Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/vmscan.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
5 | * | |
6 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
7 | * kswapd added: 7.1.96 sct | |
8 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
9 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
10 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
11 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
12 | */ | |
13 | ||
b1de0d13 MH |
14 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
15 | ||
1da177e4 LT |
16 | #include <linux/mm.h> |
17 | #include <linux/module.h> | |
5a0e3ad6 | 18 | #include <linux/gfp.h> |
1da177e4 LT |
19 | #include <linux/kernel_stat.h> |
20 | #include <linux/swap.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/init.h> | |
23 | #include <linux/highmem.h> | |
70ddf637 | 24 | #include <linux/vmpressure.h> |
e129b5c2 | 25 | #include <linux/vmstat.h> |
1da177e4 LT |
26 | #include <linux/file.h> |
27 | #include <linux/writeback.h> | |
28 | #include <linux/blkdev.h> | |
29 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
30 | buffer_heads_over_limit */ | |
31 | #include <linux/mm_inline.h> | |
1da177e4 LT |
32 | #include <linux/backing-dev.h> |
33 | #include <linux/rmap.h> | |
34 | #include <linux/topology.h> | |
35 | #include <linux/cpu.h> | |
36 | #include <linux/cpuset.h> | |
3e7d3449 | 37 | #include <linux/compaction.h> |
1da177e4 LT |
38 | #include <linux/notifier.h> |
39 | #include <linux/rwsem.h> | |
248a0301 | 40 | #include <linux/delay.h> |
3218ae14 | 41 | #include <linux/kthread.h> |
7dfb7103 | 42 | #include <linux/freezer.h> |
66e1707b | 43 | #include <linux/memcontrol.h> |
873b4771 | 44 | #include <linux/delayacct.h> |
af936a16 | 45 | #include <linux/sysctl.h> |
929bea7c | 46 | #include <linux/oom.h> |
268bb0ce | 47 | #include <linux/prefetch.h> |
b1de0d13 | 48 | #include <linux/printk.h> |
f9fe48be | 49 | #include <linux/dax.h> |
1da177e4 LT |
50 | |
51 | #include <asm/tlbflush.h> | |
52 | #include <asm/div64.h> | |
53 | ||
54 | #include <linux/swapops.h> | |
117aad1e | 55 | #include <linux/balloon_compaction.h> |
1da177e4 | 56 | |
0f8053a5 NP |
57 | #include "internal.h" |
58 | ||
33906bc5 MG |
59 | #define CREATE_TRACE_POINTS |
60 | #include <trace/events/vmscan.h> | |
61 | ||
1da177e4 | 62 | struct scan_control { |
22fba335 KM |
63 | /* How many pages shrink_list() should reclaim */ |
64 | unsigned long nr_to_reclaim; | |
65 | ||
1da177e4 | 66 | /* This context's GFP mask */ |
6daa0e28 | 67 | gfp_t gfp_mask; |
1da177e4 | 68 | |
ee814fe2 | 69 | /* Allocation order */ |
5ad333eb | 70 | int order; |
66e1707b | 71 | |
ee814fe2 JW |
72 | /* |
73 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
74 | * are scanned. | |
75 | */ | |
76 | nodemask_t *nodemask; | |
9e3b2f8c | 77 | |
f16015fb JW |
78 | /* |
79 | * The memory cgroup that hit its limit and as a result is the | |
80 | * primary target of this reclaim invocation. | |
81 | */ | |
82 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 83 | |
ee814fe2 JW |
84 | /* Scan (total_size >> priority) pages at once */ |
85 | int priority; | |
86 | ||
87 | unsigned int may_writepage:1; | |
88 | ||
89 | /* Can mapped pages be reclaimed? */ | |
90 | unsigned int may_unmap:1; | |
91 | ||
92 | /* Can pages be swapped as part of reclaim? */ | |
93 | unsigned int may_swap:1; | |
94 | ||
241994ed JW |
95 | /* Can cgroups be reclaimed below their normal consumption range? */ |
96 | unsigned int may_thrash:1; | |
97 | ||
ee814fe2 JW |
98 | unsigned int hibernation_mode:1; |
99 | ||
100 | /* One of the zones is ready for compaction */ | |
101 | unsigned int compaction_ready:1; | |
102 | ||
103 | /* Incremented by the number of inactive pages that were scanned */ | |
104 | unsigned long nr_scanned; | |
105 | ||
106 | /* Number of pages freed so far during a call to shrink_zones() */ | |
107 | unsigned long nr_reclaimed; | |
1da177e4 LT |
108 | }; |
109 | ||
1da177e4 LT |
110 | #ifdef ARCH_HAS_PREFETCH |
111 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
112 | do { \ | |
113 | if ((_page)->lru.prev != _base) { \ | |
114 | struct page *prev; \ | |
115 | \ | |
116 | prev = lru_to_page(&(_page->lru)); \ | |
117 | prefetch(&prev->_field); \ | |
118 | } \ | |
119 | } while (0) | |
120 | #else | |
121 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
122 | #endif | |
123 | ||
124 | #ifdef ARCH_HAS_PREFETCHW | |
125 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
126 | do { \ | |
127 | if ((_page)->lru.prev != _base) { \ | |
128 | struct page *prev; \ | |
129 | \ | |
130 | prev = lru_to_page(&(_page->lru)); \ | |
131 | prefetchw(&prev->_field); \ | |
132 | } \ | |
133 | } while (0) | |
134 | #else | |
135 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
136 | #endif | |
137 | ||
138 | /* | |
139 | * From 0 .. 100. Higher means more swappy. | |
140 | */ | |
141 | int vm_swappiness = 60; | |
d0480be4 WSH |
142 | /* |
143 | * The total number of pages which are beyond the high watermark within all | |
144 | * zones. | |
145 | */ | |
146 | unsigned long vm_total_pages; | |
1da177e4 LT |
147 | |
148 | static LIST_HEAD(shrinker_list); | |
149 | static DECLARE_RWSEM(shrinker_rwsem); | |
150 | ||
c255a458 | 151 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
152 | static bool global_reclaim(struct scan_control *sc) |
153 | { | |
f16015fb | 154 | return !sc->target_mem_cgroup; |
89b5fae5 | 155 | } |
97c9341f TH |
156 | |
157 | /** | |
158 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
159 | * @sc: scan_control in question | |
160 | * | |
161 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
162 | * completely broken with the legacy memcg and direct stalling in | |
163 | * shrink_page_list() is used for throttling instead, which lacks all the | |
164 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
165 | * allocation and configurability. | |
166 | * | |
167 | * This function tests whether the vmscan currently in progress can assume | |
168 | * that the normal dirty throttling mechanism is operational. | |
169 | */ | |
170 | static bool sane_reclaim(struct scan_control *sc) | |
171 | { | |
172 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
173 | ||
174 | if (!memcg) | |
175 | return true; | |
176 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 177 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
178 | return true; |
179 | #endif | |
180 | return false; | |
181 | } | |
91a45470 | 182 | #else |
89b5fae5 JW |
183 | static bool global_reclaim(struct scan_control *sc) |
184 | { | |
185 | return true; | |
186 | } | |
97c9341f TH |
187 | |
188 | static bool sane_reclaim(struct scan_control *sc) | |
189 | { | |
190 | return true; | |
191 | } | |
91a45470 KH |
192 | #endif |
193 | ||
599d0c95 MG |
194 | /* |
195 | * This misses isolated pages which are not accounted for to save counters. | |
196 | * As the data only determines if reclaim or compaction continues, it is | |
197 | * not expected that isolated pages will be a dominating factor. | |
198 | */ | |
0a0337e0 | 199 | unsigned long zone_reclaimable_pages(struct zone *zone) |
6e543d57 | 200 | { |
d031a157 | 201 | unsigned long nr; |
6e543d57 | 202 | |
599d0c95 MG |
203 | nr = zone_page_state_snapshot(zone, NR_ZONE_LRU_FILE); |
204 | if (get_nr_swap_pages() > 0) | |
205 | nr += zone_page_state_snapshot(zone, NR_ZONE_LRU_ANON); | |
206 | ||
207 | return nr; | |
208 | } | |
209 | ||
210 | unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat) | |
211 | { | |
212 | unsigned long nr; | |
213 | ||
214 | nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) + | |
215 | node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) + | |
216 | node_page_state_snapshot(pgdat, NR_ISOLATED_FILE); | |
6e543d57 LD |
217 | |
218 | if (get_nr_swap_pages() > 0) | |
599d0c95 MG |
219 | nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) + |
220 | node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) + | |
221 | node_page_state_snapshot(pgdat, NR_ISOLATED_ANON); | |
6e543d57 LD |
222 | |
223 | return nr; | |
224 | } | |
225 | ||
599d0c95 | 226 | bool pgdat_reclaimable(struct pglist_data *pgdat) |
6e543d57 | 227 | { |
599d0c95 MG |
228 | return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) < |
229 | pgdat_reclaimable_pages(pgdat) * 6; | |
6e543d57 LD |
230 | } |
231 | ||
23047a96 | 232 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru) |
c9f299d9 | 233 | { |
c3c787e8 | 234 | if (!mem_cgroup_disabled()) |
4d7dcca2 | 235 | return mem_cgroup_get_lru_size(lruvec, lru); |
a3d8e054 | 236 | |
599d0c95 | 237 | return node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); |
c9f299d9 KM |
238 | } |
239 | ||
1da177e4 | 240 | /* |
1d3d4437 | 241 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 242 | */ |
1d3d4437 | 243 | int register_shrinker(struct shrinker *shrinker) |
1da177e4 | 244 | { |
1d3d4437 GC |
245 | size_t size = sizeof(*shrinker->nr_deferred); |
246 | ||
1d3d4437 GC |
247 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
248 | size *= nr_node_ids; | |
249 | ||
250 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
251 | if (!shrinker->nr_deferred) | |
252 | return -ENOMEM; | |
253 | ||
8e1f936b RR |
254 | down_write(&shrinker_rwsem); |
255 | list_add_tail(&shrinker->list, &shrinker_list); | |
256 | up_write(&shrinker_rwsem); | |
1d3d4437 | 257 | return 0; |
1da177e4 | 258 | } |
8e1f936b | 259 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
260 | |
261 | /* | |
262 | * Remove one | |
263 | */ | |
8e1f936b | 264 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
265 | { |
266 | down_write(&shrinker_rwsem); | |
267 | list_del(&shrinker->list); | |
268 | up_write(&shrinker_rwsem); | |
ae393321 | 269 | kfree(shrinker->nr_deferred); |
1da177e4 | 270 | } |
8e1f936b | 271 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
272 | |
273 | #define SHRINK_BATCH 128 | |
1d3d4437 | 274 | |
cb731d6c VD |
275 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
276 | struct shrinker *shrinker, | |
277 | unsigned long nr_scanned, | |
278 | unsigned long nr_eligible) | |
1d3d4437 GC |
279 | { |
280 | unsigned long freed = 0; | |
281 | unsigned long long delta; | |
282 | long total_scan; | |
d5bc5fd3 | 283 | long freeable; |
1d3d4437 GC |
284 | long nr; |
285 | long new_nr; | |
286 | int nid = shrinkctl->nid; | |
287 | long batch_size = shrinker->batch ? shrinker->batch | |
288 | : SHRINK_BATCH; | |
289 | ||
d5bc5fd3 VD |
290 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
291 | if (freeable == 0) | |
1d3d4437 GC |
292 | return 0; |
293 | ||
294 | /* | |
295 | * copy the current shrinker scan count into a local variable | |
296 | * and zero it so that other concurrent shrinker invocations | |
297 | * don't also do this scanning work. | |
298 | */ | |
299 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
300 | ||
301 | total_scan = nr; | |
6b4f7799 | 302 | delta = (4 * nr_scanned) / shrinker->seeks; |
d5bc5fd3 | 303 | delta *= freeable; |
6b4f7799 | 304 | do_div(delta, nr_eligible + 1); |
1d3d4437 GC |
305 | total_scan += delta; |
306 | if (total_scan < 0) { | |
8612c663 | 307 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
a0b02131 | 308 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 309 | total_scan = freeable; |
1d3d4437 GC |
310 | } |
311 | ||
312 | /* | |
313 | * We need to avoid excessive windup on filesystem shrinkers | |
314 | * due to large numbers of GFP_NOFS allocations causing the | |
315 | * shrinkers to return -1 all the time. This results in a large | |
316 | * nr being built up so when a shrink that can do some work | |
317 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 318 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
319 | * memory. |
320 | * | |
321 | * Hence only allow the shrinker to scan the entire cache when | |
322 | * a large delta change is calculated directly. | |
323 | */ | |
d5bc5fd3 VD |
324 | if (delta < freeable / 4) |
325 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
326 | |
327 | /* | |
328 | * Avoid risking looping forever due to too large nr value: | |
329 | * never try to free more than twice the estimate number of | |
330 | * freeable entries. | |
331 | */ | |
d5bc5fd3 VD |
332 | if (total_scan > freeable * 2) |
333 | total_scan = freeable * 2; | |
1d3d4437 GC |
334 | |
335 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
6b4f7799 JW |
336 | nr_scanned, nr_eligible, |
337 | freeable, delta, total_scan); | |
1d3d4437 | 338 | |
0b1fb40a VD |
339 | /* |
340 | * Normally, we should not scan less than batch_size objects in one | |
341 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
342 | * than batch_size objects in total and we are really tight on memory, | |
343 | * we will try to reclaim all available objects, otherwise we can end | |
344 | * up failing allocations although there are plenty of reclaimable | |
345 | * objects spread over several slabs with usage less than the | |
346 | * batch_size. | |
347 | * | |
348 | * We detect the "tight on memory" situations by looking at the total | |
349 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 350 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
351 | * scanning at high prio and therefore should try to reclaim as much as |
352 | * possible. | |
353 | */ | |
354 | while (total_scan >= batch_size || | |
d5bc5fd3 | 355 | total_scan >= freeable) { |
a0b02131 | 356 | unsigned long ret; |
0b1fb40a | 357 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 358 | |
0b1fb40a | 359 | shrinkctl->nr_to_scan = nr_to_scan; |
a0b02131 DC |
360 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
361 | if (ret == SHRINK_STOP) | |
362 | break; | |
363 | freed += ret; | |
1d3d4437 | 364 | |
0b1fb40a VD |
365 | count_vm_events(SLABS_SCANNED, nr_to_scan); |
366 | total_scan -= nr_to_scan; | |
1d3d4437 GC |
367 | |
368 | cond_resched(); | |
369 | } | |
370 | ||
371 | /* | |
372 | * move the unused scan count back into the shrinker in a | |
373 | * manner that handles concurrent updates. If we exhausted the | |
374 | * scan, there is no need to do an update. | |
375 | */ | |
376 | if (total_scan > 0) | |
377 | new_nr = atomic_long_add_return(total_scan, | |
378 | &shrinker->nr_deferred[nid]); | |
379 | else | |
380 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
381 | ||
df9024a8 | 382 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 383 | return freed; |
1495f230 YH |
384 | } |
385 | ||
6b4f7799 | 386 | /** |
cb731d6c | 387 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
388 | * @gfp_mask: allocation context |
389 | * @nid: node whose slab caches to target | |
cb731d6c | 390 | * @memcg: memory cgroup whose slab caches to target |
6b4f7799 JW |
391 | * @nr_scanned: pressure numerator |
392 | * @nr_eligible: pressure denominator | |
1da177e4 | 393 | * |
6b4f7799 | 394 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 395 | * |
6b4f7799 JW |
396 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
397 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 398 | * |
cb731d6c VD |
399 | * @memcg specifies the memory cgroup to target. If it is not NULL, |
400 | * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan | |
0fc9f58a VD |
401 | * objects from the memory cgroup specified. Otherwise, only unaware |
402 | * shrinkers are called. | |
cb731d6c | 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 | |
0fc9f58a | 422 | if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) |
cb731d6c VD |
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 | |
0fc9f58a VD |
446 | /* |
447 | * If kernel memory accounting is disabled, we ignore | |
448 | * SHRINKER_MEMCG_AWARE flag and call all shrinkers | |
449 | * passing NULL for memcg. | |
450 | */ | |
451 | if (memcg_kmem_enabled() && | |
452 | !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
cb731d6c VD |
453 | continue; |
454 | ||
6b4f7799 JW |
455 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
456 | sc.nid = 0; | |
1da177e4 | 457 | |
cb731d6c | 458 | freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); |
1da177e4 | 459 | } |
6b4f7799 | 460 | |
1da177e4 | 461 | up_read(&shrinker_rwsem); |
f06590bd MK |
462 | out: |
463 | cond_resched(); | |
24f7c6b9 | 464 | return freed; |
1da177e4 LT |
465 | } |
466 | ||
cb731d6c VD |
467 | void drop_slab_node(int nid) |
468 | { | |
469 | unsigned long freed; | |
470 | ||
471 | do { | |
472 | struct mem_cgroup *memcg = NULL; | |
473 | ||
474 | freed = 0; | |
475 | do { | |
476 | freed += shrink_slab(GFP_KERNEL, nid, memcg, | |
477 | 1000, 1000); | |
478 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
479 | } while (freed > 10); | |
480 | } | |
481 | ||
482 | void drop_slab(void) | |
483 | { | |
484 | int nid; | |
485 | ||
486 | for_each_online_node(nid) | |
487 | drop_slab_node(nid); | |
488 | } | |
489 | ||
1da177e4 LT |
490 | static inline int is_page_cache_freeable(struct page *page) |
491 | { | |
ceddc3a5 JW |
492 | /* |
493 | * A freeable page cache page is referenced only by the caller | |
494 | * that isolated the page, the page cache radix tree and | |
495 | * optional buffer heads at page->private. | |
496 | */ | |
edcf4748 | 497 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
498 | } |
499 | ||
703c2708 | 500 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
1da177e4 | 501 | { |
930d9152 | 502 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 503 | return 1; |
703c2708 | 504 | if (!inode_write_congested(inode)) |
1da177e4 | 505 | return 1; |
703c2708 | 506 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
507 | return 1; |
508 | return 0; | |
509 | } | |
510 | ||
511 | /* | |
512 | * We detected a synchronous write error writing a page out. Probably | |
513 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
514 | * fsync(), msync() or close(). | |
515 | * | |
516 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
517 | * prevents it from being freed up. But we have a ref on the page and once | |
518 | * that page is locked, the mapping is pinned. | |
519 | * | |
520 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
521 | * __GFP_FS. | |
522 | */ | |
523 | static void handle_write_error(struct address_space *mapping, | |
524 | struct page *page, int error) | |
525 | { | |
7eaceacc | 526 | lock_page(page); |
3e9f45bd GC |
527 | if (page_mapping(page) == mapping) |
528 | mapping_set_error(mapping, error); | |
1da177e4 LT |
529 | unlock_page(page); |
530 | } | |
531 | ||
04e62a29 CL |
532 | /* possible outcome of pageout() */ |
533 | typedef enum { | |
534 | /* failed to write page out, page is locked */ | |
535 | PAGE_KEEP, | |
536 | /* move page to the active list, page is locked */ | |
537 | PAGE_ACTIVATE, | |
538 | /* page has been sent to the disk successfully, page is unlocked */ | |
539 | PAGE_SUCCESS, | |
540 | /* page is clean and locked */ | |
541 | PAGE_CLEAN, | |
542 | } pageout_t; | |
543 | ||
1da177e4 | 544 | /* |
1742f19f AM |
545 | * pageout is called by shrink_page_list() for each dirty page. |
546 | * Calls ->writepage(). | |
1da177e4 | 547 | */ |
c661b078 | 548 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 549 | struct scan_control *sc) |
1da177e4 LT |
550 | { |
551 | /* | |
552 | * If the page is dirty, only perform writeback if that write | |
553 | * will be non-blocking. To prevent this allocation from being | |
554 | * stalled by pagecache activity. But note that there may be | |
555 | * stalls if we need to run get_block(). We could test | |
