]>
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 | ||
14 | #include <linux/mm.h> | |
15 | #include <linux/module.h> | |
5a0e3ad6 | 16 | #include <linux/gfp.h> |
1da177e4 LT |
17 | #include <linux/kernel_stat.h> |
18 | #include <linux/swap.h> | |
19 | #include <linux/pagemap.h> | |
20 | #include <linux/init.h> | |
21 | #include <linux/highmem.h> | |
70ddf637 | 22 | #include <linux/vmpressure.h> |
e129b5c2 | 23 | #include <linux/vmstat.h> |
1da177e4 LT |
24 | #include <linux/file.h> |
25 | #include <linux/writeback.h> | |
26 | #include <linux/blkdev.h> | |
27 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
28 | buffer_heads_over_limit */ | |
29 | #include <linux/mm_inline.h> | |
1da177e4 LT |
30 | #include <linux/backing-dev.h> |
31 | #include <linux/rmap.h> | |
32 | #include <linux/topology.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/cpuset.h> | |
3e7d3449 | 35 | #include <linux/compaction.h> |
1da177e4 LT |
36 | #include <linux/notifier.h> |
37 | #include <linux/rwsem.h> | |
248a0301 | 38 | #include <linux/delay.h> |
3218ae14 | 39 | #include <linux/kthread.h> |
7dfb7103 | 40 | #include <linux/freezer.h> |
66e1707b | 41 | #include <linux/memcontrol.h> |
873b4771 | 42 | #include <linux/delayacct.h> |
af936a16 | 43 | #include <linux/sysctl.h> |
929bea7c | 44 | #include <linux/oom.h> |
268bb0ce | 45 | #include <linux/prefetch.h> |
1da177e4 LT |
46 | |
47 | #include <asm/tlbflush.h> | |
48 | #include <asm/div64.h> | |
49 | ||
50 | #include <linux/swapops.h> | |
117aad1e | 51 | #include <linux/balloon_compaction.h> |
1da177e4 | 52 | |
0f8053a5 NP |
53 | #include "internal.h" |
54 | ||
33906bc5 MG |
55 | #define CREATE_TRACE_POINTS |
56 | #include <trace/events/vmscan.h> | |
57 | ||
1da177e4 | 58 | struct scan_control { |
1da177e4 LT |
59 | /* Incremented by the number of inactive pages that were scanned */ |
60 | unsigned long nr_scanned; | |
61 | ||
a79311c1 RR |
62 | /* Number of pages freed so far during a call to shrink_zones() */ |
63 | unsigned long nr_reclaimed; | |
64 | ||
22fba335 KM |
65 | /* How many pages shrink_list() should reclaim */ |
66 | unsigned long nr_to_reclaim; | |
67 | ||
7b51755c KM |
68 | unsigned long hibernation_mode; |
69 | ||
1da177e4 | 70 | /* This context's GFP mask */ |
6daa0e28 | 71 | gfp_t gfp_mask; |
1da177e4 LT |
72 | |
73 | int may_writepage; | |
74 | ||
a6dc60f8 JW |
75 | /* Can mapped pages be reclaimed? */ |
76 | int may_unmap; | |
f1fd1067 | 77 | |
2e2e4259 KM |
78 | /* Can pages be swapped as part of reclaim? */ |
79 | int may_swap; | |
80 | ||
5ad333eb | 81 | int order; |
66e1707b | 82 | |
9e3b2f8c KK |
83 | /* Scan (total_size >> priority) pages at once */ |
84 | int priority; | |
85 | ||
f16015fb JW |
86 | /* |
87 | * The memory cgroup that hit its limit and as a result is the | |
88 | * primary target of this reclaim invocation. | |
89 | */ | |
90 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 91 | |
327c0e96 KH |
92 | /* |
93 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
94 | * are scanned. | |
95 | */ | |
96 | nodemask_t *nodemask; | |
1da177e4 LT |
97 | }; |
98 | ||
1da177e4 LT |
99 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
100 | ||
101 | #ifdef ARCH_HAS_PREFETCH | |
102 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
103 | do { \ | |
104 | if ((_page)->lru.prev != _base) { \ | |
105 | struct page *prev; \ | |
106 | \ | |
107 | prev = lru_to_page(&(_page->lru)); \ | |
108 | prefetch(&prev->_field); \ | |
109 | } \ | |
110 | } while (0) | |
111 | #else | |
112 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
113 | #endif | |
114 | ||
115 | #ifdef ARCH_HAS_PREFETCHW | |
116 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
117 | do { \ | |
118 | if ((_page)->lru.prev != _base) { \ | |
119 | struct page *prev; \ | |
120 | \ | |
121 | prev = lru_to_page(&(_page->lru)); \ | |
122 | prefetchw(&prev->_field); \ | |
123 | } \ | |
124 | } while (0) | |
125 | #else | |
126 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
127 | #endif | |
128 | ||
129 | /* | |
130 | * From 0 .. 100. Higher means more swappy. | |
131 | */ | |
132 | int vm_swappiness = 60; | |
b21e0b90 | 133 | unsigned long vm_total_pages; /* The total number of pages which the VM controls */ |
1da177e4 LT |
134 | |
135 | static LIST_HEAD(shrinker_list); | |
136 | static DECLARE_RWSEM(shrinker_rwsem); | |
137 | ||
c255a458 | 138 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
139 | static bool global_reclaim(struct scan_control *sc) |
140 | { | |
f16015fb | 141 | return !sc->target_mem_cgroup; |
89b5fae5 | 142 | } |
91a45470 | 143 | #else |
89b5fae5 JW |
144 | static bool global_reclaim(struct scan_control *sc) |
145 | { | |
146 | return true; | |
147 | } | |
91a45470 KH |
148 | #endif |
149 | ||
a1c3bfb2 | 150 | static unsigned long zone_reclaimable_pages(struct zone *zone) |
6e543d57 LD |
151 | { |
152 | int nr; | |
153 | ||
154 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
155 | zone_page_state(zone, NR_INACTIVE_FILE); | |
156 | ||
157 | if (get_nr_swap_pages() > 0) | |
158 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
159 | zone_page_state(zone, NR_INACTIVE_ANON); | |
160 | ||
161 | return nr; | |
162 | } | |
163 | ||
164 | bool zone_reclaimable(struct zone *zone) | |
165 | { | |
166 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
167 | } | |
168 | ||
4d7dcca2 | 169 | static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) |
c9f299d9 | 170 | { |
c3c787e8 | 171 | if (!mem_cgroup_disabled()) |
4d7dcca2 | 172 | return mem_cgroup_get_lru_size(lruvec, lru); |
a3d8e054 | 173 | |
074291fe | 174 | return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); |
c9f299d9 KM |
175 | } |
176 | ||
1da177e4 | 177 | /* |
1d3d4437 | 178 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 179 | */ |
1d3d4437 | 180 | int register_shrinker(struct shrinker *shrinker) |
1da177e4 | 181 | { |
1d3d4437 GC |
182 | size_t size = sizeof(*shrinker->nr_deferred); |
183 | ||
184 | /* | |
185 | * If we only have one possible node in the system anyway, save | |
186 | * ourselves the trouble and disable NUMA aware behavior. This way we | |
187 | * will save memory and some small loop time later. | |
188 | */ | |
189 | if (nr_node_ids == 1) | |
190 | shrinker->flags &= ~SHRINKER_NUMA_AWARE; | |
191 | ||
192 | if (shrinker->flags & SHRINKER_NUMA_AWARE) | |
193 | size *= nr_node_ids; | |
194 | ||
195 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
196 | if (!shrinker->nr_deferred) | |
197 | return -ENOMEM; | |
198 | ||
8e1f936b RR |
199 | down_write(&shrinker_rwsem); |
200 | list_add_tail(&shrinker->list, &shrinker_list); | |
201 | up_write(&shrinker_rwsem); | |
1d3d4437 | 202 | return 0; |
1da177e4 | 203 | } |
8e1f936b | 204 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
205 | |
206 | /* | |
207 | * Remove one | |
208 | */ | |
8e1f936b | 209 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
210 | { |
211 | down_write(&shrinker_rwsem); | |
212 | list_del(&shrinker->list); | |
213 | up_write(&shrinker_rwsem); | |
ae393321 | 214 | kfree(shrinker->nr_deferred); |
1da177e4 | 215 | } |
8e1f936b | 216 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
217 | |
218 | #define SHRINK_BATCH 128 | |
1d3d4437 GC |
219 | |
220 | static unsigned long | |
221 | shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker, | |
222 | unsigned long nr_pages_scanned, unsigned long lru_pages) | |
223 | { | |
224 | unsigned long freed = 0; | |
225 | unsigned long long delta; | |
226 | long total_scan; | |
d5bc5fd3 | 227 | long freeable; |
1d3d4437 GC |
228 | long nr; |
229 | long new_nr; | |
230 | int nid = shrinkctl->nid; | |
231 | long batch_size = shrinker->batch ? shrinker->batch | |
232 | : SHRINK_BATCH; | |
233 | ||
d5bc5fd3 VD |
234 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
235 | if (freeable == 0) | |
1d3d4437 GC |
236 | return 0; |
237 | ||
238 | /* | |
239 | * copy the current shrinker scan count into a local variable | |
240 | * and zero it so that other concurrent shrinker invocations | |
241 | * don't also do this scanning work. | |
242 | */ | |
243 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
244 | ||
245 | total_scan = nr; | |
246 | delta = (4 * nr_pages_scanned) / shrinker->seeks; | |
d5bc5fd3 | 247 | delta *= freeable; |
1d3d4437 GC |
248 | do_div(delta, lru_pages + 1); |
249 | total_scan += delta; | |
250 | if (total_scan < 0) { | |
251 | printk(KERN_ERR | |
252 | "shrink_slab: %pF negative objects to delete nr=%ld\n", | |
a0b02131 | 253 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 254 | total_scan = freeable; |
1d3d4437 GC |
255 | } |
256 | ||
257 | /* | |
258 | * We need to avoid excessive windup on filesystem shrinkers | |
259 | * due to large numbers of GFP_NOFS allocations causing the | |
260 | * shrinkers to return -1 all the time. This results in a large | |
261 | * nr being built up so when a shrink that can do some work | |
262 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 263 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
264 | * memory. |
265 | * | |
266 | * Hence only allow the shrinker to scan the entire cache when | |
267 | * a large delta change is calculated directly. | |
268 | */ | |
d5bc5fd3 VD |
269 | if (delta < freeable / 4) |
270 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
271 | |
272 | /* | |
273 | * Avoid risking looping forever due to too large nr value: | |
274 | * never try to free more than twice the estimate number of | |
275 | * freeable entries. | |
276 | */ | |
d5bc5fd3 VD |
277 | if (total_scan > freeable * 2) |
278 | total_scan = freeable * 2; | |
1d3d4437 GC |
279 | |
280 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
281 | nr_pages_scanned, lru_pages, | |
d5bc5fd3 | 282 | freeable, delta, total_scan); |
1d3d4437 | 283 | |
0b1fb40a VD |
284 | /* |
285 | * Normally, we should not scan less than batch_size objects in one | |
286 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
287 | * than batch_size objects in total and we are really tight on memory, | |
288 | * we will try to reclaim all available objects, otherwise we can end | |
289 | * up failing allocations although there are plenty of reclaimable | |
290 | * objects spread over several slabs with usage less than the | |
291 | * batch_size. | |
292 | * | |
293 | * We detect the "tight on memory" situations by looking at the total | |
294 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 295 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
296 | * scanning at high prio and therefore should try to reclaim as much as |
297 | * possible. | |
298 | */ | |
299 | while (total_scan >= batch_size || | |
d5bc5fd3 | 300 | total_scan >= freeable) { |
a0b02131 | 301 | unsigned long ret; |
0b1fb40a | 302 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 303 | |
0b1fb40a | 304 | shrinkctl->nr_to_scan = nr_to_scan; |
a0b02131 DC |
305 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
306 | if (ret == SHRINK_STOP) | |
307 | break; | |
308 | freed += ret; | |
1d3d4437 | 309 | |
0b1fb40a VD |
310 | count_vm_events(SLABS_SCANNED, nr_to_scan); |
311 | total_scan -= nr_to_scan; | |
1d3d4437 GC |
312 | |
313 | cond_resched(); | |
314 | } | |
315 | ||
316 | /* | |
317 | * move the unused scan count back into the shrinker in a | |
318 | * manner that handles concurrent updates. If we exhausted the | |
319 | * scan, there is no need to do an update. | |
320 | */ | |
321 | if (total_scan > 0) | |
322 | new_nr = atomic_long_add_return(total_scan, | |
323 | &shrinker->nr_deferred[nid]); | |
324 | else | |
325 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
326 | ||
df9024a8 | 327 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 328 | return freed; |
1495f230 YH |
329 | } |
330 | ||
1da177e4 LT |
331 | /* |
332 | * Call the shrink functions to age shrinkable caches | |
333 | * | |
334 | * Here we assume it costs one seek to replace a lru page and that it also | |
335 | * takes a seek to recreate a cache object. With this in mind we age equal | |
336 | * percentages of the lru and ageable caches. This should balance the seeks | |
337 | * generated by these structures. | |
338 | * | |
183ff22b | 339 | * If the vm encountered mapped pages on the LRU it increase the pressure on |
1da177e4 LT |
340 | * slab to avoid swapping. |
341 | * | |
342 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
343 | * | |
344 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
345 | * are eligible for the caller's allocation attempt. It is used for balancing | |
346 | * slab reclaim versus page reclaim. | |
b15e0905 | 347 | * |
348 | * Returns the number of slab objects which we shrunk. | |
1da177e4 | 349 | */ |
24f7c6b9 | 350 | unsigned long shrink_slab(struct shrink_control *shrinkctl, |
1495f230 | 351 | unsigned long nr_pages_scanned, |
a09ed5e0 | 352 | unsigned long lru_pages) |
1da177e4 LT |
353 | { |
354 | struct shrinker *shrinker; | |
24f7c6b9 | 355 | unsigned long freed = 0; |
1da177e4 | 356 | |
1495f230 YH |
357 | if (nr_pages_scanned == 0) |
358 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 359 | |
f06590bd | 360 | if (!down_read_trylock(&shrinker_rwsem)) { |
24f7c6b9 DC |
361 | /* |
362 | * If we would return 0, our callers would understand that we | |
363 | * have nothing else to shrink and give up trying. By returning | |
364 | * 1 we keep it going and assume we'll be able to shrink next | |
365 | * time. | |
366 | */ | |
367 | freed = 1; | |
f06590bd MK |
368 | goto out; |
369 | } | |
1da177e4 LT |
370 | |
371 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
ec97097b VD |
372 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) { |
373 | shrinkctl->nid = 0; | |
1d3d4437 | 374 | freed += shrink_slab_node(shrinkctl, shrinker, |
ec97097b VD |
375 | nr_pages_scanned, lru_pages); |
376 | continue; | |
377 | } | |
378 | ||
379 | for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) { | |
380 | if (node_online(shrinkctl->nid)) | |
381 | freed += shrink_slab_node(shrinkctl, shrinker, | |
382 | nr_pages_scanned, lru_pages); | |
1da177e4 | 383 | |
1da177e4 | 384 | } |
1da177e4 LT |
385 | } |
386 | up_read(&shrinker_rwsem); | |
f06590bd MK |
387 | out: |
388 | cond_resched(); | |
24f7c6b9 | 389 | return freed; |
1da177e4 LT |
390 | } |
391 | ||
1da177e4 LT |
392 | static inline int is_page_cache_freeable(struct page *page) |
393 | { | |
ceddc3a5 JW |
394 | /* |
395 | * A freeable page cache page is referenced only by the caller | |
396 | * that isolated the page, the page cache radix tree and | |
397 | * optional buffer heads at page->private. | |
398 | */ | |
edcf4748 | 399 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
400 | } |
401 | ||
7d3579e8 KM |
402 | static int may_write_to_queue(struct backing_dev_info *bdi, |
403 | struct scan_control *sc) | |
1da177e4 | 404 | { |
930d9152 | 405 | if (current->flags & PF_SWAPWRITE) |
1da177e4 LT |
406 | return 1; |
407 | if (!bdi_write_congested(bdi)) | |
408 | return 1; | |
409 | if (bdi == current->backing_dev_info) | |
410 | return 1; | |
411 | return 0; | |
412 | } | |
413 | ||
414 | /* | |
415 | * We detected a synchronous write error writing a page out. Probably | |
416 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
417 | * fsync(), msync() or close(). | |
418 | * | |
419 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
420 | * prevents it from being freed up. But we have a ref on the page and once | |
421 | * that page is locked, the mapping is pinned. | |
422 | * | |
423 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
424 | * __GFP_FS. | |
425 | */ | |
426 | static void handle_write_error(struct address_space *mapping, | |
427 | struct page *page, int error) | |
428 | { | |
7eaceacc | 429 | lock_page(page); |
3e9f45bd GC |
430 | if (page_mapping(page) == mapping) |
431 | mapping_set_error(mapping, error); | |
1da177e4 LT |
432 | unlock_page(page); |
433 | } | |
434 | ||
04e62a29 CL |
435 | /* possible outcome of pageout() */ |
436 | typedef enum { | |
437 | /* failed to write page out, page is locked */ | |
438 | PAGE_KEEP, | |
439 | /* move page to the active list, page is locked */ | |
440 | PAGE_ACTIVATE, | |
441 | /* page has been sent to the disk successfully, page is unlocked */ | |
442 | PAGE_SUCCESS, | |
443 | /* page is clean and locked */ | |
444 | PAGE_CLEAN, | |
445 | } pageout_t; | |
446 | ||
1da177e4 | 447 | /* |
1742f19f AM |
448 | * pageout is called by shrink_page_list() for each dirty page. |
449 | * Calls ->writepage(). | |
1da177e4 | 450 | */ |
c661b078 | 451 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 452 | struct scan_control *sc) |
1da177e4 LT |
453 | { |
454 | /* | |
455 | * If the page is dirty, only perform writeback if that write | |
456 | * will be non-blocking. To prevent this allocation from being | |
457 | * stalled by pagecache activity. But note that there may be | |
458 | * stalls if we need to run get_block(). We could test | |
459 | * PagePrivate for that. | |
460 | * | |
6aceb53b | 461 | * If this process is currently in __generic_file_aio_write() against |
1da177e4 LT |
462 | * this page's queue, we can perform writeback even if that |
463 | * will block. | |
464 | * | |
465 | * If the page is swapcache, write it back even if that would | |
466 | * block, for some throttling. This happens by accident, because | |
467 | * swap_backing_dev_info is bust: it doesn't reflect the | |
468 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
469 | */ |
470 | if (!is_page_cache_freeable(page)) | |
471 | return PAGE_KEEP; | |
472 | if (!mapping) { | |
473 | /* | |
474 | * Some data journaling orphaned pages can have | |
475 | * page->mapping == NULL while being dirty with clean buffers. | |
476 | */ | |
266cf658 | 477 | if (page_has_private(page)) { |
1da177e4 LT |
478 | if (try_to_free_buffers(page)) { |
479 | ClearPageDirty(page); | |
d40cee24 | 480 | printk("%s: orphaned page\n", __func__); |
1da177e4 LT |
481 | return PAGE_CLEAN; |
482 | } | |
483 | } | |
484 | return PAGE_KEEP; | |
485 | } | |
486 | if (mapping->a_ops->writepage == NULL) | |
487 | return PAGE_ACTIVATE; | |
0e093d99 | 488 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) |
1da177e4 LT |
489 | return PAGE_KEEP; |
490 | ||
491 | if (clear_page_dirty_for_io(page)) { | |
492 | int res; | |
493 | struct writeback_control wbc = { | |
494 | .sync_mode = WB_SYNC_NONE, | |
495 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
496 | .range_start = 0, |
497 | .range_end = LLONG_MAX, | |
1da177e4 LT |
498 | .for_reclaim = 1, |
499 | }; | |
500 | ||
501 | SetPageReclaim(page); | |
502 | res = mapping->a_ops->writepage(page, &wbc); | |
503 | if (res < 0) | |
504 | handle_write_error(mapping, page, res); | |
994fc28c | 505 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
506 | ClearPageReclaim(page); |
507 | return PAGE_ACTIVATE; | |
508 | } | |
c661b078 | 509 | |
1da177e4 LT |
510 | if (!PageWriteback(page)) { |
