560 | sc->compaction_ready = true; | |
2561 | continue; | |
e0887c19 | 2562 | } |
0b06496a | 2563 | |
0608f43d AM |
2564 | /* |
2565 | * This steals pages from memory cgroups over softlimit | |
2566 | * and returns the number of reclaimed pages and | |
2567 | * scanned pages. This works for global memory pressure | |
2568 | * and balancing, not for a memcg's limit. | |
2569 | */ | |
2570 | nr_soft_scanned = 0; | |
2571 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2572 | sc->order, sc->gfp_mask, | |
2573 | &nr_soft_scanned); | |
2574 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2575 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2576 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2577 | } |
408d8544 | 2578 | |
c630ec12 | 2579 | shrink_zone(zone, sc, zone_idx(zone) == classzone_idx); |
1da177e4 | 2580 | } |
e0c23279 | 2581 | |
619d0d76 WY |
2582 | /* |
2583 | * Restore to original mask to avoid the impact on the caller if we | |
2584 | * promoted it to __GFP_HIGHMEM. | |
2585 | */ | |
2586 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2587 | } |
4f98a2fe | 2588 | |
1da177e4 LT |
2589 | /* |
2590 | * This is the main entry point to direct page reclaim. | |
2591 | * | |
2592 | * If a full scan of the inactive list fails to free enough memory then we | |
2593 | * are "out of memory" and something needs to be killed. | |
2594 | * | |
2595 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2596 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2597 | * caller can't do much about. We kick the writeback threads and take explicit |
2598 | * naps in the hope that some of these pages can be written. But if the | |
2599 | * allocating task holds filesystem locks which prevent writeout this might not | |
2600 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2601 | * |
2602 | * returns: 0, if no pages reclaimed | |
2603 | * else, the number of pages reclaimed | |
1da177e4 | 2604 | */ |
dac1d27b | 2605 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2606 | struct scan_control *sc) |
1da177e4 | 2607 | { |
241994ed | 2608 | int initial_priority = sc->priority; |
69e05944 | 2609 | unsigned long total_scanned = 0; |
22fba335 | 2610 | unsigned long writeback_threshold; |
241994ed | 2611 | retry: |
873b4771 KK |
2612 | delayacct_freepages_start(); |
2613 | ||
89b5fae5 | 2614 | if (global_reclaim(sc)) |
1cfb419b | 2615 | count_vm_event(ALLOCSTALL); |
1da177e4 | 2616 | |
9e3b2f8c | 2617 | do { |
70ddf637 AV |
2618 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2619 | sc->priority); | |
66e1707b | 2620 | sc->nr_scanned = 0; |
c630ec12 | 2621 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 2622 | |
66e1707b | 2623 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2624 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
2625 | break; |
2626 | ||
2627 | if (sc->compaction_ready) | |
2628 | break; | |
1da177e4 | 2629 | |
0e50ce3b MK |
2630 | /* |
2631 | * If we're getting trouble reclaiming, start doing | |
2632 | * writepage even in laptop mode. | |
2633 | */ | |
2634 | if (sc->priority < DEF_PRIORITY - 2) | |
2635 | sc->may_writepage = 1; | |
2636 | ||
1da177e4 LT |
2637 | /* |
2638 | * Try to write back as many pages as we just scanned. This | |
2639 | * tends to cause slow streaming writers to write data to the | |
2640 | * disk smoothly, at the dirtying rate, which is nice. But | |
2641 | * that's undesirable in laptop mode, where we *want* lumpy | |
2642 | * writeout. So in laptop mode, write out the whole world. | |
2643 | */ | |
22fba335 KM |
2644 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2645 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2646 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2647 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2648 | sc->may_writepage = 1; |
1da177e4 | 2649 | } |
0b06496a | 2650 | } while (--sc->priority >= 0); |
bb21c7ce | 2651 | |
873b4771 KK |
2652 | delayacct_freepages_end(); |
2653 | ||
bb21c7ce KM |
2654 | if (sc->nr_reclaimed) |
2655 | return sc->nr_reclaimed; | |
2656 | ||
0cee34fd | 2657 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 2658 | if (sc->compaction_ready) |
7335084d MG |
2659 | return 1; |
2660 | ||
241994ed JW |
2661 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
2662 | if (!sc->may_thrash) { | |
2663 | sc->priority = initial_priority; | |
2664 | sc->may_thrash = 1; | |
2665 | goto retry; | |
2666 | } | |
2667 | ||
bb21c7ce | 2668 | return 0; |
1da177e4 LT |
2669 | } |
2670 | ||
5515061d MG |
2671 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) |
2672 | { | |
2673 | struct zone *zone; | |
2674 | unsigned long pfmemalloc_reserve = 0; | |
2675 | unsigned long free_pages = 0; | |
2676 | int i; | |
2677 | bool wmark_ok; | |
2678 | ||
2679 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2680 | zone = &pgdat->node_zones[i]; | |
f012a84a NA |
2681 | if (!populated_zone(zone) || |
2682 | zone_reclaimable_pages(zone) == 0) | |
675becce MG |
2683 | continue; |
2684 | ||
5515061d MG |
2685 | pfmemalloc_reserve += min_wmark_pages(zone); |
2686 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2687 | } | |
2688 | ||
675becce MG |
2689 | /* If there are no reserves (unexpected config) then do not throttle */ |
2690 | if (!pfmemalloc_reserve) | |
2691 | return true; | |
2692 | ||
5515061d MG |
2693 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
2694 | ||
2695 | /* kswapd must be awake if processes are being throttled */ | |
2696 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2697 | pgdat->classzone_idx = min(pgdat->classzone_idx, | |
2698 | (enum zone_type)ZONE_NORMAL); | |
2699 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2700 | } | |
2701 | ||
2702 | return wmark_ok; | |
2703 | } | |
2704 | ||
2705 | /* | |
2706 | * Throttle direct reclaimers if backing storage is backed by the network | |
2707 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2708 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2709 | * when the low watermark is reached. |
2710 | * | |
2711 | * Returns true if a fatal signal was delivered during throttling. If this | |
2712 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2713 | */ |
50694c28 | 2714 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2715 | nodemask_t *nodemask) |
2716 | { | |
675becce | 2717 | struct zoneref *z; |
5515061d | 2718 | struct zone *zone; |
675becce | 2719 | pg_data_t *pgdat = NULL; |
5515061d MG |
2720 | |
2721 | /* | |
2722 | * Kernel threads should not be throttled as they may be indirectly | |
2723 | * responsible for cleaning pages necessary for reclaim to make forward | |
2724 | * progress. kjournald for example may enter direct reclaim while | |
2725 | * committing a transaction where throttling it could forcing other | |