556 | * PagePrivate for that. | |
557 | * | |
8174202b | 558 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
559 | * this page's queue, we can perform writeback even if that |
560 | * will block. | |
561 | * | |
562 | * If the page is swapcache, write it back even if that would | |
563 | * block, for some throttling. This happens by accident, because | |
564 | * swap_backing_dev_info is bust: it doesn't reflect the | |
565 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
566 | */ |
567 | if (!is_page_cache_freeable(page)) | |
568 | return PAGE_KEEP; | |
569 | if (!mapping) { | |
570 | /* | |
571 | * Some data journaling orphaned pages can have | |
572 | * page->mapping == NULL while being dirty with clean buffers. | |
573 | */ | |
266cf658 | 574 | if (page_has_private(page)) { |
1da177e4 LT |
575 | if (try_to_free_buffers(page)) { |
576 | ClearPageDirty(page); | |
b1de0d13 | 577 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
578 | return PAGE_CLEAN; |
579 | } | |
580 | } | |
581 | return PAGE_KEEP; | |
582 | } | |
583 | if (mapping->a_ops->writepage == NULL) | |
584 | return PAGE_ACTIVATE; | |
703c2708 | 585 | if (!may_write_to_inode(mapping->host, sc)) |
1da177e4 LT |
586 | return PAGE_KEEP; |
587 | ||
588 | if (clear_page_dirty_for_io(page)) { | |
589 | int res; | |
590 | struct writeback_control wbc = { | |
591 | .sync_mode = WB_SYNC_NONE, | |
592 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
593 | .range_start = 0, |
594 | .range_end = LLONG_MAX, | |
1da177e4 LT |
595 | .for_reclaim = 1, |
596 | }; | |
597 | ||
598 | SetPageReclaim(page); | |
599 | res = mapping->a_ops->writepage(page, &wbc); | |
600 | if (res < 0) | |
601 | handle_write_error(mapping, page, res); | |
994fc28c | 602 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
603 | ClearPageReclaim(page); |
604 | return PAGE_ACTIVATE; | |
605 | } | |
c661b078 | 606 | |
1da177e4 LT |
607 | if (!PageWriteback(page)) { |
608 | /* synchronous write or broken a_ops? */ | |
609 | ClearPageReclaim(page); | |
610 | } | |
3aa23851 | 611 | trace_mm_vmscan_writepage(page); |
e129b5c2 | 612 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
613 | return PAGE_SUCCESS; |
614 | } | |
615 | ||
616 | return PAGE_CLEAN; | |
617 | } | |
618 | ||
a649fd92 | 619 | /* |
e286781d NP |
620 | * Same as remove_mapping, but if the page is removed from the mapping, it |
621 | * gets returned with a refcount of 0. | |
a649fd92 | 622 | */ |
a528910e JW |
623 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
624 | bool reclaimed) | |
49d2e9cc | 625 | { |
c4843a75 | 626 | unsigned long flags; |
c4843a75 | 627 | |
28e4d965 NP |
628 | BUG_ON(!PageLocked(page)); |
629 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 630 | |
c4843a75 | 631 | spin_lock_irqsave(&mapping->tree_lock, flags); |
49d2e9cc | 632 | /* |
0fd0e6b0 NP |
633 | * The non racy check for a busy page. |
634 | * | |
635 | * Must be careful with the order of the tests. When someone has | |
636 | * a ref to the page, it may be possible that they dirty it then | |
637 | * drop the reference. So if PageDirty is tested before page_count | |
638 | * here, then the following race may occur: | |
639 | * | |
640 | * get_user_pages(&page); | |
641 | * [user mapping goes away] | |
642 | * write_to(page); | |
643 | * !PageDirty(page) [good] | |
644 | * SetPageDirty(page); | |
645 | * put_page(page); | |
646 | * !page_count(page) [good, discard it] | |
647 | * | |
648 | * [oops, our write_to data is lost] | |
649 | * | |
650 | * Reversing the order of the tests ensures such a situation cannot | |
651 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 652 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
653 | * |
654 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
655 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 656 | */ |
fe896d18 | 657 | if (!page_ref_freeze(page, 2)) |
49d2e9cc | 658 | goto cannot_free; |
e286781d NP |
659 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
660 | if (unlikely(PageDirty(page))) { | |
fe896d18 | 661 | page_ref_unfreeze(page, 2); |
49d2e9cc | 662 | goto cannot_free; |
e286781d | 663 | } |
49d2e9cc CL |
664 | |
665 | if (PageSwapCache(page)) { | |
666 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 667 | mem_cgroup_swapout(page, swap); |
49d2e9cc | 668 | __delete_from_swap_cache(page); |
c4843a75 | 669 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
0a31bc97 | 670 | swapcache_free(swap); |
e286781d | 671 | } else { |
6072d13c | 672 | void (*freepage)(struct page *); |
a528910e | 673 | void *shadow = NULL; |
6072d13c LT |
674 | |
675 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
676 | /* |
677 | * Remember a shadow entry for reclaimed file cache in | |
678 | * order to detect refaults, thus thrashing, later on. | |
679 | * | |
680 | * But don't store shadows in an address space that is | |
681 | * already exiting. This is not just an optizimation, | |
682 | * inode reclaim needs to empty out the radix tree or | |
683 | * the nodes are lost. Don't plant shadows behind its | |
684 | * back. | |
f9fe48be RZ |
685 | * |
686 | * We also don't store shadows for DAX mappings because the | |
687 | * only page cache pages found in these are zero pages | |
688 | * covering holes, and because we don't want to mix DAX | |
689 | * exceptional entries and shadow exceptional entries in the | |
690 | * same page_tree. | |
a528910e JW |
691 | */ |
692 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 693 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
a528910e | 694 | shadow = workingset_eviction(mapping, page); |
62cccb8c | 695 | __delete_from_page_cache(page, shadow); |
c4843a75 | 696 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
6072d13c LT |
697 | |
698 | if (freepage != NULL) | |
699 | freepage(page); | |
49d2e9cc CL |
700 | } |
701 | ||
49d2e9cc CL |
702 | return 1; |
703 | ||
704 | cannot_free: | |
c4843a75 | 705 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
49d2e9cc CL |
706 | return 0; |
707 | } | |
708 | ||
e286781d NP |
709 | /* |
710 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
711 | * someone else has a ref on the page, abort and return 0. If it was | |
712 | * successfully detached, return 1. Assumes the caller has a single ref on | |
713 | * this page. | |
714 | */ | |
715 | int remove_mapping(struct address_space *mapping, struct page *page) | |
716 | { | |
a528910e | 717 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
718 | /* |
719 | * Unfreezing the refcount with 1 rather than 2 effectively | |
720 | * drops the pagecache ref for us without requiring another | |
721 | * atomic operation. | |
722 | */ | |
fe896d18 | 723 | page_ref_unfreeze(page, 1); |
e286781d NP |
724 | return 1; |
725 | } | |
726 | return 0; | |
727 | } | |
728 | ||
894bc310 LS |
729 | /** |
730 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
731 | * @page: page to be put back to appropriate lru list | |
732 | * | |
733 | * Add previously isolated @page to appropriate LRU list. | |
734 | * Page may still be unevictable for other reasons. | |
735 | * | |
736 | * lru_lock must not be held, interrupts must be enabled. | |
737 | */ | |
894bc310 LS |
738 | void putback_lru_page(struct page *page) |
739 | { | |
0ec3b74c | 740 | bool is_unevictable; |
bbfd28ee | 741 | int was_unevictable = PageUnevictable(page); |
894bc310 | 742 | |
309381fe | 743 | VM_BUG_ON_PAGE(PageLRU(page), page); |
894bc310 LS |
744 | |
745 | redo: | |
746 | ClearPageUnevictable(page); | |
747 | ||
39b5f29a | 748 | if (page_evictable(page)) { |
894bc310 LS |
749 | /* |
750 | * For evictable pages, we can use the cache. | |
751 | * In event of a race, worst case is we end up with an | |
752 | * unevictable page on [in]active list. | |
753 | * We know how to handle that. | |
754 | */ | |
0ec3b74c | 755 | is_unevictable = false; |
c53954a0 | 756 | lru_cache_add(page); |
894bc310 LS |
757 | } else { |
758 | /* | |
759 | * Put unevictable pages directly on zone's unevictable | |
760 | * list. | |
761 | */ | |
0ec3b74c | 762 | is_unevictable = true; |
894bc310 | 763 | add_page_to_unevictable_list(page); |
6a7b9548 | 764 | /* |
21ee9f39 MK |
765 | * When racing with an mlock or AS_UNEVICTABLE clearing |
766 | * (page is unlocked) make sure that if the other thread | |
767 | * does not observe our setting of PG_lru and fails | |
24513264 | 768 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 769 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
770 | * the page back to the evictable list. |
771 | * | |
21ee9f39 | 772 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
773 | */ |
774 | smp_mb(); | |
894bc310 | 775 | } |
894bc310 LS |
776 | |
777 | /* | |
778 | * page's status can change while we move it among lru. If an evictable | |
779 | * page is on unevictable list, it never be freed. To avoid that, | |
780 | * check after we added it to the list, again. | |
781 | */ | |
0ec3b74c | 782 | if (is_unevictable && page_evictable(page)) { |
894bc310 LS |
783 | if (!isolate_lru_page(page)) { |
784 | put_page(page); | |
785 | goto redo; | |
786 | } | |
787 | /* This means someone else dropped this page from LRU | |
788 | * So, it will be freed or putback to LRU again. There is | |
789 | * nothing to do here. | |
790 | */ | |
791 | } | |
792 | ||
0ec3b74c | 793 | if (was_unevictable && !is_unevictable) |
bbfd28ee | 794 | count_vm_event(UNEVICTABLE_PGRESCUED); |
0ec3b74c | 795 | else if (!was_unevictable && is_unevictable) |
bbfd28ee LS |
796 | count_vm_event(UNEVICTABLE_PGCULLED); |
797 | ||
894bc310 LS |
798 | put_page(page); /* drop ref from isolate */ |
799 | } | |
800 | ||
dfc8d636 JW |
801 | enum page_references { |
802 | PAGEREF_RECLAIM, | |
803 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 804 | PAGEREF_KEEP, |
dfc8d636 JW |
805 | PAGEREF_ACTIVATE, |
806 | }; | |
807 | ||
808 | static enum page_references page_check_references(struct page *page, | |
809 | struct scan_control *sc) | |
810 | { | |
64574746 | 811 | int referenced_ptes, referenced_page; |
dfc8d636 | 812 | unsigned long vm_flags; |
dfc8d636 | 813 | |
c3ac9a8a JW |
814 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
815 | &vm_flags); | |
64574746 | 816 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 817 | |
dfc8d636 JW |
818 | /* |
819 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
820 | * move the page to the unevictable list. | |
821 | */ | |
822 | if (vm_flags & VM_LOCKED) | |
823 | return PAGEREF_RECLAIM; | |
824 | ||
64574746 | 825 | if (referenced_ptes) { |
e4898273 | 826 | if (PageSwapBacked(page)) |
64574746 JW |
827 | return PAGEREF_ACTIVATE; |
828 | /* | |
829 | * All mapped pages start out with page table | |
830 | * references from the instantiating fault, so we need | |
831 | * to look twice if a mapped file page is used more | |
832 | * than once. | |
833 | * | |
834 | * Mark it and spare it for another trip around the | |
835 | * inactive list. Another page table reference will | |
836 | * lead to its activation. | |
837 | * | |
838 | * Note: the mark is set for activated pages as well | |
839 | * so that recently deactivated but used pages are | |
840 | * quickly recovered. | |
841 | */ | |
842 | SetPageReferenced(page); | |
843 | ||
34dbc67a | 844 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
845 | return PAGEREF_ACTIVATE; |
846 | ||
c909e993 KK |
847 | /* |
848 | * Activate file-backed executable pages after first usage. | |
849 | */ | |
850 | if (vm_flags & VM_EXEC) | |
851 | return PAGEREF_ACTIVATE; | |
852 | ||
64574746 JW |
853 | return PAGEREF_KEEP; |
854 | } | |
dfc8d636 JW |
855 | |
856 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 857 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
858 | return PAGEREF_RECLAIM_CLEAN; |
859 | ||
860 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
861 | } |
862 | ||
e2be15f6 MG |
863 | /* Check if a page is dirty or under writeback */ |
864 | static void page_check_dirty_writeback(struct page *page, | |
865 | bool *dirty, bool *writeback) | |
866 | { | |
b4597226 MG |
867 | struct address_space *mapping; |
868 | ||
e2be15f6 MG |
869 | /* |
870 | * Anonymous pages are not handled by flushers and must be written | |
871 | * from reclaim context. Do not stall reclaim based on them | |
872 | */ | |
873 | if (!page_is_file_cache(page)) { | |
874 | *dirty = false; | |
875 | *writeback = false; | |
876 | return; | |
877 | } | |
878 | ||
879 | /* By default assume that the page flags are accurate */ | |
880 | *dirty = PageDirty(page); | |
881 | *writeback = PageWriteback(page); | |
b4597226 MG |
882 | |
883 | /* Verify dirty/writeback state if the filesystem supports it */ | |
884 | if (!page_has_private(page)) | |
885 | return; | |
886 | ||
887 | mapping = page_mapping(page); | |
888 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
889 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
890 | } |
891 | ||
1da177e4 | 892 | /* |
1742f19f | 893 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 894 | */ |
1742f19f | 895 | static unsigned long shrink_page_list(struct list_head *page_list, |
599d0c95 | 896 | struct pglist_data *pgdat, |
f84f6e2b | 897 | struct scan_control *sc, |
02c6de8d | 898 | enum ttu_flags ttu_flags, |
8e950282 | 899 | unsigned long *ret_nr_dirty, |
d43006d5 | 900 | unsigned long *ret_nr_unqueued_dirty, |
8e950282 | 901 | unsigned long *ret_nr_congested, |
02c6de8d | 902 | unsigned long *ret_nr_writeback, |
b1a6f21e | 903 | unsigned long *ret_nr_immediate, |
02c6de8d | 904 | bool force_reclaim) |
1da177e4 LT |
905 | { |
906 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 907 | LIST_HEAD(free_pages); |
1da177e4 | 908 | int pgactivate = 0; |
d43006d5 | 909 | unsigned long nr_unqueued_dirty = 0; |
0e093d99 MG |
910 | unsigned long nr_dirty = 0; |
911 | unsigned long nr_congested = 0; | |
05ff5137 | 912 | unsigned long nr_reclaimed = 0; |
92df3a72 | 913 | unsigned long nr_writeback = 0; |
b1a6f21e | 914 | unsigned long nr_immediate = 0; |
1da177e4 LT |
915 | |
916 | cond_resched(); | |
917 | ||
1da177e4 LT |
918 | while (!list_empty(page_list)) { |
919 | struct address_space *mapping; | |
920 | struct page *page; | |
921 | int may_enter_fs; | |
02c6de8d | 922 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
e2be15f6 | 923 | bool dirty, writeback; |
854e9ed0 MK |
924 | bool lazyfree = false; |
925 | int ret = SWAP_SUCCESS; | |
1da177e4 LT |
926 | |
927 | cond_resched(); | |
928 | ||
929 | page = lru_to_page(page_list); | |
930 | list_del(&page->lru); | |
931 | ||
529ae9aa | 932 | if (!trylock_page(page)) |
1da177e4 LT |
933 | goto keep; |
934 | ||
309381fe | 935 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
936 | |
937 | sc->nr_scanned++; | |
80e43426 | 938 | |
39b5f29a | 939 | if (unlikely(!page_evictable(page))) |
b291f000 | 940 | goto cull_mlocked; |
894bc310 | 941 | |
a6dc60f8 | 942 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
943 | goto keep_locked; |
944 | ||
1da177e4 LT |
945 | /* Double the slab pressure for mapped and swapcache pages */ |
946 | if (page_mapped(page) || PageSwapCache(page)) | |
947 | sc->nr_scanned++; | |
948 | ||
c661b078 AW |
949 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
950 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
951 | ||
e2be15f6 MG |
952 | /* |
953 | * The number of dirty pages determines if a zone is marked | |
954 | * reclaim_congested which affects wait_iff_congested. kswapd | |
955 | * will stall and start writing pages if the tail of the LRU | |
956 | * is all dirty unqueued pages. | |
957 | */ | |
958 | page_check_dirty_writeback(page, &dirty, &writeback); | |
959 | if (dirty || writeback) | |
960 | nr_dirty++; | |
961 | ||
962 | if (dirty && !writeback) | |
963 | nr_unqueued_dirty++; | |
964 | ||
d04e8acd MG |
965 | /* |
966 | * Treat this page as congested if the underlying BDI is or if | |
967 | * pages are cycling through the LRU so quickly that the | |
968 | * pages marked for immediate reclaim are making it to the | |
969 | * end of the LRU a second time. | |
970 | */ | |
e2be15f6 | 971 | mapping = page_mapping(page); |
1da58ee2 | 972 | if (((dirty || writeback) && mapping && |
703c2708 | 973 | inode_write_congested(mapping->host)) || |
d04e8acd | 974 | (writeback && PageReclaim(page))) |
e2be15f6 MG |
975 | nr_congested++; |
976 | ||
283aba9f MG |
977 | /* |
978 | * If a page at the tail of the LRU is under writeback, there | |
979 | * are three cases to consider. | |
980 | * | |
981 | * 1) If reclaim is encountering an excessive number of pages | |
982 | * under writeback and this page is both under writeback and | |
983 | * PageReclaim then it indicates that pages are being queued | |
984 | * for IO but are being recycled through the LRU before the | |
985 | * IO can complete. Waiting on the page itself risks an | |
986 | * indefinite stall if it is impossible to writeback the | |
987 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
988 | * note that the LRU is being scanned too quickly and the |
989 | * caller can stall after page list has been processed. | |
283aba9f | 990 | * |
97c9341f | 991 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
992 | * not marked for immediate reclaim, or the caller does not |
993 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
994 | * not to fs). In this case mark the page for immediate | |
97c9341f | 995 | * reclaim and continue scanning. |
283aba9f | 996 | * |
ecf5fc6e MH |
997 | * Require may_enter_fs because we would wait on fs, which |
998 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
999 | * enter reclaim, and deadlock if it waits on a page for |
1000 | * which it is needed to do the write (loop masks off | |
1001 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1002 | * would probably show more reasons. | |
1003 | * | |
7fadc820 | 1004 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1005 | * PageReclaim. memcg does not have any dirty pages |
1006 | * throttling so we could easily OOM just because too many | |
1007 | * pages are in writeback and there is nothing else to | |
1008 | * reclaim. Wait for the writeback to complete. | |
1009 | */ | |
c661b078 | 1010 | if (PageWriteback(page)) { |
283aba9f MG |
1011 | /* Case 1 above */ |
1012 | if (current_is_kswapd() && | |
1013 | PageReclaim(page) && | |
599d0c95 | 1014 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
b1a6f21e MG |
1015 | nr_immediate++; |
1016 | goto keep_locked; | |
283aba9f MG |
1017 | |
1018 | /* Case 2 above */ | |
97c9341f | 1019 | } else if (sane_reclaim(sc) || |
ecf5fc6e | 1020 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1021 | /* |
1022 | * This is slightly racy - end_page_writeback() | |
1023 | * might have just cleared PageReclaim, then | |
1024 | * setting PageReclaim here end up interpreted | |
1025 | * as PageReadahead - but that does not matter | |
1026 | * enough to care. What we do want is for this | |
1027 | * page to have PageReclaim set next time memcg | |
1028 | * reclaim reaches the tests above, so it will | |
1029 | * then wait_on_page_writeback() to avoid OOM; | |
1030 | * and it's also appropriate in global reclaim. | |
1031 | */ | |
1032 | SetPageReclaim(page); | |
e62e384e | 1033 | nr_writeback++; |
c3b94f44 | 1034 | goto keep_locked; |
283aba9f MG |
1035 | |
1036 | /* Case 3 above */ | |
1037 | } else { | |
7fadc820 | 1038 | unlock_page(page); |
283aba9f | 1039 | wait_on_page_writeback(page); |
7fadc820 HD |
1040 | /* then go back and try same page again */ |