511 | /* synchronous write or broken a_ops? */ | |
512 | ClearPageReclaim(page); | |
513 | } | |
23b9da55 | 514 | trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); |
e129b5c2 | 515 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
516 | return PAGE_SUCCESS; |
517 | } | |
518 | ||
519 | return PAGE_CLEAN; | |
520 | } | |
521 | ||
a649fd92 | 522 | /* |
e286781d NP |
523 | * Same as remove_mapping, but if the page is removed from the mapping, it |
524 | * gets returned with a refcount of 0. | |
a649fd92 | 525 | */ |
a528910e JW |
526 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
527 | bool reclaimed) | |
49d2e9cc | 528 | { |
28e4d965 NP |
529 | BUG_ON(!PageLocked(page)); |
530 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 531 | |
19fd6231 | 532 | spin_lock_irq(&mapping->tree_lock); |
49d2e9cc | 533 | /* |
0fd0e6b0 NP |
534 | * The non racy check for a busy page. |
535 | * | |
536 | * Must be careful with the order of the tests. When someone has | |
537 | * a ref to the page, it may be possible that they dirty it then | |
538 | * drop the reference. So if PageDirty is tested before page_count | |
539 | * here, then the following race may occur: | |
540 | * | |
541 | * get_user_pages(&page); | |
542 | * [user mapping goes away] | |
543 | * write_to(page); | |
544 | * !PageDirty(page) [good] | |
545 | * SetPageDirty(page); | |
546 | * put_page(page); | |
547 | * !page_count(page) [good, discard it] | |
548 | * | |
549 | * [oops, our write_to data is lost] | |
550 | * | |
551 | * Reversing the order of the tests ensures such a situation cannot | |
552 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
553 | * load is not satisfied before that of page->_count. | |
554 | * | |
555 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
556 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 557 | */ |
e286781d | 558 | if (!page_freeze_refs(page, 2)) |
49d2e9cc | 559 | goto cannot_free; |
e286781d NP |
560 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
561 | if (unlikely(PageDirty(page))) { | |
562 | page_unfreeze_refs(page, 2); | |
49d2e9cc | 563 | goto cannot_free; |
e286781d | 564 | } |
49d2e9cc CL |
565 | |
566 | if (PageSwapCache(page)) { | |
567 | swp_entry_t swap = { .val = page_private(page) }; | |
568 | __delete_from_swap_cache(page); | |
19fd6231 | 569 | spin_unlock_irq(&mapping->tree_lock); |
cb4b86ba | 570 | swapcache_free(swap, page); |
e286781d | 571 | } else { |
6072d13c | 572 | void (*freepage)(struct page *); |
a528910e | 573 | void *shadow = NULL; |
6072d13c LT |
574 | |
575 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
576 | /* |
577 | * Remember a shadow entry for reclaimed file cache in | |
578 | * order to detect refaults, thus thrashing, later on. | |
579 | * | |
580 | * But don't store shadows in an address space that is | |
581 | * already exiting. This is not just an optizimation, | |
582 | * inode reclaim needs to empty out the radix tree or | |
583 | * the nodes are lost. Don't plant shadows behind its | |
584 | * back. | |
585 | */ | |
586 | if (reclaimed && page_is_file_cache(page) && | |
587 | !mapping_exiting(mapping)) | |
588 | shadow = workingset_eviction(mapping, page); | |
589 | __delete_from_page_cache(page, shadow); | |
19fd6231 | 590 | spin_unlock_irq(&mapping->tree_lock); |
e767e056 | 591 | mem_cgroup_uncharge_cache_page(page); |
6072d13c LT |
592 | |
593 | if (freepage != NULL) | |
594 | freepage(page); | |
49d2e9cc CL |
595 | } |
596 | ||
49d2e9cc CL |
597 | return 1; |
598 | ||
599 | cannot_free: | |
19fd6231 | 600 | spin_unlock_irq(&mapping->tree_lock); |
49d2e9cc CL |
601 | return 0; |
602 | } | |
603 | ||
e286781d NP |
604 | /* |
605 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
606 | * someone else has a ref on the page, abort and return 0. If it was | |
607 | * successfully detached, return 1. Assumes the caller has a single ref on | |
608 | * this page. | |
609 | */ | |
610 | int remove_mapping(struct address_space *mapping, struct page *page) | |
611 | { | |
a528910e | 612 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
613 | /* |
614 | * Unfreezing the refcount with 1 rather than 2 effectively | |
615 | * drops the pagecache ref for us without requiring another | |
616 | * atomic operation. | |
617 | */ | |
618 | page_unfreeze_refs(page, 1); | |
619 | return 1; | |
620 | } | |
621 | return 0; | |
622 | } | |
623 | ||
894bc310 LS |
624 | /** |
625 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
626 | * @page: page to be put back to appropriate lru list | |
627 | * | |
628 | * Add previously isolated @page to appropriate LRU list. | |
629 | * Page may still be unevictable for other reasons. | |
630 | * | |
631 | * lru_lock must not be held, interrupts must be enabled. | |
632 | */ | |
894bc310 LS |
633 | void putback_lru_page(struct page *page) |
634 | { | |
0ec3b74c | 635 | bool is_unevictable; |
bbfd28ee | 636 | int was_unevictable = PageUnevictable(page); |
894bc310 | 637 | |
309381fe | 638 | VM_BUG_ON_PAGE(PageLRU(page), page); |
894bc310 LS |
639 | |
640 | redo: | |
641 | ClearPageUnevictable(page); | |
642 | ||
39b5f29a | 643 | if (page_evictable(page)) { |
894bc310 LS |
644 | /* |
645 | * For evictable pages, we can use the cache. | |
646 | * In event of a race, worst case is we end up with an | |
647 | * unevictable page on [in]active list. | |
648 | * We know how to handle that. | |
649 | */ | |
0ec3b74c | 650 | is_unevictable = false; |
c53954a0 | 651 | lru_cache_add(page); |
894bc310 LS |
652 | } else { |
653 | /* | |
654 | * Put unevictable pages directly on zone's unevictable | |
655 | * list. | |
656 | */ | |
0ec3b74c | 657 | is_unevictable = true; |
894bc310 | 658 | add_page_to_unevictable_list(page); |
6a7b9548 | 659 | /* |
21ee9f39 MK |
660 | * When racing with an mlock or AS_UNEVICTABLE clearing |
661 | * (page is unlocked) make sure that if the other thread | |
662 | * does not observe our setting of PG_lru and fails | |
24513264 | 663 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 664 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
665 | * the page back to the evictable list. |
666 | * | |
21ee9f39 | 667 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
668 | */ |
669 | smp_mb(); | |
894bc310 | 670 | } |
894bc310 LS |
671 | |
672 | /* | |
673 | * page's status can change while we move it among lru. If an evictable | |
674 | * page is on unevictable list, it never be freed. To avoid that, | |
675 | * check after we added it to the list, again. | |
676 | */ | |
0ec3b74c | 677 | if (is_unevictable && page_evictable(page)) { |
894bc310 LS |
678 | if (!isolate_lru_page(page)) { |
679 | put_page(page); | |
680 | goto redo; | |
681 | } | |
682 | /* This means someone else dropped this page from LRU | |
683 | * So, it will be freed or putback to LRU again. There is | |
684 | * nothing to do here. | |
685 | */ | |
686 | } | |
687 | ||
0ec3b74c | 688 | if (was_unevictable && !is_unevictable) |
bbfd28ee | 689 | count_vm_event(UNEVICTABLE_PGRESCUED); |
0ec3b74c | 690 | else if (!was_unevictable && is_unevictable) |
bbfd28ee LS |
691 | count_vm_event(UNEVICTABLE_PGCULLED); |
692 | ||
894bc310 LS |
693 | put_page(page); /* drop ref from isolate */ |
694 | } | |
695 | ||
dfc8d636 JW |
696 | enum page_references { |
697 | PAGEREF_RECLAIM, | |
698 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 699 | PAGEREF_KEEP, |
dfc8d636 JW |
700 | PAGEREF_ACTIVATE, |
701 | }; | |
702 | ||
703 | static enum page_references page_check_references(struct page *page, | |
704 | struct scan_control *sc) | |
705 | { | |
64574746 | 706 | int referenced_ptes, referenced_page; |
dfc8d636 | 707 | unsigned long vm_flags; |
dfc8d636 | 708 | |
c3ac9a8a JW |
709 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
710 | &vm_flags); | |
64574746 | 711 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 712 | |
dfc8d636 JW |
713 | /* |
714 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
715 | * move the page to the unevictable list. | |
716 | */ | |
717 | if (vm_flags & VM_LOCKED) | |
718 | return PAGEREF_RECLAIM; | |
719 | ||
64574746 | 720 | if (referenced_ptes) { |
e4898273 | 721 | if (PageSwapBacked(page)) |
64574746 JW |
722 | return PAGEREF_ACTIVATE; |
723 | /* | |
724 | * All mapped pages start out with page table | |
725 | * references from the instantiating fault, so we need | |
726 | * to look twice if a mapped file page is used more | |
727 | * than once. | |
728 | * | |
729 | * Mark it and spare it for another trip around the | |
730 | * inactive list. Another page table reference will | |
731 | * lead to its activation. | |
732 | * | |
733 | * Note: the mark is set for activated pages as well | |
734 | * so that recently deactivated but used pages are | |
735 | * quickly recovered. | |
736 | */ | |
737 | SetPageReferenced(page); | |
738 | ||
34dbc67a | 739 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
740 | return PAGEREF_ACTIVATE; |
741 | ||
c909e993 KK |
742 | /* |
743 | * Activate file-backed executable pages after first usage. | |
744 | */ | |
745 | if (vm_flags & VM_EXEC) | |
746 | return PAGEREF_ACTIVATE; | |
747 | ||
64574746 JW |
748 | return PAGEREF_KEEP; |
749 | } | |
dfc8d636 JW |
750 | |
751 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 752 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
753 | return PAGEREF_RECLAIM_CLEAN; |
754 | ||
755 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
756 | } |
757 | ||
e2be15f6 MG |
758 | /* Check if a page is dirty or under writeback */ |
759 | static void page_check_dirty_writeback(struct page *page, | |
760 | bool *dirty, bool *writeback) | |
761 | { | |
b4597226 MG |
762 | struct address_space *mapping; |
763 | ||
e2be15f6 MG |
764 | /* |
765 | * Anonymous pages are not handled by flushers and must be written | |
766 | * from reclaim context. Do not stall reclaim based on them | |
767 | */ | |
768 | if (!page_is_file_cache(page)) { | |
769 | *dirty = false; | |
770 | *writeback = false; | |
771 | return; | |
772 | } | |
773 | ||
774 | /* By default assume that the page flags are accurate */ | |
775 | *dirty = PageDirty(page); | |
776 | *writeback = PageWriteback(page); | |
b4597226 MG |
777 | |
778 | /* Verify dirty/writeback state if the filesystem supports it */ | |
779 | if (!page_has_private(page)) | |
780 | return; | |
781 | ||
782 | mapping = page_mapping(page); | |
783 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
784 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
785 | } |
786 | ||
1da177e4 | 787 | /* |
1742f19f | 788 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 789 | */ |
1742f19f | 790 | static unsigned long shrink_page_list(struct list_head *page_list, |
6a18adb3 | 791 | struct zone *zone, |
f84f6e2b | 792 | struct scan_control *sc, |
02c6de8d | 793 | enum ttu_flags ttu_flags, |
8e950282 | 794 | unsigned long *ret_nr_dirty, |
d43006d5 | 795 | unsigned long *ret_nr_unqueued_dirty, |
8e950282 | 796 | unsigned long *ret_nr_congested, |
02c6de8d | 797 | unsigned long *ret_nr_writeback, |
b1a6f21e | 798 | unsigned long *ret_nr_immediate, |
02c6de8d | 799 | bool force_reclaim) |
1da177e4 LT |
800 | { |
801 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 802 | LIST_HEAD(free_pages); |
1da177e4 | 803 | int pgactivate = 0; |
d43006d5 | 804 | unsigned long nr_unqueued_dirty = 0; |
0e093d99 MG |
805 | unsigned long nr_dirty = 0; |
806 | unsigned long nr_congested = 0; | |
05ff5137 | 807 | unsigned long nr_reclaimed = 0; |
92df3a72 | 808 | unsigned long nr_writeback = 0; |
b1a6f21e | 809 | unsigned long nr_immediate = 0; |
1da177e4 LT |
810 | |
811 | cond_resched(); | |
812 | ||
69980e31 | 813 | mem_cgroup_uncharge_start(); |
1da177e4 LT |
814 | while (!list_empty(page_list)) { |
815 | struct address_space *mapping; | |
816 | struct page *page; | |
817 | int may_enter_fs; | |
02c6de8d | 818 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
e2be15f6 | 819 | bool dirty, writeback; |
1da177e4 LT |
820 | |
821 | cond_resched(); | |
822 | ||
823 | page = lru_to_page(page_list); | |
824 | list_del(&page->lru); | |
825 | ||
529ae9aa | 826 | if (!trylock_page(page)) |
1da177e4 LT |
827 | goto keep; |
828 | ||
309381fe SL |
829 | VM_BUG_ON_PAGE(PageActive(page), page); |
830 | VM_BUG_ON_PAGE(page_zone(page) != zone, page); | |
1da177e4 LT |
831 | |
832 | sc->nr_scanned++; | |
80e43426 | 833 | |
39b5f29a | 834 | if (unlikely(!page_evictable(page))) |
b291f000 | 835 | goto cull_mlocked; |
894bc310 | 836 | |
a6dc60f8 | 837 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
838 | goto keep_locked; |
839 | ||
1da177e4 LT |
840 | /* Double the slab pressure for mapped and swapcache pages */ |
841 | if (page_mapped(page) || PageSwapCache(page)) | |
842 | sc->nr_scanned++; | |
843 | ||
c661b078 AW |
844 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
845 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
846 | ||
e2be15f6 MG |
847 | /* |
848 | * The number of dirty pages determines if a zone is marked | |
849 | * reclaim_congested which affects wait_iff_congested. kswapd | |
850 | * will stall and start writing pages if the tail of the LRU | |
851 | * is all dirty unqueued pages. | |
852 | */ | |
853 | page_check_dirty_writeback(page, &dirty, &writeback); | |
854 | if (dirty || writeback) | |
855 | nr_dirty++; | |
856 | ||
857 | if (dirty && !writeback) | |
858 | nr_unqueued_dirty++; | |
859 | ||
d04e8acd MG |
860 | /* |
861 | * Treat this page as congested if the underlying BDI is or if | |
862 | * pages are cycling through the LRU so quickly that the | |
863 | * pages marked for immediate reclaim are making it to the | |
864 | * end of the LRU a second time. | |
865 | */ | |
e2be15f6 | 866 | mapping = page_mapping(page); |
d04e8acd MG |
867 | if ((mapping && bdi_write_congested(mapping->backing_dev_info)) || |
868 | (writeback && PageReclaim(page))) | |
e2be15f6 MG |
869 | nr_congested++; |
870 | ||
283aba9f MG |
871 | /* |
872 | * If a page at the tail of the LRU is under writeback, there | |
873 | * are three cases to consider. | |
874 | * | |
875 | * 1) If reclaim is encountering an excessive number of pages | |
876 | * under writeback and this page is both under writeback and | |
877 | * PageReclaim then it indicates that pages are being queued | |
878 | * for IO but are being recycled through the LRU before the | |
879 | * IO can complete. Waiting on the page itself risks an | |
880 | * indefinite stall if it is impossible to writeback the | |
881 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
882 | * note that the LRU is being scanned too quickly and the |
883 | * caller can stall after page list has been processed. | |
283aba9f MG |
884 | * |
885 | * 2) Global reclaim encounters a page, memcg encounters a | |
886 | * page that is not marked for immediate reclaim or | |
887 | * the caller does not have __GFP_IO. In this case mark | |
888 | * the page for immediate reclaim and continue scanning. | |
889 | * | |
890 | * __GFP_IO is checked because a loop driver thread might | |
891 | * enter reclaim, and deadlock if it waits on a page for | |
892 | * which it is needed to do the write (loop masks off | |
893 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
894 | * would probably show more reasons. | |
895 | * | |
896 | * Don't require __GFP_FS, since we're not going into the | |
897 | * FS, just waiting on its writeback completion. Worryingly, | |
898 | * ext4 gfs2 and xfs allocate pages with | |
899 | * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing | |
900 | * may_enter_fs here is liable to OOM on them. | |
901 | * | |
902 | * 3) memcg encounters a page that is not already marked | |
903 | * PageReclaim. memcg does not have any dirty pages | |
904 | * throttling so we could easily OOM just because too many | |
905 | * pages are in writeback and there is nothing else to | |
906 | * reclaim. Wait for the writeback to complete. | |
907 | */ | |
c661b078 | 908 | if (PageWriteback(page)) { |
283aba9f MG |
909 | /* Case 1 above */ |
910 | if (current_is_kswapd() && | |
911 | PageReclaim(page) && | |
912 | zone_is_reclaim_writeback(zone)) { | |
b1a6f21e MG |
913 | nr_immediate++; |
914 | goto keep_locked; | |
283aba9f MG |
915 | |
916 | /* Case 2 above */ | |
917 | } else if (global_reclaim(sc) || | |
c3b94f44 HD |
918 | !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) { |
919 | /* | |
920 | * This is slightly racy - end_page_writeback() | |
921 | * might have just cleared PageReclaim, then | |
922 | * setting PageReclaim here end up interpreted | |
923 | * as PageReadahead - but that does not matter | |
924 | * enough to care. What we do want is for this | |
925 | * page to have PageReclaim set next time memcg | |
926 | * reclaim reaches the tests above, so it will | |
927 | * then wait_on_page_writeback() to avoid OOM; | |
928 | * and it's also appropriate in global reclaim. | |
929 | */ | |
930 | SetPageReclaim(page); | |
e62e384e | 931 | nr_writeback++; |
283aba9f | 932 | |
c3b94f44 | 933 | goto keep_locked; |
283aba9f MG |
934 | |
935 | /* Case 3 above */ | |
936 | } else { | |
937 | wait_on_page_writeback(page); | |
e62e384e | 938 | } |
c661b078 | 939 | } |
1da177e4 | 940 | |
02c6de8d MK |
941 | if (!force_reclaim) |
942 | references = page_check_references(page, sc); | |
943 | ||
dfc8d636 JW |
944 | switch (references) { |
945 | case PAGEREF_ACTIVATE: | |
1da177e4 | 946 | goto activate_locked; |
64574746 JW |
947 | case PAGEREF_KEEP: |
948 | goto keep_locked; | |
dfc8d636 JW |
949 | case PAGEREF_RECLAIM: |
950 | case PAGEREF_RECLAIM_CLEAN: | |
951 | ; /* try to reclaim the page below */ | |
952 | } | |
1da177e4 | 953 | |
1da177e4 LT |
954 | /* |
955 | * Anonymous process memory has backing store? | |
956 | * Try to allocate it some swap space here. | |
957 | */ | |
b291f000 | 958 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
959 | if (!(sc->gfp_mask & __GFP_IO)) |
960 | goto keep_locked; | |
5bc7b8ac | 961 | if (!add_to_swap(page, page_list)) |
1da177e4 | 962 | goto activate_locked; |
63eb6b93 | 963 | may_enter_fs = 1; |
1da177e4 | 964 | |
e2be15f6 MG |
965 | /* Adding to swap updated mapping */ |
966 | mapping = page_mapping(page); | |
967 | } | |
1da177e4 LT |
968 | |
969 | /* | |
970 | * The page is mapped into the page tables of one or more | |
971 | * processes. Try to unmap it here. | |
972 | */ | |
973 | if (page_mapped(page) && mapping) { | |
02c6de8d | 974 | switch (try_to_unmap(page, ttu_flags)) { |
1da177e4 LT |
975 | case SWAP_FAIL: |
976 | goto activate_locked; | |
977 | case SWAP_AGAIN: | |
978 | goto keep_locked; | |
b291f000 NP |
979 | case SWAP_MLOCK: |
980 | goto cull_mlocked; | |
1da177e4 LT |
981 | case SWAP_SUCCESS: |
982 | ; /* try to free the page below */ | |
983 | } | |
984 | } | |
985 | ||
986 | if (PageDirty(page)) { | |
ee72886d MG |
987 | /* |
988 | * Only kswapd can writeback filesystem pages to | |
d43006d5 MG |