2726 | * processes to block on log_wait_commit(). | |
2727 | */ | |
2728 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2729 | goto out; |
2730 | ||
2731 | /* | |
2732 | * If a fatal signal is pending, this process should not throttle. | |
2733 | * It should return quickly so it can exit and free its memory | |
2734 | */ | |
2735 | if (fatal_signal_pending(current)) | |
2736 | goto out; | |
5515061d | 2737 | |
675becce MG |
2738 | /* |
2739 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2740 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2741 | * GFP_KERNEL will be required for allocating network buffers when | |
2742 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2743 | * | |
2744 | * Throttling is based on the first usable node and throttled processes | |
2745 | * wait on a queue until kswapd makes progress and wakes them. There | |
2746 | * is an affinity then between processes waking up and where reclaim | |
2747 | * progress has been made assuming the process wakes on the same node. | |
2748 | * More importantly, processes running on remote nodes will not compete | |
2749 | * for remote pfmemalloc reserves and processes on different nodes | |
2750 | * should make reasonable progress. | |
2751 | */ | |
2752 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 2753 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
2754 | if (zone_idx(zone) > ZONE_NORMAL) |
2755 | continue; | |
2756 | ||
2757 | /* Throttle based on the first usable node */ | |
2758 | pgdat = zone->zone_pgdat; | |
2759 | if (pfmemalloc_watermark_ok(pgdat)) | |
2760 | goto out; | |
2761 | break; | |
2762 | } | |
2763 | ||
2764 | /* If no zone was usable by the allocation flags then do not throttle */ | |
2765 | if (!pgdat) | |
50694c28 | 2766 | goto out; |
5515061d | 2767 | |
68243e76 MG |
2768 | /* Account for the throttling */ |
2769 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2770 | ||
5515061d MG |
2771 | /* |
2772 | * If the caller cannot enter the filesystem, it's possible that it | |
2773 | * is due to the caller holding an FS lock or performing a journal | |
2774 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2775 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2776 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2777 | * second before continuing. | |
2778 | */ | |
2779 | if (!(gfp_mask & __GFP_FS)) { | |
2780 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2781 | pfmemalloc_watermark_ok(pgdat), HZ); | |
50694c28 MG |
2782 | |
2783 | goto check_pending; | |
5515061d MG |
2784 | } |
2785 | ||
2786 | /* Throttle until kswapd wakes the process */ | |
2787 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2788 | pfmemalloc_watermark_ok(pgdat)); | |
50694c28 MG |
2789 | |
2790 | check_pending: | |
2791 | if (fatal_signal_pending(current)) | |
2792 | return true; | |
2793 | ||
2794 | out: | |
2795 | return false; | |
5515061d MG |
2796 | } |
2797 | ||
dac1d27b | 2798 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2799 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2800 | { |
33906bc5 | 2801 | unsigned long nr_reclaimed; |
66e1707b | 2802 | struct scan_control sc = { |
ee814fe2 | 2803 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
21caf2fc | 2804 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
ee814fe2 JW |
2805 | .order = order, |
2806 | .nodemask = nodemask, | |
2807 | .priority = DEF_PRIORITY, | |
66e1707b | 2808 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2809 | .may_unmap = 1, |
2e2e4259 | 2810 | .may_swap = 1, |
66e1707b BS |
2811 | }; |
2812 | ||
5515061d | 2813 | /* |
50694c28 MG |
2814 | * Do not enter reclaim if fatal signal was delivered while throttled. |
2815 | * 1 is returned so that the page allocator does not OOM kill at this | |
2816 | * point. | |
5515061d | 2817 | */ |
50694c28 | 2818 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) |
5515061d MG |
2819 | return 1; |
2820 | ||
33906bc5 MG |
2821 | trace_mm_vmscan_direct_reclaim_begin(order, |
2822 | sc.may_writepage, | |
2823 | gfp_mask); | |
2824 | ||
3115cd91 | 2825 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
2826 | |
2827 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2828 | ||
2829 | return nr_reclaimed; | |
66e1707b BS |
2830 | } |
2831 | ||
c255a458 | 2832 | #ifdef CONFIG_MEMCG |
66e1707b | 2833 | |
72835c86 | 2834 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, |
4e416953 | 2835 | gfp_t gfp_mask, bool noswap, |
0ae5e89c YH |
2836 | struct zone *zone, |
2837 | unsigned long *nr_scanned) | |
4e416953 BS |
2838 | { |
2839 | struct scan_control sc = { | |
b8f5c566 | 2840 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 2841 | .target_mem_cgroup = memcg, |
4e416953 BS |
2842 | .may_writepage = !laptop_mode, |
2843 | .may_unmap = 1, | |
2844 | .may_swap = !noswap, | |
4e416953 | 2845 | }; |
f9be23d6 | 2846 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
02695175 | 2847 | int swappiness = mem_cgroup_swappiness(memcg); |
6b4f7799 | 2848 | unsigned long lru_pages; |
0ae5e89c | 2849 | |
4e416953 BS |
2850 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2851 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 2852 | |
9e3b2f8c | 2853 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e KM |
2854 | sc.may_writepage, |
2855 | sc.gfp_mask); | |
2856 | ||
4e416953 BS |
2857 | /* |
2858 | * NOTE: Although we can get the priority field, using it | |
2859 | * here is not a good idea, since it limits the pages we can scan. | |
2860 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2861 | * will pick up pages from other mem cgroup's as well. We hack | |
2862 | * the priority and make it zero. | |
2863 | */ | |
6b4f7799 | 2864 | shrink_lruvec(lruvec, swappiness, &sc, &lru_pages); |
bdce6d9e KM |
2865 | |
2866 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2867 | ||
0ae5e89c | 2868 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2869 | return sc.nr_reclaimed; |
2870 | } | |
2871 | ||
72835c86 | 2872 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 2873 | unsigned long nr_pages, |
a7885eb8 | 2874 | gfp_t gfp_mask, |
b70a2a21 | 2875 | bool may_swap) |
66e1707b | 2876 | { |
4e416953 | 2877 | struct zonelist *zonelist; |
bdce6d9e | 2878 | unsigned long nr_reclaimed; |
889976db | 2879 | int nid; |
66e1707b | 2880 | struct scan_control sc = { |
b70a2a21 | 2881 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
a09ed5e0 YH |
2882 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2883 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
ee814fe2 JW |
2884 | .target_mem_cgroup = memcg, |
2885 | .priority = DEF_PRIORITY, | |
2886 | .may_writepage = !laptop_mode, | |
2887 | .may_unmap = 1, | |
b70a2a21 | 2888 | .may_swap = may_swap, |
a09ed5e0 | 2889 | }; |
66e1707b | 2890 | |