1041 | list_add_tail(&page->lru, page_list); | |
1042 | continue; | |
e62e384e | 1043 | } |
c661b078 | 1044 | } |
1da177e4 | 1045 | |
02c6de8d MK |
1046 | if (!force_reclaim) |
1047 | references = page_check_references(page, sc); | |
1048 | ||
dfc8d636 JW |
1049 | switch (references) { |
1050 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1051 | goto activate_locked; |
64574746 JW |
1052 | case PAGEREF_KEEP: |
1053 | goto keep_locked; | |
dfc8d636 JW |
1054 | case PAGEREF_RECLAIM: |
1055 | case PAGEREF_RECLAIM_CLEAN: | |
1056 | ; /* try to reclaim the page below */ | |
1057 | } | |
1da177e4 | 1058 | |
1da177e4 LT |
1059 | /* |
1060 | * Anonymous process memory has backing store? | |
1061 | * Try to allocate it some swap space here. | |
1062 | */ | |
b291f000 | 1063 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
1064 | if (!(sc->gfp_mask & __GFP_IO)) |
1065 | goto keep_locked; | |
5bc7b8ac | 1066 | if (!add_to_swap(page, page_list)) |
1da177e4 | 1067 | goto activate_locked; |
854e9ed0 | 1068 | lazyfree = true; |
63eb6b93 | 1069 | may_enter_fs = 1; |
1da177e4 | 1070 | |
e2be15f6 MG |
1071 | /* Adding to swap updated mapping */ |
1072 | mapping = page_mapping(page); | |
7751b2da KS |
1073 | } else if (unlikely(PageTransHuge(page))) { |
1074 | /* Split file THP */ | |
1075 | if (split_huge_page_to_list(page, page_list)) | |
1076 | goto keep_locked; | |
e2be15f6 | 1077 | } |
1da177e4 | 1078 | |
7751b2da KS |
1079 | VM_BUG_ON_PAGE(PageTransHuge(page), page); |
1080 | ||
1da177e4 LT |
1081 | /* |
1082 | * The page is mapped into the page tables of one or more | |
1083 | * processes. Try to unmap it here. | |
1084 | */ | |
1085 | if (page_mapped(page) && mapping) { | |
854e9ed0 MK |
1086 | switch (ret = try_to_unmap(page, lazyfree ? |
1087 | (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) : | |
1088 | (ttu_flags | TTU_BATCH_FLUSH))) { | |
1da177e4 LT |
1089 | case SWAP_FAIL: |
1090 | goto activate_locked; | |
1091 | case SWAP_AGAIN: | |
1092 | goto keep_locked; | |
b291f000 NP |
1093 | case SWAP_MLOCK: |
1094 | goto cull_mlocked; | |
854e9ed0 MK |
1095 | case SWAP_LZFREE: |
1096 | goto lazyfree; | |
1da177e4 LT |
1097 | case SWAP_SUCCESS: |
1098 | ; /* try to free the page below */ | |
1099 | } | |
1100 | } | |
1101 | ||
1102 | if (PageDirty(page)) { | |
ee72886d MG |
1103 | /* |
1104 | * Only kswapd can writeback filesystem pages to | |
d43006d5 MG |
1105 | * avoid risk of stack overflow but only writeback |
1106 | * if many dirty pages have been encountered. | |
ee72886d | 1107 | */ |
f84f6e2b | 1108 | if (page_is_file_cache(page) && |
9e3b2f8c | 1109 | (!current_is_kswapd() || |
599d0c95 | 1110 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { |
49ea7eb6 MG |
1111 | /* |
1112 | * Immediately reclaim when written back. | |
1113 | * Similar in principal to deactivate_page() | |
1114 | * except we already have the page isolated | |
1115 | * and know it's dirty | |
1116 | */ | |
1117 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
1118 | SetPageReclaim(page); | |
1119 | ||
ee72886d MG |
1120 | goto keep_locked; |
1121 | } | |
1122 | ||
dfc8d636 | 1123 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1124 | goto keep_locked; |
4dd4b920 | 1125 | if (!may_enter_fs) |
1da177e4 | 1126 | goto keep_locked; |
52a8363e | 1127 | if (!sc->may_writepage) |
1da177e4 LT |
1128 | goto keep_locked; |
1129 | ||
d950c947 MG |
1130 | /* |
1131 | * Page is dirty. Flush the TLB if a writable entry | |
1132 | * potentially exists to avoid CPU writes after IO | |
1133 | * starts and then write it out here. | |
1134 | */ | |
1135 | try_to_unmap_flush_dirty(); | |
7d3579e8 | 1136 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1137 | case PAGE_KEEP: |
1138 | goto keep_locked; | |
1139 | case PAGE_ACTIVATE: | |
1140 | goto activate_locked; | |
1141 | case PAGE_SUCCESS: | |
7d3579e8 | 1142 | if (PageWriteback(page)) |
41ac1999 | 1143 | goto keep; |
7d3579e8 | 1144 | if (PageDirty(page)) |
1da177e4 | 1145 | goto keep; |
7d3579e8 | 1146 | |
1da177e4 LT |
1147 | /* |
1148 | * A synchronous write - probably a ramdisk. Go | |
1149 | * ahead and try to reclaim the page. | |
1150 | */ | |
529ae9aa | 1151 | if (!trylock_page(page)) |
1da177e4 LT |
1152 | goto keep; |
1153 | if (PageDirty(page) || PageWriteback(page)) | |
1154 | goto keep_locked; | |
1155 | mapping = page_mapping(page); | |
1156 | case PAGE_CLEAN: | |
1157 | ; /* try to free the page below */ | |
1158 | } | |
1159 | } | |
1160 | ||
1161 | /* | |
1162 | * If the page has buffers, try to free the buffer mappings | |
1163 | * associated with this page. If we succeed we try to free | |
1164 | * the page as well. | |
1165 | * | |
1166 | * We do this even if the page is PageDirty(). | |
1167 | * try_to_release_page() does not perform I/O, but it is | |
1168 | * possible for a page to have PageDirty set, but it is actually | |
1169 | * clean (all its buffers are clean). This happens if the | |
1170 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1171 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1172 | * try_to_release_page() will discover that cleanness and will |
1173 | * drop the buffers and mark the page clean - it can be freed. | |
1174 | * | |
1175 | * Rarely, pages can have buffers and no ->mapping. These are | |
1176 | * the pages which were not successfully invalidated in | |
1177 | * truncate_complete_page(). We try to drop those buffers here | |
1178 | * and if that worked, and the page is no longer mapped into | |
1179 | * process address space (page_count == 1) it can be freed. | |
1180 | * Otherwise, leave the page on the LRU so it is swappable. | |
1181 | */ | |
266cf658 | 1182 | if (page_has_private(page)) { |
1da177e4 LT |
1183 | if (!try_to_release_page(page, sc->gfp_mask)) |
1184 | goto activate_locked; | |
e286781d NP |
1185 | if (!mapping && page_count(page) == 1) { |
1186 | unlock_page(page); | |
1187 | if (put_page_testzero(page)) | |
1188 | goto free_it; | |
1189 | else { | |
1190 | /* | |
1191 | * rare race with speculative reference. | |
1192 | * the speculative reference will free | |
1193 | * this page shortly, so we may | |
1194 | * increment nr_reclaimed here (and | |
1195 | * leave it off the LRU). | |
1196 | */ | |
1197 | nr_reclaimed++; | |
1198 | continue; | |
1199 | } | |
1200 | } | |
1da177e4 LT |
1201 | } |
1202 | ||
854e9ed0 | 1203 | lazyfree: |
a528910e | 1204 | if (!mapping || !__remove_mapping(mapping, page, true)) |
49d2e9cc | 1205 | goto keep_locked; |
1da177e4 | 1206 | |
a978d6f5 NP |
1207 | /* |
1208 | * At this point, we have no other references and there is | |
1209 | * no way to pick any more up (removed from LRU, removed | |
1210 | * from pagecache). Can use non-atomic bitops now (and | |
1211 | * we obviously don't have to worry about waking up a process | |
1212 | * waiting on the page lock, because there are no references. | |
1213 | */ | |
48c935ad | 1214 | __ClearPageLocked(page); |
e286781d | 1215 | free_it: |
854e9ed0 MK |
1216 | if (ret == SWAP_LZFREE) |
1217 | count_vm_event(PGLAZYFREED); | |
1218 | ||
05ff5137 | 1219 | nr_reclaimed++; |
abe4c3b5 MG |
1220 | |
1221 | /* | |
1222 | * Is there need to periodically free_page_list? It would | |
1223 | * appear not as the counts should be low | |
1224 | */ | |
1225 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1226 | continue; |
1227 | ||
b291f000 | 1228 | cull_mlocked: |
63d6c5ad HD |
1229 | if (PageSwapCache(page)) |
1230 | try_to_free_swap(page); | |
b291f000 | 1231 | unlock_page(page); |
c54839a7 | 1232 | list_add(&page->lru, &ret_pages); |
b291f000 NP |
1233 | continue; |
1234 | ||
1da177e4 | 1235 | activate_locked: |
68a22394 | 1236 | /* Not a candidate for swapping, so reclaim swap space. */ |
5ccc5aba | 1237 | if (PageSwapCache(page) && mem_cgroup_swap_full(page)) |
a2c43eed | 1238 | try_to_free_swap(page); |
309381fe | 1239 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
1240 | SetPageActive(page); |
1241 | pgactivate++; | |
1242 | keep_locked: | |
1243 | unlock_page(page); | |
1244 | keep: | |
1245 | list_add(&page->lru, &ret_pages); | |
309381fe | 1246 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1247 | } |
abe4c3b5 | 1248 | |
747db954 | 1249 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1250 | try_to_unmap_flush(); |
b745bc85 | 1251 | free_hot_cold_page_list(&free_pages, true); |
abe4c3b5 | 1252 | |
1da177e4 | 1253 | list_splice(&ret_pages, page_list); |
f8891e5e | 1254 | count_vm_events(PGACTIVATE, pgactivate); |
0a31bc97 | 1255 | |
8e950282 MG |
1256 | *ret_nr_dirty += nr_dirty; |
1257 | *ret_nr_congested += nr_congested; | |
d43006d5 | 1258 | *ret_nr_unqueued_dirty += nr_unqueued_dirty; |
92df3a72 | 1259 | *ret_nr_writeback += nr_writeback; |
b1a6f21e | 1260 | *ret_nr_immediate += nr_immediate; |
05ff5137 | 1261 | return nr_reclaimed; |
1da177e4 LT |
1262 | } |
1263 | ||
02c6de8d MK |
1264 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1265 | struct list_head *page_list) | |
1266 | { | |
1267 | struct scan_control sc = { | |
1268 | .gfp_mask = GFP_KERNEL, | |
1269 | .priority = DEF_PRIORITY, | |
1270 | .may_unmap = 1, | |
1271 | }; | |
8e950282 | 1272 | unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; |
02c6de8d MK |
1273 | struct page *page, *next; |
1274 | LIST_HEAD(clean_pages); | |
1275 | ||
1276 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1277 | if (page_is_file_cache(page) && !PageDirty(page) && |
b1123ea6 | 1278 | !__PageMovable(page)) { |
02c6de8d MK |
1279 | ClearPageActive(page); |
1280 | list_move(&page->lru, &clean_pages); | |
1281 | } | |
1282 | } | |
1283 | ||
599d0c95 | 1284 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
8e950282 MG |
1285 | TTU_UNMAP|TTU_IGNORE_ACCESS, |
1286 | &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); | |
02c6de8d | 1287 | list_splice(&clean_pages, page_list); |
599d0c95 | 1288 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1289 | return ret; |
1290 | } | |
1291 | ||
5ad333eb AW |
1292 | /* |
1293 | * Attempt to remove the specified page from its LRU. Only take this page | |
1294 | * if it is of the appropriate PageActive status. Pages which are being | |
1295 | * freed elsewhere are also ignored. | |
1296 | * | |
1297 | * page: page to consider | |
1298 | * mode: one of the LRU isolation modes defined above | |
1299 | * | |
1300 | * returns 0 on success, -ve errno on failure. | |
1301 | */ | |
f3fd4a61 | 1302 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1303 | { |
1304 | int ret = -EINVAL; | |
1305 | ||
1306 | /* Only take pages on the LRU. */ | |
1307 | if (!PageLRU(page)) | |
1308 | return ret; | |
1309 | ||
e46a2879 MK |
1310 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1311 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1312 | return ret; |
1313 | ||
5ad333eb | 1314 | ret = -EBUSY; |
08e552c6 | 1315 | |
c8244935 MG |
1316 | /* |
1317 | * To minimise LRU disruption, the caller can indicate that it only | |
1318 | * wants to isolate pages it will be able to operate on without | |
1319 | * blocking - clean pages for the most part. | |
1320 | * | |
1321 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1322 | * is used by reclaim when it is cannot write to backing storage | |
1323 | * | |
1324 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1325 | * that it is possible to migrate without blocking | |
1326 | */ | |
1327 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1328 | /* All the caller can do on PageWriteback is block */ | |
1329 | if (PageWriteback(page)) | |
1330 | return ret; | |
1331 | ||
1332 | if (PageDirty(page)) { | |
1333 | struct address_space *mapping; | |
1334 | ||
1335 | /* ISOLATE_CLEAN means only clean pages */ | |
1336 | if (mode & ISOLATE_CLEAN) | |
1337 | return ret; | |
1338 | ||
1339 | /* | |
1340 | * Only pages without mappings or that have a | |
1341 | * ->migratepage callback are possible to migrate | |
1342 | * without blocking | |
1343 | */ | |
1344 | mapping = page_mapping(page); | |
1345 | if (mapping && !mapping->a_ops->migratepage) | |
1346 | return ret; | |
1347 | } | |
1348 | } | |
39deaf85 | 1349 | |
f80c0673 MK |
1350 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1351 | return ret; | |
1352 | ||
5ad333eb AW |
1353 | if (likely(get_page_unless_zero(page))) { |
1354 | /* | |
1355 | * Be careful not to clear PageLRU until after we're | |
1356 | * sure the page is not being freed elsewhere -- the | |
1357 | * page release code relies on it. | |
1358 | */ | |
1359 | ClearPageLRU(page); | |
1360 | ret = 0; | |
1361 | } | |
1362 | ||
1363 | return ret; | |
1364 | } | |
1365 | ||
1da177e4 | 1366 | /* |
a52633d8 | 1367 | * zone_lru_lock is heavily contended. Some of the functions that |
1da177e4 LT |
1368 | * shrink the lists perform better by taking out a batch of pages |
1369 | * and working on them outside the LRU lock. | |
1370 | * | |
1371 | * For pagecache intensive workloads, this function is the hottest | |
1372 | * spot in the kernel (apart from copy_*_user functions). | |
1373 | * | |
1374 | * Appropriate locks must be held before calling this function. | |
1375 | * | |
1376 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1377 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1378 | * @dst: The temp list to put pages on to. |
f626012d | 1379 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1380 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1381 | * @mode: One of the LRU isolation modes |
3cb99451 | 1382 | * @lru: LRU list id for isolating |
1da177e4 LT |
1383 | * |
1384 | * returns how many pages were moved onto *@dst. | |
1385 | */ | |
69e05944 | 1386 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1387 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1388 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1389 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1390 | { |
75b00af7 | 1391 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1392 | unsigned long nr_taken = 0; |
599d0c95 MG |
1393 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
1394 | unsigned long scan, nr_pages; | |
1da177e4 | 1395 | |
0b802f10 VD |
1396 | for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan && |
1397 | !list_empty(src); scan++) { | |
5ad333eb | 1398 | struct page *page; |
5ad333eb | 1399 | |
1da177e4 LT |
1400 | page = lru_to_page(src); |
1401 | prefetchw_prev_lru_page(page, src, flags); | |
1402 | ||
309381fe | 1403 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1404 | |
f3fd4a61 | 1405 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1406 | case 0: |
599d0c95 MG |
1407 | nr_pages = hpage_nr_pages(page); |
1408 | nr_taken += nr_pages; | |
1409 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1410 | list_move(&page->lru, dst); |
5ad333eb AW |
1411 | break; |
1412 | ||
1413 | case -EBUSY: | |
1414 | /* else it is being freed elsewhere */ | |
1415 | list_move(&page->lru, src); | |
1416 | continue; | |
46453a6e | 1417 | |
5ad333eb AW |
1418 | default: |
1419 | BUG(); | |
1420 | } | |
1da177e4 LT |
1421 | } |
1422 | ||
f626012d | 1423 | *nr_scanned = scan; |
75b00af7 HD |
1424 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, |
1425 | nr_taken, mode, is_file_lru(lru)); | |
599d0c95 MG |
1426 | for (scan = 0; scan < MAX_NR_ZONES; scan++) { |
1427 | nr_pages = nr_zone_taken[scan]; | |
1428 | if (!nr_pages) | |
1429 | continue; | |
1430 | ||
1431 | update_lru_size(lruvec, lru, scan, -nr_pages); | |
1432 | } | |
1da177e4 LT |
1433 | return nr_taken; |
1434 | } | |
1435 | ||
62695a84 NP |
1436 | /** |
1437 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1438 | * @page: page to isolate from its LRU list | |
1439 | * | |
1440 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1441 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1442 | * | |
1443 | * Returns 0 if the page was removed from an LRU list. | |
1444 | * Returns -EBUSY if the page was not on an LRU list. | |
1445 | * | |
1446 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1447 | * the active list, it will have PageActive set. If it was found on |
1448 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1449 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1450 | * |
1451 | * The vmstat statistic corresponding to the list on which the page was | |
1452 | * found will be decremented. | |
1453 | * | |
1454 | * Restrictions: | |
1455 | * (1) Must be called with an elevated refcount on the page. This is a | |
1456 | * fundamentnal difference from isolate_lru_pages (which is called | |
1457 | * without a stable reference). | |
1458 | * (2) the lru_lock must not be held. | |
1459 | * (3) interrupts must be enabled. | |
1460 | */ | |
1461 | int isolate_lru_page(struct page *page) | |
1462 | { | |
1463 | int ret = -EBUSY; | |
1464 | ||
309381fe | 1465 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1466 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1467 | |
62695a84 NP |
1468 | if (PageLRU(page)) { |
1469 | struct zone *zone = page_zone(page); | |
fa9add64 | 1470 | struct lruvec *lruvec; |
62695a84 | 1471 | |
a52633d8 | 1472 | spin_lock_irq(zone_lru_lock(zone)); |
599d0c95 | 1473 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
0c917313 | 1474 | if (PageLRU(page)) { |
894bc310 | 1475 | int lru = page_lru(page); |
0c917313 | 1476 | get_page(page); |
62695a84 | 1477 | ClearPageLRU(page); |
fa9add64 HD |
1478 | del_page_from_lru_list(page, lruvec, lru); |
1479 | ret = 0; | |
62695a84 | 1480 | } |
a52633d8 | 1481 | spin_unlock_irq(zone_lru_lock(zone)); |
62695a84 NP |
1482 | } |
1483 | return ret; | |
1484 | } | |
1485 | ||
35cd7815 | 1486 | /* |
d37dd5dc FW |
1487 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1488 | * then get resheduled. When there are massive number of tasks doing page | |
1489 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1490 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1491 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1492 | */ |
599d0c95 | 1493 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1494 | struct scan_control *sc) |
1495 | { | |
1496 | unsigned long inactive, isolated; | |
1497 | ||
1498 | if (current_is_kswapd()) | |
1499 | return 0; | |
1500 | ||
97c9341f | 1501 | if (!sane_reclaim(sc)) |
35cd7815 RR |
1502 | return 0; |
1503 | ||
1504 | if (file) { | |
599d0c95 MG |
1505 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1506 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1507 | } else { |
599d0c95 MG |
1508 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1509 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1510 | } |
1511 | ||
3cf23841 FW |
1512 | /* |
1513 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1514 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1515 | * deadlock. | |
1516 | */ | |
d0164adc | 1517 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1518 | inactive >>= 3; |
1519 | ||
35cd7815 RR |
1520 | return isolated > inactive; |
1521 | } | |
1522 | ||
66635629 | 1523 | static noinline_for_stack void |
75b00af7 | 1524 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1525 | { |
27ac81d8 | 1526 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
599d0c95 | 1527 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3f79768f | 1528 | LIST_HEAD(pages_to_free); |
66635629 | 1529 | |
66635629 MG |
1530 | /* |
1531 | * Put back any unfreeable pages. | |
1532 | */ | |
66635629 | 1533 | while (!list_empty(page_list)) { |
3f79768f | 1534 | struct page *page = lru_to_page(page_list); |