989 | * avoid risk of stack overflow but only writeback |
990 | * if many dirty pages have been encountered. | |
ee72886d | 991 | */ |
f84f6e2b | 992 | if (page_is_file_cache(page) && |
9e3b2f8c | 993 | (!current_is_kswapd() || |
d43006d5 | 994 | !zone_is_reclaim_dirty(zone))) { |
49ea7eb6 MG |
995 | /* |
996 | * Immediately reclaim when written back. | |
997 | * Similar in principal to deactivate_page() | |
998 | * except we already have the page isolated | |
999 | * and know it's dirty | |
1000 | */ | |
1001 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
1002 | SetPageReclaim(page); | |
1003 | ||
ee72886d MG |
1004 | goto keep_locked; |
1005 | } | |
1006 | ||
dfc8d636 | 1007 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1008 | goto keep_locked; |
4dd4b920 | 1009 | if (!may_enter_fs) |
1da177e4 | 1010 | goto keep_locked; |
52a8363e | 1011 | if (!sc->may_writepage) |
1da177e4 LT |
1012 | goto keep_locked; |
1013 | ||
1014 | /* Page is dirty, try to write it out here */ | |
7d3579e8 | 1015 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1016 | case PAGE_KEEP: |
1017 | goto keep_locked; | |
1018 | case PAGE_ACTIVATE: | |
1019 | goto activate_locked; | |
1020 | case PAGE_SUCCESS: | |
7d3579e8 | 1021 | if (PageWriteback(page)) |
41ac1999 | 1022 | goto keep; |
7d3579e8 | 1023 | if (PageDirty(page)) |
1da177e4 | 1024 | goto keep; |
7d3579e8 | 1025 | |
1da177e4 LT |
1026 | /* |
1027 | * A synchronous write - probably a ramdisk. Go | |
1028 | * ahead and try to reclaim the page. | |
1029 | */ | |
529ae9aa | 1030 | if (!trylock_page(page)) |
1da177e4 LT |
1031 | goto keep; |
1032 | if (PageDirty(page) || PageWriteback(page)) | |
1033 | goto keep_locked; | |
1034 | mapping = page_mapping(page); | |
1035 | case PAGE_CLEAN: | |
1036 | ; /* try to free the page below */ | |
1037 | } | |
1038 | } | |
1039 | ||
1040 | /* | |
1041 | * If the page has buffers, try to free the buffer mappings | |
1042 | * associated with this page. If we succeed we try to free | |
1043 | * the page as well. | |
1044 | * | |
1045 | * We do this even if the page is PageDirty(). | |
1046 | * try_to_release_page() does not perform I/O, but it is | |
1047 | * possible for a page to have PageDirty set, but it is actually | |
1048 | * clean (all its buffers are clean). This happens if the | |
1049 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1050 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1051 | * try_to_release_page() will discover that cleanness and will |
1052 | * drop the buffers and mark the page clean - it can be freed. | |
1053 | * | |
1054 | * Rarely, pages can have buffers and no ->mapping. These are | |
1055 | * the pages which were not successfully invalidated in | |
1056 | * truncate_complete_page(). We try to drop those buffers here | |
1057 | * and if that worked, and the page is no longer mapped into | |
1058 | * process address space (page_count == 1) it can be freed. | |
1059 | * Otherwise, leave the page on the LRU so it is swappable. | |
1060 | */ | |
266cf658 | 1061 | if (page_has_private(page)) { |
1da177e4 LT |
1062 | if (!try_to_release_page(page, sc->gfp_mask)) |
1063 | goto activate_locked; | |
e286781d NP |
1064 | if (!mapping && page_count(page) == 1) { |
1065 | unlock_page(page); | |
1066 | if (put_page_testzero(page)) | |
1067 | goto free_it; | |
1068 | else { | |
1069 | /* | |
1070 | * rare race with speculative reference. | |
1071 | * the speculative reference will free | |
1072 | * this page shortly, so we may | |
1073 | * increment nr_reclaimed here (and | |
1074 | * leave it off the LRU). | |
1075 | */ | |
1076 | nr_reclaimed++; | |
1077 | continue; | |
1078 | } | |
1079 | } | |
1da177e4 LT |
1080 | } |
1081 | ||
a528910e | 1082 | if (!mapping || !__remove_mapping(mapping, page, true)) |
49d2e9cc | 1083 | goto keep_locked; |
1da177e4 | 1084 | |
a978d6f5 NP |
1085 | /* |
1086 | * At this point, we have no other references and there is | |
1087 | * no way to pick any more up (removed from LRU, removed | |
1088 | * from pagecache). Can use non-atomic bitops now (and | |
1089 | * we obviously don't have to worry about waking up a process | |
1090 | * waiting on the page lock, because there are no references. | |
1091 | */ | |
1092 | __clear_page_locked(page); | |
e286781d | 1093 | free_it: |
05ff5137 | 1094 | nr_reclaimed++; |
abe4c3b5 MG |
1095 | |
1096 | /* | |
1097 | * Is there need to periodically free_page_list? It would | |
1098 | * appear not as the counts should be low | |
1099 | */ | |
1100 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1101 | continue; |
1102 | ||
b291f000 | 1103 | cull_mlocked: |
63d6c5ad HD |
1104 | if (PageSwapCache(page)) |
1105 | try_to_free_swap(page); | |
b291f000 NP |
1106 | unlock_page(page); |
1107 | putback_lru_page(page); | |
1108 | continue; | |
1109 | ||
1da177e4 | 1110 | activate_locked: |
68a22394 RR |
1111 | /* Not a candidate for swapping, so reclaim swap space. */ |
1112 | if (PageSwapCache(page) && vm_swap_full()) | |
a2c43eed | 1113 | try_to_free_swap(page); |
309381fe | 1114 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
1115 | SetPageActive(page); |
1116 | pgactivate++; | |
1117 | keep_locked: | |
1118 | unlock_page(page); | |
1119 | keep: | |
1120 | list_add(&page->lru, &ret_pages); | |
309381fe | 1121 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1122 | } |
abe4c3b5 | 1123 | |
b745bc85 | 1124 | free_hot_cold_page_list(&free_pages, true); |
abe4c3b5 | 1125 | |
1da177e4 | 1126 | list_splice(&ret_pages, page_list); |
f8891e5e | 1127 | count_vm_events(PGACTIVATE, pgactivate); |
69980e31 | 1128 | mem_cgroup_uncharge_end(); |
8e950282 MG |
1129 | *ret_nr_dirty += nr_dirty; |
1130 | *ret_nr_congested += nr_congested; | |
d43006d5 | 1131 | *ret_nr_unqueued_dirty += nr_unqueued_dirty; |
92df3a72 | 1132 | *ret_nr_writeback += nr_writeback; |
b1a6f21e | 1133 | *ret_nr_immediate += nr_immediate; |
05ff5137 | 1134 | return nr_reclaimed; |
1da177e4 LT |
1135 | } |
1136 | ||
02c6de8d MK |
1137 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1138 | struct list_head *page_list) | |
1139 | { | |
1140 | struct scan_control sc = { | |
1141 | .gfp_mask = GFP_KERNEL, | |
1142 | .priority = DEF_PRIORITY, | |
1143 | .may_unmap = 1, | |
1144 | }; | |
8e950282 | 1145 | unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; |
02c6de8d MK |
1146 | struct page *page, *next; |
1147 | LIST_HEAD(clean_pages); | |
1148 | ||
1149 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e RA |
1150 | if (page_is_file_cache(page) && !PageDirty(page) && |
1151 | !isolated_balloon_page(page)) { | |
02c6de8d MK |
1152 | ClearPageActive(page); |
1153 | list_move(&page->lru, &clean_pages); | |
1154 | } | |
1155 | } | |
1156 | ||
1157 | ret = shrink_page_list(&clean_pages, zone, &sc, | |
8e950282 MG |
1158 | TTU_UNMAP|TTU_IGNORE_ACCESS, |
1159 | &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); | |
02c6de8d | 1160 | list_splice(&clean_pages, page_list); |
83da7510 | 1161 | mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1162 | return ret; |
1163 | } | |
1164 | ||
5ad333eb AW |
1165 | /* |
1166 | * Attempt to remove the specified page from its LRU. Only take this page | |
1167 | * if it is of the appropriate PageActive status. Pages which are being | |
1168 | * freed elsewhere are also ignored. | |
1169 | * | |
1170 | * page: page to consider | |
1171 | * mode: one of the LRU isolation modes defined above | |
1172 | * | |
1173 | * returns 0 on success, -ve errno on failure. | |
1174 | */ | |
f3fd4a61 | 1175 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1176 | { |
1177 | int ret = -EINVAL; | |
1178 | ||
1179 | /* Only take pages on the LRU. */ | |
1180 | if (!PageLRU(page)) | |
1181 | return ret; | |
1182 | ||
e46a2879 MK |
1183 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1184 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1185 | return ret; |
1186 | ||
5ad333eb | 1187 | ret = -EBUSY; |
08e552c6 | 1188 | |
c8244935 MG |
1189 | /* |
1190 | * To minimise LRU disruption, the caller can indicate that it only | |
1191 | * wants to isolate pages it will be able to operate on without | |
1192 | * blocking - clean pages for the most part. | |
1193 | * | |
1194 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1195 | * is used by reclaim when it is cannot write to backing storage | |
1196 | * | |
1197 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1198 | * that it is possible to migrate without blocking | |
1199 | */ | |
1200 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1201 | /* All the caller can do on PageWriteback is block */ | |
1202 | if (PageWriteback(page)) | |
1203 | return ret; | |
1204 | ||
1205 | if (PageDirty(page)) { | |
1206 | struct address_space *mapping; | |
1207 | ||
1208 | /* ISOLATE_CLEAN means only clean pages */ | |
1209 | if (mode & ISOLATE_CLEAN) | |
1210 | return ret; | |
1211 | ||
1212 | /* | |
1213 | * Only pages without mappings or that have a | |
1214 | * ->migratepage callback are possible to migrate | |
1215 | * without blocking | |
1216 | */ | |
1217 | mapping = page_mapping(page); | |
1218 | if (mapping && !mapping->a_ops->migratepage) | |
1219 | return ret; | |
1220 | } | |
1221 | } | |
39deaf85 | 1222 | |
f80c0673 MK |
1223 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1224 | return ret; | |
1225 | ||
5ad333eb AW |
1226 | if (likely(get_page_unless_zero(page))) { |
1227 | /* | |
1228 | * Be careful not to clear PageLRU until after we're | |
1229 | * sure the page is not being freed elsewhere -- the | |
1230 | * page release code relies on it. | |
1231 | */ | |
1232 | ClearPageLRU(page); | |
1233 | ret = 0; | |
1234 | } | |
1235 | ||
1236 | return ret; | |
1237 | } | |
1238 | ||
1da177e4 LT |
1239 | /* |
1240 | * zone->lru_lock is heavily contended. Some of the functions that | |
1241 | * shrink the lists perform better by taking out a batch of pages | |
1242 | * and working on them outside the LRU lock. | |
1243 | * | |
1244 | * For pagecache intensive workloads, this function is the hottest | |
1245 | * spot in the kernel (apart from copy_*_user functions). | |
1246 | * | |
1247 | * Appropriate locks must be held before calling this function. | |
1248 | * | |
1249 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1250 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1251 | * @dst: The temp list to put pages on to. |
f626012d | 1252 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1253 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1254 | * @mode: One of the LRU isolation modes |
3cb99451 | 1255 | * @lru: LRU list id for isolating |
1da177e4 LT |
1256 | * |
1257 | * returns how many pages were moved onto *@dst. | |
1258 | */ | |
69e05944 | 1259 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1260 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1261 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1262 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1263 | { |
75b00af7 | 1264 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1265 | unsigned long nr_taken = 0; |
c9b02d97 | 1266 | unsigned long scan; |
1da177e4 | 1267 | |
c9b02d97 | 1268 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
5ad333eb | 1269 | struct page *page; |
fa9add64 | 1270 | int nr_pages; |
5ad333eb | 1271 | |
1da177e4 LT |
1272 | page = lru_to_page(src); |
1273 | prefetchw_prev_lru_page(page, src, flags); | |
1274 | ||
309381fe | 1275 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1276 | |
f3fd4a61 | 1277 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1278 | case 0: |
fa9add64 HD |
1279 | nr_pages = hpage_nr_pages(page); |
1280 | mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); | |
5ad333eb | 1281 | list_move(&page->lru, dst); |
fa9add64 | 1282 | nr_taken += nr_pages; |
5ad333eb AW |
1283 | break; |
1284 | ||
1285 | case -EBUSY: | |
1286 | /* else it is being freed elsewhere */ | |
1287 | list_move(&page->lru, src); | |
1288 | continue; | |
46453a6e | 1289 | |
5ad333eb AW |
1290 | default: |
1291 | BUG(); | |
1292 | } | |
1da177e4 LT |
1293 | } |
1294 | ||
f626012d | 1295 | *nr_scanned = scan; |
75b00af7 HD |
1296 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, |
1297 | nr_taken, mode, is_file_lru(lru)); | |
1da177e4 LT |
1298 | return nr_taken; |
1299 | } | |
1300 | ||
62695a84 NP |
1301 | /** |
1302 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1303 | * @page: page to isolate from its LRU list | |
1304 | * | |
1305 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1306 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1307 | * | |
1308 | * Returns 0 if the page was removed from an LRU list. | |
1309 | * Returns -EBUSY if the page was not on an LRU list. | |
1310 | * | |
1311 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1312 | * the active list, it will have PageActive set. If it was found on |
1313 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1314 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1315 | * |
1316 | * The vmstat statistic corresponding to the list on which the page was | |
1317 | * found will be decremented. | |
1318 | * | |
1319 | * Restrictions: | |
1320 | * (1) Must be called with an elevated refcount on the page. This is a | |
1321 | * fundamentnal difference from isolate_lru_pages (which is called | |
1322 | * without a stable reference). | |
1323 | * (2) the lru_lock must not be held. | |
1324 | * (3) interrupts must be enabled. | |
1325 | */ | |
1326 | int isolate_lru_page(struct page *page) | |
1327 | { | |
1328 | int ret = -EBUSY; | |
1329 | ||
309381fe | 1330 | VM_BUG_ON_PAGE(!page_count(page), page); |
0c917313 | 1331 | |
62695a84 NP |
1332 | if (PageLRU(page)) { |
1333 | struct zone *zone = page_zone(page); | |
fa9add64 | 1334 | struct lruvec *lruvec; |
62695a84 NP |
1335 | |
1336 | spin_lock_irq(&zone->lru_lock); | |
fa9add64 | 1337 | lruvec = mem_cgroup_page_lruvec(page, zone); |
0c917313 | 1338 | if (PageLRU(page)) { |
894bc310 | 1339 | int lru = page_lru(page); |
0c917313 | 1340 | get_page(page); |
62695a84 | 1341 | ClearPageLRU(page); |
fa9add64 HD |
1342 | del_page_from_lru_list(page, lruvec, lru); |
1343 | ret = 0; | |
62695a84 NP |
1344 | } |
1345 | spin_unlock_irq(&zone->lru_lock); | |
1346 | } | |
1347 | return ret; | |
1348 | } | |
1349 | ||
35cd7815 | 1350 | /* |
d37dd5dc FW |
1351 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1352 | * then get resheduled. When there are massive number of tasks doing page | |
1353 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1354 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1355 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 RR |
1356 | */ |
1357 | static int too_many_isolated(struct zone *zone, int file, | |
1358 | struct scan_control *sc) | |
1359 | { | |
1360 | unsigned long inactive, isolated; | |
1361 | ||
1362 | if (current_is_kswapd()) | |
1363 | return 0; | |
1364 | ||
89b5fae5 | 1365 | if (!global_reclaim(sc)) |
35cd7815 RR |
1366 | return 0; |
1367 | ||
1368 | if (file) { | |
1369 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1370 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1371 | } else { | |
1372 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1373 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1374 | } | |
1375 | ||
3cf23841 FW |
1376 | /* |
1377 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1378 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1379 | * deadlock. | |
1380 | */ | |
1381 | if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS) | |
1382 | inactive >>= 3; | |
1383 | ||
35cd7815 RR |
1384 | return isolated > inactive; |
1385 | } | |
1386 | ||
66635629 | 1387 | static noinline_for_stack void |
75b00af7 | 1388 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1389 | { |
27ac81d8 KK |
1390 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1391 | struct zone *zone = lruvec_zone(lruvec); | |
3f79768f | 1392 | LIST_HEAD(pages_to_free); |
66635629 | 1393 | |
66635629 MG |
1394 | /* |
1395 | * Put back any unfreeable pages. | |
1396 | */ | |
66635629 | 1397 | while (!list_empty(page_list)) { |
3f79768f | 1398 | struct page *page = lru_to_page(page_list); |
66635629 | 1399 | int lru; |
3f79768f | 1400 | |
309381fe | 1401 | VM_BUG_ON_PAGE(PageLRU(page), page); |
66635629 | 1402 | list_del(&page->lru); |
39b5f29a | 1403 | if (unlikely(!page_evictable(page))) { |
66635629 MG |
1404 | spin_unlock_irq(&zone->lru_lock); |
1405 | putback_lru_page(page); | |
1406 | spin_lock_irq(&zone->lru_lock); | |
1407 | continue; | |
1408 | } | |
fa9add64 HD |
1409 | |
1410 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1411 | ||
7a608572 | 1412 | SetPageLRU(page); |
66635629 | 1413 | lru = page_lru(page); |
fa9add64 HD |
1414 | add_page_to_lru_list(page, lruvec, lru); |
1415 | ||
66635629 MG |
1416 | if (is_active_lru(lru)) { |
1417 | int file = is_file_lru(lru); | |
9992af10 RR |
1418 | int numpages = hpage_nr_pages(page); |
1419 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1420 | } |
2bcf8879 HD |
1421 | if (put_page_testzero(page)) { |
1422 | __ClearPageLRU(page); | |
1423 | __ClearPageActive(page); | |
fa9add64 | 1424 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1425 | |
1426 | if (unlikely(PageCompound(page))) { | |
1427 | spin_unlock_irq(&zone->lru_lock); | |
1428 | (*get_compound_page_dtor(page))(page); | |
1429 | spin_lock_irq(&zone->lru_lock); | |
1430 | } else | |
1431 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1432 | } |
1433 | } | |
66635629 | 1434 | |
3f79768f HD |
1435 | /* |
1436 | * To save our caller's stack, now use input list for pages to free. | |
1437 | */ | |
1438 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1439 | } |
1440 | ||
399ba0b9 N |
1441 | /* |
1442 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1443 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1444 | * In that case we should only throttle if the backing device it is | |
1445 | * writing to is congested. In other cases it is safe to throttle. | |
1446 | */ | |
1447 | static int current_may_throttle(void) | |
1448 | { | |
1449 | return !(current->flags & PF_LESS_THROTTLE) || | |
1450 | current->backing_dev_info == NULL || | |
1451 | bdi_write_congested(current->backing_dev_info); | |
1452 | } | |
1453 | ||
1da177e4 | 1454 | /* |
1742f19f AM |
1455 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1456 | * of reclaimed pages | |
1da177e4 | 1457 | */ |
66635629 | 1458 | static noinline_for_stack unsigned long |
1a93be0e | 1459 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1460 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1461 | { |
1462 | LIST_HEAD(page_list); | |
e247dbce | 1463 | unsigned long nr_scanned; |
05ff5137 | 1464 | unsigned long nr_reclaimed = 0; |
e247dbce | 1465 | unsigned long nr_taken; |
8e950282 MG |
1466 | unsigned long nr_dirty = 0; |
1467 | unsigned long nr_congested = 0; | |
e2be15f6 | 1468 | unsigned long nr_unqueued_dirty = 0; |