889976db YH |
2891 | /* |
2892 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2893 | * take care of from where we get pages. So the node where we start the | |
2894 | * scan does not need to be the current node. | |
2895 | */ | |
72835c86 | 2896 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2897 | |
2898 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2899 | |
2900 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2901 | sc.may_writepage, | |
2902 | sc.gfp_mask); | |
2903 | ||
dac5bdef | 2904 | current->flags |= PF_MEMALLOC; |
3115cd91 | 2905 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
dac5bdef | 2906 | current->flags &= ~PF_MEMALLOC; |
bdce6d9e KM |
2907 | |
2908 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2909 | ||
2910 | return nr_reclaimed; | |
66e1707b BS |
2911 | } |
2912 | #endif | |
2913 | ||
9e3b2f8c | 2914 | static void age_active_anon(struct zone *zone, struct scan_control *sc) |
f16015fb | 2915 | { |
b95a2f2d | 2916 | struct mem_cgroup *memcg; |
f16015fb | 2917 | |
b95a2f2d JW |
2918 | if (!total_swap_pages) |
2919 | return; | |
2920 | ||
2921 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2922 | do { | |
c56d5c7d | 2923 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
b95a2f2d | 2924 | |
c56d5c7d | 2925 | if (inactive_anon_is_low(lruvec)) |
1a93be0e | 2926 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 2927 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
2928 | |
2929 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2930 | } while (memcg); | |
f16015fb JW |
2931 | } |
2932 | ||
60cefed4 JW |
2933 | static bool zone_balanced(struct zone *zone, int order, |
2934 | unsigned long balance_gap, int classzone_idx) | |
2935 | { | |
08556c97 GG |
2936 | /* |
2937 | * if zone is so small that watermarks are the same, don't bother trying | |
2938 | * to balance; kswapd would just spin continuously trying to balance it. | |
2939 | */ | |
2940 | if (low_wmark_pages(zone) == high_wmark_pages(zone)) | |
2941 | return true; | |
2942 | ||
60cefed4 | 2943 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + |
e2b19197 | 2944 | balance_gap, classzone_idx)) |
60cefed4 JW |
2945 | return false; |
2946 | ||
ebff3980 VB |
2947 | if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone, |
2948 | order, 0, classzone_idx) == COMPACT_SKIPPED) | |
60cefed4 JW |
2949 | return false; |
2950 | ||
2951 | return true; | |
2952 | } | |
2953 | ||
1741c877 | 2954 | /* |
4ae0a48b ZC |
2955 | * pgdat_balanced() is used when checking if a node is balanced. |
2956 | * | |
2957 | * For order-0, all zones must be balanced! | |
2958 | * | |
2959 | * For high-order allocations only zones that meet watermarks and are in a | |
2960 | * zone allowed by the callers classzone_idx are added to balanced_pages. The | |
2961 | * total of balanced pages must be at least 25% of the zones allowed by | |
2962 | * classzone_idx for the node to be considered balanced. Forcing all zones to | |
2963 | * be balanced for high orders can cause excessive reclaim when there are | |
2964 | * imbalanced zones. | |
1741c877 MG |
2965 | * The choice of 25% is due to |
2966 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2967 | * reasonable sized machine | |
2968 | * o On all other machines, the top zone must be at least a reasonable | |
25985edc | 2969 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
1741c877 MG |
2970 | * would need to be at least 256M for it to be balance a whole node. |
2971 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2972 | * to balance a node on its own. These seemed like reasonable ratios. | |
2973 | */ | |
4ae0a48b | 2974 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) |
1741c877 | 2975 | { |
b40da049 | 2976 | unsigned long managed_pages = 0; |
4ae0a48b | 2977 | unsigned long balanced_pages = 0; |
1741c877 MG |
2978 | int i; |
2979 | ||
4ae0a48b ZC |
2980 | /* Check the watermark levels */ |
2981 | for (i = 0; i <= classzone_idx; i++) { | |
2982 | struct zone *zone = pgdat->node_zones + i; | |
1741c877 | 2983 | |
4ae0a48b ZC |
2984 | if (!populated_zone(zone)) |
2985 | continue; | |
2986 | ||
b40da049 | 2987 | managed_pages += zone->managed_pages; |
4ae0a48b ZC |
2988 | |
2989 | /* | |
2990 | * A special case here: | |
2991 | * | |
2992 | * balance_pgdat() skips over all_unreclaimable after | |
2993 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2994 | * they must be considered balanced here as well! | |
2995 | */ | |
6e543d57 | 2996 | if (!zone_reclaimable(zone)) { |
b40da049 | 2997 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
2998 | continue; |
2999 | } | |
3000 | ||
3001 | if (zone_balanced(zone, order, 0, i)) | |
b40da049 | 3002 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
3003 | else if (!order) |
3004 | return false; | |
3005 | } | |
3006 | ||
3007 | if (order) | |
b40da049 | 3008 | return balanced_pages >= (managed_pages >> 2); |
4ae0a48b ZC |
3009 | else |
3010 | return true; | |
1741c877 MG |
3011 | } |
3012 | ||
5515061d MG |
3013 | /* |
3014 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3015 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3016 | * | |
3017 | * Returns true if kswapd is ready to sleep | |
3018 | */ | |
3019 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, | |
dc83edd9 | 3020 | int classzone_idx) |
f50de2d3 | 3021 | { |
f50de2d3 MG |
3022 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ |
3023 | if (remaining) | |
5515061d MG |
3024 | return false; |
3025 | ||
3026 | /* | |
9e5e3661 VB |
3027 | * The throttled processes are normally woken up in balance_pgdat() as |
3028 | * soon as pfmemalloc_watermark_ok() is true. But there is a potential | |
3029 | * race between when kswapd checks the watermarks and a process gets | |
3030 | * throttled. There is also a potential race if processes get | |
3031 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3032 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3033 | * the wake up checks. If kswapd is going to sleep, no process should | |
3034 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3035 | * the wake up is premature, processes will wake kswapd and get | |
3036 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3037 | * that here we are under prepare_to_wait(). | |
5515061d | 3038 | */ |
9e5e3661 VB |
3039 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3040 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3041 | |
4ae0a48b | 3042 | return pgdat_balanced(pgdat, order, classzone_idx); |
f50de2d3 MG |
3043 | } |
3044 | ||
75485363 MG |
3045 | /* |
3046 | * kswapd shrinks the zone by the number of pages required to reach | |
3047 | * the high watermark. | |
b8e83b94 MG |
3048 | * |
3049 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3050 | * reclaim or if the lack of progress was due to pages under writeback. |