66635629 | 1535 | int lru; |
3f79768f | 1536 | |
309381fe | 1537 | VM_BUG_ON_PAGE(PageLRU(page), page); |
66635629 | 1538 | list_del(&page->lru); |
39b5f29a | 1539 | if (unlikely(!page_evictable(page))) { |
599d0c95 | 1540 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1541 | putback_lru_page(page); |
599d0c95 | 1542 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1543 | continue; |
1544 | } | |
fa9add64 | 1545 | |
599d0c95 | 1546 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1547 | |
7a608572 | 1548 | SetPageLRU(page); |
66635629 | 1549 | lru = page_lru(page); |
fa9add64 HD |
1550 | add_page_to_lru_list(page, lruvec, lru); |
1551 | ||
66635629 MG |
1552 | if (is_active_lru(lru)) { |
1553 | int file = is_file_lru(lru); | |
9992af10 RR |
1554 | int numpages = hpage_nr_pages(page); |
1555 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1556 | } |
2bcf8879 HD |
1557 | if (put_page_testzero(page)) { |
1558 | __ClearPageLRU(page); | |
1559 | __ClearPageActive(page); | |
fa9add64 | 1560 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1561 | |
1562 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1563 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1564 | mem_cgroup_uncharge(page); |
2bcf8879 | 1565 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1566 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1567 | } else |
1568 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1569 | } |
1570 | } | |
66635629 | 1571 | |
3f79768f HD |
1572 | /* |
1573 | * To save our caller's stack, now use input list for pages to free. | |
1574 | */ | |
1575 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1576 | } |
1577 | ||
399ba0b9 N |
1578 | /* |
1579 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1580 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1581 | * In that case we should only throttle if the backing device it is | |
1582 | * writing to is congested. In other cases it is safe to throttle. | |
1583 | */ | |
1584 | static int current_may_throttle(void) | |
1585 | { | |
1586 | return !(current->flags & PF_LESS_THROTTLE) || | |
1587 | current->backing_dev_info == NULL || | |
1588 | bdi_write_congested(current->backing_dev_info); | |
1589 | } | |
1590 | ||
1da177e4 | 1591 | /* |
1742f19f AM |
1592 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1593 | * of reclaimed pages | |
1da177e4 | 1594 | */ |
66635629 | 1595 | static noinline_for_stack unsigned long |
1a93be0e | 1596 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1597 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1598 | { |
1599 | LIST_HEAD(page_list); | |
e247dbce | 1600 | unsigned long nr_scanned; |
05ff5137 | 1601 | unsigned long nr_reclaimed = 0; |
e247dbce | 1602 | unsigned long nr_taken; |
8e950282 MG |
1603 | unsigned long nr_dirty = 0; |
1604 | unsigned long nr_congested = 0; | |
e2be15f6 | 1605 | unsigned long nr_unqueued_dirty = 0; |
92df3a72 | 1606 | unsigned long nr_writeback = 0; |
b1a6f21e | 1607 | unsigned long nr_immediate = 0; |
f3fd4a61 | 1608 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1609 | int file = is_file_lru(lru); |
599d0c95 | 1610 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1611 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
78dc583d | 1612 | |
599d0c95 | 1613 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
58355c78 | 1614 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1615 | |
1616 | /* We are about to die and free our memory. Return now. */ | |
1617 | if (fatal_signal_pending(current)) | |
1618 | return SWAP_CLUSTER_MAX; | |
1619 | } | |
1620 | ||
1da177e4 | 1621 | lru_add_drain(); |
f80c0673 MK |
1622 | |
1623 | if (!sc->may_unmap) | |
61317289 | 1624 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1625 | if (!sc->may_writepage) |
61317289 | 1626 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1627 | |
599d0c95 | 1628 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1629 | |
5dc35979 KK |
1630 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1631 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc | 1632 | |
599d0c95 | 1633 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1634 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1635 | |
89b5fae5 | 1636 | if (global_reclaim(sc)) { |
599d0c95 | 1637 | __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned); |
e247dbce | 1638 | if (current_is_kswapd()) |
599d0c95 | 1639 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
e247dbce | 1640 | else |
599d0c95 | 1641 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
e247dbce | 1642 | } |
599d0c95 | 1643 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1644 | |
d563c050 | 1645 | if (nr_taken == 0) |
66635629 | 1646 | return 0; |
5ad333eb | 1647 | |
599d0c95 | 1648 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP, |
8e950282 MG |
1649 | &nr_dirty, &nr_unqueued_dirty, &nr_congested, |
1650 | &nr_writeback, &nr_immediate, | |
1651 | false); | |
c661b078 | 1652 | |
599d0c95 | 1653 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 1654 | |
904249aa YH |
1655 | if (global_reclaim(sc)) { |
1656 | if (current_is_kswapd()) | |
599d0c95 | 1657 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
904249aa | 1658 | else |
599d0c95 | 1659 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
904249aa | 1660 | } |
a74609fa | 1661 | |
27ac81d8 | 1662 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1663 | |
599d0c95 | 1664 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 1665 | |
599d0c95 | 1666 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 1667 | |
747db954 | 1668 | mem_cgroup_uncharge_list(&page_list); |
b745bc85 | 1669 | free_hot_cold_page_list(&page_list, true); |
e11da5b4 | 1670 | |
92df3a72 MG |
1671 | /* |
1672 | * If reclaim is isolating dirty pages under writeback, it implies | |
1673 | * that the long-lived page allocation rate is exceeding the page | |
1674 | * laundering rate. Either the global limits are not being effective | |
1675 | * at throttling processes due to the page distribution throughout | |
1676 | * zones or there is heavy usage of a slow backing device. The | |
1677 | * only option is to throttle from reclaim context which is not ideal | |
1678 | * as there is no guarantee the dirtying process is throttled in the | |
1679 | * same way balance_dirty_pages() manages. | |
1680 | * | |
8e950282 MG |
1681 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1682 | * of pages under pages flagged for immediate reclaim and stall if any | |
1683 | * are encountered in the nr_immediate check below. | |
92df3a72 | 1684 | */ |
918fc718 | 1685 | if (nr_writeback && nr_writeback == nr_taken) |
599d0c95 | 1686 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); |
92df3a72 | 1687 | |
d43006d5 | 1688 | /* |
97c9341f TH |
1689 | * Legacy memcg will stall in page writeback so avoid forcibly |
1690 | * stalling here. | |
d43006d5 | 1691 | */ |
97c9341f | 1692 | if (sane_reclaim(sc)) { |
8e950282 MG |
1693 | /* |
1694 | * Tag a zone as congested if all the dirty pages scanned were | |
1695 | * backed by a congested BDI and wait_iff_congested will stall. | |
1696 | */ | |
1697 | if (nr_dirty && nr_dirty == nr_congested) | |
599d0c95 | 1698 | set_bit(PGDAT_CONGESTED, &pgdat->flags); |
8e950282 | 1699 | |
b1a6f21e MG |
1700 | /* |
1701 | * If dirty pages are scanned that are not queued for IO, it | |
1702 | * implies that flushers are not keeping up. In this case, flag | |
599d0c95 | 1703 | * the pgdat PGDAT_DIRTY and kswapd will start writing pages from |
57054651 | 1704 | * reclaim context. |
b1a6f21e MG |
1705 | */ |
1706 | if (nr_unqueued_dirty == nr_taken) | |
599d0c95 | 1707 | set_bit(PGDAT_DIRTY, &pgdat->flags); |
b1a6f21e MG |
1708 | |
1709 | /* | |
b738d764 LT |
1710 | * If kswapd scans pages marked marked for immediate |
1711 | * reclaim and under writeback (nr_immediate), it implies | |
1712 | * that pages are cycling through the LRU faster than | |
b1a6f21e MG |
1713 | * they are written so also forcibly stall. |
1714 | */ | |
b738d764 | 1715 | if (nr_immediate && current_may_throttle()) |
b1a6f21e | 1716 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
e2be15f6 | 1717 | } |
d43006d5 | 1718 | |
8e950282 MG |
1719 | /* |
1720 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1721 | * is congested. Allow kswapd to continue until it starts encountering | |
1722 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1723 | */ | |
399ba0b9 N |
1724 | if (!sc->hibernation_mode && !current_is_kswapd() && |
1725 | current_may_throttle()) | |
599d0c95 | 1726 | wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); |
8e950282 | 1727 | |
599d0c95 MG |
1728 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
1729 | nr_scanned, nr_reclaimed, | |
ba5e9579 | 1730 | sc->priority, file); |
05ff5137 | 1731 | return nr_reclaimed; |
1da177e4 LT |
1732 | } |
1733 | ||
1734 | /* | |
1735 | * This moves pages from the active list to the inactive list. | |
1736 | * | |
1737 | * We move them the other way if the page is referenced by one or more | |
1738 | * processes, from rmap. | |
1739 | * | |
1740 | * If the pages are mostly unmapped, the processing is fast and it is | |
a52633d8 | 1741 | * appropriate to hold zone_lru_lock across the whole operation. But if |
1da177e4 | 1742 | * the pages are mapped, the processing is slow (page_referenced()) so we |
a52633d8 | 1743 | * should drop zone_lru_lock around each page. It's impossible to balance |
1da177e4 LT |
1744 | * this, so instead we remove the pages from the LRU while processing them. |
1745 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1746 | * nobody will play with that bit on a non-LRU page. | |
1747 | * | |
0139aa7b | 1748 | * The downside is that we have to touch page->_refcount against each page. |
1da177e4 LT |
1749 | * But we had to alter page->flags anyway. |
1750 | */ | |
1cfb419b | 1751 | |
fa9add64 | 1752 | static void move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1753 | struct list_head *list, |
2bcf8879 | 1754 | struct list_head *pages_to_free, |
3eb4140f WF |
1755 | enum lru_list lru) |
1756 | { | |
599d0c95 | 1757 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3eb4140f | 1758 | unsigned long pgmoved = 0; |
3eb4140f | 1759 | struct page *page; |
fa9add64 | 1760 | int nr_pages; |
3eb4140f | 1761 | |
3eb4140f WF |
1762 | while (!list_empty(list)) { |
1763 | page = lru_to_page(list); | |
599d0c95 | 1764 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
3eb4140f | 1765 | |
309381fe | 1766 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3eb4140f WF |
1767 | SetPageLRU(page); |
1768 | ||
fa9add64 | 1769 | nr_pages = hpage_nr_pages(page); |
599d0c95 | 1770 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
925b7673 | 1771 | list_move(&page->lru, &lruvec->lists[lru]); |
fa9add64 | 1772 | pgmoved += nr_pages; |
3eb4140f | 1773 | |
2bcf8879 HD |
1774 | if (put_page_testzero(page)) { |
1775 | __ClearPageLRU(page); | |
1776 | __ClearPageActive(page); | |
fa9add64 | 1777 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1778 | |
1779 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1780 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1781 | mem_cgroup_uncharge(page); |
2bcf8879 | 1782 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1783 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1784 | } else |
1785 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1786 | } |
1787 | } | |
9d5e6a9f | 1788 | |
3eb4140f WF |
1789 | if (!is_active_lru(lru)) |
1790 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1791 | } | |
1cfb419b | 1792 | |
f626012d | 1793 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1794 | struct lruvec *lruvec, |
f16015fb | 1795 | struct scan_control *sc, |
9e3b2f8c | 1796 | enum lru_list lru) |
1da177e4 | 1797 | { |
44c241f1 | 1798 | unsigned long nr_taken; |
f626012d | 1799 | unsigned long nr_scanned; |
6fe6b7e3 | 1800 | unsigned long vm_flags; |
1da177e4 | 1801 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1802 | LIST_HEAD(l_active); |
b69408e8 | 1803 | LIST_HEAD(l_inactive); |
1da177e4 | 1804 | struct page *page; |
1a93be0e | 1805 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
44c241f1 | 1806 | unsigned long nr_rotated = 0; |
f3fd4a61 | 1807 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1808 | int file = is_file_lru(lru); |
599d0c95 | 1809 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
1810 | |
1811 | lru_add_drain(); | |
f80c0673 MK |
1812 | |
1813 | if (!sc->may_unmap) | |
61317289 | 1814 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1815 | if (!sc->may_writepage) |
61317289 | 1816 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1817 | |
599d0c95 | 1818 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 1819 | |
5dc35979 KK |
1820 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1821 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1822 | |
599d0c95 | 1823 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 1824 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1825 | |
9d5e6a9f | 1826 | if (global_reclaim(sc)) |
599d0c95 MG |
1827 | __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned); |
1828 | __count_vm_events(PGREFILL, nr_scanned); | |
9d5e6a9f | 1829 | |
599d0c95 | 1830 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 1831 | |
1da177e4 LT |
1832 | while (!list_empty(&l_hold)) { |
1833 | cond_resched(); | |
1834 | page = lru_to_page(&l_hold); | |
1835 | list_del(&page->lru); | |
7e9cd484 | 1836 | |
39b5f29a | 1837 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
1838 | putback_lru_page(page); |
1839 | continue; | |
1840 | } | |
1841 | ||
cc715d99 MG |
1842 | if (unlikely(buffer_heads_over_limit)) { |
1843 | if (page_has_private(page) && trylock_page(page)) { | |
1844 | if (page_has_private(page)) | |
1845 | try_to_release_page(page, 0); | |
1846 | unlock_page(page); | |
1847 | } | |
1848 | } | |
1849 | ||
c3ac9a8a JW |
1850 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1851 | &vm_flags)) { | |
9992af10 | 1852 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1853 | /* |
1854 | * Identify referenced, file-backed active pages and | |
1855 | * give them one more trip around the active list. So | |
1856 | * that executable code get better chances to stay in | |
1857 | * memory under moderate memory pressure. Anon pages | |
1858 | * are not likely to be evicted by use-once streaming | |
1859 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1860 | * so we ignore them here. | |
1861 | */ | |
41e20983 | 1862 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1863 | list_add(&page->lru, &l_active); |
1864 | continue; | |
1865 | } | |
1866 | } | |
7e9cd484 | 1867 | |
5205e56e | 1868 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1869 | list_add(&page->lru, &l_inactive); |
1870 | } | |
1871 | ||
b555749a | 1872 | /* |
8cab4754 | 1873 | * Move pages back to the lru list. |
b555749a | 1874 | */ |
599d0c95 | 1875 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 1876 | /* |
8cab4754 WF |
1877 | * Count referenced pages from currently used mappings as rotated, |
1878 | * even though only some of them are actually re-activated. This | |
1879 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 1880 | * get_scan_count. |
7e9cd484 | 1881 | */ |
b7c46d15 | 1882 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1883 | |
fa9add64 HD |
1884 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
1885 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
599d0c95 MG |
1886 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
1887 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 1888 | |
747db954 | 1889 | mem_cgroup_uncharge_list(&l_hold); |
b745bc85 | 1890 | free_hot_cold_page_list(&l_hold, true); |
1da177e4 LT |
1891 | } |
1892 | ||
59dc76b0 RR |
1893 | /* |
1894 | * The inactive anon list should be small enough that the VM never has | |
1895 | * to do too much work. | |
14797e23 | 1896 | * |
59dc76b0 RR |
1897 | * The inactive file list should be small enough to leave most memory |
1898 | * to the established workingset on the scan-resistant active list, | |
1899 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 1900 | * |
59dc76b0 RR |
1901 | * Both inactive lists should also be large enough that each inactive |
1902 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 1903 | * |
59dc76b0 RR |
1904 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
1905 | * on this LRU, maintained by the pageout code. A zone->inactive_ratio | |
1906 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. | |
56e49d21 | 1907 | * |
59dc76b0 RR |
1908 | * total target max |
1909 | * memory ratio inactive | |
1910 | * ------------------------------------- | |
1911 | * 10MB 1 5MB | |
1912 | * 100MB 1 50MB | |
1913 | * 1GB 3 250MB | |
1914 | * 10GB 10 0.9GB | |
1915 | * 100GB 31 3GB | |
1916 | * 1TB 101 10GB | |
1917 | * 10TB 320 32GB | |
56e49d21 | 1918 | */ |
59dc76b0 | 1919 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file) |
56e49d21 | 1920 | { |
59dc76b0 | 1921 | unsigned long inactive_ratio; |
e3790144 JW |
1922 | unsigned long inactive; |
1923 | unsigned long active; | |
59dc76b0 | 1924 | unsigned long gb; |
e3790144 | 1925 | |
59dc76b0 RR |
1926 | /* |
1927 | * If we don't have swap space, anonymous page deactivation | |
1928 | * is pointless. | |
1929 | */ | |
1930 | if (!file && !total_swap_pages) | |
1931 | return false; | |
56e49d21 | 1932 | |
59dc76b0 RR |
1933 | inactive = lruvec_lru_size(lruvec, file * LRU_FILE); |
1934 | active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE); | |
56e49d21 | 1935 | |
59dc76b0 RR |
1936 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
1937 | if (gb) | |
1938 | inactive_ratio = int_sqrt(10 * gb); | |
b39415b2 | 1939 | else |
59dc76b0 RR |
1940 | inactive_ratio = 1; |
1941 | ||
1942 | return inactive * inactive_ratio < active; | |
b39415b2 RR |
1943 | } |
1944 | ||
4f98a2fe | 1945 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
1a93be0e | 1946 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 1947 | { |
b39415b2 | 1948 | if (is_active_lru(lru)) { |
59dc76b0 | 1949 | if (inactive_list_is_low(lruvec, is_file_lru(lru))) |
1a93be0e | 1950 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
1951 | return 0; |
1952 | } | |
1953 | ||
1a93be0e | 1954 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
1955 | } |
1956 | ||
9a265114 JW |
1957 | enum scan_balance { |
1958 | SCAN_EQUAL, | |
1959 | SCAN_FRACT, | |
1960 | SCAN_ANON, | |
1961 | SCAN_FILE, | |
1962 | }; | |
1963 | ||
4f98a2fe RR |
1964 | /* |
1965 | * Determine how aggressively the anon and file LRU lists should be | |
1966 | * scanned. The relative value of each set of LRU lists is determined | |
1967 | * by looking at the fraction of the pages scanned we did rotate back | |
1968 | * onto the active list instead of evict. | |
1969 | * | |
be7bd59d WL |
1970 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
1971 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 1972 | */ |
33377678 | 1973 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
6b4f7799 JW |
1974 | struct scan_control *sc, unsigned long *nr, |
1975 | unsigned long *lru_pages) | |
4f98a2fe | 1976 | { |
33377678 | 1977 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
1978 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1979 | u64 fraction[2]; | |