92df3a72 | 1469 | unsigned long nr_writeback = 0; |
b1a6f21e | 1470 | unsigned long nr_immediate = 0; |
f3fd4a61 | 1471 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1472 | int file = is_file_lru(lru); |
1a93be0e KK |
1473 | struct zone *zone = lruvec_zone(lruvec); |
1474 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
78dc583d | 1475 | |
35cd7815 | 1476 | while (unlikely(too_many_isolated(zone, file, sc))) { |
58355c78 | 1477 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1478 | |
1479 | /* We are about to die and free our memory. Return now. */ | |
1480 | if (fatal_signal_pending(current)) | |
1481 | return SWAP_CLUSTER_MAX; | |
1482 | } | |
1483 | ||
1da177e4 | 1484 | lru_add_drain(); |
f80c0673 MK |
1485 | |
1486 | if (!sc->may_unmap) | |
61317289 | 1487 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1488 | if (!sc->may_writepage) |
61317289 | 1489 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1490 | |
1da177e4 | 1491 | spin_lock_irq(&zone->lru_lock); |
b35ea17b | 1492 | |
5dc35979 KK |
1493 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1494 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc KK |
1495 | |
1496 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1497 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1498 | ||
89b5fae5 | 1499 | if (global_reclaim(sc)) { |
e247dbce KM |
1500 | zone->pages_scanned += nr_scanned; |
1501 | if (current_is_kswapd()) | |
75b00af7 | 1502 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); |
e247dbce | 1503 | else |
75b00af7 | 1504 | __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); |
e247dbce | 1505 | } |
d563c050 | 1506 | spin_unlock_irq(&zone->lru_lock); |
b35ea17b | 1507 | |
d563c050 | 1508 | if (nr_taken == 0) |
66635629 | 1509 | return 0; |
5ad333eb | 1510 | |
02c6de8d | 1511 | nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, |
8e950282 MG |
1512 | &nr_dirty, &nr_unqueued_dirty, &nr_congested, |
1513 | &nr_writeback, &nr_immediate, | |
1514 | false); | |
c661b078 | 1515 | |
3f79768f HD |
1516 | spin_lock_irq(&zone->lru_lock); |
1517 | ||
95d918fc | 1518 | reclaim_stat->recent_scanned[file] += nr_taken; |
d563c050 | 1519 | |
904249aa YH |
1520 | if (global_reclaim(sc)) { |
1521 | if (current_is_kswapd()) | |
1522 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1523 | nr_reclaimed); | |
1524 | else | |
1525 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1526 | nr_reclaimed); | |
1527 | } | |
a74609fa | 1528 | |
27ac81d8 | 1529 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1530 | |
95d918fc | 1531 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f HD |
1532 | |
1533 | spin_unlock_irq(&zone->lru_lock); | |
1534 | ||
b745bc85 | 1535 | free_hot_cold_page_list(&page_list, true); |
e11da5b4 | 1536 | |
92df3a72 MG |
1537 | /* |
1538 | * If reclaim is isolating dirty pages under writeback, it implies | |
1539 | * that the long-lived page allocation rate is exceeding the page | |
1540 | * laundering rate. Either the global limits are not being effective | |
1541 | * at throttling processes due to the page distribution throughout | |
1542 | * zones or there is heavy usage of a slow backing device. The | |
1543 | * only option is to throttle from reclaim context which is not ideal | |
1544 | * as there is no guarantee the dirtying process is throttled in the | |
1545 | * same way balance_dirty_pages() manages. | |
1546 | * | |
8e950282 MG |
1547 | * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number |
1548 | * of pages under pages flagged for immediate reclaim and stall if any | |
1549 | * are encountered in the nr_immediate check below. | |
92df3a72 | 1550 | */ |
918fc718 | 1551 | if (nr_writeback && nr_writeback == nr_taken) |
283aba9f | 1552 | zone_set_flag(zone, ZONE_WRITEBACK); |
92df3a72 | 1553 | |
d43006d5 | 1554 | /* |
b1a6f21e MG |
1555 | * memcg will stall in page writeback so only consider forcibly |
1556 | * stalling for global reclaim | |
d43006d5 | 1557 | */ |
b1a6f21e | 1558 | if (global_reclaim(sc)) { |
8e950282 MG |
1559 | /* |
1560 | * Tag a zone as congested if all the dirty pages scanned were | |
1561 | * backed by a congested BDI and wait_iff_congested will stall. | |
1562 | */ | |
1563 | if (nr_dirty && nr_dirty == nr_congested) | |
1564 | zone_set_flag(zone, ZONE_CONGESTED); | |
1565 | ||
b1a6f21e MG |
1566 | /* |
1567 | * If dirty pages are scanned that are not queued for IO, it | |
1568 | * implies that flushers are not keeping up. In this case, flag | |
1569 | * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing | |
1570 | * pages from reclaim context. It will forcibly stall in the | |
1571 | * next check. | |
1572 | */ | |
1573 | if (nr_unqueued_dirty == nr_taken) | |
1574 | zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY); | |
1575 | ||
1576 | /* | |
1577 | * In addition, if kswapd scans pages marked marked for | |
1578 | * immediate reclaim and under writeback (nr_immediate), it | |
1579 | * implies that pages are cycling through the LRU faster than | |
1580 | * they are written so also forcibly stall. | |
1581 | */ | |
399ba0b9 N |
1582 | if ((nr_unqueued_dirty == nr_taken || nr_immediate) && |
1583 | current_may_throttle()) | |
b1a6f21e | 1584 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
e2be15f6 | 1585 | } |
d43006d5 | 1586 | |
8e950282 MG |
1587 | /* |
1588 | * Stall direct reclaim for IO completions if underlying BDIs or zone | |
1589 | * is congested. Allow kswapd to continue until it starts encountering | |
1590 | * unqueued dirty pages or cycling through the LRU too quickly. | |
1591 | */ | |
399ba0b9 N |
1592 | if (!sc->hibernation_mode && !current_is_kswapd() && |
1593 | current_may_throttle()) | |
8e950282 MG |
1594 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); |
1595 | ||
e11da5b4 MG |
1596 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, |
1597 | zone_idx(zone), | |
1598 | nr_scanned, nr_reclaimed, | |
9e3b2f8c | 1599 | sc->priority, |
23b9da55 | 1600 | trace_shrink_flags(file)); |
05ff5137 | 1601 | return nr_reclaimed; |
1da177e4 LT |
1602 | } |
1603 | ||
1604 | /* | |
1605 | * This moves pages from the active list to the inactive list. | |
1606 | * | |
1607 | * We move them the other way if the page is referenced by one or more | |
1608 | * processes, from rmap. | |
1609 | * | |
1610 | * If the pages are mostly unmapped, the processing is fast and it is | |
1611 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1612 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1613 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1614 | * this, so instead we remove the pages from the LRU while processing them. | |
1615 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1616 | * nobody will play with that bit on a non-LRU page. | |
1617 | * | |
1618 | * The downside is that we have to touch page->_count against each page. | |
1619 | * But we had to alter page->flags anyway. | |
1620 | */ | |
1cfb419b | 1621 | |
fa9add64 | 1622 | static void move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1623 | struct list_head *list, |
2bcf8879 | 1624 | struct list_head *pages_to_free, |
3eb4140f WF |
1625 | enum lru_list lru) |
1626 | { | |
fa9add64 | 1627 | struct zone *zone = lruvec_zone(lruvec); |
3eb4140f | 1628 | unsigned long pgmoved = 0; |
3eb4140f | 1629 | struct page *page; |
fa9add64 | 1630 | int nr_pages; |
3eb4140f | 1631 | |
3eb4140f WF |
1632 | while (!list_empty(list)) { |
1633 | page = lru_to_page(list); | |
fa9add64 | 1634 | lruvec = mem_cgroup_page_lruvec(page, zone); |
3eb4140f | 1635 | |
309381fe | 1636 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3eb4140f WF |
1637 | SetPageLRU(page); |
1638 | ||
fa9add64 HD |
1639 | nr_pages = hpage_nr_pages(page); |
1640 | mem_cgroup_update_lru_size(lruvec, lru, nr_pages); | |
925b7673 | 1641 | list_move(&page->lru, &lruvec->lists[lru]); |
fa9add64 | 1642 | pgmoved += nr_pages; |
3eb4140f | 1643 | |
2bcf8879 HD |
1644 | if (put_page_testzero(page)) { |
1645 | __ClearPageLRU(page); | |
1646 | __ClearPageActive(page); | |
fa9add64 | 1647 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1648 | |
1649 | if (unlikely(PageCompound(page))) { | |
1650 | spin_unlock_irq(&zone->lru_lock); | |
1651 | (*get_compound_page_dtor(page))(page); | |
1652 | spin_lock_irq(&zone->lru_lock); | |
1653 | } else | |
1654 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1655 | } |
1656 | } | |
1657 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1658 | if (!is_active_lru(lru)) | |
1659 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1660 | } | |
1cfb419b | 1661 | |
f626012d | 1662 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1663 | struct lruvec *lruvec, |
f16015fb | 1664 | struct scan_control *sc, |
9e3b2f8c | 1665 | enum lru_list lru) |
1da177e4 | 1666 | { |
44c241f1 | 1667 | unsigned long nr_taken; |
f626012d | 1668 | unsigned long nr_scanned; |
6fe6b7e3 | 1669 | unsigned long vm_flags; |
1da177e4 | 1670 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1671 | LIST_HEAD(l_active); |
b69408e8 | 1672 | LIST_HEAD(l_inactive); |
1da177e4 | 1673 | struct page *page; |
1a93be0e | 1674 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
44c241f1 | 1675 | unsigned long nr_rotated = 0; |
f3fd4a61 | 1676 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1677 | int file = is_file_lru(lru); |
1a93be0e | 1678 | struct zone *zone = lruvec_zone(lruvec); |
1da177e4 LT |
1679 | |
1680 | lru_add_drain(); | |
f80c0673 MK |
1681 | |
1682 | if (!sc->may_unmap) | |
61317289 | 1683 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1684 | if (!sc->may_writepage) |
61317289 | 1685 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1686 | |
1da177e4 | 1687 | spin_lock_irq(&zone->lru_lock); |
925b7673 | 1688 | |
5dc35979 KK |
1689 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1690 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1691 | if (global_reclaim(sc)) |
f626012d | 1692 | zone->pages_scanned += nr_scanned; |
89b5fae5 | 1693 | |
b7c46d15 | 1694 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1695 | |
f626012d | 1696 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); |
3cb99451 | 1697 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); |
a731286d | 1698 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
1da177e4 LT |
1699 | spin_unlock_irq(&zone->lru_lock); |
1700 | ||
1da177e4 LT |
1701 | while (!list_empty(&l_hold)) { |
1702 | cond_resched(); | |
1703 | page = lru_to_page(&l_hold); | |
1704 | list_del(&page->lru); | |
7e9cd484 | 1705 | |
39b5f29a | 1706 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
1707 | putback_lru_page(page); |
1708 | continue; | |
1709 | } | |
1710 | ||
cc715d99 MG |
1711 | if (unlikely(buffer_heads_over_limit)) { |
1712 | if (page_has_private(page) && trylock_page(page)) { | |
1713 | if (page_has_private(page)) | |
1714 | try_to_release_page(page, 0); | |
1715 | unlock_page(page); | |
1716 | } | |
1717 | } | |
1718 | ||
c3ac9a8a JW |
1719 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1720 | &vm_flags)) { | |
9992af10 | 1721 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1722 | /* |
1723 | * Identify referenced, file-backed active pages and | |
1724 | * give them one more trip around the active list. So | |
1725 | * that executable code get better chances to stay in | |
1726 | * memory under moderate memory pressure. Anon pages | |
1727 | * are not likely to be evicted by use-once streaming | |
1728 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1729 | * so we ignore them here. | |
1730 | */ | |
41e20983 | 1731 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1732 | list_add(&page->lru, &l_active); |
1733 | continue; | |
1734 | } | |
1735 | } | |
7e9cd484 | 1736 | |
5205e56e | 1737 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1738 | list_add(&page->lru, &l_inactive); |
1739 | } | |
1740 | ||
b555749a | 1741 | /* |
8cab4754 | 1742 | * Move pages back to the lru list. |
b555749a | 1743 | */ |
2a1dc509 | 1744 | spin_lock_irq(&zone->lru_lock); |
556adecb | 1745 | /* |
8cab4754 WF |
1746 | * Count referenced pages from currently used mappings as rotated, |
1747 | * even though only some of them are actually re-activated. This | |
1748 | * helps balance scan pressure between file and anonymous pages in | |
1749 | * get_scan_ratio. | |
7e9cd484 | 1750 | */ |
b7c46d15 | 1751 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1752 | |
fa9add64 HD |
1753 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
1754 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
a731286d | 1755 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
f8891e5e | 1756 | spin_unlock_irq(&zone->lru_lock); |
2bcf8879 | 1757 | |
b745bc85 | 1758 | free_hot_cold_page_list(&l_hold, true); |
1da177e4 LT |
1759 | } |
1760 | ||
74e3f3c3 | 1761 | #ifdef CONFIG_SWAP |
14797e23 | 1762 | static int inactive_anon_is_low_global(struct zone *zone) |
f89eb90e KM |
1763 | { |
1764 | unsigned long active, inactive; | |
1765 | ||
1766 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1767 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1768 | ||
1769 | if (inactive * zone->inactive_ratio < active) | |
1770 | return 1; | |
1771 | ||
1772 | return 0; | |
1773 | } | |
1774 | ||
14797e23 KM |
1775 | /** |
1776 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
c56d5c7d | 1777 | * @lruvec: LRU vector to check |
14797e23 KM |
1778 | * |
1779 | * Returns true if the zone does not have enough inactive anon pages, | |
1780 | * meaning some active anon pages need to be deactivated. | |
1781 | */ | |
c56d5c7d | 1782 | static int inactive_anon_is_low(struct lruvec *lruvec) |
14797e23 | 1783 | { |
74e3f3c3 MK |
1784 | /* |
1785 | * If we don't have swap space, anonymous page deactivation | |
1786 | * is pointless. | |
1787 | */ | |
1788 | if (!total_swap_pages) | |
1789 | return 0; | |
1790 | ||
c3c787e8 | 1791 | if (!mem_cgroup_disabled()) |
c56d5c7d | 1792 | return mem_cgroup_inactive_anon_is_low(lruvec); |
f16015fb | 1793 | |
c56d5c7d | 1794 | return inactive_anon_is_low_global(lruvec_zone(lruvec)); |
14797e23 | 1795 | } |
74e3f3c3 | 1796 | #else |
c56d5c7d | 1797 | static inline int inactive_anon_is_low(struct lruvec *lruvec) |
74e3f3c3 MK |
1798 | { |
1799 | return 0; | |
1800 | } | |
1801 | #endif | |
14797e23 | 1802 | |
56e49d21 RR |
1803 | /** |
1804 | * inactive_file_is_low - check if file pages need to be deactivated | |
c56d5c7d | 1805 | * @lruvec: LRU vector to check |
56e49d21 RR |
1806 | * |
1807 | * When the system is doing streaming IO, memory pressure here | |
1808 | * ensures that active file pages get deactivated, until more | |
1809 | * than half of the file pages are on the inactive list. | |
1810 | * | |
1811 | * Once we get to that situation, protect the system's working | |
1812 | * set from being evicted by disabling active file page aging. | |
1813 | * | |
1814 | * This uses a different ratio than the anonymous pages, because | |
1815 | * the page cache uses a use-once replacement algorithm. | |
1816 | */ | |
c56d5c7d | 1817 | static int inactive_file_is_low(struct lruvec *lruvec) |
56e49d21 | 1818 | { |
e3790144 JW |
1819 | unsigned long inactive; |
1820 | unsigned long active; | |
1821 | ||
1822 | inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
1823 | active = get_lru_size(lruvec, LRU_ACTIVE_FILE); | |
56e49d21 | 1824 | |
e3790144 | 1825 | return active > inactive; |
56e49d21 RR |
1826 | } |
1827 | ||
75b00af7 | 1828 | static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) |
b39415b2 | 1829 | { |
75b00af7 | 1830 | if (is_file_lru(lru)) |
c56d5c7d | 1831 | return inactive_file_is_low(lruvec); |
b39415b2 | 1832 | else |
c56d5c7d | 1833 | return inactive_anon_is_low(lruvec); |
b39415b2 RR |
1834 | } |
1835 | ||
4f98a2fe | 1836 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
1a93be0e | 1837 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 1838 | { |
b39415b2 | 1839 | if (is_active_lru(lru)) { |
75b00af7 | 1840 | if (inactive_list_is_low(lruvec, lru)) |
1a93be0e | 1841 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
1842 | return 0; |
1843 | } | |
1844 | ||
1a93be0e | 1845 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
1846 | } |
1847 | ||
3d58ab5c | 1848 | static int vmscan_swappiness(struct scan_control *sc) |
1f4c025b | 1849 | { |
89b5fae5 | 1850 | if (global_reclaim(sc)) |
1f4c025b | 1851 | return vm_swappiness; |
3d58ab5c | 1852 | return mem_cgroup_swappiness(sc->target_mem_cgroup); |
1f4c025b KH |
1853 | } |
1854 | ||
9a265114 JW |
1855 | enum scan_balance { |
1856 | SCAN_EQUAL, | |
1857 | SCAN_FRACT, | |
1858 | SCAN_ANON, | |
1859 | SCAN_FILE, | |
1860 | }; | |
1861 | ||
4f98a2fe RR |
1862 | /* |
1863 | * Determine how aggressively the anon and file LRU lists should be | |
1864 | * scanned. The relative value of each set of LRU lists is determined | |
1865 | * by looking at the fraction of the pages scanned we did rotate back | |
1866 | * onto the active list instead of evict. | |
1867 | * | |
be7bd59d WL |
1868 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
1869 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 1870 | */ |
90126375 | 1871 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
9e3b2f8c | 1872 | unsigned long *nr) |
4f98a2fe | 1873 | { |
9a265114 JW |
1874 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1875 | u64 fraction[2]; | |
1876 | u64 denominator = 0; /* gcc */ | |
1877 | struct zone *zone = lruvec_zone(lruvec); | |
4f98a2fe | 1878 | unsigned long anon_prio, file_prio; |
9a265114 | 1879 | enum scan_balance scan_balance; |
0bf1457f | 1880 | unsigned long anon, file; |
9a265114 | 1881 | bool force_scan = false; |
4f98a2fe | 1882 | unsigned long ap, fp; |
4111304d | 1883 | enum lru_list lru; |
6f04f48d SS |
1884 | bool some_scanned; |
1885 | int pass; | |
246e87a9 | 1886 | |
f11c0ca5 JW |
1887 | /* |
1888 | * If the zone or memcg is small, nr[l] can be 0. This | |
1889 | * results in no scanning on this priority and a potential | |
1890 | * priority drop. Global direct reclaim can go to the next | |
1891 | * zone and tends to have no problems. Global kswapd is for | |
1892 | * zone balancing and it needs to scan a minimum amount. When | |
1893 | * reclaiming for a memcg, a priority drop can cause high | |
1894 | * latencies, so it's better to scan a minimum amount there as | |
1895 | * well. | |
1896 | */ | |
6e543d57 | 1897 | if (current_is_kswapd() && !zone_reclaimable(zone)) |
a4d3e9e7 | 1898 | force_scan = true; |
89b5fae5 | 1899 | if (!global_reclaim(sc)) |
a4d3e9e7 | 1900 | force_scan = true; |
76a33fc3 SL |
1901 | |
1902 | /* If we have no swap space, do not bother scanning anon pages. */ | |
ec8acf20 | 1903 | if (!sc->may_swap || (get_nr_swap_pages() <= 0)) { |
9a265114 | 1904 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
1905 | goto out; |
1906 | } | |
4f98a2fe | 1907 | |
10316b31 JW |
1908 | /* |