3051 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3052 | */ |
b8e83b94 | 3053 | static bool kswapd_shrink_zone(struct zone *zone, |
7c954f6d | 3054 | int classzone_idx, |
75485363 | 3055 | struct scan_control *sc, |
2ab44f43 | 3056 | unsigned long *nr_attempted) |
75485363 | 3057 | { |
7c954f6d MG |
3058 | int testorder = sc->order; |
3059 | unsigned long balance_gap; | |
7c954f6d | 3060 | bool lowmem_pressure; |
75485363 MG |
3061 | |
3062 | /* Reclaim above the high watermark. */ | |
3063 | sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone)); | |
7c954f6d MG |
3064 | |
3065 | /* | |
3066 | * Kswapd reclaims only single pages with compaction enabled. Trying | |
3067 | * too hard to reclaim until contiguous free pages have become | |
3068 | * available can hurt performance by evicting too much useful data | |
3069 | * from memory. Do not reclaim more than needed for compaction. | |
3070 | */ | |
3071 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && | |
ebff3980 VB |
3072 | compaction_suitable(zone, sc->order, 0, classzone_idx) |
3073 | != COMPACT_SKIPPED) | |
7c954f6d MG |
3074 | testorder = 0; |
3075 | ||
3076 | /* | |
3077 | * We put equal pressure on every zone, unless one zone has way too | |
3078 | * many pages free already. The "too many pages" is defined as the | |
3079 | * high wmark plus a "gap" where the gap is either the low | |
3080 | * watermark or 1% of the zone, whichever is smaller. | |
3081 | */ | |
4be89a34 JZ |
3082 | balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( |
3083 | zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); | |
7c954f6d MG |
3084 | |
3085 | /* | |
3086 | * If there is no low memory pressure or the zone is balanced then no | |
3087 | * reclaim is necessary | |
3088 | */ | |
3089 | lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone)); | |
3090 | if (!lowmem_pressure && zone_balanced(zone, testorder, | |
3091 | balance_gap, classzone_idx)) | |
3092 | return true; | |
3093 | ||
6b4f7799 | 3094 | shrink_zone(zone, sc, zone_idx(zone) == classzone_idx); |
75485363 | 3095 | |
2ab44f43 MG |
3096 | /* Account for the number of pages attempted to reclaim */ |
3097 | *nr_attempted += sc->nr_to_reclaim; | |
3098 | ||
57054651 | 3099 | clear_bit(ZONE_WRITEBACK, &zone->flags); |
283aba9f | 3100 | |
7c954f6d MG |
3101 | /* |
3102 | * If a zone reaches its high watermark, consider it to be no longer | |
3103 | * congested. It's possible there are dirty pages backed by congested | |
3104 | * BDIs but as pressure is relieved, speculatively avoid congestion | |
3105 | * waits. | |
3106 | */ | |
6e543d57 | 3107 | if (zone_reclaimable(zone) && |
7c954f6d | 3108 | zone_balanced(zone, testorder, 0, classzone_idx)) { |
57054651 JW |
3109 | clear_bit(ZONE_CONGESTED, &zone->flags); |
3110 | clear_bit(ZONE_DIRTY, &zone->flags); | |
7c954f6d MG |
3111 | } |
3112 | ||
b8e83b94 | 3113 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3114 | } |
3115 | ||
1da177e4 LT |
3116 | /* |
3117 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 3118 | * they are all at high_wmark_pages(zone). |
1da177e4 | 3119 | * |
0abdee2b | 3120 | * Returns the final order kswapd was reclaiming at |
1da177e4 LT |
3121 | * |
3122 | * There is special handling here for zones which are full of pinned pages. | |
3123 | * This can happen if the pages are all mlocked, or if they are all used by | |
3124 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
3125 | * What we do is to detect the case where all pages in the zone have been | |
3126 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
3127 | * dead and from now on, only perform a short scan. Basically we're polling | |
3128 | * the zone for when the problem goes away. | |
3129 | * | |
3130 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
3131 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
3132 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
3133 | * lower zones regardless of the number of free pages in the lower zones. This | |
3134 | * interoperates with the page allocator fallback scheme to ensure that aging | |
3135 | * of pages is balanced across the zones. | |
1da177e4 | 3136 | */ |
99504748 | 3137 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, |
dc83edd9 | 3138 | int *classzone_idx) |
1da177e4 | 3139 | { |
1da177e4 | 3140 | int i; |
99504748 | 3141 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
0608f43d AM |
3142 | unsigned long nr_soft_reclaimed; |
3143 | unsigned long nr_soft_scanned; | |
179e9639 AM |
3144 | struct scan_control sc = { |
3145 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3146 | .order = order, |
b8e83b94 | 3147 | .priority = DEF_PRIORITY, |
ee814fe2 | 3148 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3149 | .may_unmap = 1, |
2e2e4259 | 3150 | .may_swap = 1, |
179e9639 | 3151 | }; |
f8891e5e | 3152 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3153 | |
9e3b2f8c | 3154 | do { |
2ab44f43 | 3155 | unsigned long nr_attempted = 0; |
b8e83b94 | 3156 | bool raise_priority = true; |
2ab44f43 | 3157 | bool pgdat_needs_compaction = (order > 0); |
b8e83b94 MG |
3158 | |
3159 | sc.nr_reclaimed = 0; | |
1da177e4 | 3160 | |
d6277db4 RW |
3161 | /* |
3162 | * Scan in the highmem->dma direction for the highest | |
3163 | * zone which needs scanning | |
3164 | */ | |
3165 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
3166 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 3167 | |
d6277db4 RW |
3168 | if (!populated_zone(zone)) |
3169 | continue; | |
1da177e4 | 3170 | |
6e543d57 LD |
3171 | if (sc.priority != DEF_PRIORITY && |
3172 | !zone_reclaimable(zone)) | |
d6277db4 | 3173 | continue; |
1da177e4 | 3174 | |
556adecb RR |
3175 | /* |
3176 | * Do some background aging of the anon list, to give | |
3177 | * pages a chance to be referenced before reclaiming. | |
3178 | */ | |
9e3b2f8c | 3179 | age_active_anon(zone, &sc); |
556adecb | 3180 | |
cc715d99 MG |
3181 | /* |
3182 | * If the number of buffer_heads in the machine | |
3183 | * exceeds the maximum allowed level and this node | |
3184 | * has a highmem zone, force kswapd to reclaim from | |
3185 | * it to relieve lowmem pressure. | |
3186 | */ | |
3187 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
3188 | end_zone = i; | |
3189 | break; | |
3190 | } | |
3191 | ||
60cefed4 | 3192 | if (!zone_balanced(zone, order, 0, 0)) { |
d6277db4 | 3193 | end_zone = i; |
e1dbeda6 | 3194 | break; |
439423f6 | 3195 | } else { |
d43006d5 MG |
3196 | /* |
3197 | * If balanced, clear the dirty and congested | |
3198 | * flags | |
3199 | */ | |
57054651 JW |
3200 | clear_bit(ZONE_CONGESTED, &zone->flags); |
3201 | clear_bit(ZONE_DIRTY, &zone->flags); | |
1da177e4 | 3202 | } |
1da177e4 | 3203 | } |