1980 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 1981 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 1982 | unsigned long anon_prio, file_prio; |
9a265114 | 1983 | enum scan_balance scan_balance; |
0bf1457f | 1984 | unsigned long anon, file; |
9a265114 | 1985 | bool force_scan = false; |
4f98a2fe | 1986 | unsigned long ap, fp; |
4111304d | 1987 | enum lru_list lru; |
6f04f48d SS |
1988 | bool some_scanned; |
1989 | int pass; | |
246e87a9 | 1990 | |
f11c0ca5 JW |
1991 | /* |
1992 | * If the zone or memcg is small, nr[l] can be 0. This | |
1993 | * results in no scanning on this priority and a potential | |
1994 | * priority drop. Global direct reclaim can go to the next | |
1995 | * zone and tends to have no problems. Global kswapd is for | |
1996 | * zone balancing and it needs to scan a minimum amount. When | |
1997 | * reclaiming for a memcg, a priority drop can cause high | |
1998 | * latencies, so it's better to scan a minimum amount there as | |
1999 | * well. | |
2000 | */ | |
90cbc250 | 2001 | if (current_is_kswapd()) { |
599d0c95 | 2002 | if (!pgdat_reclaimable(pgdat)) |
90cbc250 | 2003 | force_scan = true; |
eb01aaab | 2004 | if (!mem_cgroup_online(memcg)) |
90cbc250 VD |
2005 | force_scan = true; |
2006 | } | |
89b5fae5 | 2007 | if (!global_reclaim(sc)) |
a4d3e9e7 | 2008 | force_scan = true; |
76a33fc3 SL |
2009 | |
2010 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2011 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2012 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2013 | goto out; |
2014 | } | |
4f98a2fe | 2015 | |
10316b31 JW |
2016 | /* |
2017 | * Global reclaim will swap to prevent OOM even with no | |
2018 | * swappiness, but memcg users want to use this knob to | |
2019 | * disable swapping for individual groups completely when | |
2020 | * using the memory controller's swap limit feature would be | |
2021 | * too expensive. | |
2022 | */ | |
02695175 | 2023 | if (!global_reclaim(sc) && !swappiness) { |
9a265114 | 2024 | scan_balance = SCAN_FILE; |
10316b31 JW |
2025 | goto out; |
2026 | } | |
2027 | ||
2028 | /* | |
2029 | * Do not apply any pressure balancing cleverness when the | |
2030 | * system is close to OOM, scan both anon and file equally | |
2031 | * (unless the swappiness setting disagrees with swapping). | |
2032 | */ | |
02695175 | 2033 | if (!sc->priority && swappiness) { |
9a265114 | 2034 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2035 | goto out; |
2036 | } | |
2037 | ||
62376251 JW |
2038 | /* |
2039 | * Prevent the reclaimer from falling into the cache trap: as | |
2040 | * cache pages start out inactive, every cache fault will tip | |
2041 | * the scan balance towards the file LRU. And as the file LRU | |
2042 | * shrinks, so does the window for rotation from references. | |
2043 | * This means we have a runaway feedback loop where a tiny | |
2044 | * thrashing file LRU becomes infinitely more attractive than | |
2045 | * anon pages. Try to detect this based on file LRU size. | |
2046 | */ | |
2047 | if (global_reclaim(sc)) { | |
599d0c95 MG |
2048 | unsigned long pgdatfile; |
2049 | unsigned long pgdatfree; | |
2050 | int z; | |
2051 | unsigned long total_high_wmark = 0; | |
2ab051e1 | 2052 | |
599d0c95 MG |
2053 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2054 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2055 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2056 | ||
2057 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2058 | struct zone *zone = &pgdat->node_zones[z]; | |
2059 | if (!populated_zone(zone)) | |
2060 | continue; | |
2061 | ||
2062 | total_high_wmark += high_wmark_pages(zone); | |
2063 | } | |
62376251 | 2064 | |
599d0c95 | 2065 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
62376251 JW |
2066 | scan_balance = SCAN_ANON; |
2067 | goto out; | |
2068 | } | |
2069 | } | |
2070 | ||
7c5bd705 | 2071 | /* |
316bda0e VD |
2072 | * If there is enough inactive page cache, i.e. if the size of the |
2073 | * inactive list is greater than that of the active list *and* the | |
2074 | * inactive list actually has some pages to scan on this priority, we | |
2075 | * do not reclaim anything from the anonymous working set right now. | |
2076 | * Without the second condition we could end up never scanning an | |
2077 | * lruvec even if it has plenty of old anonymous pages unless the | |
2078 | * system is under heavy pressure. | |
7c5bd705 | 2079 | */ |
59dc76b0 | 2080 | if (!inactive_list_is_low(lruvec, true) && |
23047a96 | 2081 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) { |
9a265114 | 2082 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2083 | goto out; |
2084 | } | |
2085 | ||
9a265114 JW |
2086 | scan_balance = SCAN_FRACT; |
2087 | ||
58c37f6e KM |
2088 | /* |
2089 | * With swappiness at 100, anonymous and file have the same priority. | |
2090 | * This scanning priority is essentially the inverse of IO cost. | |
2091 | */ | |
02695175 | 2092 | anon_prio = swappiness; |
75b00af7 | 2093 | file_prio = 200 - anon_prio; |
58c37f6e | 2094 | |
4f98a2fe RR |
2095 | /* |
2096 | * OK, so we have swap space and a fair amount of page cache | |
2097 | * pages. We use the recently rotated / recently scanned | |
2098 | * ratios to determine how valuable each cache is. | |
2099 | * | |
2100 | * Because workloads change over time (and to avoid overflow) | |
2101 | * we keep these statistics as a floating average, which ends | |
2102 | * up weighing recent references more than old ones. | |
2103 | * | |
2104 | * anon in [0], file in [1] | |
2105 | */ | |
2ab051e1 | 2106 | |
23047a96 JW |
2107 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) + |
2108 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON); | |
2109 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
2110 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE); | |
2ab051e1 | 2111 | |
599d0c95 | 2112 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2113 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2114 | reclaim_stat->recent_scanned[0] /= 2; |
2115 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2116 | } |
2117 | ||
6e901571 | 2118 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2119 | reclaim_stat->recent_scanned[1] /= 2; |
2120 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2121 | } |
2122 | ||
4f98a2fe | 2123 | /* |
00d8089c RR |
2124 | * The amount of pressure on anon vs file pages is inversely |
2125 | * proportional to the fraction of recently scanned pages on | |
2126 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2127 | */ |
fe35004f | 2128 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2129 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2130 | |
fe35004f | 2131 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2132 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2133 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2134 | |
76a33fc3 SL |
2135 | fraction[0] = ap; |
2136 | fraction[1] = fp; | |
2137 | denominator = ap + fp + 1; | |
2138 | out: | |
6f04f48d SS |
2139 | some_scanned = false; |
2140 | /* Only use force_scan on second pass. */ | |
2141 | for (pass = 0; !some_scanned && pass < 2; pass++) { | |
6b4f7799 | 2142 | *lru_pages = 0; |
6f04f48d SS |
2143 | for_each_evictable_lru(lru) { |
2144 | int file = is_file_lru(lru); | |
2145 | unsigned long size; | |
2146 | unsigned long scan; | |
6e08a369 | 2147 | |
23047a96 | 2148 | size = lruvec_lru_size(lruvec, lru); |
6f04f48d | 2149 | scan = size >> sc->priority; |
9a265114 | 2150 | |
6f04f48d SS |
2151 | if (!scan && pass && force_scan) |
2152 | scan = min(size, SWAP_CLUSTER_MAX); | |
9a265114 | 2153 | |
6f04f48d SS |
2154 | switch (scan_balance) { |
2155 | case SCAN_EQUAL: | |
2156 | /* Scan lists relative to size */ | |
2157 | break; | |
2158 | case SCAN_FRACT: | |
2159 | /* | |
2160 | * Scan types proportional to swappiness and | |
2161 | * their relative recent reclaim efficiency. | |
2162 | */ | |
2163 | scan = div64_u64(scan * fraction[file], | |
2164 | denominator); | |
2165 | break; | |
2166 | case SCAN_FILE: | |
2167 | case SCAN_ANON: | |
2168 | /* Scan one type exclusively */ | |
6b4f7799 JW |
2169 | if ((scan_balance == SCAN_FILE) != file) { |
2170 | size = 0; | |
6f04f48d | 2171 | scan = 0; |
6b4f7799 | 2172 | } |
6f04f48d SS |
2173 | break; |
2174 | default: | |
2175 | /* Look ma, no brain */ | |
2176 | BUG(); | |
2177 | } | |
6b4f7799 JW |
2178 | |
2179 | *lru_pages += size; | |
6f04f48d | 2180 | nr[lru] = scan; |
6b4f7799 | 2181 | |
9a265114 | 2182 | /* |
6f04f48d SS |
2183 | * Skip the second pass and don't force_scan, |
2184 | * if we found something to scan. | |
9a265114 | 2185 | */ |
6f04f48d | 2186 | some_scanned |= !!scan; |
9a265114 | 2187 | } |
76a33fc3 | 2188 | } |
6e08a369 | 2189 | } |
4f98a2fe | 2190 | |
72b252ae MG |
2191 | #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH |
2192 | static void init_tlb_ubc(void) | |
2193 | { | |
2194 | /* | |
2195 | * This deliberately does not clear the cpumask as it's expensive | |
2196 | * and unnecessary. If there happens to be data in there then the | |
2197 | * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and | |
2198 | * then will be cleared. | |
2199 | */ | |
2200 | current->tlb_ubc.flush_required = false; | |
2201 | } | |
2202 | #else | |
2203 | static inline void init_tlb_ubc(void) | |
2204 | { | |
2205 | } | |
2206 | #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ | |
2207 | ||
9b4f98cd JW |
2208 | /* |
2209 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
2210 | */ | |
33377678 VD |
2211 | static void shrink_zone_memcg(struct zone *zone, struct mem_cgroup *memcg, |
2212 | struct scan_control *sc, unsigned long *lru_pages) | |
9b4f98cd | 2213 | { |
33377678 | 2214 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
9b4f98cd | 2215 | unsigned long nr[NR_LRU_LISTS]; |
e82e0561 | 2216 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2217 | unsigned long nr_to_scan; |
2218 | enum lru_list lru; | |
2219 | unsigned long nr_reclaimed = 0; | |
2220 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2221 | struct blk_plug plug; | |
1a501907 | 2222 | bool scan_adjusted; |
9b4f98cd | 2223 | |
33377678 | 2224 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
9b4f98cd | 2225 | |
e82e0561 MG |
2226 | /* Record the original scan target for proportional adjustments later */ |
2227 | memcpy(targets, nr, sizeof(nr)); | |
2228 | ||
1a501907 MG |
2229 | /* |
2230 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2231 | * event that can occur when there is little memory pressure e.g. | |
2232 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2233 | * when the requested number of pages are reclaimed when scanning at | |
2234 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2235 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2236 | * do a batch of work at once. For memcg reclaim one check is made to | |
2237 | * abort proportional reclaim if either the file or anon lru has already | |
2238 | * dropped to zero at the first pass. | |
2239 | */ | |
2240 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2241 | sc->priority == DEF_PRIORITY); | |
2242 | ||
72b252ae MG |
2243 | init_tlb_ubc(); |
2244 | ||
9b4f98cd JW |
2245 | blk_start_plug(&plug); |
2246 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2247 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2248 | unsigned long nr_anon, nr_file, percentage; |
2249 | unsigned long nr_scanned; | |
2250 | ||
9b4f98cd JW |
2251 | for_each_evictable_lru(lru) { |
2252 | if (nr[lru]) { | |
2253 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2254 | nr[lru] -= nr_to_scan; | |
2255 | ||
2256 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2257 | lruvec, sc); | |
2258 | } | |
2259 | } | |
e82e0561 MG |
2260 | |
2261 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) | |
2262 | continue; | |
2263 | ||
e82e0561 MG |
2264 | /* |
2265 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2266 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2267 | * proportionally what was requested by get_scan_count(). We |
2268 | * stop reclaiming one LRU and reduce the amount scanning | |
2269 | * proportional to the original scan target. | |
2270 | */ | |
2271 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2272 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2273 | ||
1a501907 MG |
2274 | /* |
2275 | * It's just vindictive to attack the larger once the smaller | |
2276 | * has gone to zero. And given the way we stop scanning the | |
2277 | * smaller below, this makes sure that we only make one nudge | |
2278 | * towards proportionality once we've got nr_to_reclaim. | |
2279 | */ | |
2280 | if (!nr_file || !nr_anon) | |
2281 | break; | |
2282 | ||
e82e0561 MG |
2283 | if (nr_file > nr_anon) { |
2284 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2285 | targets[LRU_ACTIVE_ANON] + 1; | |
2286 | lru = LRU_BASE; | |
2287 | percentage = nr_anon * 100 / scan_target; | |
2288 | } else { | |
2289 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2290 | targets[LRU_ACTIVE_FILE] + 1; | |
2291 | lru = LRU_FILE; | |
2292 | percentage = nr_file * 100 / scan_target; | |
2293 | } | |
2294 | ||
2295 | /* Stop scanning the smaller of the LRU */ | |
2296 | nr[lru] = 0; | |
2297 | nr[lru + LRU_ACTIVE] = 0; | |
2298 | ||
2299 | /* | |
2300 | * Recalculate the other LRU scan count based on its original | |
2301 | * scan target and the percentage scanning already complete | |
2302 | */ | |
2303 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2304 | nr_scanned = targets[lru] - nr[lru]; | |
2305 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2306 | nr[lru] -= min(nr[lru], nr_scanned); | |
2307 | ||
2308 | lru += LRU_ACTIVE; | |
2309 | nr_scanned = targets[lru] - nr[lru]; | |
2310 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2311 | nr[lru] -= min(nr[lru], nr_scanned); | |
2312 | ||
2313 | scan_adjusted = true; | |
9b4f98cd JW |
2314 | } |
2315 | blk_finish_plug(&plug); | |
2316 | sc->nr_reclaimed += nr_reclaimed; | |
2317 | ||
2318 | /* | |
2319 | * Even if we did not try to evict anon pages at all, we want to | |
2320 | * rebalance the anon lru active/inactive ratio. | |
2321 | */ | |
59dc76b0 | 2322 | if (inactive_list_is_low(lruvec, false)) |
9b4f98cd JW |
2323 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2324 | sc, LRU_ACTIVE_ANON); | |
2325 | ||
2326 | throttle_vm_writeout(sc->gfp_mask); | |
2327 | } | |
2328 | ||
23b9da55 | 2329 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2330 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2331 | { |
d84da3f9 | 2332 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2333 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2334 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2335 | return true; |
2336 | ||
2337 | return false; | |
2338 | } | |
2339 | ||
3e7d3449 | 2340 | /* |
23b9da55 MG |
2341 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2342 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2343 | * true if more pages should be reclaimed such that when the page allocator | |
2344 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2345 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2346 | */ |
9b4f98cd | 2347 | static inline bool should_continue_reclaim(struct zone *zone, |
3e7d3449 MG |
2348 | unsigned long nr_reclaimed, |
2349 | unsigned long nr_scanned, | |
2350 | struct scan_control *sc) | |
2351 | { | |
2352 | unsigned long pages_for_compaction; | |
2353 | unsigned long inactive_lru_pages; | |
2354 | ||
2355 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2356 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2357 | return false; |
2358 | ||
2876592f MG |
2359 | /* Consider stopping depending on scan and reclaim activity */ |
2360 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2361 | /* | |
2362 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2363 | * full LRU list has been scanned and we are still failing | |
2364 | * to reclaim pages. This full LRU scan is potentially | |
2365 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2366 | */ | |
2367 | if (!nr_reclaimed && !nr_scanned) | |
2368 | return false; | |
2369 | } else { | |
2370 | /* | |
2371 | * For non-__GFP_REPEAT allocations which can presumably | |
2372 | * fail without consequence, stop if we failed to reclaim | |
2373 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2374 | * pages that were scanned. This will return to the | |
2375 | * caller faster at the risk reclaim/compaction and | |
2376 | * the resulting allocation attempt fails | |
2377 | */ | |
2378 | if (!nr_reclaimed) | |
2379 | return false; | |
2380 | } | |
3e7d3449 MG |
2381 | |
2382 | /* | |
2383 | * If we have not reclaimed enough pages for compaction and the | |
2384 | * inactive lists are large enough, continue reclaiming | |
2385 | */ | |
2386 | pages_for_compaction = (2UL << sc->order); | |
599d0c95 | 2387 | inactive_lru_pages = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE); |
ec8acf20 | 2388 | if (get_nr_swap_pages() > 0) |
599d0c95 | 2389 | inactive_lru_pages += node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON); |
3e7d3449 MG |
2390 | if (sc->nr_reclaimed < pages_for_compaction && |
2391 | inactive_lru_pages > pages_for_compaction) | |
2392 | return true; | |
2393 | ||
2394 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
ebff3980 | 2395 | switch (compaction_suitable(zone, sc->order, 0, 0)) { |
3e7d3449 MG |
2396 | case COMPACT_PARTIAL: |
2397 | case COMPACT_CONTINUE: | |
2398 | return false; | |
2399 | default: | |
2400 | return true; | |
2401 | } | |
2402 | } | |
2403 | ||
6b4f7799 JW |
2404 | static bool shrink_zone(struct zone *zone, struct scan_control *sc, |
2405 | bool is_classzone) | |
1da177e4 | 2406 | { |
cb731d6c | 2407 | struct reclaim_state *reclaim_state = current->reclaim_state; |
f0fdc5e8 | 2408 | unsigned long nr_reclaimed, nr_scanned; |
2344d7e4 | 2409 | bool reclaimable = false; |
1da177e4 | 2410 | |
9b4f98cd JW |
2411 | do { |
2412 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2413 | struct mem_cgroup_reclaim_cookie reclaim = { | |
2414 | .zone = zone, | |
2415 | .priority = sc->priority, | |
2416 | }; | |
6b4f7799 | 2417 | unsigned long zone_lru_pages = 0; |
694fbc0f | 2418 | struct mem_cgroup *memcg; |
3e7d3449 | 2419 | |
9b4f98cd JW |
2420 | nr_reclaimed = sc->nr_reclaimed; |
2421 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2422 | |
694fbc0f AM |
2423 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2424 | do { | |
6b4f7799 | 2425 | unsigned long lru_pages; |
8e8ae645 | 2426 | unsigned long reclaimed; |
cb731d6c | 2427 | unsigned long scanned; |
5660048c | 2428 | |
241994ed JW |
2429 | if (mem_cgroup_low(root, memcg)) { |
2430 | if (!sc->may_thrash) | |
2431 | continue; | |
2432 | mem_cgroup_events(memcg, MEMCG_LOW, 1); | |
2433 | } | |
2434 | ||
8e8ae645 | 2435 | reclaimed = sc->nr_reclaimed; |
cb731d6c | 2436 | scanned = sc->nr_scanned; |
f9be23d6 | 2437 | |
33377678 | 2438 | shrink_zone_memcg(zone, memcg, sc, &lru_pages); |
6b4f7799 | 2439 | zone_lru_pages += lru_pages; |
f16015fb | 2440 | |
cb731d6c VD |
2441 | if (memcg && is_classzone) |
2442 | shrink_slab(sc->gfp_mask, zone_to_nid(zone), | |
2443 | memcg, sc->nr_scanned - scanned, | |
2444 | lru_pages); | |
2445 | ||
8e8ae645 JW |
2446 | /* Record the group's reclaim efficiency */ |
2447 | vmpressure(sc->gfp_mask, memcg, false, | |
2448 | sc->nr_scanned - scanned, | |
2449 | sc->nr_reclaimed - reclaimed); | |