1909 | * Global reclaim will swap to prevent OOM even with no | |
1910 | * swappiness, but memcg users want to use this knob to | |
1911 | * disable swapping for individual groups completely when | |
1912 | * using the memory controller's swap limit feature would be | |
1913 | * too expensive. | |
1914 | */ | |
1915 | if (!global_reclaim(sc) && !vmscan_swappiness(sc)) { | |
9a265114 | 1916 | scan_balance = SCAN_FILE; |
10316b31 JW |
1917 | goto out; |
1918 | } | |
1919 | ||
1920 | /* | |
1921 | * Do not apply any pressure balancing cleverness when the | |
1922 | * system is close to OOM, scan both anon and file equally | |
1923 | * (unless the swappiness setting disagrees with swapping). | |
1924 | */ | |
1925 | if (!sc->priority && vmscan_swappiness(sc)) { | |
9a265114 | 1926 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
1927 | goto out; |
1928 | } | |
1929 | ||
4d7dcca2 HD |
1930 | anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + |
1931 | get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
1932 | file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
1933 | get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
a4d3e9e7 | 1934 | |
62376251 JW |
1935 | /* |
1936 | * Prevent the reclaimer from falling into the cache trap: as | |
1937 | * cache pages start out inactive, every cache fault will tip | |
1938 | * the scan balance towards the file LRU. And as the file LRU | |
1939 | * shrinks, so does the window for rotation from references. | |
1940 | * This means we have a runaway feedback loop where a tiny | |
1941 | * thrashing file LRU becomes infinitely more attractive than | |
1942 | * anon pages. Try to detect this based on file LRU size. | |
1943 | */ | |
1944 | if (global_reclaim(sc)) { | |
1945 | unsigned long free = zone_page_state(zone, NR_FREE_PAGES); | |
1946 | ||
1947 | if (unlikely(file + free <= high_wmark_pages(zone))) { | |
1948 | scan_balance = SCAN_ANON; | |
1949 | goto out; | |
1950 | } | |
1951 | } | |
1952 | ||
7c5bd705 JW |
1953 | /* |
1954 | * There is enough inactive page cache, do not reclaim | |
1955 | * anything from the anonymous working set right now. | |
1956 | */ | |
1957 | if (!inactive_file_is_low(lruvec)) { | |
9a265114 | 1958 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
1959 | goto out; |
1960 | } | |
1961 | ||
9a265114 JW |
1962 | scan_balance = SCAN_FRACT; |
1963 | ||
58c37f6e KM |
1964 | /* |
1965 | * With swappiness at 100, anonymous and file have the same priority. | |
1966 | * This scanning priority is essentially the inverse of IO cost. | |
1967 | */ | |
3d58ab5c | 1968 | anon_prio = vmscan_swappiness(sc); |
75b00af7 | 1969 | file_prio = 200 - anon_prio; |
58c37f6e | 1970 | |
4f98a2fe RR |
1971 | /* |
1972 | * OK, so we have swap space and a fair amount of page cache | |
1973 | * pages. We use the recently rotated / recently scanned | |
1974 | * ratios to determine how valuable each cache is. | |
1975 | * | |
1976 | * Because workloads change over time (and to avoid overflow) | |
1977 | * we keep these statistics as a floating average, which ends | |
1978 | * up weighing recent references more than old ones. | |
1979 | * | |
1980 | * anon in [0], file in [1] | |
1981 | */ | |
90126375 | 1982 | spin_lock_irq(&zone->lru_lock); |
6e901571 | 1983 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
1984 | reclaim_stat->recent_scanned[0] /= 2; |
1985 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
1986 | } |
1987 | ||
6e901571 | 1988 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
1989 | reclaim_stat->recent_scanned[1] /= 2; |
1990 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
1991 | } |
1992 | ||
4f98a2fe | 1993 | /* |
00d8089c RR |
1994 | * The amount of pressure on anon vs file pages is inversely |
1995 | * proportional to the fraction of recently scanned pages on | |
1996 | * each list that were recently referenced and in active use. | |
4f98a2fe | 1997 | */ |
fe35004f | 1998 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 1999 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2000 | |
fe35004f | 2001 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2002 | fp /= reclaim_stat->recent_rotated[1] + 1; |
90126375 | 2003 | spin_unlock_irq(&zone->lru_lock); |
4f98a2fe | 2004 | |
76a33fc3 SL |
2005 | fraction[0] = ap; |
2006 | fraction[1] = fp; | |
2007 | denominator = ap + fp + 1; | |
2008 | out: | |
6f04f48d SS |
2009 | some_scanned = false; |
2010 | /* Only use force_scan on second pass. */ | |
2011 | for (pass = 0; !some_scanned && pass < 2; pass++) { | |
2012 | for_each_evictable_lru(lru) { | |
2013 | int file = is_file_lru(lru); | |
2014 | unsigned long size; | |
2015 | unsigned long scan; | |
6e08a369 | 2016 | |
6f04f48d SS |
2017 | size = get_lru_size(lruvec, lru); |
2018 | scan = size >> sc->priority; | |
9a265114 | 2019 | |
6f04f48d SS |
2020 | if (!scan && pass && force_scan) |
2021 | scan = min(size, SWAP_CLUSTER_MAX); | |
9a265114 | 2022 | |
6f04f48d SS |
2023 | switch (scan_balance) { |
2024 | case SCAN_EQUAL: | |
2025 | /* Scan lists relative to size */ | |
2026 | break; | |
2027 | case SCAN_FRACT: | |
2028 | /* | |
2029 | * Scan types proportional to swappiness and | |
2030 | * their relative recent reclaim efficiency. | |
2031 | */ | |
2032 | scan = div64_u64(scan * fraction[file], | |
2033 | denominator); | |
2034 | break; | |
2035 | case SCAN_FILE: | |
2036 | case SCAN_ANON: | |
2037 | /* Scan one type exclusively */ | |
2038 | if ((scan_balance == SCAN_FILE) != file) | |
2039 | scan = 0; | |
2040 | break; | |
2041 | default: | |
2042 | /* Look ma, no brain */ | |
2043 | BUG(); | |
2044 | } | |
2045 | nr[lru] = scan; | |
9a265114 | 2046 | /* |
6f04f48d SS |
2047 | * Skip the second pass and don't force_scan, |
2048 | * if we found something to scan. | |
9a265114 | 2049 | */ |
6f04f48d | 2050 | some_scanned |= !!scan; |
9a265114 | 2051 | } |
76a33fc3 | 2052 | } |
6e08a369 | 2053 | } |
4f98a2fe | 2054 | |
9b4f98cd JW |
2055 | /* |
2056 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
2057 | */ | |
2058 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) | |
2059 | { | |
2060 | unsigned long nr[NR_LRU_LISTS]; | |
e82e0561 | 2061 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2062 | unsigned long nr_to_scan; |
2063 | enum lru_list lru; | |
2064 | unsigned long nr_reclaimed = 0; | |
2065 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2066 | struct blk_plug plug; | |
1a501907 | 2067 | bool scan_adjusted; |
9b4f98cd JW |
2068 | |
2069 | get_scan_count(lruvec, sc, nr); | |
2070 | ||
e82e0561 MG |
2071 | /* Record the original scan target for proportional adjustments later */ |
2072 | memcpy(targets, nr, sizeof(nr)); | |
2073 | ||
1a501907 MG |
2074 | /* |
2075 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2076 | * event that can occur when there is little memory pressure e.g. | |
2077 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2078 | * when the requested number of pages are reclaimed when scanning at | |
2079 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2080 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2081 | * do a batch of work at once. For memcg reclaim one check is made to | |
2082 | * abort proportional reclaim if either the file or anon lru has already | |
2083 | * dropped to zero at the first pass. | |
2084 | */ | |
2085 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2086 | sc->priority == DEF_PRIORITY); | |
2087 | ||
9b4f98cd JW |
2088 | blk_start_plug(&plug); |
2089 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2090 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2091 | unsigned long nr_anon, nr_file, percentage; |
2092 | unsigned long nr_scanned; | |
2093 | ||
9b4f98cd JW |
2094 | for_each_evictable_lru(lru) { |
2095 | if (nr[lru]) { | |
2096 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2097 | nr[lru] -= nr_to_scan; | |
2098 | ||
2099 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2100 | lruvec, sc); | |
2101 | } | |
2102 | } | |
e82e0561 MG |
2103 | |
2104 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) | |
2105 | continue; | |
2106 | ||
e82e0561 MG |
2107 | /* |
2108 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2109 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2110 | * proportionally what was requested by get_scan_count(). We |
2111 | * stop reclaiming one LRU and reduce the amount scanning | |
2112 | * proportional to the original scan target. | |
2113 | */ | |
2114 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2115 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2116 | ||
1a501907 MG |
2117 | /* |
2118 | * It's just vindictive to attack the larger once the smaller | |
2119 | * has gone to zero. And given the way we stop scanning the | |
2120 | * smaller below, this makes sure that we only make one nudge | |
2121 | * towards proportionality once we've got nr_to_reclaim. | |
2122 | */ | |
2123 | if (!nr_file || !nr_anon) | |
2124 | break; | |
2125 | ||
e82e0561 MG |
2126 | if (nr_file > nr_anon) { |
2127 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2128 | targets[LRU_ACTIVE_ANON] + 1; | |
2129 | lru = LRU_BASE; | |
2130 | percentage = nr_anon * 100 / scan_target; | |
2131 | } else { | |
2132 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2133 | targets[LRU_ACTIVE_FILE] + 1; | |
2134 | lru = LRU_FILE; | |
2135 | percentage = nr_file * 100 / scan_target; | |
2136 | } | |
2137 | ||
2138 | /* Stop scanning the smaller of the LRU */ | |
2139 | nr[lru] = 0; | |
2140 | nr[lru + LRU_ACTIVE] = 0; | |
2141 | ||
2142 | /* | |
2143 | * Recalculate the other LRU scan count based on its original | |
2144 | * scan target and the percentage scanning already complete | |
2145 | */ | |
2146 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2147 | nr_scanned = targets[lru] - nr[lru]; | |
2148 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2149 | nr[lru] -= min(nr[lru], nr_scanned); | |
2150 | ||
2151 | lru += LRU_ACTIVE; | |
2152 | nr_scanned = targets[lru] - nr[lru]; | |
2153 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2154 | nr[lru] -= min(nr[lru], nr_scanned); | |
2155 | ||
2156 | scan_adjusted = true; | |
9b4f98cd JW |
2157 | } |
2158 | blk_finish_plug(&plug); | |
2159 | sc->nr_reclaimed += nr_reclaimed; | |
2160 | ||
2161 | /* | |
2162 | * Even if we did not try to evict anon pages at all, we want to | |
2163 | * rebalance the anon lru active/inactive ratio. | |
2164 | */ | |
2165 | if (inactive_anon_is_low(lruvec)) | |
2166 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
2167 | sc, LRU_ACTIVE_ANON); | |
2168 | ||
2169 | throttle_vm_writeout(sc->gfp_mask); | |
2170 | } | |
2171 | ||
23b9da55 | 2172 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2173 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2174 | { |
d84da3f9 | 2175 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2176 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2177 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2178 | return true; |
2179 | ||
2180 | return false; | |
2181 | } | |
2182 | ||
3e7d3449 | 2183 | /* |
23b9da55 MG |
2184 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2185 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2186 | * true if more pages should be reclaimed such that when the page allocator | |
2187 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2188 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2189 | */ |
9b4f98cd | 2190 | static inline bool should_continue_reclaim(struct zone *zone, |
3e7d3449 MG |
2191 | unsigned long nr_reclaimed, |
2192 | unsigned long nr_scanned, | |
2193 | struct scan_control *sc) | |
2194 | { | |
2195 | unsigned long pages_for_compaction; | |
2196 | unsigned long inactive_lru_pages; | |
2197 | ||
2198 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2199 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2200 | return false; |
2201 | ||
2876592f MG |
2202 | /* Consider stopping depending on scan and reclaim activity */ |
2203 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2204 | /* | |
2205 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2206 | * full LRU list has been scanned and we are still failing | |
2207 | * to reclaim pages. This full LRU scan is potentially | |
2208 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2209 | */ | |
2210 | if (!nr_reclaimed && !nr_scanned) | |
2211 | return false; | |
2212 | } else { | |
2213 | /* | |
2214 | * For non-__GFP_REPEAT allocations which can presumably | |
2215 | * fail without consequence, stop if we failed to reclaim | |
2216 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2217 | * pages that were scanned. This will return to the | |
2218 | * caller faster at the risk reclaim/compaction and | |
2219 | * the resulting allocation attempt fails | |
2220 | */ | |
2221 | if (!nr_reclaimed) | |
2222 | return false; | |
2223 | } | |
3e7d3449 MG |
2224 | |
2225 | /* | |
2226 | * If we have not reclaimed enough pages for compaction and the | |
2227 | * inactive lists are large enough, continue reclaiming | |
2228 | */ | |
2229 | pages_for_compaction = (2UL << sc->order); | |
9b4f98cd | 2230 | inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); |
ec8acf20 | 2231 | if (get_nr_swap_pages() > 0) |
9b4f98cd | 2232 | inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); |
3e7d3449 MG |
2233 | if (sc->nr_reclaimed < pages_for_compaction && |
2234 | inactive_lru_pages > pages_for_compaction) | |
2235 | return true; | |
2236 | ||
2237 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
9b4f98cd | 2238 | switch (compaction_suitable(zone, sc->order)) { |
3e7d3449 MG |
2239 | case COMPACT_PARTIAL: |
2240 | case COMPACT_CONTINUE: | |
2241 | return false; | |
2242 | default: | |
2243 | return true; | |
2244 | } | |
2245 | } | |
2246 | ||
0608f43d | 2247 | static void shrink_zone(struct zone *zone, struct scan_control *sc) |
1da177e4 | 2248 | { |
f0fdc5e8 | 2249 | unsigned long nr_reclaimed, nr_scanned; |
1da177e4 | 2250 | |
9b4f98cd JW |
2251 | do { |
2252 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2253 | struct mem_cgroup_reclaim_cookie reclaim = { | |
2254 | .zone = zone, | |
2255 | .priority = sc->priority, | |
2256 | }; | |
694fbc0f | 2257 | struct mem_cgroup *memcg; |
3e7d3449 | 2258 | |
9b4f98cd JW |
2259 | nr_reclaimed = sc->nr_reclaimed; |
2260 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2261 | |
694fbc0f AM |
2262 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2263 | do { | |
9b4f98cd | 2264 | struct lruvec *lruvec; |
5660048c | 2265 | |
9b4f98cd | 2266 | lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
f9be23d6 | 2267 | |
9b4f98cd | 2268 | shrink_lruvec(lruvec, sc); |
f16015fb | 2269 | |
9b4f98cd | 2270 | /* |
a394cb8e MH |
2271 | * Direct reclaim and kswapd have to scan all memory |
2272 | * cgroups to fulfill the overall scan target for the | |
9b4f98cd | 2273 | * zone. |
a394cb8e MH |
2274 | * |
2275 | * Limit reclaim, on the other hand, only cares about | |
2276 | * nr_to_reclaim pages to be reclaimed and it will | |
2277 | * retry with decreasing priority if one round over the | |
2278 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2279 | */ |
a394cb8e MH |
2280 | if (!global_reclaim(sc) && |
2281 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2282 | mem_cgroup_iter_break(root, memcg); |
2283 | break; | |
2284 | } | |
694fbc0f AM |
2285 | memcg = mem_cgroup_iter(root, memcg, &reclaim); |
2286 | } while (memcg); | |
70ddf637 AV |
2287 | |
2288 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, | |
2289 | sc->nr_scanned - nr_scanned, | |
2290 | sc->nr_reclaimed - nr_reclaimed); | |
2291 | ||
9b4f98cd JW |
2292 | } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, |
2293 | sc->nr_scanned - nr_scanned, sc)); | |
f16015fb JW |
2294 | } |
2295 | ||
fe4b1b24 MG |
2296 | /* Returns true if compaction should go ahead for a high-order request */ |
2297 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
2298 | { | |
2299 | unsigned long balance_gap, watermark; | |
2300 | bool watermark_ok; | |
2301 | ||
2302 | /* Do not consider compaction for orders reclaim is meant to satisfy */ | |
2303 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
2304 | return false; | |
2305 | ||
2306 | /* | |
2307 | * Compaction takes time to run and there are potentially other | |
2308 | * callers using the pages just freed. Continue reclaiming until | |
2309 | * there is a buffer of free pages available to give compaction | |
2310 | * a reasonable chance of completing and allocating the page | |
2311 | */ | |
4be89a34 JZ |
2312 | balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( |
2313 | zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); | |
fe4b1b24 MG |
2314 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order); |
2315 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); | |
2316 | ||
2317 | /* | |
2318 | * If compaction is deferred, reclaim up to a point where | |
2319 | * compaction will have a chance of success when re-enabled | |
2320 | */ | |
aff62249 | 2321 | if (compaction_deferred(zone, sc->order)) |
fe4b1b24 MG |
2322 | return watermark_ok; |
2323 | ||
2324 | /* If compaction is not ready to start, keep reclaiming */ | |
2325 | if (!compaction_suitable(zone, sc->order)) | |
2326 | return false; | |
2327 | ||
2328 | return watermark_ok; | |
2329 | } | |
2330 | ||
1da177e4 LT |
2331 | /* |
2332 | * This is the direct reclaim path, for page-allocating processes. We only | |
2333 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2334 | * request. | |
2335 | * | |
41858966 MG |
2336 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
2337 | * Because: | |
1da177e4 LT |
2338 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
2339 | * allocation or | |
41858966 MG |
2340 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
2341 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2342 | * zone defense algorithm. | |
1da177e4 | 2343 | * |
1da177e4 LT |
2344 | * If a zone is deemed to be full of pinned pages then just give it a light |
2345 | * scan then give up on it. | |
e0c23279 MG |
2346 | * |
2347 | * This function returns true if a zone is being reclaimed for a costly | |
fe4b1b24 | 2348 | * high-order allocation and compaction is ready to begin. This indicates to |
0cee34fd MG |
2349 | * the caller that it should consider retrying the allocation instead of |
2350 | * further reclaim. | |
1da177e4 | 2351 | */ |
3115cd91 | 2352 | static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2353 | { |
dd1a239f | 2354 | struct zoneref *z; |
54a6eb5c | 2355 | struct zone *zone; |
0608f43d AM |
2356 | unsigned long nr_soft_reclaimed; |
2357 | unsigned long nr_soft_scanned; | |
65ec02cb | 2358 | unsigned long lru_pages = 0; |
0cee34fd | 2359 | bool aborted_reclaim = false; |
65ec02cb | 2360 | struct reclaim_state *reclaim_state = current->reclaim_state; |