dafcb73e | 3204 | |
b8e83b94 | 3205 | if (i < 0) |
e1dbeda6 AM |
3206 | goto out; |
3207 | ||
1da177e4 LT |
3208 | for (i = 0; i <= end_zone; i++) { |
3209 | struct zone *zone = pgdat->node_zones + i; | |
3210 | ||
2ab44f43 MG |
3211 | if (!populated_zone(zone)) |
3212 | continue; | |
3213 | ||
2ab44f43 MG |
3214 | /* |
3215 | * If any zone is currently balanced then kswapd will | |
3216 | * not call compaction as it is expected that the | |
3217 | * necessary pages are already available. | |
3218 | */ | |
3219 | if (pgdat_needs_compaction && | |
3220 | zone_watermark_ok(zone, order, | |
3221 | low_wmark_pages(zone), | |
3222 | *classzone_idx, 0)) | |
3223 | pgdat_needs_compaction = false; | |
1da177e4 LT |
3224 | } |
3225 | ||
b7ea3c41 MG |
3226 | /* |
3227 | * If we're getting trouble reclaiming, start doing writepage | |
3228 | * even in laptop mode. | |
3229 | */ | |
3230 | if (sc.priority < DEF_PRIORITY - 2) | |
3231 | sc.may_writepage = 1; | |
3232 | ||
1da177e4 LT |
3233 | /* |
3234 | * Now scan the zone in the dma->highmem direction, stopping | |
3235 | * at the last zone which needs scanning. | |
3236 | * | |
3237 | * We do this because the page allocator works in the opposite | |
3238 | * direction. This prevents the page allocator from allocating | |
3239 | * pages behind kswapd's direction of progress, which would | |
3240 | * cause too much scanning of the lower zones. | |
3241 | */ | |
3242 | for (i = 0; i <= end_zone; i++) { | |
3243 | struct zone *zone = pgdat->node_zones + i; | |
3244 | ||
f3fe6512 | 3245 | if (!populated_zone(zone)) |
1da177e4 LT |
3246 | continue; |
3247 | ||
6e543d57 LD |
3248 | if (sc.priority != DEF_PRIORITY && |
3249 | !zone_reclaimable(zone)) | |
1da177e4 LT |
3250 | continue; |
3251 | ||
1da177e4 | 3252 | sc.nr_scanned = 0; |
4e416953 | 3253 | |
0608f43d AM |
3254 | nr_soft_scanned = 0; |
3255 | /* | |
3256 | * Call soft limit reclaim before calling shrink_zone. | |
3257 | */ | |
3258 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
3259 | order, sc.gfp_mask, | |
3260 | &nr_soft_scanned); | |
3261 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3262 | ||
32a4330d | 3263 | /* |
7c954f6d MG |
3264 | * There should be no need to raise the scanning |
3265 | * priority if enough pages are already being scanned | |
3266 | * that that high watermark would be met at 100% | |
3267 | * efficiency. | |
fe2c2a10 | 3268 | */ |
6b4f7799 JW |
3269 | if (kswapd_shrink_zone(zone, end_zone, |
3270 | &sc, &nr_attempted)) | |
7c954f6d | 3271 | raise_priority = false; |
1da177e4 | 3272 | } |
5515061d MG |
3273 | |
3274 | /* | |
3275 | * If the low watermark is met there is no need for processes | |
3276 | * to be throttled on pfmemalloc_wait as they should not be | |
3277 | * able to safely make forward progress. Wake them | |
3278 | */ | |
3279 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
3280 | pfmemalloc_watermark_ok(pgdat)) | |
cfc51155 | 3281 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3282 | |
1da177e4 | 3283 | /* |
b8e83b94 MG |
3284 | * Fragmentation may mean that the system cannot be rebalanced |
3285 | * for high-order allocations in all zones. If twice the | |
3286 | * allocation size has been reclaimed and the zones are still | |
3287 | * not balanced then recheck the watermarks at order-0 to | |
3288 | * prevent kswapd reclaiming excessively. Assume that a | |
3289 | * process requested a high-order can direct reclaim/compact. | |
1da177e4 | 3290 | */ |
b8e83b94 MG |
3291 | if (order && sc.nr_reclaimed >= 2UL << order) |
3292 | order = sc.order = 0; | |
8357376d | 3293 | |
b8e83b94 MG |
3294 | /* Check if kswapd should be suspending */ |
3295 | if (try_to_freeze() || kthread_should_stop()) | |
3296 | break; | |
8357376d | 3297 | |
2ab44f43 MG |
3298 | /* |
3299 | * Compact if necessary and kswapd is reclaiming at least the | |
3300 | * high watermark number of pages as requsted | |
3301 | */ | |
3302 | if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted) | |
3303 | compact_pgdat(pgdat, order); | |
3304 | ||
73ce02e9 | 3305 | /* |
b8e83b94 MG |
3306 | * Raise priority if scanning rate is too low or there was no |
3307 | * progress in reclaiming pages | |
73ce02e9 | 3308 | */ |
b8e83b94 MG |
3309 | if (raise_priority || !sc.nr_reclaimed) |
3310 | sc.priority--; | |
9aa41348 | 3311 | } while (sc.priority >= 1 && |
b8e83b94 | 3312 | !pgdat_balanced(pgdat, order, *classzone_idx)); |
1da177e4 | 3313 | |
b8e83b94 | 3314 | out: |
0abdee2b | 3315 | /* |
5515061d | 3316 | * Return the order we were reclaiming at so prepare_kswapd_sleep() |
0abdee2b MG |
3317 | * makes a decision on the order we were last reclaiming at. However, |
3318 | * if another caller entered the allocator slow path while kswapd | |
3319 | * was awake, order will remain at the higher level | |
3320 | */ | |
dc83edd9 | 3321 | *classzone_idx = end_zone; |
0abdee2b | 3322 | return order; |
1da177e4 LT |
3323 | } |
3324 | ||
dc83edd9 | 3325 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f0bc0a60 KM |
3326 | { |
3327 | long remaining = 0; | |
3328 | DEFINE_WAIT(wait); | |
3329 | ||
3330 | if (freezing(current) || kthread_should_stop()) | |
3331 | return; | |
3332 | ||
3333 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3334 | ||
3335 | /* Try to sleep for a short interval */ | |
5515061d | 3336 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
3337 | remaining = schedule_timeout(HZ/10); |
3338 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3339 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3340 | } | |
3341 | ||
3342 | /* | |
3343 | * After a short sleep, check if it was a premature sleep. If not, then | |
3344 | * go fully to sleep until explicitly woken up. | |
3345 | */ | |
5515061d | 3346 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
3347 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3348 | ||
3349 | /* | |
3350 | * vmstat counters are not perfectly accurate and the estimated | |
3351 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3352 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3353 | * watermarks being breached while under pressure, we reduce the | |
3354 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3355 | * them before going back to sleep. | |
3356 | */ | |
3357 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c | 3358 | |
62997027 MG |
3359 | /* |
3360 | * Compaction records what page blocks it recently failed to | |
3361 | * isolate pages from and skips them in the future scanning. | |
3362 | * When kswapd is going to sleep, it is reasonable to assume | |
3363 | * that pages and compaction may succeed so reset the cache. | |
3364 | */ | |
3365 | reset_isolation_suitable(pgdat); | |
3366 | ||
1c7e7f6c AK |
3367 | if (!kthread_should_stop()) |