2450 | ||
9b4f98cd | 2451 | /* |
a394cb8e MH |
2452 | * Direct reclaim and kswapd have to scan all memory |
2453 | * cgroups to fulfill the overall scan target for the | |
9b4f98cd | 2454 | * zone. |
a394cb8e MH |
2455 | * |
2456 | * Limit reclaim, on the other hand, only cares about | |
2457 | * nr_to_reclaim pages to be reclaimed and it will | |
2458 | * retry with decreasing priority if one round over the | |
2459 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2460 | */ |
a394cb8e MH |
2461 | if (!global_reclaim(sc) && |
2462 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2463 | mem_cgroup_iter_break(root, memcg); |
2464 | break; | |
2465 | } | |
241994ed | 2466 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
70ddf637 | 2467 | |
6b4f7799 JW |
2468 | /* |
2469 | * Shrink the slab caches in the same proportion that | |
2470 | * the eligible LRU pages were scanned. | |
2471 | */ | |
cb731d6c VD |
2472 | if (global_reclaim(sc) && is_classzone) |
2473 | shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL, | |
2474 | sc->nr_scanned - nr_scanned, | |
2475 | zone_lru_pages); | |
2476 | ||
2477 | if (reclaim_state) { | |
2478 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2479 | reclaim_state->reclaimed_slab = 0; | |
6b4f7799 JW |
2480 | } |
2481 | ||
8e8ae645 JW |
2482 | /* Record the subtree's reclaim efficiency */ |
2483 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
70ddf637 AV |
2484 | sc->nr_scanned - nr_scanned, |
2485 | sc->nr_reclaimed - nr_reclaimed); | |
2486 | ||
2344d7e4 JW |
2487 | if (sc->nr_reclaimed - nr_reclaimed) |
2488 | reclaimable = true; | |
2489 | ||
9b4f98cd JW |
2490 | } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, |
2491 | sc->nr_scanned - nr_scanned, sc)); | |
2344d7e4 JW |
2492 | |
2493 | return reclaimable; | |
f16015fb JW |
2494 | } |
2495 | ||
53853e2d VB |
2496 | /* |
2497 | * Returns true if compaction should go ahead for a high-order request, or | |
2498 | * the high-order allocation would succeed without compaction. | |
2499 | */ | |
b6459cc1 | 2500 | static inline bool compaction_ready(struct zone *zone, int order, int classzone_idx) |
fe4b1b24 MG |
2501 | { |
2502 | unsigned long balance_gap, watermark; | |
2503 | bool watermark_ok; | |
2504 | ||
fe4b1b24 MG |
2505 | /* |
2506 | * Compaction takes time to run and there are potentially other | |
2507 | * callers using the pages just freed. Continue reclaiming until | |
2508 | * there is a buffer of free pages available to give compaction | |
2509 | * a reasonable chance of completing and allocating the page | |
2510 | */ | |
4be89a34 JZ |
2511 | balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( |
2512 | zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); | |
0b06496a | 2513 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << order); |
b6459cc1 | 2514 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, classzone_idx); |
fe4b1b24 MG |
2515 | |
2516 | /* | |
2517 | * If compaction is deferred, reclaim up to a point where | |
2518 | * compaction will have a chance of success when re-enabled | |
2519 | */ | |
0b06496a | 2520 | if (compaction_deferred(zone, order)) |
fe4b1b24 MG |
2521 | return watermark_ok; |
2522 | ||
53853e2d VB |
2523 | /* |
2524 | * If compaction is not ready to start and allocation is not likely | |
2525 | * to succeed without it, then keep reclaiming. | |
2526 | */ | |
b6459cc1 | 2527 | if (compaction_suitable(zone, order, 0, classzone_idx) == COMPACT_SKIPPED) |
fe4b1b24 MG |
2528 | return false; |
2529 | ||
2530 | return watermark_ok; | |
2531 | } | |
2532 | ||
1da177e4 LT |
2533 | /* |
2534 | * This is the direct reclaim path, for page-allocating processes. We only | |
2535 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2536 | * request. | |
2537 | * | |
41858966 MG |
2538 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
2539 | * Because: | |
1da177e4 LT |
2540 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
2541 | * allocation or | |
41858966 MG |
2542 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
2543 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2544 | * zone defense algorithm. | |
1da177e4 | 2545 | * |
1da177e4 LT |
2546 | * If a zone is deemed to be full of pinned pages then just give it a light |
2547 | * scan then give up on it. | |
2548 | */ | |
0a0337e0 | 2549 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2550 | { |
dd1a239f | 2551 | struct zoneref *z; |
54a6eb5c | 2552 | struct zone *zone; |
0608f43d AM |
2553 | unsigned long nr_soft_reclaimed; |
2554 | unsigned long nr_soft_scanned; | |
619d0d76 | 2555 | gfp_t orig_mask; |
9bbc04ee | 2556 | enum zone_type requested_highidx = gfp_zone(sc->gfp_mask); |
1cfb419b | 2557 | |
cc715d99 MG |
2558 | /* |
2559 | * If the number of buffer_heads in the machine exceeds the maximum | |
2560 | * allowed level, force direct reclaim to scan the highmem zone as | |
2561 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2562 | */ | |
619d0d76 | 2563 | orig_mask = sc->gfp_mask; |
cc715d99 MG |
2564 | if (buffer_heads_over_limit) |
2565 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2566 | ||
d4debc66 | 2567 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
7bf52fb8 | 2568 | gfp_zone(sc->gfp_mask), sc->nodemask) { |
6b4f7799 JW |
2569 | enum zone_type classzone_idx; |
2570 | ||
f3fe6512 | 2571 | if (!populated_zone(zone)) |
1da177e4 | 2572 | continue; |
6b4f7799 JW |
2573 | |
2574 | classzone_idx = requested_highidx; | |
2575 | while (!populated_zone(zone->zone_pgdat->node_zones + | |
2576 | classzone_idx)) | |
2577 | classzone_idx--; | |
2578 | ||
1cfb419b KH |
2579 | /* |
2580 | * Take care memory controller reclaiming has small influence | |
2581 | * to global LRU. | |
2582 | */ | |
89b5fae5 | 2583 | if (global_reclaim(sc)) { |
344736f2 VD |
2584 | if (!cpuset_zone_allowed(zone, |
2585 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2586 | continue; |
65ec02cb | 2587 | |
6e543d57 | 2588 | if (sc->priority != DEF_PRIORITY && |
599d0c95 | 2589 | !pgdat_reclaimable(zone->zone_pgdat)) |
1cfb419b | 2590 | continue; /* Let kswapd poll it */ |
0b06496a JW |
2591 | |
2592 | /* | |
2593 | * If we already have plenty of memory free for | |
2594 | * compaction in this zone, don't free any more. | |
2595 | * Even though compaction is invoked for any | |
2596 | * non-zero order, only frequent costly order | |
2597 | * reclamation is disruptive enough to become a | |
2598 | * noticeable problem, like transparent huge | |
2599 | * page allocations. | |
2600 | */ | |
2601 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2602 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
2603 | zonelist_zone_idx(z) <= requested_highidx && | |
b6459cc1 | 2604 | compaction_ready(zone, sc->order, requested_highidx)) { |
0b06496a JW |
2605 | sc->compaction_ready = true; |
2606 | continue; | |
e0887c19 | 2607 | } |
0b06496a | 2608 | |
0608f43d AM |
2609 | /* |
2610 | * This steals pages from memory cgroups over softlimit | |
2611 | * and returns the number of reclaimed pages and | |
2612 | * scanned pages. This works for global memory pressure | |
2613 | * and balancing, not for a memcg's limit. | |
2614 | */ | |
2615 | nr_soft_scanned = 0; | |
2616 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2617 | sc->order, sc->gfp_mask, | |
2618 | &nr_soft_scanned); | |
2619 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2620 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2621 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2622 | } |
408d8544 | 2623 | |
0a0337e0 | 2624 | shrink_zone(zone, sc, zone_idx(zone) == classzone_idx); |
1da177e4 | 2625 | } |
e0c23279 | 2626 | |
619d0d76 WY |
2627 | /* |
2628 | * Restore to original mask to avoid the impact on the caller if we | |
2629 | * promoted it to __GFP_HIGHMEM. | |
2630 | */ | |
2631 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2632 | } |
4f98a2fe | 2633 | |
1da177e4 LT |
2634 | /* |
2635 | * This is the main entry point to direct page reclaim. | |
2636 | * | |
2637 | * If a full scan of the inactive list fails to free enough memory then we | |
2638 | * are "out of memory" and something needs to be killed. | |
2639 | * | |
2640 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2641 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2642 | * caller can't do much about. We kick the writeback threads and take explicit |
2643 | * naps in the hope that some of these pages can be written. But if the | |
2644 | * allocating task holds filesystem locks which prevent writeout this might not | |
2645 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2646 | * |
2647 | * returns: 0, if no pages reclaimed | |
2648 | * else, the number of pages reclaimed | |
1da177e4 | 2649 | */ |
dac1d27b | 2650 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2651 | struct scan_control *sc) |
1da177e4 | 2652 | { |
241994ed | 2653 | int initial_priority = sc->priority; |
69e05944 | 2654 | unsigned long total_scanned = 0; |
22fba335 | 2655 | unsigned long writeback_threshold; |
241994ed | 2656 | retry: |
873b4771 KK |
2657 | delayacct_freepages_start(); |
2658 | ||
89b5fae5 | 2659 | if (global_reclaim(sc)) |
1cfb419b | 2660 | count_vm_event(ALLOCSTALL); |
1da177e4 | 2661 | |
9e3b2f8c | 2662 | do { |
70ddf637 AV |
2663 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2664 | sc->priority); | |
66e1707b | 2665 | sc->nr_scanned = 0; |
0a0337e0 | 2666 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 2667 | |
66e1707b | 2668 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2669 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
2670 | break; |
2671 | ||
2672 | if (sc->compaction_ready) | |
2673 | break; | |
1da177e4 | 2674 | |
0e50ce3b MK |
2675 | /* |
2676 | * If we're getting trouble reclaiming, start doing | |
2677 | * writepage even in laptop mode. | |
2678 | */ | |
2679 | if (sc->priority < DEF_PRIORITY - 2) | |
2680 | sc->may_writepage = 1; | |
2681 | ||
1da177e4 LT |
2682 | /* |
2683 | * Try to write back as many pages as we just scanned. This | |
2684 | * tends to cause slow streaming writers to write data to the | |
2685 | * disk smoothly, at the dirtying rate, which is nice. But | |
2686 | * that's undesirable in laptop mode, where we *want* lumpy | |
2687 | * writeout. So in laptop mode, write out the whole world. | |
2688 | */ | |
22fba335 KM |
2689 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2690 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2691 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2692 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2693 | sc->may_writepage = 1; |
1da177e4 | 2694 | } |
0b06496a | 2695 | } while (--sc->priority >= 0); |
bb21c7ce | 2696 | |
873b4771 KK |
2697 | delayacct_freepages_end(); |
2698 | ||
bb21c7ce KM |
2699 | if (sc->nr_reclaimed) |
2700 | return sc->nr_reclaimed; | |
2701 | ||
0cee34fd | 2702 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 2703 | if (sc->compaction_ready) |
7335084d MG |
2704 | return 1; |
2705 | ||
241994ed JW |
2706 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
2707 | if (!sc->may_thrash) { | |
2708 | sc->priority = initial_priority; | |
2709 | sc->may_thrash = 1; | |
2710 | goto retry; | |
2711 | } | |
2712 | ||
bb21c7ce | 2713 | return 0; |
1da177e4 LT |
2714 | } |
2715 | ||
5515061d MG |
2716 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) |
2717 | { | |
2718 | struct zone *zone; | |
2719 | unsigned long pfmemalloc_reserve = 0; | |
2720 | unsigned long free_pages = 0; | |
2721 | int i; | |
2722 | bool wmark_ok; | |
2723 | ||
2724 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2725 | zone = &pgdat->node_zones[i]; | |
f012a84a | 2726 | if (!populated_zone(zone) || |
599d0c95 | 2727 | pgdat_reclaimable_pages(pgdat) == 0) |
675becce MG |
2728 | continue; |
2729 | ||
5515061d MG |
2730 | pfmemalloc_reserve += min_wmark_pages(zone); |
2731 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2732 | } | |
2733 | ||
675becce MG |
2734 | /* If there are no reserves (unexpected config) then do not throttle */ |
2735 | if (!pfmemalloc_reserve) | |
2736 | return true; | |
2737 | ||
5515061d MG |
2738 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
2739 | ||
2740 | /* kswapd must be awake if processes are being throttled */ | |
2741 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2742 | pgdat->classzone_idx = min(pgdat->classzone_idx, | |
2743 | (enum zone_type)ZONE_NORMAL); | |
2744 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2745 | } | |
2746 | ||
2747 | return wmark_ok; | |
2748 | } | |
2749 | ||
2750 | /* | |
2751 | * Throttle direct reclaimers if backing storage is backed by the network | |
2752 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2753 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2754 | * when the low watermark is reached. |
2755 | * | |
2756 | * Returns true if a fatal signal was delivered during throttling. If this | |
2757 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2758 | */ |
50694c28 | 2759 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2760 | nodemask_t *nodemask) |
2761 | { | |
675becce | 2762 | struct zoneref *z; |
5515061d | 2763 | struct zone *zone; |
675becce | 2764 | pg_data_t *pgdat = NULL; |
5515061d MG |
2765 | |
2766 | /* | |
2767 | * Kernel threads should not be throttled as they may be indirectly | |
2768 | * responsible for cleaning pages necessary for reclaim to make forward | |
2769 | * progress. kjournald for example may enter direct reclaim while | |
2770 | * committing a transaction where throttling it could forcing other | |
2771 | * processes to block on log_wait_commit(). | |
2772 | */ | |
2773 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2774 | goto out; |
2775 | ||
2776 | /* | |
2777 | * If a fatal signal is pending, this process should not throttle. | |
2778 | * It should return quickly so it can exit and free its memory | |
2779 | */ | |
2780 | if (fatal_signal_pending(current)) | |
2781 | goto out; | |
5515061d | 2782 | |
675becce MG |
2783 | /* |
2784 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2785 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2786 | * GFP_KERNEL will be required for allocating network buffers when | |
2787 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2788 | * | |
2789 | * Throttling is based on the first usable node and throttled processes | |
2790 | * wait on a queue until kswapd makes progress and wakes them. There | |
2791 | * is an affinity then between processes waking up and where reclaim | |
2792 | * progress has been made assuming the process wakes on the same node. | |
2793 | * More importantly, processes running on remote nodes will not compete | |
2794 | * for remote pfmemalloc reserves and processes on different nodes | |
2795 | * should make reasonable progress. | |
2796 | */ | |
2797 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 2798 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
2799 | if (zone_idx(zone) > ZONE_NORMAL) |
2800 | continue; | |
2801 | ||
2802 | /* Throttle based on the first usable node */ | |
2803 | pgdat = zone->zone_pgdat; | |
2804 | if (pfmemalloc_watermark_ok(pgdat)) | |
2805 | goto out; | |
2806 | break; | |
2807 | } | |
2808 | ||
2809 | /* If no zone was usable by the allocation flags then do not throttle */ | |
2810 | if (!pgdat) | |
50694c28 | 2811 | goto out; |
5515061d | 2812 | |
68243e76 MG |
2813 | /* Account for the throttling */ |
2814 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2815 | ||
5515061d MG |
2816 | /* |
2817 | * If the caller cannot enter the filesystem, it's possible that it | |
2818 | * is due to the caller holding an FS lock or performing a journal | |
2819 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2820 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2821 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2822 | * second before continuing. | |
2823 | */ | |
2824 | if (!(gfp_mask & __GFP_FS)) { | |
2825 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2826 | pfmemalloc_watermark_ok(pgdat), HZ); | |
50694c28 MG |
2827 | |
2828 | goto check_pending; | |
5515061d MG |
2829 | } |
2830 | ||
2831 | /* Throttle until kswapd wakes the process */ | |
2832 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2833 | pfmemalloc_watermark_ok(pgdat)); | |
50694c28 MG |
2834 | |
2835 | check_pending: | |
2836 | if (fatal_signal_pending(current)) | |
2837 | return true; | |
2838 | ||
2839 | out: | |
2840 | return false; | |
5515061d MG |
2841 | } |
2842 | ||
dac1d27b | 2843 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2844 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2845 | { |
33906bc5 | 2846 | unsigned long nr_reclaimed; |
66e1707b | 2847 | struct scan_control sc = { |
ee814fe2 | 2848 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
21caf2fc | 2849 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
ee814fe2 JW |
2850 | .order = order, |
2851 | .nodemask = nodemask, | |
2852 | .priority = DEF_PRIORITY, | |
66e1707b | 2853 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2854 | .may_unmap = 1, |
2e2e4259 | 2855 | .may_swap = 1, |
66e1707b BS |
2856 | }; |
2857 | ||
5515061d | 2858 | /* |
50694c28 MG |
2859 | * Do not enter reclaim if fatal signal was delivered while throttled. |
2860 | * 1 is returned so that the page allocator does not OOM kill at this | |
2861 | * point. | |
5515061d | 2862 | */ |
50694c28 | 2863 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) |
5515061d MG |
2864 | return 1; |
2865 | ||
33906bc5 MG |
2866 | trace_mm_vmscan_direct_reclaim_begin(order, |
2867 | sc.may_writepage, | |
2868 | gfp_mask); | |
2869 | ||
3115cd91 | 2870 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
2871 | |
2872 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2873 | ||
2874 | return nr_reclaimed; | |
66e1707b BS |
2875 | } |
2876 | ||
c255a458 | 2877 | #ifdef CONFIG_MEMCG |
66e1707b | 2878 | |
72835c86 | 2879 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, |
4e416953 | 2880 | gfp_t gfp_mask, bool noswap, |
0ae5e89c YH |
2881 | struct zone *zone, |
2882 | unsigned long *nr_scanned) | |
4e416953 BS |
2883 | { |
2884 | struct scan_control sc = { | |
b8f5c566 | 2885 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 2886 | .target_mem_cgroup = memcg, |
4e416953 BS |
2887 | .may_writepage = !laptop_mode, |
2888 | .may_unmap = 1, | |
2889 | .may_swap = !noswap, | |
4e416953 | 2890 | }; |
6b4f7799 | 2891 | unsigned long lru_pages; |
0ae5e89c | 2892 | |
4e416953 BS |
2893 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2894 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 2895 | |
9e3b2f8c | 2896 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e KM |
2897 | sc.may_writepage, |
2898 | sc.gfp_mask); | |
2899 | ||
4e416953 BS |
2900 | /* |
2901 | * NOTE: Although we can get the priority field, using it | |
2902 | * here is not a good idea, since it limits the pages we can scan. | |
2903 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2904 | * will pick up pages from other mem cgroup's as well. We hack | |
2905 | * the priority and make it zero. | |
2906 | */ | |
33377678 | 2907 | shrink_zone_memcg(zone, memcg, &sc, &lru_pages); |
bdce6d9e KM |
2908 | |
2909 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2910 | ||
0ae5e89c | 2911 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2912 | return sc.nr_reclaimed; |
2913 | } | |
2914 | ||