619d0d76 | 2361 | gfp_t orig_mask; |
3115cd91 VD |
2362 | struct shrink_control shrink = { |
2363 | .gfp_mask = sc->gfp_mask, | |
2364 | }; | |
9bbc04ee | 2365 | enum zone_type requested_highidx = gfp_zone(sc->gfp_mask); |
1cfb419b | 2366 | |
cc715d99 MG |
2367 | /* |
2368 | * If the number of buffer_heads in the machine exceeds the maximum | |
2369 | * allowed level, force direct reclaim to scan the highmem zone as | |
2370 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2371 | */ | |
619d0d76 | 2372 | orig_mask = sc->gfp_mask; |
cc715d99 MG |
2373 | if (buffer_heads_over_limit) |
2374 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2375 | ||
3115cd91 | 2376 | nodes_clear(shrink.nodes_to_scan); |
65ec02cb | 2377 | |
d4debc66 MG |
2378 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
2379 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
f3fe6512 | 2380 | if (!populated_zone(zone)) |
1da177e4 | 2381 | continue; |
1cfb419b KH |
2382 | /* |
2383 | * Take care memory controller reclaiming has small influence | |
2384 | * to global LRU. | |
2385 | */ | |
89b5fae5 | 2386 | if (global_reclaim(sc)) { |
1cfb419b KH |
2387 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2388 | continue; | |
65ec02cb VD |
2389 | |
2390 | lru_pages += zone_reclaimable_pages(zone); | |
3115cd91 | 2391 | node_set(zone_to_nid(zone), shrink.nodes_to_scan); |
65ec02cb | 2392 | |
6e543d57 LD |
2393 | if (sc->priority != DEF_PRIORITY && |
2394 | !zone_reclaimable(zone)) | |
1cfb419b | 2395 | continue; /* Let kswapd poll it */ |
d84da3f9 | 2396 | if (IS_ENABLED(CONFIG_COMPACTION)) { |
e0887c19 | 2397 | /* |
e0c23279 MG |
2398 | * If we already have plenty of memory free for |
2399 | * compaction in this zone, don't free any more. | |
2400 | * Even though compaction is invoked for any | |
2401 | * non-zero order, only frequent costly order | |
2402 | * reclamation is disruptive enough to become a | |
c7cfa37b CA |
2403 | * noticeable problem, like transparent huge |
2404 | * page allocations. | |
e0887c19 | 2405 | */ |
9bbc04ee WY |
2406 | if ((zonelist_zone_idx(z) <= requested_highidx) |
2407 | && compaction_ready(zone, sc)) { | |
0cee34fd | 2408 | aborted_reclaim = true; |
e0887c19 | 2409 | continue; |
e0c23279 | 2410 | } |
e0887c19 | 2411 | } |
0608f43d AM |
2412 | /* |
2413 | * This steals pages from memory cgroups over softlimit | |
2414 | * and returns the number of reclaimed pages and | |
2415 | * scanned pages. This works for global memory pressure | |
2416 | * and balancing, not for a memcg's limit. | |
2417 | */ | |
2418 | nr_soft_scanned = 0; | |
2419 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2420 | sc->order, sc->gfp_mask, | |
2421 | &nr_soft_scanned); | |
2422 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2423 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2424 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2425 | } |
408d8544 | 2426 | |
9e3b2f8c | 2427 | shrink_zone(zone, sc); |
1da177e4 | 2428 | } |
e0c23279 | 2429 | |
65ec02cb VD |
2430 | /* |
2431 | * Don't shrink slabs when reclaiming memory from over limit cgroups | |
2432 | * but do shrink slab at least once when aborting reclaim for | |
2433 | * compaction to avoid unevenly scanning file/anon LRU pages over slab | |
2434 | * pages. | |
2435 | */ | |
2436 | if (global_reclaim(sc)) { | |
3115cd91 | 2437 | shrink_slab(&shrink, sc->nr_scanned, lru_pages); |
65ec02cb VD |
2438 | if (reclaim_state) { |
2439 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2440 | reclaim_state->reclaimed_slab = 0; | |
2441 | } | |
2442 | } | |
2443 | ||
619d0d76 WY |
2444 | /* |
2445 | * Restore to original mask to avoid the impact on the caller if we | |
2446 | * promoted it to __GFP_HIGHMEM. | |
2447 | */ | |
2448 | sc->gfp_mask = orig_mask; | |
2449 | ||
0cee34fd | 2450 | return aborted_reclaim; |
d1908362 MK |
2451 | } |
2452 | ||
929bea7c | 2453 | /* All zones in zonelist are unreclaimable? */ |
d1908362 MK |
2454 | static bool all_unreclaimable(struct zonelist *zonelist, |
2455 | struct scan_control *sc) | |
2456 | { | |
2457 | struct zoneref *z; | |
2458 | struct zone *zone; | |
d1908362 MK |
2459 | |
2460 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2461 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2462 | if (!populated_zone(zone)) | |
2463 | continue; | |
2464 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2465 | continue; | |
6e543d57 | 2466 | if (zone_reclaimable(zone)) |
929bea7c | 2467 | return false; |
d1908362 MK |
2468 | } |
2469 | ||
929bea7c | 2470 | return true; |
1da177e4 | 2471 | } |
4f98a2fe | 2472 | |
1da177e4 LT |
2473 | /* |
2474 | * This is the main entry point to direct page reclaim. | |
2475 | * | |
2476 | * If a full scan of the inactive list fails to free enough memory then we | |
2477 | * are "out of memory" and something needs to be killed. | |
2478 | * | |
2479 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2480 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2481 | * caller can't do much about. We kick the writeback threads and take explicit |
2482 | * naps in the hope that some of these pages can be written. But if the | |
2483 | * allocating task holds filesystem locks which prevent writeout this might not | |
2484 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2485 | * |
2486 | * returns: 0, if no pages reclaimed | |
2487 | * else, the number of pages reclaimed | |
1da177e4 | 2488 | */ |
dac1d27b | 2489 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2490 | struct scan_control *sc) |
1da177e4 | 2491 | { |
69e05944 | 2492 | unsigned long total_scanned = 0; |
22fba335 | 2493 | unsigned long writeback_threshold; |
0cee34fd | 2494 | bool aborted_reclaim; |
1da177e4 | 2495 | |
873b4771 KK |
2496 | delayacct_freepages_start(); |
2497 | ||
89b5fae5 | 2498 | if (global_reclaim(sc)) |
1cfb419b | 2499 | count_vm_event(ALLOCSTALL); |
1da177e4 | 2500 | |
9e3b2f8c | 2501 | do { |
70ddf637 AV |
2502 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2503 | sc->priority); | |
66e1707b | 2504 | sc->nr_scanned = 0; |
3115cd91 | 2505 | aborted_reclaim = shrink_zones(zonelist, sc); |
c6a8a8c5 | 2506 | |
66e1707b | 2507 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2508 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
1da177e4 | 2509 | goto out; |
1da177e4 | 2510 | |
0e50ce3b MK |
2511 | /* |
2512 | * If we're getting trouble reclaiming, start doing | |
2513 | * writepage even in laptop mode. | |
2514 | */ | |
2515 | if (sc->priority < DEF_PRIORITY - 2) | |
2516 | sc->may_writepage = 1; | |
2517 | ||
1da177e4 LT |
2518 | /* |
2519 | * Try to write back as many pages as we just scanned. This | |
2520 | * tends to cause slow streaming writers to write data to the | |
2521 | * disk smoothly, at the dirtying rate, which is nice. But | |
2522 | * that's undesirable in laptop mode, where we *want* lumpy | |
2523 | * writeout. So in laptop mode, write out the whole world. | |
2524 | */ | |
22fba335 KM |
2525 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2526 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2527 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2528 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2529 | sc->may_writepage = 1; |
1da177e4 | 2530 | } |
5a1c9cbc | 2531 | } while (--sc->priority >= 0 && !aborted_reclaim); |
bb21c7ce | 2532 | |
1da177e4 | 2533 | out: |
873b4771 KK |
2534 | delayacct_freepages_end(); |
2535 | ||
bb21c7ce KM |
2536 | if (sc->nr_reclaimed) |
2537 | return sc->nr_reclaimed; | |
2538 | ||
929bea7c KM |
2539 | /* |
2540 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2541 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2542 | * check. | |
2543 | */ | |
2544 | if (oom_killer_disabled) | |
2545 | return 0; | |
2546 | ||
0cee34fd MG |
2547 | /* Aborted reclaim to try compaction? don't OOM, then */ |
2548 | if (aborted_reclaim) | |
7335084d MG |
2549 | return 1; |
2550 | ||
bb21c7ce | 2551 | /* top priority shrink_zones still had more to do? don't OOM, then */ |
89b5fae5 | 2552 | if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc)) |
bb21c7ce KM |
2553 | return 1; |
2554 | ||
2555 | return 0; | |
1da177e4 LT |
2556 | } |
2557 | ||
5515061d MG |
2558 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) |
2559 | { | |
2560 | struct zone *zone; | |
2561 | unsigned long pfmemalloc_reserve = 0; | |
2562 | unsigned long free_pages = 0; | |
2563 | int i; | |
2564 | bool wmark_ok; | |
2565 | ||
2566 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2567 | zone = &pgdat->node_zones[i]; | |
675becce MG |
2568 | if (!populated_zone(zone)) |
2569 | continue; | |
2570 | ||
5515061d MG |
2571 | pfmemalloc_reserve += min_wmark_pages(zone); |
2572 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2573 | } | |
2574 | ||
675becce MG |
2575 | /* If there are no reserves (unexpected config) then do not throttle */ |
2576 | if (!pfmemalloc_reserve) | |
2577 | return true; | |
2578 | ||
5515061d MG |
2579 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
2580 | ||
2581 | /* kswapd must be awake if processes are being throttled */ | |
2582 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2583 | pgdat->classzone_idx = min(pgdat->classzone_idx, | |
2584 | (enum zone_type)ZONE_NORMAL); | |
2585 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2586 | } | |
2587 | ||
2588 | return wmark_ok; | |
2589 | } | |
2590 | ||
2591 | /* | |
2592 | * Throttle direct reclaimers if backing storage is backed by the network | |
2593 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2594 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2595 | * when the low watermark is reached. |
2596 | * | |
2597 | * Returns true if a fatal signal was delivered during throttling. If this | |
2598 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2599 | */ |
50694c28 | 2600 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2601 | nodemask_t *nodemask) |
2602 | { | |
675becce | 2603 | struct zoneref *z; |
5515061d | 2604 | struct zone *zone; |
675becce | 2605 | pg_data_t *pgdat = NULL; |
5515061d MG |
2606 | |
2607 | /* | |
2608 | * Kernel threads should not be throttled as they may be indirectly | |
2609 | * responsible for cleaning pages necessary for reclaim to make forward | |
2610 | * progress. kjournald for example may enter direct reclaim while | |
2611 | * committing a transaction where throttling it could forcing other | |
2612 | * processes to block on log_wait_commit(). | |
2613 | */ | |
2614 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2615 | goto out; |
2616 | ||
2617 | /* | |
2618 | * If a fatal signal is pending, this process should not throttle. | |
2619 | * It should return quickly so it can exit and free its memory | |
2620 | */ | |
2621 | if (fatal_signal_pending(current)) | |
2622 | goto out; | |
5515061d | 2623 | |
675becce MG |
2624 | /* |
2625 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2626 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2627 | * GFP_KERNEL will be required for allocating network buffers when | |
2628 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2629 | * | |
2630 | * Throttling is based on the first usable node and throttled processes | |
2631 | * wait on a queue until kswapd makes progress and wakes them. There | |
2632 | * is an affinity then between processes waking up and where reclaim | |
2633 | * progress has been made assuming the process wakes on the same node. | |
2634 | * More importantly, processes running on remote nodes will not compete | |
2635 | * for remote pfmemalloc reserves and processes on different nodes | |
2636 | * should make reasonable progress. | |
2637 | */ | |
2638 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2639 | gfp_mask, nodemask) { | |
2640 | if (zone_idx(zone) > ZONE_NORMAL) | |
2641 | continue; | |
2642 | ||
2643 | /* Throttle based on the first usable node */ | |
2644 | pgdat = zone->zone_pgdat; | |
2645 | if (pfmemalloc_watermark_ok(pgdat)) | |
2646 | goto out; | |
2647 | break; | |
2648 | } | |
2649 | ||
2650 | /* If no zone was usable by the allocation flags then do not throttle */ | |
2651 | if (!pgdat) | |
50694c28 | 2652 | goto out; |
5515061d | 2653 | |
68243e76 MG |
2654 | /* Account for the throttling */ |
2655 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2656 | ||
5515061d MG |
2657 | /* |
2658 | * If the caller cannot enter the filesystem, it's possible that it | |
2659 | * is due to the caller holding an FS lock or performing a journal | |
2660 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2661 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2662 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2663 | * second before continuing. | |
2664 | */ | |
2665 | if (!(gfp_mask & __GFP_FS)) { | |
2666 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2667 | pfmemalloc_watermark_ok(pgdat), HZ); | |
50694c28 MG |
2668 | |
2669 | goto check_pending; | |
5515061d MG |
2670 | } |
2671 | ||
2672 | /* Throttle until kswapd wakes the process */ | |
2673 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2674 | pfmemalloc_watermark_ok(pgdat)); | |
50694c28 MG |
2675 | |
2676 | check_pending: | |
2677 | if (fatal_signal_pending(current)) | |
2678 | return true; | |
2679 | ||
2680 | out: | |
2681 | return false; | |
5515061d MG |
2682 | } |
2683 | ||
dac1d27b | 2684 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2685 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2686 | { |
33906bc5 | 2687 | unsigned long nr_reclaimed; |
66e1707b | 2688 | struct scan_control sc = { |
21caf2fc | 2689 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
66e1707b | 2690 | .may_writepage = !laptop_mode, |
22fba335 | 2691 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
a6dc60f8 | 2692 | .may_unmap = 1, |
2e2e4259 | 2693 | .may_swap = 1, |
66e1707b | 2694 | .order = order, |
9e3b2f8c | 2695 | .priority = DEF_PRIORITY, |
f16015fb | 2696 | .target_mem_cgroup = NULL, |
327c0e96 | 2697 | .nodemask = nodemask, |
66e1707b BS |
2698 | }; |
2699 | ||
5515061d | 2700 | /* |
50694c28 MG |
2701 | * Do not enter reclaim if fatal signal was delivered while throttled. |
2702 | * 1 is returned so that the page allocator does not OOM kill at this | |
2703 | * point. | |
5515061d | 2704 | */ |
50694c28 | 2705 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) |
5515061d MG |
2706 | return 1; |
2707 | ||
33906bc5 MG |
2708 | trace_mm_vmscan_direct_reclaim_begin(order, |
2709 | sc.may_writepage, | |
2710 | gfp_mask); | |
2711 | ||
3115cd91 | 2712 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
2713 | |
2714 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2715 | ||
2716 | return nr_reclaimed; | |
66e1707b BS |
2717 | } |
2718 | ||
c255a458 | 2719 | #ifdef CONFIG_MEMCG |
66e1707b | 2720 | |
72835c86 | 2721 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, |
4e416953 | 2722 | gfp_t gfp_mask, bool noswap, |
0ae5e89c YH |
2723 | struct zone *zone, |
2724 | unsigned long *nr_scanned) | |
4e416953 BS |
2725 | { |
2726 | struct scan_control sc = { | |
0ae5e89c | 2727 | .nr_scanned = 0, |
b8f5c566 | 2728 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
4e416953 BS |
2729 | .may_writepage = !laptop_mode, |
2730 | .may_unmap = 1, | |
2731 | .may_swap = !noswap, | |
4e416953 | 2732 | .order = 0, |
9e3b2f8c | 2733 | .priority = 0, |
72835c86 | 2734 | .target_mem_cgroup = memcg, |
4e416953 | 2735 | }; |
f9be23d6 | 2736 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
0ae5e89c | 2737 | |
4e416953 BS |
2738 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2739 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 2740 | |
9e3b2f8c | 2741 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e KM |
2742 | sc.may_writepage, |
2743 | sc.gfp_mask); | |
2744 | ||
4e416953 BS |
2745 | /* |
2746 | * NOTE: Although we can get the priority field, using it | |
2747 | * here is not a good idea, since it limits the pages we can scan. | |
2748 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2749 | * will pick up pages from other mem cgroup's as well. We hack | |
2750 | * the priority and make it zero. | |
2751 | */ | |
f9be23d6 | 2752 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
2753 | |
2754 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2755 | ||
0ae5e89c | 2756 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2757 | return sc.nr_reclaimed; |
2758 | } | |
2759 | ||
72835c86 | 2760 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
a7885eb8 | 2761 | gfp_t gfp_mask, |
185efc0f | 2762 | bool noswap) |
66e1707b | 2763 | { |
4e416953 | 2764 | struct zonelist *zonelist; |
bdce6d9e | 2765 | unsigned long nr_reclaimed; |
889976db | 2766 | int nid; |
66e1707b | 2767 | struct scan_control sc = { |
66e1707b | 2768 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2769 | .may_unmap = 1, |
2e2e4259 | 2770 | .may_swap = !noswap, |
22fba335 | 2771 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
66e1707b | 2772 | .order = 0, |
9e3b2f8c | 2773 | .priority = DEF_PRIORITY, |
72835c86 | 2774 | .target_mem_cgroup = memcg, |
327c0e96 | 2775 | .nodemask = NULL, /* we don't care the placement */ |
a09ed5e0 YH |
2776 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2777 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2778 | }; | |
66e1707b | 2779 | |
889976db YH |
2780 | /* |
2781 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2782 | * take care of from where we get pages. So the node where we start the | |
2783 | * scan does not need to be the current node. | |
2784 | */ | |
72835c86 | 2785 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2786 | |
2787 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2788 | |
2789 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2790 | sc.may_writepage, | |
2791 | sc.gfp_mask); | |
2792 | ||
3115cd91 | 2793 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
bdce6d9e KM |
2794 | |
2795 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2796 | ||
2797 | return nr_reclaimed; | |
66e1707b BS |
2798 | } |
2799 | #endif | |
2800 | ||
9e3b2f8c | 2801 | static void age_active_anon(struct zone *zone, struct scan_control *sc) |
f16015fb | 2802 | { |
b95a2f2d | 2803 | struct mem_cgroup *memcg; |
f16015fb | 2804 | |
b95a2f2d JW |
2805 | if (!total_swap_pages) |
2806 | return; | |
2807 | ||
2808 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2809 | do { | |
c56d5c7d | 2810 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
b95a2f2d | 2811 | |
c56d5c7d | 2812 | if (inactive_anon_is_low(lruvec)) |
1a93be0e | 2813 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 2814 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
2815 | |
2816 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2817 | } while (memcg); | |
f16015fb JW |
2818 | } |
2819 | ||
60cefed4 JW |
2820 | static bool zone_balanced(struct zone *zone, int order, |
2821 | unsigned long balance_gap, int classzone_idx) | |