3368 | schedule(); | |
3369 | ||
f0bc0a60 KM |
3370 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3371 | } else { | |
3372 | if (remaining) | |
3373 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3374 | else | |
3375 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3376 | } | |
3377 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3378 | } | |
3379 | ||
1da177e4 LT |
3380 | /* |
3381 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3382 | * from the init process. |
1da177e4 LT |
3383 | * |
3384 | * This basically trickles out pages so that we have _some_ | |
3385 | * free memory available even if there is no other activity | |
3386 | * that frees anything up. This is needed for things like routing | |
3387 | * etc, where we otherwise might have all activity going on in | |
3388 | * asynchronous contexts that cannot page things out. | |
3389 | * | |
3390 | * If there are applications that are active memory-allocators | |
3391 | * (most normal use), this basically shouldn't matter. | |
3392 | */ | |
3393 | static int kswapd(void *p) | |
3394 | { | |
215ddd66 | 3395 | unsigned long order, new_order; |
d2ebd0f6 | 3396 | unsigned balanced_order; |
215ddd66 | 3397 | int classzone_idx, new_classzone_idx; |
d2ebd0f6 | 3398 | int balanced_classzone_idx; |
1da177e4 LT |
3399 | pg_data_t *pgdat = (pg_data_t*)p; |
3400 | struct task_struct *tsk = current; | |
f0bc0a60 | 3401 | |
1da177e4 LT |
3402 | struct reclaim_state reclaim_state = { |
3403 | .reclaimed_slab = 0, | |
3404 | }; | |
a70f7302 | 3405 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3406 | |
cf40bd16 NP |
3407 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3408 | ||
174596a0 | 3409 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3410 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3411 | current->reclaim_state = &reclaim_state; |
3412 | ||
3413 | /* | |
3414 | * Tell the memory management that we're a "memory allocator", | |
3415 | * and that if we need more memory we should get access to it | |
3416 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3417 | * never get caught in the normal page freeing logic. | |
3418 | * | |
3419 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3420 | * you need a small amount of memory in order to be able to | |
3421 | * page out something else, and this flag essentially protects | |
3422 | * us from recursively trying to free more memory as we're | |
3423 | * trying to free the first piece of memory in the first place). | |
3424 | */ | |
930d9152 | 3425 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3426 | set_freezable(); |
1da177e4 | 3427 | |
215ddd66 | 3428 | order = new_order = 0; |
d2ebd0f6 | 3429 | balanced_order = 0; |
215ddd66 | 3430 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; |
d2ebd0f6 | 3431 | balanced_classzone_idx = classzone_idx; |
1da177e4 | 3432 | for ( ; ; ) { |
6f6313d4 | 3433 | bool ret; |
3e1d1d28 | 3434 | |
215ddd66 MG |
3435 | /* |
3436 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
3437 | * new request of a similar or harder type will succeed soon | |
3438 | * so consider going to sleep on the basis we reclaimed at | |
3439 | */ | |
d2ebd0f6 AS |
3440 | if (balanced_classzone_idx >= new_classzone_idx && |
3441 | balanced_order == new_order) { | |
215ddd66 MG |
3442 | new_order = pgdat->kswapd_max_order; |
3443 | new_classzone_idx = pgdat->classzone_idx; | |
3444 | pgdat->kswapd_max_order = 0; | |
3445 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3446 | } | |
3447 | ||
99504748 | 3448 | if (order < new_order || classzone_idx > new_classzone_idx) { |
1da177e4 LT |
3449 | /* |
3450 | * Don't sleep if someone wants a larger 'order' | |
99504748 | 3451 | * allocation or has tigher zone constraints |
1da177e4 LT |
3452 | */ |
3453 | order = new_order; | |
99504748 | 3454 | classzone_idx = new_classzone_idx; |
1da177e4 | 3455 | } else { |
d2ebd0f6 AS |
3456 | kswapd_try_to_sleep(pgdat, balanced_order, |
3457 | balanced_classzone_idx); | |
1da177e4 | 3458 | order = pgdat->kswapd_max_order; |
99504748 | 3459 | classzone_idx = pgdat->classzone_idx; |
f0dfcde0 AS |
3460 | new_order = order; |
3461 | new_classzone_idx = classzone_idx; | |
4d40502e | 3462 | pgdat->kswapd_max_order = 0; |
215ddd66 | 3463 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
1da177e4 | 3464 | } |
1da177e4 | 3465 | |
8fe23e05 DR |
3466 | ret = try_to_freeze(); |
3467 | if (kthread_should_stop()) | |
3468 | break; | |
3469 | ||
3470 | /* | |
3471 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3472 | * after returning from the refrigerator | |
3473 | */ | |
33906bc5 MG |
3474 | if (!ret) { |
3475 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
d2ebd0f6 AS |
3476 | balanced_classzone_idx = classzone_idx; |
3477 | balanced_order = balance_pgdat(pgdat, order, | |
3478 | &balanced_classzone_idx); | |
33906bc5 | 3479 | } |
1da177e4 | 3480 | } |
b0a8cc58 | 3481 | |
71abdc15 | 3482 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
b0a8cc58 | 3483 | current->reclaim_state = NULL; |
71abdc15 JW |
3484 | lockdep_clear_current_reclaim_state(); |
3485 | ||
1da177e4 LT |
3486 | return 0; |
3487 | } | |
3488 | ||
3489 | /* | |
3490 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3491 | */ | |
99504748 | 3492 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3493 | { |
3494 | pg_data_t *pgdat; | |
3495 | ||
f3fe6512 | 3496 | if (!populated_zone(zone)) |
1da177e4 LT |
3497 | return; |
3498 | ||
344736f2 | 3499 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) |
1da177e4 | 3500 | return; |
88f5acf8 | 3501 | pgdat = zone->zone_pgdat; |
99504748 | 3502 | if (pgdat->kswapd_max_order < order) { |
1da177e4 | 3503 | pgdat->kswapd_max_order = order; |
99504748 MG |
3504 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
3505 | } | |
8d0986e2 | 3506 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3507 | return; |
892f795d | 3508 | if (zone_balanced(zone, order, 0, 0)) |
88f5acf8 MG |
3509 | return; |
3510 | ||
3511 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3512 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3513 | } |
3514 | ||
c6f37f12 | 3515 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3516 | /* |
7b51755c | 3517 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3518 | * freed pages. |
3519 | * | |
3520 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3521 | * LRU order by reclaiming preferentially | |
3522 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3523 | */ |
7b51755c | 3524 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3525 | { |
d6277db4 | 3526 | struct reclaim_state reclaim_state; |
d6277db4 | 3527 | struct scan_control sc = { |
ee814fe2 | 3528 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3529 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