72835c86 | 2915 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 2916 | unsigned long nr_pages, |
a7885eb8 | 2917 | gfp_t gfp_mask, |
b70a2a21 | 2918 | bool may_swap) |
66e1707b | 2919 | { |
4e416953 | 2920 | struct zonelist *zonelist; |
bdce6d9e | 2921 | unsigned long nr_reclaimed; |
889976db | 2922 | int nid; |
66e1707b | 2923 | struct scan_control sc = { |
b70a2a21 | 2924 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
a09ed5e0 YH |
2925 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2926 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
ee814fe2 JW |
2927 | .target_mem_cgroup = memcg, |
2928 | .priority = DEF_PRIORITY, | |
2929 | .may_writepage = !laptop_mode, | |
2930 | .may_unmap = 1, | |
b70a2a21 | 2931 | .may_swap = may_swap, |
a09ed5e0 | 2932 | }; |
66e1707b | 2933 | |
889976db YH |
2934 | /* |
2935 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2936 | * take care of from where we get pages. So the node where we start the | |
2937 | * scan does not need to be the current node. | |
2938 | */ | |
72835c86 | 2939 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2940 | |
2941 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2942 | |
2943 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2944 | sc.may_writepage, | |
2945 | sc.gfp_mask); | |
2946 | ||
3115cd91 | 2947 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
bdce6d9e KM |
2948 | |
2949 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2950 | ||
2951 | return nr_reclaimed; | |
66e1707b BS |
2952 | } |
2953 | #endif | |
2954 | ||
9e3b2f8c | 2955 | static void age_active_anon(struct zone *zone, struct scan_control *sc) |
f16015fb | 2956 | { |
b95a2f2d | 2957 | struct mem_cgroup *memcg; |
f16015fb | 2958 | |
b95a2f2d JW |
2959 | if (!total_swap_pages) |
2960 | return; | |
2961 | ||
2962 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2963 | do { | |
c56d5c7d | 2964 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
b95a2f2d | 2965 | |
59dc76b0 | 2966 | if (inactive_list_is_low(lruvec, false)) |
1a93be0e | 2967 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 2968 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
2969 | |
2970 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2971 | } while (memcg); | |
f16015fb JW |
2972 | } |
2973 | ||
accf6242 VB |
2974 | static bool zone_balanced(struct zone *zone, int order, bool highorder, |
2975 | unsigned long balance_gap, int classzone_idx) | |
60cefed4 | 2976 | { |
accf6242 | 2977 | unsigned long mark = high_wmark_pages(zone) + balance_gap; |
60cefed4 | 2978 | |
accf6242 VB |
2979 | /* |
2980 | * When checking from pgdat_balanced(), kswapd should stop and sleep | |
2981 | * when it reaches the high order-0 watermark and let kcompactd take | |
2982 | * over. Other callers such as wakeup_kswapd() want to determine the | |
2983 | * true high-order watermark. | |
2984 | */ | |
2985 | if (IS_ENABLED(CONFIG_COMPACTION) && !highorder) { | |
2986 | mark += (1UL << order); | |
2987 | order = 0; | |
2988 | } | |
60cefed4 | 2989 | |
accf6242 | 2990 | return zone_watermark_ok_safe(zone, order, mark, classzone_idx); |
60cefed4 JW |
2991 | } |
2992 | ||
1741c877 | 2993 | /* |
4ae0a48b ZC |
2994 | * pgdat_balanced() is used when checking if a node is balanced. |
2995 | * | |
2996 | * For order-0, all zones must be balanced! | |
2997 | * | |
2998 | * For high-order allocations only zones that meet watermarks and are in a | |
2999 | * zone allowed by the callers classzone_idx are added to balanced_pages. The | |
3000 | * total of balanced pages must be at least 25% of the zones allowed by | |
3001 | * classzone_idx for the node to be considered balanced. Forcing all zones to | |
3002 | * be balanced for high orders can cause excessive reclaim when there are | |
3003 | * imbalanced zones. | |
1741c877 MG |
3004 | * The choice of 25% is due to |
3005 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
3006 | * reasonable sized machine | |
3007 | * o On all other machines, the top zone must be at least a reasonable | |
25985edc | 3008 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
1741c877 MG |
3009 | * would need to be at least 256M for it to be balance a whole node. |
3010 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
3011 | * to balance a node on its own. These seemed like reasonable ratios. | |
3012 | */ | |
4ae0a48b | 3013 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) |
1741c877 | 3014 | { |
b40da049 | 3015 | unsigned long managed_pages = 0; |
4ae0a48b | 3016 | unsigned long balanced_pages = 0; |
1741c877 MG |
3017 | int i; |
3018 | ||
4ae0a48b ZC |
3019 | /* Check the watermark levels */ |
3020 | for (i = 0; i <= classzone_idx; i++) { | |
3021 | struct zone *zone = pgdat->node_zones + i; | |
1741c877 | 3022 | |
4ae0a48b ZC |
3023 | if (!populated_zone(zone)) |
3024 | continue; | |
3025 | ||
b40da049 | 3026 | managed_pages += zone->managed_pages; |
4ae0a48b ZC |
3027 | |
3028 | /* | |
3029 | * A special case here: | |
3030 | * | |
3031 | * balance_pgdat() skips over all_unreclaimable after | |
3032 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
3033 | * they must be considered balanced here as well! | |
3034 | */ | |
599d0c95 | 3035 | if (!pgdat_reclaimable(zone->zone_pgdat)) { |
b40da049 | 3036 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
3037 | continue; |
3038 | } | |
3039 | ||
accf6242 | 3040 | if (zone_balanced(zone, order, false, 0, i)) |
b40da049 | 3041 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
3042 | else if (!order) |
3043 | return false; | |
3044 | } | |
3045 | ||
3046 | if (order) | |
b40da049 | 3047 | return balanced_pages >= (managed_pages >> 2); |
4ae0a48b ZC |
3048 | else |
3049 | return true; | |
1741c877 MG |
3050 | } |
3051 | ||
5515061d MG |
3052 | /* |
3053 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3054 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3055 | * | |
3056 | * Returns true if kswapd is ready to sleep | |
3057 | */ | |
3058 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, | |
dc83edd9 | 3059 | int classzone_idx) |
f50de2d3 | 3060 | { |
f50de2d3 MG |
3061 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ |
3062 | if (remaining) | |
5515061d MG |
3063 | return false; |
3064 | ||
3065 | /* | |
9e5e3661 VB |
3066 | * The throttled processes are normally woken up in balance_pgdat() as |
3067 | * soon as pfmemalloc_watermark_ok() is true. But there is a potential | |
3068 | * race between when kswapd checks the watermarks and a process gets | |
3069 | * throttled. There is also a potential race if processes get | |
3070 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3071 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3072 | * the wake up checks. If kswapd is going to sleep, no process should | |
3073 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3074 | * the wake up is premature, processes will wake kswapd and get | |
3075 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3076 | * that here we are under prepare_to_wait(). | |
5515061d | 3077 | */ |
9e5e3661 VB |
3078 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3079 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3080 | |
4ae0a48b | 3081 | return pgdat_balanced(pgdat, order, classzone_idx); |
f50de2d3 MG |
3082 | } |
3083 | ||
75485363 MG |
3084 | /* |
3085 | * kswapd shrinks the zone by the number of pages required to reach | |
3086 | * the high watermark. | |
b8e83b94 MG |
3087 | * |
3088 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3089 | * reclaim or if the lack of progress was due to pages under writeback. |
3090 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3091 | */ |
b8e83b94 | 3092 | static bool kswapd_shrink_zone(struct zone *zone, |
7c954f6d | 3093 | int classzone_idx, |
accf6242 | 3094 | struct scan_control *sc) |
75485363 | 3095 | { |
7c954f6d | 3096 | unsigned long balance_gap; |
7c954f6d | 3097 | bool lowmem_pressure; |
599d0c95 | 3098 | struct pglist_data *pgdat = zone->zone_pgdat; |
75485363 MG |
3099 | |
3100 | /* Reclaim above the high watermark. */ | |
3101 | sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone)); | |
7c954f6d | 3102 | |
7c954f6d MG |
3103 | /* |
3104 | * We put equal pressure on every zone, unless one zone has way too | |
3105 | * many pages free already. The "too many pages" is defined as the | |
3106 | * high wmark plus a "gap" where the gap is either the low | |
3107 | * watermark or 1% of the zone, whichever is smaller. | |
3108 | */ | |
4be89a34 JZ |
3109 | balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( |
3110 | zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); | |
7c954f6d MG |
3111 | |
3112 | /* | |
3113 | * If there is no low memory pressure or the zone is balanced then no | |
3114 | * reclaim is necessary | |
3115 | */ | |
3116 | lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone)); | |
accf6242 | 3117 | if (!lowmem_pressure && zone_balanced(zone, sc->order, false, |
7c954f6d MG |
3118 | balance_gap, classzone_idx)) |
3119 | return true; | |
3120 | ||
6b4f7799 | 3121 | shrink_zone(zone, sc, zone_idx(zone) == classzone_idx); |
75485363 | 3122 | |
599d0c95 MG |
3123 | /* TODO: ANOMALY */ |
3124 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
283aba9f | 3125 | |
7c954f6d MG |
3126 | /* |
3127 | * If a zone reaches its high watermark, consider it to be no longer | |
3128 | * congested. It's possible there are dirty pages backed by congested | |
3129 | * BDIs but as pressure is relieved, speculatively avoid congestion | |
3130 | * waits. | |
3131 | */ | |
599d0c95 | 3132 | if (pgdat_reclaimable(zone->zone_pgdat) && |
accf6242 | 3133 | zone_balanced(zone, sc->order, false, 0, classzone_idx)) { |
599d0c95 MG |
3134 | clear_bit(PGDAT_CONGESTED, &pgdat->flags); |
3135 | clear_bit(PGDAT_DIRTY, &pgdat->flags); | |
7c954f6d MG |
3136 | } |
3137 | ||
b8e83b94 | 3138 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3139 | } |
3140 | ||
1da177e4 LT |
3141 | /* |
3142 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 3143 | * they are all at high_wmark_pages(zone). |
1da177e4 | 3144 | * |
accf6242 | 3145 | * Returns the highest zone idx kswapd was reclaiming at |
1da177e4 LT |
3146 | * |
3147 | * There is special handling here for zones which are full of pinned pages. | |
3148 | * This can happen if the pages are all mlocked, or if they are all used by | |
3149 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
3150 | * What we do is to detect the case where all pages in the zone have been | |
3151 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
3152 | * dead and from now on, only perform a short scan. Basically we're polling | |
3153 | * the zone for when the problem goes away. | |
3154 | * | |
3155 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
3156 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
3157 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
3158 | * lower zones regardless of the number of free pages in the lower zones. This | |
3159 | * interoperates with the page allocator fallback scheme to ensure that aging | |
3160 | * of pages is balanced across the zones. | |
1da177e4 | 3161 | */ |
accf6242 | 3162 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3163 | { |
1da177e4 | 3164 | int i; |
99504748 | 3165 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
0608f43d AM |
3166 | unsigned long nr_soft_reclaimed; |
3167 | unsigned long nr_soft_scanned; | |
179e9639 AM |
3168 | struct scan_control sc = { |
3169 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3170 | .order = order, |
b8e83b94 | 3171 | .priority = DEF_PRIORITY, |
ee814fe2 | 3172 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3173 | .may_unmap = 1, |
2e2e4259 | 3174 | .may_swap = 1, |
179e9639 | 3175 | }; |
f8891e5e | 3176 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3177 | |
9e3b2f8c | 3178 | do { |
b8e83b94 MG |
3179 | bool raise_priority = true; |
3180 | ||
3181 | sc.nr_reclaimed = 0; | |
1da177e4 | 3182 | |
d6277db4 RW |
3183 | /* |
3184 | * Scan in the highmem->dma direction for the highest | |
3185 | * zone which needs scanning | |
3186 | */ | |
3187 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
3188 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 3189 | |
d6277db4 RW |
3190 | if (!populated_zone(zone)) |
3191 | continue; | |
1da177e4 | 3192 | |
6e543d57 | 3193 | if (sc.priority != DEF_PRIORITY && |
599d0c95 | 3194 | !pgdat_reclaimable(zone->zone_pgdat)) |
d6277db4 | 3195 | continue; |
1da177e4 | 3196 | |
556adecb RR |
3197 | /* |
3198 | * Do some background aging of the anon list, to give | |
3199 | * pages a chance to be referenced before reclaiming. | |
3200 | */ | |
9e3b2f8c | 3201 | age_active_anon(zone, &sc); |
556adecb | 3202 | |
cc715d99 MG |
3203 | /* |
3204 | * If the number of buffer_heads in the machine | |
3205 | * exceeds the maximum allowed level and this node | |
3206 | * has a highmem zone, force kswapd to reclaim from | |
3207 | * it to relieve lowmem pressure. | |
3208 | */ | |
3209 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
3210 | end_zone = i; | |
3211 | break; | |
3212 | } | |
3213 | ||
accf6242 | 3214 | if (!zone_balanced(zone, order, false, 0, 0)) { |
d6277db4 | 3215 | end_zone = i; |
e1dbeda6 | 3216 | break; |
439423f6 | 3217 | } else { |
d43006d5 MG |
3218 | /* |
3219 | * If balanced, clear the dirty and congested | |
3220 | * flags | |
599d0c95 MG |
3221 | * |
3222 | * TODO: ANOMALY | |
d43006d5 | 3223 | */ |
599d0c95 MG |
3224 | clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags); |
3225 | clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags); | |
1da177e4 | 3226 | } |
1da177e4 | 3227 | } |
dafcb73e | 3228 | |
b8e83b94 | 3229 | if (i < 0) |
e1dbeda6 AM |
3230 | goto out; |
3231 | ||
b7ea3c41 MG |
3232 | /* |
3233 | * If we're getting trouble reclaiming, start doing writepage | |
3234 | * even in laptop mode. | |
3235 | */ | |
3236 | if (sc.priority < DEF_PRIORITY - 2) | |
3237 | sc.may_writepage = 1; | |
3238 | ||
1da177e4 LT |
3239 | /* |
3240 | * Now scan the zone in the dma->highmem direction, stopping | |
3241 | * at the last zone which needs scanning. | |
3242 | * | |
3243 | * We do this because the page allocator works in the opposite | |
3244 | * direction. This prevents the page allocator from allocating | |
3245 | * pages behind kswapd's direction of progress, which would | |
3246 | * cause too much scanning of the lower zones. | |
3247 | */ | |
3248 | for (i = 0; i <= end_zone; i++) { | |
3249 | struct zone *zone = pgdat->node_zones + i; | |
3250 | ||
f3fe6512 | 3251 | if (!populated_zone(zone)) |
1da177e4 LT |
3252 | continue; |
3253 | ||
6e543d57 | 3254 | if (sc.priority != DEF_PRIORITY && |
599d0c95 | 3255 | !pgdat_reclaimable(zone->zone_pgdat)) |
1da177e4 LT |
3256 | continue; |
3257 | ||
1da177e4 | 3258 | sc.nr_scanned = 0; |
4e416953 | 3259 | |
0608f43d AM |
3260 | nr_soft_scanned = 0; |
3261 | /* | |
3262 | * Call soft limit reclaim before calling shrink_zone. | |
3263 | */ | |
3264 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
3265 | order, sc.gfp_mask, | |
3266 | &nr_soft_scanned); | |
3267 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3268 | ||
32a4330d | 3269 | /* |
7c954f6d MG |
3270 | * There should be no need to raise the scanning |
3271 | * priority if enough pages are already being scanned | |
3272 | * that that high watermark would be met at 100% | |
3273 | * efficiency. | |
fe2c2a10 | 3274 | */ |
accf6242 | 3275 | if (kswapd_shrink_zone(zone, end_zone, &sc)) |
7c954f6d | 3276 | raise_priority = false; |
1da177e4 | 3277 | } |
5515061d MG |
3278 | |
3279 | /* | |
3280 | * If the low watermark is met there is no need for processes | |
3281 | * to be throttled on pfmemalloc_wait as they should not be | |
3282 | * able to safely make forward progress. Wake them | |
3283 | */ | |
3284 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
3285 | pfmemalloc_watermark_ok(pgdat)) | |
cfc51155 | 3286 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3287 | |
b8e83b94 MG |
3288 | /* Check if kswapd should be suspending */ |
3289 | if (try_to_freeze() || kthread_should_stop()) | |
3290 | break; | |
8357376d | 3291 | |
73ce02e9 | 3292 | /* |
b8e83b94 MG |
3293 | * Raise priority if scanning rate is too low or there was no |
3294 | * progress in reclaiming pages | |
73ce02e9 | 3295 | */ |
b8e83b94 MG |
3296 | if (raise_priority || !sc.nr_reclaimed) |
3297 | sc.priority--; | |
9aa41348 | 3298 | } while (sc.priority >= 1 && |
accf6242 | 3299 | !pgdat_balanced(pgdat, order, classzone_idx)); |
1da177e4 | 3300 | |
b8e83b94 | 3301 | out: |
0abdee2b | 3302 | /* |
accf6242 VB |
3303 | * Return the highest zone idx we were reclaiming at so |
3304 | * prepare_kswapd_sleep() makes the same decisions as here. | |
0abdee2b | 3305 | */ |
accf6242 | 3306 | return end_zone; |
1da177e4 LT |
3307 | } |
3308 | ||
accf6242 VB |
3309 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, |
3310 | int classzone_idx, int balanced_classzone_idx) | |
f0bc0a60 KM |
3311 | { |
3312 | long remaining = 0; | |
3313 | DEFINE_WAIT(wait); | |
3314 | ||
3315 | if (freezing(current) || kthread_should_stop()) | |
3316 | return; | |
3317 | ||
3318 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3319 | ||
3320 | /* Try to sleep for a short interval */ | |
accf6242 VB |
3321 | if (prepare_kswapd_sleep(pgdat, order, remaining, |
3322 | balanced_classzone_idx)) { | |
fd901c95 VB |
3323 | /* |
3324 | * Compaction records what page blocks it recently failed to | |
3325 | * isolate pages from and skips them in the future scanning. | |
3326 | * When kswapd is going to sleep, it is reasonable to assume | |
3327 | * that pages and compaction may succeed so reset the cache. | |
3328 | */ | |
3329 | reset_isolation_suitable(pgdat); | |
3330 | ||
3331 | /* | |
3332 | * We have freed the memory, now we should compact it to make | |
3333 | * allocation of the requested order possible. | |
3334 | */ | |
3335 | wakeup_kcompactd(pgdat, order, classzone_idx); | |
3336 | ||
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 | */ | |
accf6242 VB |
3346 | if (prepare_kswapd_sleep(pgdat, order, remaining, |
3347 | balanced_classzone_idx)) { | |
f0bc0a60 KM |
3348 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3349 | ||
3350 | /* | |
3351 | * vmstat counters are not perfectly accurate and the estimated | |
3352 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3353 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3354 | * watermarks being breached while under pressure, we reduce the | |
3355 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3356 | * them before going back to sleep. | |
3357 | */ | |
3358 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3359 | |
3360 | if (!kthread_should_stop()) | |
3361 | schedule(); | |
3362 | ||
f0bc0a60 KM |
3363 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3364 | } else { | |
3365 | if (remaining) | |
3366 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3367 | else | |
3368 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3369 | } | |
3370 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3371 | } | |
3372 | ||
1da177e4 LT |
3373 | /* |
3374 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3375 | * from the init process. |
1da177e4 LT |
3376 | * |
3377 | * This basically trickles out pages so that we have _some_ | |
3378 | * free memory available even if there is no other activity | |
3379 | * that frees anything up. This is needed for things like routing | |