2822 | { | |
2823 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + | |
2824 | balance_gap, classzone_idx, 0)) | |
2825 | return false; | |
2826 | ||
d84da3f9 KS |
2827 | if (IS_ENABLED(CONFIG_COMPACTION) && order && |
2828 | !compaction_suitable(zone, order)) | |
60cefed4 JW |
2829 | return false; |
2830 | ||
2831 | return true; | |
2832 | } | |
2833 | ||
1741c877 | 2834 | /* |
4ae0a48b ZC |
2835 | * pgdat_balanced() is used when checking if a node is balanced. |
2836 | * | |
2837 | * For order-0, all zones must be balanced! | |
2838 | * | |
2839 | * For high-order allocations only zones that meet watermarks and are in a | |
2840 | * zone allowed by the callers classzone_idx are added to balanced_pages. The | |
2841 | * total of balanced pages must be at least 25% of the zones allowed by | |
2842 | * classzone_idx for the node to be considered balanced. Forcing all zones to | |
2843 | * be balanced for high orders can cause excessive reclaim when there are | |
2844 | * imbalanced zones. | |
1741c877 MG |
2845 | * The choice of 25% is due to |
2846 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2847 | * reasonable sized machine | |
2848 | * o On all other machines, the top zone must be at least a reasonable | |
25985edc | 2849 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
1741c877 MG |
2850 | * would need to be at least 256M for it to be balance a whole node. |
2851 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2852 | * to balance a node on its own. These seemed like reasonable ratios. | |
2853 | */ | |
4ae0a48b | 2854 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) |
1741c877 | 2855 | { |
b40da049 | 2856 | unsigned long managed_pages = 0; |
4ae0a48b | 2857 | unsigned long balanced_pages = 0; |
1741c877 MG |
2858 | int i; |
2859 | ||
4ae0a48b ZC |
2860 | /* Check the watermark levels */ |
2861 | for (i = 0; i <= classzone_idx; i++) { | |
2862 | struct zone *zone = pgdat->node_zones + i; | |
1741c877 | 2863 | |
4ae0a48b ZC |
2864 | if (!populated_zone(zone)) |
2865 | continue; | |
2866 | ||
b40da049 | 2867 | managed_pages += zone->managed_pages; |
4ae0a48b ZC |
2868 | |
2869 | /* | |
2870 | * A special case here: | |
2871 | * | |
2872 | * balance_pgdat() skips over all_unreclaimable after | |
2873 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2874 | * they must be considered balanced here as well! | |
2875 | */ | |
6e543d57 | 2876 | if (!zone_reclaimable(zone)) { |
b40da049 | 2877 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
2878 | continue; |
2879 | } | |
2880 | ||
2881 | if (zone_balanced(zone, order, 0, i)) | |
b40da049 | 2882 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
2883 | else if (!order) |
2884 | return false; | |
2885 | } | |
2886 | ||
2887 | if (order) | |
b40da049 | 2888 | return balanced_pages >= (managed_pages >> 2); |
4ae0a48b ZC |
2889 | else |
2890 | return true; | |
1741c877 MG |
2891 | } |
2892 | ||
5515061d MG |
2893 | /* |
2894 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
2895 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
2896 | * | |
2897 | * Returns true if kswapd is ready to sleep | |
2898 | */ | |
2899 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, | |
dc83edd9 | 2900 | int classzone_idx) |
f50de2d3 | 2901 | { |
f50de2d3 MG |
2902 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ |
2903 | if (remaining) | |
5515061d MG |
2904 | return false; |
2905 | ||
2906 | /* | |
2907 | * There is a potential race between when kswapd checks its watermarks | |
2908 | * and a process gets throttled. There is also a potential race if | |
2909 | * processes get throttled, kswapd wakes, a large process exits therby | |
2910 | * balancing the zones that causes kswapd to miss a wakeup. If kswapd | |
2911 | * is going to sleep, no process should be sleeping on pfmemalloc_wait | |
2912 | * so wake them now if necessary. If necessary, processes will wake | |
2913 | * kswapd and get throttled again | |
2914 | */ | |
2915 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) { | |
2916 | wake_up(&pgdat->pfmemalloc_wait); | |
2917 | return false; | |
2918 | } | |
f50de2d3 | 2919 | |
4ae0a48b | 2920 | return pgdat_balanced(pgdat, order, classzone_idx); |
f50de2d3 MG |
2921 | } |
2922 | ||
75485363 MG |
2923 | /* |
2924 | * kswapd shrinks the zone by the number of pages required to reach | |
2925 | * the high watermark. | |
b8e83b94 MG |
2926 | * |
2927 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
2928 | * reclaim or if the lack of progress was due to pages under writeback. |
2929 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 2930 | */ |
b8e83b94 | 2931 | static bool kswapd_shrink_zone(struct zone *zone, |
7c954f6d | 2932 | int classzone_idx, |
75485363 | 2933 | struct scan_control *sc, |
2ab44f43 MG |
2934 | unsigned long lru_pages, |
2935 | unsigned long *nr_attempted) | |
75485363 | 2936 | { |
7c954f6d MG |
2937 | int testorder = sc->order; |
2938 | unsigned long balance_gap; | |
75485363 MG |
2939 | struct reclaim_state *reclaim_state = current->reclaim_state; |
2940 | struct shrink_control shrink = { | |
2941 | .gfp_mask = sc->gfp_mask, | |
2942 | }; | |
7c954f6d | 2943 | bool lowmem_pressure; |
75485363 MG |
2944 | |
2945 | /* Reclaim above the high watermark. */ | |
2946 | sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone)); | |
7c954f6d MG |
2947 | |
2948 | /* | |
2949 | * Kswapd reclaims only single pages with compaction enabled. Trying | |
2950 | * too hard to reclaim until contiguous free pages have become | |
2951 | * available can hurt performance by evicting too much useful data | |
2952 | * from memory. Do not reclaim more than needed for compaction. | |
2953 | */ | |
2954 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && | |
2955 | compaction_suitable(zone, sc->order) != | |
2956 | COMPACT_SKIPPED) | |
2957 | testorder = 0; | |
2958 | ||
2959 | /* | |
2960 | * We put equal pressure on every zone, unless one zone has way too | |
2961 | * many pages free already. The "too many pages" is defined as the | |
2962 | * high wmark plus a "gap" where the gap is either the low | |
2963 | * watermark or 1% of the zone, whichever is smaller. | |
2964 | */ | |
4be89a34 JZ |
2965 | balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( |
2966 | zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); | |
7c954f6d MG |
2967 | |
2968 | /* | |
2969 | * If there is no low memory pressure or the zone is balanced then no | |
2970 | * reclaim is necessary | |
2971 | */ | |
2972 | lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone)); | |
2973 | if (!lowmem_pressure && zone_balanced(zone, testorder, | |
2974 | balance_gap, classzone_idx)) | |
2975 | return true; | |
2976 | ||
75485363 | 2977 | shrink_zone(zone, sc); |
0ce3d744 DC |
2978 | nodes_clear(shrink.nodes_to_scan); |
2979 | node_set(zone_to_nid(zone), shrink.nodes_to_scan); | |
75485363 MG |
2980 | |
2981 | reclaim_state->reclaimed_slab = 0; | |
6e543d57 | 2982 | shrink_slab(&shrink, sc->nr_scanned, lru_pages); |
75485363 MG |
2983 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
2984 | ||
2ab44f43 MG |
2985 | /* Account for the number of pages attempted to reclaim */ |
2986 | *nr_attempted += sc->nr_to_reclaim; | |
2987 | ||
283aba9f MG |
2988 | zone_clear_flag(zone, ZONE_WRITEBACK); |
2989 | ||
7c954f6d MG |
2990 | /* |
2991 | * If a zone reaches its high watermark, consider it to be no longer | |
2992 | * congested. It's possible there are dirty pages backed by congested | |
2993 | * BDIs but as pressure is relieved, speculatively avoid congestion | |
2994 | * waits. | |
2995 | */ | |
6e543d57 | 2996 | if (zone_reclaimable(zone) && |
7c954f6d MG |
2997 | zone_balanced(zone, testorder, 0, classzone_idx)) { |
2998 | zone_clear_flag(zone, ZONE_CONGESTED); | |
2999 | zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY); | |
3000 | } | |
3001 | ||
b8e83b94 | 3002 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3003 | } |
3004 | ||
1da177e4 LT |
3005 | /* |
3006 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 3007 | * they are all at high_wmark_pages(zone). |
1da177e4 | 3008 | * |
0abdee2b | 3009 | * Returns the final order kswapd was reclaiming at |
1da177e4 LT |
3010 | * |
3011 | * There is special handling here for zones which are full of pinned pages. | |
3012 | * This can happen if the pages are all mlocked, or if they are all used by | |
3013 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
3014 | * What we do is to detect the case where all pages in the zone have been | |
3015 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
3016 | * dead and from now on, only perform a short scan. Basically we're polling | |
3017 | * the zone for when the problem goes away. | |
3018 | * | |
3019 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
3020 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
3021 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
3022 | * lower zones regardless of the number of free pages in the lower zones. This | |
3023 | * interoperates with the page allocator fallback scheme to ensure that aging | |
3024 | * of pages is balanced across the zones. | |
1da177e4 | 3025 | */ |
99504748 | 3026 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, |
dc83edd9 | 3027 | int *classzone_idx) |
1da177e4 | 3028 | { |
1da177e4 | 3029 | int i; |
99504748 | 3030 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
0608f43d AM |
3031 | unsigned long nr_soft_reclaimed; |
3032 | unsigned long nr_soft_scanned; | |
179e9639 AM |
3033 | struct scan_control sc = { |
3034 | .gfp_mask = GFP_KERNEL, | |
b8e83b94 | 3035 | .priority = DEF_PRIORITY, |
a6dc60f8 | 3036 | .may_unmap = 1, |
2e2e4259 | 3037 | .may_swap = 1, |
b8e83b94 | 3038 | .may_writepage = !laptop_mode, |
5ad333eb | 3039 | .order = order, |
f16015fb | 3040 | .target_mem_cgroup = NULL, |
179e9639 | 3041 | }; |
f8891e5e | 3042 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3043 | |
9e3b2f8c | 3044 | do { |
1da177e4 | 3045 | unsigned long lru_pages = 0; |
2ab44f43 | 3046 | unsigned long nr_attempted = 0; |
b8e83b94 | 3047 | bool raise_priority = true; |
2ab44f43 | 3048 | bool pgdat_needs_compaction = (order > 0); |
b8e83b94 MG |
3049 | |
3050 | sc.nr_reclaimed = 0; | |
1da177e4 | 3051 | |
d6277db4 RW |
3052 | /* |
3053 | * Scan in the highmem->dma direction for the highest | |
3054 | * zone which needs scanning | |
3055 | */ | |
3056 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
3057 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 3058 | |
d6277db4 RW |
3059 | if (!populated_zone(zone)) |
3060 | continue; | |
1da177e4 | 3061 | |
6e543d57 LD |
3062 | if (sc.priority != DEF_PRIORITY && |
3063 | !zone_reclaimable(zone)) | |
d6277db4 | 3064 | continue; |
1da177e4 | 3065 | |
556adecb RR |
3066 | /* |
3067 | * Do some background aging of the anon list, to give | |
3068 | * pages a chance to be referenced before reclaiming. | |
3069 | */ | |
9e3b2f8c | 3070 | age_active_anon(zone, &sc); |
556adecb | 3071 | |
cc715d99 MG |
3072 | /* |
3073 | * If the number of buffer_heads in the machine | |
3074 | * exceeds the maximum allowed level and this node | |
3075 | * has a highmem zone, force kswapd to reclaim from | |
3076 | * it to relieve lowmem pressure. | |
3077 | */ | |
3078 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
3079 | end_zone = i; | |
3080 | break; | |
3081 | } | |
3082 | ||
60cefed4 | 3083 | if (!zone_balanced(zone, order, 0, 0)) { |
d6277db4 | 3084 | end_zone = i; |
e1dbeda6 | 3085 | break; |
439423f6 | 3086 | } else { |
d43006d5 MG |
3087 | /* |
3088 | * If balanced, clear the dirty and congested | |
3089 | * flags | |
3090 | */ | |
439423f6 | 3091 | zone_clear_flag(zone, ZONE_CONGESTED); |
d43006d5 | 3092 | zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY); |
1da177e4 | 3093 | } |
1da177e4 | 3094 | } |
dafcb73e | 3095 | |
b8e83b94 | 3096 | if (i < 0) |
e1dbeda6 AM |
3097 | goto out; |
3098 | ||
1da177e4 LT |
3099 | for (i = 0; i <= end_zone; i++) { |
3100 | struct zone *zone = pgdat->node_zones + i; | |
3101 | ||
2ab44f43 MG |
3102 | if (!populated_zone(zone)) |
3103 | continue; | |
3104 | ||
adea02a1 | 3105 | lru_pages += zone_reclaimable_pages(zone); |
2ab44f43 MG |
3106 | |
3107 | /* | |
3108 | * If any zone is currently balanced then kswapd will | |
3109 | * not call compaction as it is expected that the | |
3110 | * necessary pages are already available. | |
3111 | */ | |
3112 | if (pgdat_needs_compaction && | |
3113 | zone_watermark_ok(zone, order, | |
3114 | low_wmark_pages(zone), | |
3115 | *classzone_idx, 0)) | |
3116 | pgdat_needs_compaction = false; | |
1da177e4 LT |
3117 | } |
3118 | ||
b7ea3c41 MG |
3119 | /* |
3120 | * If we're getting trouble reclaiming, start doing writepage | |
3121 | * even in laptop mode. | |
3122 | */ | |
3123 | if (sc.priority < DEF_PRIORITY - 2) | |
3124 | sc.may_writepage = 1; | |
3125 | ||
1da177e4 LT |
3126 | /* |
3127 | * Now scan the zone in the dma->highmem direction, stopping | |
3128 | * at the last zone which needs scanning. | |
3129 | * | |
3130 | * We do this because the page allocator works in the opposite | |
3131 | * direction. This prevents the page allocator from allocating | |
3132 | * pages behind kswapd's direction of progress, which would | |
3133 | * cause too much scanning of the lower zones. | |
3134 | */ | |
3135 | for (i = 0; i <= end_zone; i++) { | |
3136 | struct zone *zone = pgdat->node_zones + i; | |
3137 | ||
f3fe6512 | 3138 | if (!populated_zone(zone)) |
1da177e4 LT |
3139 | continue; |
3140 | ||
6e543d57 LD |
3141 | if (sc.priority != DEF_PRIORITY && |
3142 | !zone_reclaimable(zone)) | |
1da177e4 LT |
3143 | continue; |
3144 | ||
1da177e4 | 3145 | sc.nr_scanned = 0; |
4e416953 | 3146 | |
0608f43d AM |
3147 | nr_soft_scanned = 0; |
3148 | /* | |
3149 | * Call soft limit reclaim before calling shrink_zone. | |
3150 | */ | |
3151 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
3152 | order, sc.gfp_mask, | |
3153 | &nr_soft_scanned); | |
3154 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3155 | ||
32a4330d | 3156 | /* |
7c954f6d MG |
3157 | * There should be no need to raise the scanning |
3158 | * priority if enough pages are already being scanned | |
3159 | * that that high watermark would be met at 100% | |
3160 | * efficiency. | |
fe2c2a10 | 3161 | */ |
7c954f6d MG |
3162 | if (kswapd_shrink_zone(zone, end_zone, &sc, |
3163 | lru_pages, &nr_attempted)) | |
3164 | raise_priority = false; | |
1da177e4 | 3165 | } |
5515061d MG |
3166 | |
3167 | /* | |
3168 | * If the low watermark is met there is no need for processes | |
3169 | * to be throttled on pfmemalloc_wait as they should not be | |
3170 | * able to safely make forward progress. Wake them | |
3171 | */ | |
3172 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
3173 | pfmemalloc_watermark_ok(pgdat)) | |
3174 | wake_up(&pgdat->pfmemalloc_wait); | |
3175 | ||
1da177e4 | 3176 | /* |
b8e83b94 MG |
3177 | * Fragmentation may mean that the system cannot be rebalanced |
3178 | * for high-order allocations in all zones. If twice the | |
3179 | * allocation size has been reclaimed and the zones are still | |
3180 | * not balanced then recheck the watermarks at order-0 to | |
3181 | * prevent kswapd reclaiming excessively. Assume that a | |
3182 | * process requested a high-order can direct reclaim/compact. | |
1da177e4 | 3183 | */ |
b8e83b94 MG |
3184 | if (order && sc.nr_reclaimed >= 2UL << order) |
3185 | order = sc.order = 0; | |
8357376d | 3186 | |
b8e83b94 MG |
3187 | /* Check if kswapd should be suspending */ |
3188 | if (try_to_freeze() || kthread_should_stop()) | |
3189 | break; | |
8357376d | 3190 | |
2ab44f43 MG |
3191 | /* |
3192 | * Compact if necessary and kswapd is reclaiming at least the | |
3193 | * high watermark number of pages as requsted | |
3194 | */ | |
3195 | if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted) | |
3196 | compact_pgdat(pgdat, order); | |
3197 | ||
73ce02e9 | 3198 | /* |
b8e83b94 MG |
3199 | * Raise priority if scanning rate is too low or there was no |
3200 | * progress in reclaiming pages | |
73ce02e9 | 3201 | */ |
b8e83b94 MG |
3202 | if (raise_priority || !sc.nr_reclaimed) |
3203 | sc.priority--; | |
9aa41348 | 3204 | } while (sc.priority >= 1 && |
b8e83b94 | 3205 | !pgdat_balanced(pgdat, order, *classzone_idx)); |
1da177e4 | 3206 | |
b8e83b94 | 3207 | out: |
0abdee2b | 3208 | /* |
5515061d | 3209 | * Return the order we were reclaiming at so prepare_kswapd_sleep() |
0abdee2b MG |
3210 | * makes a decision on the order we were last reclaiming at. However, |
3211 | * if another caller entered the allocator slow path while kswapd | |
3212 | * was awake, order will remain at the higher level | |
3213 | */ | |
dc83edd9 | 3214 | *classzone_idx = end_zone; |
0abdee2b | 3215 | return order; |
1da177e4 LT |
3216 | } |
3217 | ||
dc83edd9 | 3218 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f0bc0a60 KM |
3219 | { |
3220 | long remaining = 0; | |
3221 | DEFINE_WAIT(wait); | |
3222 | ||
3223 | if (freezing(current) || kthread_should_stop()) | |
3224 | return; | |
3225 | ||
3226 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3227 | ||
3228 | /* Try to sleep for a short interval */ | |
5515061d | 3229 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
3230 | remaining = schedule_timeout(HZ/10); |
3231 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3232 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3233 | } | |
3234 | ||
3235 | /* | |
3236 | * After a short sleep, check if it was a premature sleep. If not, then | |
3237 | * go fully to sleep until explicitly woken up. | |
3238 | */ | |
5515061d | 3239 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
3240 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3241 | ||
3242 | /* | |
3243 | * vmstat counters are not perfectly accurate and the estimated | |
3244 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3245 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3246 | * watermarks being breached while under pressure, we reduce the | |
3247 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3248 | * them before going back to sleep. | |
3249 | */ | |
3250 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c | 3251 | |
62997027 MG |
3252 | /* |
3253 | * Compaction records what page blocks it recently failed to | |
3254 | * isolate pages from and skips them in the future scanning. | |
3255 | * When kswapd is going to sleep, it is reasonable to assume | |
3256 | * that pages and compaction may succeed so reset the cache. | |
3257 | */ | |
3258 | reset_isolation_suitable(pgdat); | |
3259 | ||
1c7e7f6c AK |
3260 | if (!kthread_should_stop()) |
3261 | schedule(); | |
3262 | ||
f0bc0a60 KM |
3263 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3264 | } else { | |
3265 | if (remaining) | |
3266 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3267 | else | |