ee814fe2 | 3530 | .priority = DEF_PRIORITY, |
d6277db4 | 3531 | .may_writepage = 1, |
ee814fe2 JW |
3532 | .may_unmap = 1, |
3533 | .may_swap = 1, | |
7b51755c | 3534 | .hibernation_mode = 1, |
1da177e4 | 3535 | }; |
a09ed5e0 | 3536 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c KM |
3537 | struct task_struct *p = current; |
3538 | unsigned long nr_reclaimed; | |
1da177e4 | 3539 | |
7b51755c KM |
3540 | p->flags |= PF_MEMALLOC; |
3541 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3542 | reclaim_state.reclaimed_slab = 0; | |
3543 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3544 | |
3115cd91 | 3545 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3546 | |
7b51755c KM |
3547 | p->reclaim_state = NULL; |
3548 | lockdep_clear_current_reclaim_state(); | |
3549 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3550 | |
7b51755c | 3551 | return nr_reclaimed; |
1da177e4 | 3552 | } |
c6f37f12 | 3553 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3554 | |
1da177e4 LT |
3555 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3556 | not required for correctness. So if the last cpu in a node goes | |
3557 | away, we get changed to run anywhere: as the first one comes back, | |
3558 | restore their cpu bindings. */ | |
fcb35a9b GKH |
3559 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
3560 | void *hcpu) | |
1da177e4 | 3561 | { |
58c0a4a7 | 3562 | int nid; |
1da177e4 | 3563 | |
8bb78442 | 3564 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
48fb2e24 | 3565 | for_each_node_state(nid, N_MEMORY) { |
c5f59f08 | 3566 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3567 | const struct cpumask *mask; |
3568 | ||
3569 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3570 | |
3e597945 | 3571 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3572 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3573 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3574 | } |
3575 | } | |
3576 | return NOTIFY_OK; | |
3577 | } | |
1da177e4 | 3578 | |
3218ae14 YG |
3579 | /* |
3580 | * This kswapd start function will be called by init and node-hot-add. | |
3581 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3582 | */ | |
3583 | int kswapd_run(int nid) | |
3584 | { | |
3585 | pg_data_t *pgdat = NODE_DATA(nid); | |
3586 | int ret = 0; | |
3587 | ||
3588 | if (pgdat->kswapd) | |
3589 | return 0; | |
3590 | ||
3591 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3592 | if (IS_ERR(pgdat->kswapd)) { | |
3593 | /* failure at boot is fatal */ | |
3594 | BUG_ON(system_state == SYSTEM_BOOTING); | |
d5dc0ad9 GS |
3595 | pr_err("Failed to start kswapd on node %d\n", nid); |
3596 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3597 | pgdat->kswapd = NULL; |
3218ae14 YG |
3598 | } |
3599 | return ret; | |
3600 | } | |
3601 | ||
8fe23e05 | 3602 | /* |
d8adde17 | 3603 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 3604 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
3605 | */ |
3606 | void kswapd_stop(int nid) | |
3607 | { | |
3608 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3609 | ||
d8adde17 | 3610 | if (kswapd) { |
8fe23e05 | 3611 | kthread_stop(kswapd); |
d8adde17 JL |
3612 | NODE_DATA(nid)->kswapd = NULL; |
3613 | } | |
8fe23e05 DR |
3614 | } |
3615 | ||
1da177e4 LT |
3616 | static int __init kswapd_init(void) |
3617 | { | |
3218ae14 | 3618 | int nid; |
69e05944 | 3619 | |
1da177e4 | 3620 | swap_setup(); |
48fb2e24 | 3621 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3622 | kswapd_run(nid); |
1da177e4 LT |
3623 | hotcpu_notifier(cpu_callback, 0); |
3624 | return 0; | |
3625 | } | |
3626 | ||
3627 | module_init(kswapd_init) | |
9eeff239 CL |
3628 | |
3629 | #ifdef CONFIG_NUMA | |
3630 | /* | |
3631 | * Zone reclaim mode | |
3632 | * | |
3633 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3634 | * the watermarks. | |
9eeff239 CL |
3635 | */ |
3636 | int zone_reclaim_mode __read_mostly; | |
3637 | ||
1b2ffb78 | 3638 | #define RECLAIM_OFF 0 |
7d03431c | 3639 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 3640 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 3641 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 3642 | |
a92f7126 CL |
3643 | /* |
3644 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3645 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3646 | * a zone. | |
3647 | */ | |
3648 | #define ZONE_RECLAIM_PRIORITY 4 | |
3649 | ||
9614634f CL |
3650 | /* |
3651 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3652 | * occur. | |
3653 | */ | |
3654 | int sysctl_min_unmapped_ratio = 1; | |
3655 | ||
0ff38490 CL |
3656 | /* |
3657 | * If the number of slab pages in a zone grows beyond this percentage then | |
3658 | * slab reclaim needs to occur. | |
3659 | */ | |
3660 | int sysctl_min_slab_ratio = 5; | |
3661 | ||
90afa5de MG |
3662 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
3663 | { | |
3664 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3665 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3666 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3667 | ||
3668 | /* | |
3669 | * It's possible for there to be more file mapped pages than | |
3670 | * accounted for by the pages on the file LRU lists because | |
3671 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3672 | */ | |
3673 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3674 | } | |
3675 | ||
3676 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
d031a157 | 3677 | static unsigned long zone_pagecache_reclaimable(struct zone *zone) |
90afa5de | 3678 | { |
d031a157 AM |
3679 | unsigned long nr_pagecache_reclaimable; |
3680 | unsigned long delta = 0; | |
90afa5de MG |
3681 | |
3682 | /* | |
95bbc0c7 | 3683 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de MG |
3684 | * potentially reclaimable. Otherwise, we have to worry about |
3685 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3686 | * a better estimate | |
3687 | */ | |
95bbc0c7 | 3688 | if (zone_reclaim_mode & RECLAIM_UNMAP) |
90afa5de MG |
3689 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); |
3690 | else | |
3691 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3692 | ||
3693 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3694 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3695 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3696 | ||
3697 | /* Watch for any possible underflows due to delta */ | |
3698 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3699 | delta = nr_pagecache_reclaimable; | |
3700 | ||
3701 | return nr_pagecache_reclaimable - delta; | |
3702 | } | |
3703 | ||
9eeff239 CL |
3704 | /* |
3705 | * Try to free up some pages from this zone through reclaim. | |
3706 | */ | |
179e9639 | 3707 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3708 | { |