3380 | * etc, where we otherwise might have all activity going on in | |
3381 | * asynchronous contexts that cannot page things out. | |
3382 | * | |
3383 | * If there are applications that are active memory-allocators | |
3384 | * (most normal use), this basically shouldn't matter. | |
3385 | */ | |
3386 | static int kswapd(void *p) | |
3387 | { | |
215ddd66 MG |
3388 | unsigned long order, new_order; |
3389 | int classzone_idx, new_classzone_idx; | |
d2ebd0f6 | 3390 | int balanced_classzone_idx; |
1da177e4 LT |
3391 | pg_data_t *pgdat = (pg_data_t*)p; |
3392 | struct task_struct *tsk = current; | |
f0bc0a60 | 3393 | |
1da177e4 LT |
3394 | struct reclaim_state reclaim_state = { |
3395 | .reclaimed_slab = 0, | |
3396 | }; | |
a70f7302 | 3397 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3398 | |
cf40bd16 NP |
3399 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3400 | ||
174596a0 | 3401 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3402 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3403 | current->reclaim_state = &reclaim_state; |
3404 | ||
3405 | /* | |
3406 | * Tell the memory management that we're a "memory allocator", | |
3407 | * and that if we need more memory we should get access to it | |
3408 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3409 | * never get caught in the normal page freeing logic. | |
3410 | * | |
3411 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3412 | * you need a small amount of memory in order to be able to | |
3413 | * page out something else, and this flag essentially protects | |
3414 | * us from recursively trying to free more memory as we're | |
3415 | * trying to free the first piece of memory in the first place). | |
3416 | */ | |
930d9152 | 3417 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3418 | set_freezable(); |
1da177e4 | 3419 | |
215ddd66 MG |
3420 | order = new_order = 0; |
3421 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; | |
d2ebd0f6 | 3422 | balanced_classzone_idx = classzone_idx; |
1da177e4 | 3423 | for ( ; ; ) { |
6f6313d4 | 3424 | bool ret; |
3e1d1d28 | 3425 | |
215ddd66 | 3426 | /* |
accf6242 VB |
3427 | * While we were reclaiming, there might have been another |
3428 | * wakeup, so check the values. | |
215ddd66 | 3429 | */ |
accf6242 VB |
3430 | new_order = pgdat->kswapd_max_order; |
3431 | new_classzone_idx = pgdat->classzone_idx; | |
3432 | pgdat->kswapd_max_order = 0; | |
3433 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
215ddd66 | 3434 | |
99504748 | 3435 | if (order < new_order || classzone_idx > new_classzone_idx) { |
1da177e4 LT |
3436 | /* |
3437 | * Don't sleep if someone wants a larger 'order' | |
99504748 | 3438 | * allocation or has tigher zone constraints |
1da177e4 LT |
3439 | */ |
3440 | order = new_order; | |
99504748 | 3441 | classzone_idx = new_classzone_idx; |
1da177e4 | 3442 | } else { |
accf6242 | 3443 | kswapd_try_to_sleep(pgdat, order, classzone_idx, |
d2ebd0f6 | 3444 | balanced_classzone_idx); |
1da177e4 | 3445 | order = pgdat->kswapd_max_order; |
99504748 | 3446 | classzone_idx = pgdat->classzone_idx; |
f0dfcde0 AS |
3447 | new_order = order; |
3448 | new_classzone_idx = classzone_idx; | |
4d40502e | 3449 | pgdat->kswapd_max_order = 0; |
215ddd66 | 3450 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
1da177e4 | 3451 | } |
1da177e4 | 3452 | |
8fe23e05 DR |
3453 | ret = try_to_freeze(); |
3454 | if (kthread_should_stop()) | |
3455 | break; | |
3456 | ||
3457 | /* | |
3458 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3459 | * after returning from the refrigerator | |
3460 | */ | |
33906bc5 MG |
3461 | if (!ret) { |
3462 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
accf6242 VB |
3463 | balanced_classzone_idx = balance_pgdat(pgdat, order, |
3464 | classzone_idx); | |
33906bc5 | 3465 | } |
1da177e4 | 3466 | } |
b0a8cc58 | 3467 | |
71abdc15 | 3468 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
b0a8cc58 | 3469 | current->reclaim_state = NULL; |
71abdc15 JW |
3470 | lockdep_clear_current_reclaim_state(); |
3471 | ||
1da177e4 LT |
3472 | return 0; |
3473 | } | |
3474 | ||
3475 | /* | |
3476 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3477 | */ | |
99504748 | 3478 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3479 | { |
3480 | pg_data_t *pgdat; | |
3481 | ||
f3fe6512 | 3482 | if (!populated_zone(zone)) |
1da177e4 LT |
3483 | return; |
3484 | ||
344736f2 | 3485 | if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) |
1da177e4 | 3486 | return; |
88f5acf8 | 3487 | pgdat = zone->zone_pgdat; |
99504748 | 3488 | if (pgdat->kswapd_max_order < order) { |
1da177e4 | 3489 | pgdat->kswapd_max_order = order; |
99504748 MG |
3490 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
3491 | } | |
8d0986e2 | 3492 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3493 | return; |
accf6242 | 3494 | if (zone_balanced(zone, order, true, 0, 0)) |
88f5acf8 MG |
3495 | return; |
3496 | ||
3497 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3498 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3499 | } |
3500 | ||
c6f37f12 | 3501 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3502 | /* |
7b51755c | 3503 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3504 | * freed pages. |
3505 | * | |
3506 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3507 | * LRU order by reclaiming preferentially | |
3508 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3509 | */ |
7b51755c | 3510 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3511 | { |
d6277db4 | 3512 | struct reclaim_state reclaim_state; |
d6277db4 | 3513 | struct scan_control sc = { |
ee814fe2 | 3514 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3515 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
ee814fe2 | 3516 | .priority = DEF_PRIORITY, |
d6277db4 | 3517 | .may_writepage = 1, |
ee814fe2 JW |
3518 | .may_unmap = 1, |
3519 | .may_swap = 1, | |
7b51755c | 3520 | .hibernation_mode = 1, |
1da177e4 | 3521 | }; |
a09ed5e0 | 3522 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c KM |
3523 | struct task_struct *p = current; |
3524 | unsigned long nr_reclaimed; | |
1da177e4 | 3525 | |
7b51755c KM |
3526 | p->flags |= PF_MEMALLOC; |
3527 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3528 | reclaim_state.reclaimed_slab = 0; | |
3529 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3530 | |
3115cd91 | 3531 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3532 | |
7b51755c KM |
3533 | p->reclaim_state = NULL; |
3534 | lockdep_clear_current_reclaim_state(); | |
3535 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3536 | |
7b51755c | 3537 | return nr_reclaimed; |
1da177e4 | 3538 | } |
c6f37f12 | 3539 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3540 | |
1da177e4 LT |
3541 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3542 | not required for correctness. So if the last cpu in a node goes | |
3543 | away, we get changed to run anywhere: as the first one comes back, | |
3544 | restore their cpu bindings. */ | |
fcb35a9b GKH |
3545 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
3546 | void *hcpu) | |
1da177e4 | 3547 | { |
58c0a4a7 | 3548 | int nid; |
1da177e4 | 3549 | |
8bb78442 | 3550 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
48fb2e24 | 3551 | for_each_node_state(nid, N_MEMORY) { |
c5f59f08 | 3552 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3553 | const struct cpumask *mask; |
3554 | ||
3555 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3556 | |
3e597945 | 3557 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3558 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3559 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3560 | } |
3561 | } | |
3562 | return NOTIFY_OK; | |
3563 | } | |
1da177e4 | 3564 | |
3218ae14 YG |
3565 | /* |
3566 | * This kswapd start function will be called by init and node-hot-add. | |
3567 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3568 | */ | |
3569 | int kswapd_run(int nid) | |
3570 | { | |
3571 | pg_data_t *pgdat = NODE_DATA(nid); | |
3572 | int ret = 0; | |
3573 | ||
3574 | if (pgdat->kswapd) | |
3575 | return 0; | |
3576 | ||
3577 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3578 | if (IS_ERR(pgdat->kswapd)) { | |
3579 | /* failure at boot is fatal */ | |
3580 | BUG_ON(system_state == SYSTEM_BOOTING); | |
d5dc0ad9 GS |
3581 | pr_err("Failed to start kswapd on node %d\n", nid); |
3582 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3583 | pgdat->kswapd = NULL; |
3218ae14 YG |
3584 | } |
3585 | return ret; | |
3586 | } | |
3587 | ||
8fe23e05 | 3588 | /* |
d8adde17 | 3589 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 3590 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
3591 | */ |
3592 | void kswapd_stop(int nid) | |
3593 | { | |
3594 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3595 | ||
d8adde17 | 3596 | if (kswapd) { |
8fe23e05 | 3597 | kthread_stop(kswapd); |
d8adde17 JL |
3598 | NODE_DATA(nid)->kswapd = NULL; |
3599 | } | |
8fe23e05 DR |
3600 | } |
3601 | ||
1da177e4 LT |
3602 | static int __init kswapd_init(void) |
3603 | { | |
3218ae14 | 3604 | int nid; |
69e05944 | 3605 | |
1da177e4 | 3606 | swap_setup(); |
48fb2e24 | 3607 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3608 | kswapd_run(nid); |
1da177e4 LT |
3609 | hotcpu_notifier(cpu_callback, 0); |
3610 | return 0; | |
3611 | } | |
3612 | ||
3613 | module_init(kswapd_init) | |
9eeff239 CL |
3614 | |
3615 | #ifdef CONFIG_NUMA | |
3616 | /* | |
3617 | * Zone reclaim mode | |
3618 | * | |
3619 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3620 | * the watermarks. | |
9eeff239 CL |
3621 | */ |
3622 | int zone_reclaim_mode __read_mostly; | |
3623 | ||
1b2ffb78 | 3624 | #define RECLAIM_OFF 0 |
7d03431c | 3625 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 3626 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 3627 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 3628 | |
a92f7126 CL |
3629 | /* |
3630 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3631 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3632 | * a zone. | |
3633 | */ | |
3634 | #define ZONE_RECLAIM_PRIORITY 4 | |
3635 | ||
9614634f CL |
3636 | /* |
3637 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3638 | * occur. | |
3639 | */ | |
3640 | int sysctl_min_unmapped_ratio = 1; | |
3641 | ||
0ff38490 CL |
3642 | /* |
3643 | * If the number of slab pages in a zone grows beyond this percentage then | |
3644 | * slab reclaim needs to occur. | |
3645 | */ | |
3646 | int sysctl_min_slab_ratio = 5; | |
3647 | ||
90afa5de MG |
3648 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
3649 | { | |
3650 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
599d0c95 MG |
3651 | unsigned long file_lru = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) + |
3652 | node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
3653 | |
3654 | /* | |
3655 | * It's possible for there to be more file mapped pages than | |
3656 | * accounted for by the pages on the file LRU lists because | |
3657 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3658 | */ | |
3659 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3660 | } | |
3661 | ||
3662 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
d031a157 | 3663 | static unsigned long zone_pagecache_reclaimable(struct zone *zone) |
90afa5de | 3664 | { |
d031a157 AM |
3665 | unsigned long nr_pagecache_reclaimable; |
3666 | unsigned long delta = 0; | |
90afa5de MG |
3667 | |
3668 | /* | |
95bbc0c7 | 3669 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de MG |
3670 | * potentially reclaimable. Otherwise, we have to worry about |
3671 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3672 | * a better estimate | |
3673 | */ | |
95bbc0c7 | 3674 | if (zone_reclaim_mode & RECLAIM_UNMAP) |
90afa5de MG |
3675 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); |
3676 | else | |
3677 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3678 | ||
3679 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3680 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3681 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3682 | ||
3683 | /* Watch for any possible underflows due to delta */ | |
3684 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3685 | delta = nr_pagecache_reclaimable; | |
3686 | ||
3687 | return nr_pagecache_reclaimable - delta; | |
3688 | } | |
3689 | ||
9eeff239 CL |
3690 | /* |
3691 | * Try to free up some pages from this zone through reclaim. | |
3692 | */ | |
179e9639 | 3693 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3694 | { |
7fb2d46d | 3695 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3696 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3697 | struct task_struct *p = current; |
3698 | struct reclaim_state reclaim_state; | |
179e9639 | 3699 | struct scan_control sc = { |
62b726c1 | 3700 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
21caf2fc | 3701 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
bd2f6199 | 3702 | .order = order, |
9e3b2f8c | 3703 | .priority = ZONE_RECLAIM_PRIORITY, |
ee814fe2 | 3704 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), |
95bbc0c7 | 3705 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP), |
ee814fe2 | 3706 | .may_swap = 1, |
179e9639 | 3707 | }; |
9eeff239 | 3708 | |
9eeff239 | 3709 | cond_resched(); |
d4f7796e | 3710 | /* |
95bbc0c7 | 3711 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 3712 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 3713 | * and RECLAIM_UNMAP. |
d4f7796e CL |
3714 | */ |
3715 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3716 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3717 | reclaim_state.reclaimed_slab = 0; |
3718 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3719 | |
90afa5de | 3720 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
3721 | /* |
3722 | * Free memory by calling shrink zone with increasing | |
3723 | * priorities until we have enough memory freed. | |
3724 | */ | |
0ff38490 | 3725 | do { |
6b4f7799 | 3726 | shrink_zone(zone, &sc, true); |
9e3b2f8c | 3727 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 3728 | } |
c84db23c | 3729 | |
9eeff239 | 3730 | p->reclaim_state = NULL; |
d4f7796e | 3731 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3732 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3733 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3734 | } |
179e9639 AM |
3735 | |
3736 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3737 | { | |
179e9639 | 3738 | int node_id; |
d773ed6b | 3739 | int ret; |
179e9639 AM |
3740 | |
3741 | /* | |
0ff38490 CL |
3742 | * Zone reclaim reclaims unmapped file backed pages and |
3743 | * slab pages if we are over the defined limits. | |
34aa1330 | 3744 | * |
9614634f CL |
3745 | * A small portion of unmapped file backed pages is needed for |
3746 | * file I/O otherwise pages read by file I/O will be immediately | |
3747 | * thrown out if the zone is overallocated. So we do not reclaim | |
3748 | * if less than a specified percentage of the zone is used by | |
3749 | * unmapped file backed pages. | |
179e9639 | 3750 | */ |
90afa5de MG |
3751 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
3752 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 3753 | return ZONE_RECLAIM_FULL; |
179e9639 | 3754 | |
599d0c95 | 3755 | if (!pgdat_reclaimable(zone->zone_pgdat)) |
fa5e084e | 3756 | return ZONE_RECLAIM_FULL; |
d773ed6b | 3757 | |
179e9639 | 3758 | /* |
d773ed6b | 3759 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3760 | */ |
d0164adc | 3761 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 3762 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
3763 | |
3764 | /* | |
3765 | * Only run zone reclaim on the local zone or on zones that do not | |
3766 | * have associated processors. This will favor the local processor | |
3767 | * over remote processors and spread off node memory allocations | |
3768 | * as wide as possible. | |
3769 | */ | |
89fa3024 | 3770 | node_id = zone_to_nid(zone); |
37c0708d | 3771 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 3772 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b | 3773 | |
57054651 | 3774 | if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags)) |
fa5e084e MG |
3775 | return ZONE_RECLAIM_NOSCAN; |
3776 | ||
d773ed6b | 3777 | ret = __zone_reclaim(zone, gfp_mask, order); |
57054651 | 3778 | clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags); |
d773ed6b | 3779 | |
24cf7251 MG |
3780 | if (!ret) |
3781 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3782 | ||
d773ed6b | 3783 | return ret; |
179e9639 | 3784 | } |
9eeff239 | 3785 | #endif |
894bc310 | 3786 | |
894bc310 LS |
3787 | /* |
3788 | * page_evictable - test whether a page is evictable | |
3789 | * @page: the page to test | |
894bc310 LS |
3790 | * |
3791 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3792 | * lists vs unevictable list. |
894bc310 LS |
3793 | * |
3794 | * Reasons page might not be evictable: | |
ba9ddf49 | 3795 | * (1) page's mapping marked unevictable |
b291f000 | 3796 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3797 | * |
894bc310 | 3798 | */ |
39b5f29a | 3799 | int page_evictable(struct page *page) |
894bc310 | 3800 | { |
39b5f29a | 3801 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
894bc310 | 3802 | } |
89e004ea | 3803 | |
85046579 | 3804 | #ifdef CONFIG_SHMEM |
89e004ea | 3805 | /** |
24513264 HD |
3806 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3807 | * @pages: array of pages to check | |
3808 | * @nr_pages: number of pages to check | |
89e004ea | 3809 | * |
24513264 | 3810 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3811 | * |
3812 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3813 | */ |
24513264 | 3814 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3815 | { |
925b7673 | 3816 | struct lruvec *lruvec; |
24513264 HD |
3817 | struct zone *zone = NULL; |
3818 | int pgscanned = 0; | |
3819 | int pgrescued = 0; | |
3820 | int i; | |
89e004ea | 3821 | |
24513264 HD |
3822 | for (i = 0; i < nr_pages; i++) { |
3823 | struct page *page = pages[i]; | |
3824 | struct zone *pagezone; | |
89e004ea | 3825 | |
24513264 HD |
3826 | pgscanned++; |
3827 | pagezone = page_zone(page); | |
3828 | if (pagezone != zone) { | |
3829 | if (zone) | |
a52633d8 | 3830 | spin_unlock_irq(zone_lru_lock(zone)); |
24513264 | 3831 | zone = pagezone; |
a52633d8 | 3832 | spin_lock_irq(zone_lru_lock(zone)); |
24513264 | 3833 | } |
599d0c95 | 3834 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
89e004ea | 3835 | |
24513264 HD |
3836 | if (!PageLRU(page) || !PageUnevictable(page)) |
3837 | continue; | |
89e004ea | 3838 | |
39b5f29a | 3839 | if (page_evictable(page)) { |
24513264 HD |
3840 | enum lru_list lru = page_lru_base_type(page); |
3841 | ||
309381fe | 3842 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 3843 | ClearPageUnevictable(page); |
fa9add64 HD |
3844 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
3845 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 3846 | pgrescued++; |
89e004ea | 3847 | } |
24513264 | 3848 | } |
89e004ea | 3849 | |
24513264 HD |
3850 | if (zone) { |
3851 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3852 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
a52633d8 | 3853 | spin_unlock_irq(zone_lru_lock(zone)); |
89e004ea | 3854 | } |
89e004ea | 3855 | } |
85046579 | 3856 | #endif /* CONFIG_SHMEM */ |