7fb2d46d | 3709 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3710 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3711 | struct task_struct *p = current; |
3712 | struct reclaim_state reclaim_state; | |
179e9639 | 3713 | struct scan_control sc = { |
62b726c1 | 3714 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
21caf2fc | 3715 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
bd2f6199 | 3716 | .order = order, |
9e3b2f8c | 3717 | .priority = ZONE_RECLAIM_PRIORITY, |
ee814fe2 | 3718 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), |
95bbc0c7 | 3719 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP), |
ee814fe2 | 3720 | .may_swap = 1, |
179e9639 | 3721 | }; |
9eeff239 | 3722 | |
9eeff239 | 3723 | cond_resched(); |
d4f7796e | 3724 | /* |
95bbc0c7 | 3725 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 3726 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 3727 | * and RECLAIM_UNMAP. |
d4f7796e CL |
3728 | */ |
3729 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3730 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3731 | reclaim_state.reclaimed_slab = 0; |
3732 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3733 | |
90afa5de | 3734 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
3735 | /* |
3736 | * Free memory by calling shrink zone with increasing | |
3737 | * priorities until we have enough memory freed. | |
3738 | */ | |
0ff38490 | 3739 | do { |
6b4f7799 | 3740 | shrink_zone(zone, &sc, true); |
9e3b2f8c | 3741 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 3742 | } |
c84db23c | 3743 | |
9eeff239 | 3744 | p->reclaim_state = NULL; |
d4f7796e | 3745 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3746 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3747 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3748 | } |
179e9639 AM |
3749 | |
3750 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3751 | { | |
179e9639 | 3752 | int node_id; |
d773ed6b | 3753 | int ret; |
179e9639 AM |
3754 | |
3755 | /* | |
0ff38490 CL |
3756 | * Zone reclaim reclaims unmapped file backed pages and |
3757 | * slab pages if we are over the defined limits. | |
34aa1330 | 3758 | * |
9614634f CL |
3759 | * A small portion of unmapped file backed pages is needed for |
3760 | * file I/O otherwise pages read by file I/O will be immediately | |
3761 | * thrown out if the zone is overallocated. So we do not reclaim | |
3762 | * if less than a specified percentage of the zone is used by | |
3763 | * unmapped file backed pages. | |
179e9639 | 3764 | */ |
90afa5de MG |
3765 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
3766 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 3767 | return ZONE_RECLAIM_FULL; |
179e9639 | 3768 | |
6e543d57 | 3769 | if (!zone_reclaimable(zone)) |
fa5e084e | 3770 | return ZONE_RECLAIM_FULL; |
d773ed6b | 3771 | |
179e9639 | 3772 | /* |
d773ed6b | 3773 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3774 | */ |
d0164adc | 3775 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 3776 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
3777 | |
3778 | /* | |
3779 | * Only run zone reclaim on the local zone or on zones that do not | |
3780 | * have associated processors. This will favor the local processor | |
3781 | * over remote processors and spread off node memory allocations | |
3782 | * as wide as possible. | |
3783 | */ | |
89fa3024 | 3784 | node_id = zone_to_nid(zone); |
37c0708d | 3785 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 3786 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b | 3787 | |
57054651 | 3788 | if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags)) |
fa5e084e MG |
3789 | return ZONE_RECLAIM_NOSCAN; |
3790 | ||
d773ed6b | 3791 | ret = __zone_reclaim(zone, gfp_mask, order); |
57054651 | 3792 | clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags); |
d773ed6b | 3793 | |
24cf7251 MG |
3794 | if (!ret) |
3795 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3796 | ||
d773ed6b | 3797 | return ret; |
179e9639 | 3798 | } |
9eeff239 | 3799 | #endif |
894bc310 | 3800 | |
894bc310 LS |
3801 | /* |
3802 | * page_evictable - test whether a page is evictable | |
3803 | * @page: the page to test | |
894bc310 LS |
3804 | * |
3805 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3806 | * lists vs unevictable list. |
894bc310 LS |
3807 | * |
3808 | * Reasons page might not be evictable: | |
ba9ddf49 | 3809 | * (1) page's mapping marked unevictable |
b291f000 | 3810 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3811 | * |
894bc310 | 3812 | */ |
39b5f29a | 3813 | int page_evictable(struct page *page) |
894bc310 | 3814 | { |
39b5f29a | 3815 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
894bc310 | 3816 | } |
89e004ea | 3817 | |
85046579 | 3818 | #ifdef CONFIG_SHMEM |
89e004ea | 3819 | /** |
24513264 HD |
3820 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3821 | * @pages: array of pages to check | |
3822 | * @nr_pages: number of pages to check | |
89e004ea | 3823 | * |
24513264 | 3824 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3825 | * |
3826 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3827 | */ |
24513264 | 3828 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3829 | { |
925b7673 | 3830 | struct lruvec *lruvec; |
24513264 HD |
3831 | struct zone *zone = NULL; |
3832 | int pgscanned = 0; | |
3833 | int pgrescued = 0; | |
3834 | int i; | |
89e004ea | 3835 | |
24513264 HD |
3836 | for (i = 0; i < nr_pages; i++) { |
3837 | struct page *page = pages[i]; | |
3838 | struct zone *pagezone; | |
89e004ea | 3839 | |
24513264 HD |
3840 | pgscanned++; |
3841 | pagezone = page_zone(page); | |
3842 | if (pagezone != zone) { | |
3843 | if (zone) | |
3844 | spin_unlock_irq(&zone->lru_lock); | |
3845 | zone = pagezone; | |
3846 | spin_lock_irq(&zone->lru_lock); | |
3847 | } | |
fa9add64 | 3848 | lruvec = mem_cgroup_page_lruvec(page, zone); |
89e004ea | 3849 | |
24513264 HD |
3850 | if (!PageLRU(page) || !PageUnevictable(page)) |
3851 | continue; | |
89e004ea | 3852 | |
39b5f29a | 3853 | if (page_evictable(page)) { |
24513264 HD |
3854 | enum lru_list lru = page_lru_base_type(page); |
3855 | ||
309381fe | 3856 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 3857 | ClearPageUnevictable(page); |
fa9add64 HD |
3858 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
3859 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 3860 | pgrescued++; |
89e004ea | 3861 | } |
24513264 | 3862 | } |
89e004ea | 3863 | |
24513264 HD |
3864 | if (zone) { |
3865 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3866 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3867 | spin_unlock_irq(&zone->lru_lock); | |
89e004ea | 3868 | } |
89e004ea | 3869 | } |
85046579 | 3870 | #endif /* CONFIG_SHMEM */ |