3268 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3269 | } | |
3270 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3271 | } | |
3272 | ||
1da177e4 LT |
3273 | /* |
3274 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3275 | * from the init process. |
1da177e4 LT |
3276 | * |
3277 | * This basically trickles out pages so that we have _some_ | |
3278 | * free memory available even if there is no other activity | |
3279 | * that frees anything up. This is needed for things like routing | |
3280 | * etc, where we otherwise might have all activity going on in | |
3281 | * asynchronous contexts that cannot page things out. | |
3282 | * | |
3283 | * If there are applications that are active memory-allocators | |
3284 | * (most normal use), this basically shouldn't matter. | |
3285 | */ | |
3286 | static int kswapd(void *p) | |
3287 | { | |
215ddd66 | 3288 | unsigned long order, new_order; |
d2ebd0f6 | 3289 | unsigned balanced_order; |
215ddd66 | 3290 | int classzone_idx, new_classzone_idx; |
d2ebd0f6 | 3291 | int balanced_classzone_idx; |
1da177e4 LT |
3292 | pg_data_t *pgdat = (pg_data_t*)p; |
3293 | struct task_struct *tsk = current; | |
f0bc0a60 | 3294 | |
1da177e4 LT |
3295 | struct reclaim_state reclaim_state = { |
3296 | .reclaimed_slab = 0, | |
3297 | }; | |
a70f7302 | 3298 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3299 | |
cf40bd16 NP |
3300 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3301 | ||
174596a0 | 3302 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3303 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3304 | current->reclaim_state = &reclaim_state; |
3305 | ||
3306 | /* | |
3307 | * Tell the memory management that we're a "memory allocator", | |
3308 | * and that if we need more memory we should get access to it | |
3309 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3310 | * never get caught in the normal page freeing logic. | |
3311 | * | |
3312 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3313 | * you need a small amount of memory in order to be able to | |
3314 | * page out something else, and this flag essentially protects | |
3315 | * us from recursively trying to free more memory as we're | |
3316 | * trying to free the first piece of memory in the first place). | |
3317 | */ | |
930d9152 | 3318 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3319 | set_freezable(); |
1da177e4 | 3320 | |
215ddd66 | 3321 | order = new_order = 0; |
d2ebd0f6 | 3322 | balanced_order = 0; |
215ddd66 | 3323 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; |
d2ebd0f6 | 3324 | balanced_classzone_idx = classzone_idx; |
1da177e4 | 3325 | for ( ; ; ) { |
6f6313d4 | 3326 | bool ret; |
3e1d1d28 | 3327 | |
215ddd66 MG |
3328 | /* |
3329 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
3330 | * new request of a similar or harder type will succeed soon | |
3331 | * so consider going to sleep on the basis we reclaimed at | |
3332 | */ | |
d2ebd0f6 AS |
3333 | if (balanced_classzone_idx >= new_classzone_idx && |
3334 | balanced_order == new_order) { | |
215ddd66 MG |
3335 | new_order = pgdat->kswapd_max_order; |
3336 | new_classzone_idx = pgdat->classzone_idx; | |
3337 | pgdat->kswapd_max_order = 0; | |
3338 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3339 | } | |
3340 | ||
99504748 | 3341 | if (order < new_order || classzone_idx > new_classzone_idx) { |
1da177e4 LT |
3342 | /* |
3343 | * Don't sleep if someone wants a larger 'order' | |
99504748 | 3344 | * allocation or has tigher zone constraints |
1da177e4 LT |
3345 | */ |
3346 | order = new_order; | |
99504748 | 3347 | classzone_idx = new_classzone_idx; |
1da177e4 | 3348 | } else { |
d2ebd0f6 AS |
3349 | kswapd_try_to_sleep(pgdat, balanced_order, |
3350 | balanced_classzone_idx); | |
1da177e4 | 3351 | order = pgdat->kswapd_max_order; |
99504748 | 3352 | classzone_idx = pgdat->classzone_idx; |
f0dfcde0 AS |
3353 | new_order = order; |
3354 | new_classzone_idx = classzone_idx; | |
4d40502e | 3355 | pgdat->kswapd_max_order = 0; |
215ddd66 | 3356 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
1da177e4 | 3357 | } |
1da177e4 | 3358 | |
8fe23e05 DR |
3359 | ret = try_to_freeze(); |
3360 | if (kthread_should_stop()) | |
3361 | break; | |
3362 | ||
3363 | /* | |
3364 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3365 | * after returning from the refrigerator | |
3366 | */ | |
33906bc5 MG |
3367 | if (!ret) { |
3368 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
d2ebd0f6 AS |
3369 | balanced_classzone_idx = classzone_idx; |
3370 | balanced_order = balance_pgdat(pgdat, order, | |
3371 | &balanced_classzone_idx); | |
33906bc5 | 3372 | } |
1da177e4 | 3373 | } |
b0a8cc58 TY |
3374 | |
3375 | current->reclaim_state = NULL; | |
1da177e4 LT |
3376 | return 0; |
3377 | } | |
3378 | ||
3379 | /* | |
3380 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3381 | */ | |
99504748 | 3382 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3383 | { |
3384 | pg_data_t *pgdat; | |
3385 | ||
f3fe6512 | 3386 | if (!populated_zone(zone)) |
1da177e4 LT |
3387 | return; |
3388 | ||
88f5acf8 | 3389 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1da177e4 | 3390 | return; |
88f5acf8 | 3391 | pgdat = zone->zone_pgdat; |
99504748 | 3392 | if (pgdat->kswapd_max_order < order) { |
1da177e4 | 3393 | pgdat->kswapd_max_order = order; |
99504748 MG |
3394 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
3395 | } | |
8d0986e2 | 3396 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3397 | return; |
892f795d | 3398 | if (zone_balanced(zone, order, 0, 0)) |
88f5acf8 MG |
3399 | return; |
3400 | ||
3401 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3402 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3403 | } |
3404 | ||
c6f37f12 | 3405 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3406 | /* |
7b51755c | 3407 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3408 | * freed pages. |
3409 | * | |
3410 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3411 | * LRU order by reclaiming preferentially | |
3412 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3413 | */ |
7b51755c | 3414 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3415 | { |
d6277db4 | 3416 | struct reclaim_state reclaim_state; |
d6277db4 | 3417 | struct scan_control sc = { |
7b51755c KM |
3418 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
3419 | .may_swap = 1, | |
3420 | .may_unmap = 1, | |
d6277db4 | 3421 | .may_writepage = 1, |
7b51755c KM |
3422 | .nr_to_reclaim = nr_to_reclaim, |
3423 | .hibernation_mode = 1, | |
7b51755c | 3424 | .order = 0, |
9e3b2f8c | 3425 | .priority = DEF_PRIORITY, |
1da177e4 | 3426 | }; |
a09ed5e0 | 3427 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c KM |
3428 | struct task_struct *p = current; |
3429 | unsigned long nr_reclaimed; | |
1da177e4 | 3430 | |
7b51755c KM |
3431 | p->flags |= PF_MEMALLOC; |
3432 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3433 | reclaim_state.reclaimed_slab = 0; | |
3434 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3435 | |
3115cd91 | 3436 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3437 | |
7b51755c KM |
3438 | p->reclaim_state = NULL; |
3439 | lockdep_clear_current_reclaim_state(); | |
3440 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3441 | |
7b51755c | 3442 | return nr_reclaimed; |
1da177e4 | 3443 | } |
c6f37f12 | 3444 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3445 | |
1da177e4 LT |
3446 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3447 | not required for correctness. So if the last cpu in a node goes | |
3448 | away, we get changed to run anywhere: as the first one comes back, | |
3449 | restore their cpu bindings. */ | |
fcb35a9b GKH |
3450 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
3451 | void *hcpu) | |
1da177e4 | 3452 | { |
58c0a4a7 | 3453 | int nid; |
1da177e4 | 3454 | |
8bb78442 | 3455 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
48fb2e24 | 3456 | for_each_node_state(nid, N_MEMORY) { |
c5f59f08 | 3457 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3458 | const struct cpumask *mask; |
3459 | ||
3460 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3461 | |
3e597945 | 3462 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3463 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3464 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3465 | } |
3466 | } | |
3467 | return NOTIFY_OK; | |
3468 | } | |
1da177e4 | 3469 | |
3218ae14 YG |
3470 | /* |
3471 | * This kswapd start function will be called by init and node-hot-add. | |
3472 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3473 | */ | |
3474 | int kswapd_run(int nid) | |
3475 | { | |
3476 | pg_data_t *pgdat = NODE_DATA(nid); | |
3477 | int ret = 0; | |
3478 | ||
3479 | if (pgdat->kswapd) | |
3480 | return 0; | |
3481 | ||
3482 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3483 | if (IS_ERR(pgdat->kswapd)) { | |
3484 | /* failure at boot is fatal */ | |
3485 | BUG_ON(system_state == SYSTEM_BOOTING); | |
d5dc0ad9 GS |
3486 | pr_err("Failed to start kswapd on node %d\n", nid); |
3487 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3488 | pgdat->kswapd = NULL; |
3218ae14 YG |
3489 | } |
3490 | return ret; | |
3491 | } | |
3492 | ||
8fe23e05 | 3493 | /* |
d8adde17 | 3494 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 3495 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
3496 | */ |
3497 | void kswapd_stop(int nid) | |
3498 | { | |
3499 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3500 | ||
d8adde17 | 3501 | if (kswapd) { |
8fe23e05 | 3502 | kthread_stop(kswapd); |
d8adde17 JL |
3503 | NODE_DATA(nid)->kswapd = NULL; |
3504 | } | |
8fe23e05 DR |
3505 | } |
3506 | ||
1da177e4 LT |
3507 | static int __init kswapd_init(void) |
3508 | { | |
3218ae14 | 3509 | int nid; |
69e05944 | 3510 | |
1da177e4 | 3511 | swap_setup(); |
48fb2e24 | 3512 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3513 | kswapd_run(nid); |
1da177e4 LT |
3514 | hotcpu_notifier(cpu_callback, 0); |
3515 | return 0; | |
3516 | } | |
3517 | ||
3518 | module_init(kswapd_init) | |
9eeff239 CL |
3519 | |
3520 | #ifdef CONFIG_NUMA | |
3521 | /* | |
3522 | * Zone reclaim mode | |
3523 | * | |
3524 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3525 | * the watermarks. | |
9eeff239 CL |
3526 | */ |
3527 | int zone_reclaim_mode __read_mostly; | |
3528 | ||
1b2ffb78 | 3529 | #define RECLAIM_OFF 0 |
7d03431c | 3530 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 CL |
3531 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
3532 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3533 | ||
a92f7126 CL |
3534 | /* |
3535 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3536 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3537 | * a zone. | |
3538 | */ | |
3539 | #define ZONE_RECLAIM_PRIORITY 4 | |
3540 | ||
9614634f CL |
3541 | /* |
3542 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3543 | * occur. | |
3544 | */ | |
3545 | int sysctl_min_unmapped_ratio = 1; | |
3546 | ||
0ff38490 CL |
3547 | /* |
3548 | * If the number of slab pages in a zone grows beyond this percentage then | |
3549 | * slab reclaim needs to occur. | |
3550 | */ | |
3551 | int sysctl_min_slab_ratio = 5; | |
3552 | ||
90afa5de MG |
3553 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
3554 | { | |
3555 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3556 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3557 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3558 | ||
3559 | /* | |
3560 | * It's possible for there to be more file mapped pages than | |
3561 | * accounted for by the pages on the file LRU lists because | |
3562 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3563 | */ | |
3564 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3565 | } | |
3566 | ||
3567 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3568 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3569 | { | |
3570 | long nr_pagecache_reclaimable; | |
3571 | long delta = 0; | |
3572 | ||
3573 | /* | |
3574 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3575 | * potentially reclaimable. Otherwise, we have to worry about | |
3576 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3577 | * a better estimate | |
3578 | */ | |
3579 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3580 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3581 | else | |
3582 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3583 | ||
3584 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3585 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3586 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3587 | ||
3588 | /* Watch for any possible underflows due to delta */ | |
3589 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3590 | delta = nr_pagecache_reclaimable; | |
3591 | ||
3592 | return nr_pagecache_reclaimable - delta; | |
3593 | } | |
3594 | ||
9eeff239 CL |
3595 | /* |
3596 | * Try to free up some pages from this zone through reclaim. | |
3597 | */ | |
179e9639 | 3598 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3599 | { |
7fb2d46d | 3600 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3601 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3602 | struct task_struct *p = current; |
3603 | struct reclaim_state reclaim_state; | |
179e9639 AM |
3604 | struct scan_control sc = { |
3605 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
a6dc60f8 | 3606 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), |
2e2e4259 | 3607 | .may_swap = 1, |
62b726c1 | 3608 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
21caf2fc | 3609 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
bd2f6199 | 3610 | .order = order, |
9e3b2f8c | 3611 | .priority = ZONE_RECLAIM_PRIORITY, |
179e9639 | 3612 | }; |
a09ed5e0 YH |
3613 | struct shrink_control shrink = { |
3614 | .gfp_mask = sc.gfp_mask, | |
3615 | }; | |
15748048 | 3616 | unsigned long nr_slab_pages0, nr_slab_pages1; |
9eeff239 | 3617 | |
9eeff239 | 3618 | cond_resched(); |
d4f7796e CL |
3619 | /* |
3620 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3621 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3622 | * and RECLAIM_SWAP. | |
3623 | */ | |
3624 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3625 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3626 | reclaim_state.reclaimed_slab = 0; |
3627 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3628 | |
90afa5de | 3629 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
3630 | /* |
3631 | * Free memory by calling shrink zone with increasing | |
3632 | * priorities until we have enough memory freed. | |
3633 | */ | |
0ff38490 | 3634 | do { |
9e3b2f8c KK |
3635 | shrink_zone(zone, &sc); |
3636 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); | |
0ff38490 | 3637 | } |
c84db23c | 3638 | |
15748048 KM |
3639 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3640 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
2a16e3f4 | 3641 | /* |
7fb2d46d | 3642 | * shrink_slab() does not currently allow us to determine how |
0ff38490 CL |
3643 | * many pages were freed in this zone. So we take the current |
3644 | * number of slab pages and shake the slab until it is reduced | |
3645 | * by the same nr_pages that we used for reclaiming unmapped | |
3646 | * pages. | |
2a16e3f4 | 3647 | */ |
0ce3d744 DC |
3648 | nodes_clear(shrink.nodes_to_scan); |
3649 | node_set(zone_to_nid(zone), shrink.nodes_to_scan); | |
4dc4b3d9 KM |
3650 | for (;;) { |
3651 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3652 | ||
3653 | /* No reclaimable slab or very low memory pressure */ | |
1495f230 | 3654 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) |
4dc4b3d9 KM |
3655 | break; |
3656 | ||
3657 | /* Freed enough memory */ | |
3658 | nr_slab_pages1 = zone_page_state(zone, | |
3659 | NR_SLAB_RECLAIMABLE); | |
3660 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3661 | break; | |
3662 | } | |
83e33a47 CL |
3663 | |
3664 | /* | |
3665 | * Update nr_reclaimed by the number of slab pages we | |
3666 | * reclaimed from this zone. | |
3667 | */ | |
15748048 KM |
3668 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3669 | if (nr_slab_pages1 < nr_slab_pages0) | |
3670 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
2a16e3f4 CL |
3671 | } |
3672 | ||
9eeff239 | 3673 | p->reclaim_state = NULL; |
d4f7796e | 3674 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3675 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3676 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3677 | } |
179e9639 AM |
3678 | |
3679 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3680 | { | |
179e9639 | 3681 | int node_id; |
d773ed6b | 3682 | int ret; |
179e9639 AM |
3683 | |
3684 | /* | |
0ff38490 CL |
3685 | * Zone reclaim reclaims unmapped file backed pages and |
3686 | * slab pages if we are over the defined limits. | |
34aa1330 | 3687 | * |
9614634f CL |
3688 | * A small portion of unmapped file backed pages is needed for |
3689 | * file I/O otherwise pages read by file I/O will be immediately | |
3690 | * thrown out if the zone is overallocated. So we do not reclaim | |
3691 | * if less than a specified percentage of the zone is used by | |
3692 | * unmapped file backed pages. | |
179e9639 | 3693 | */ |
90afa5de MG |
3694 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
3695 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 3696 | return ZONE_RECLAIM_FULL; |
179e9639 | 3697 | |
6e543d57 | 3698 | if (!zone_reclaimable(zone)) |
fa5e084e | 3699 | return ZONE_RECLAIM_FULL; |
d773ed6b | 3700 | |
179e9639 | 3701 | /* |
d773ed6b | 3702 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3703 | */ |
d773ed6b | 3704 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 3705 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
3706 | |
3707 | /* | |
3708 | * Only run zone reclaim on the local zone or on zones that do not | |
3709 | * have associated processors. This will favor the local processor | |
3710 | * over remote processors and spread off node memory allocations | |
3711 | * as wide as possible. | |
3712 | */ | |
89fa3024 | 3713 | node_id = zone_to_nid(zone); |
37c0708d | 3714 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 3715 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b DR |
3716 | |
3717 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
fa5e084e MG |
3718 | return ZONE_RECLAIM_NOSCAN; |
3719 | ||
d773ed6b DR |
3720 | ret = __zone_reclaim(zone, gfp_mask, order); |
3721 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3722 | ||
24cf7251 MG |
3723 | if (!ret) |
3724 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3725 | ||
d773ed6b | 3726 | return ret; |
179e9639 | 3727 | } |
9eeff239 | 3728 | #endif |
894bc310 | 3729 | |
894bc310 LS |
3730 | /* |
3731 | * page_evictable - test whether a page is evictable | |
3732 | * @page: the page to test | |
894bc310 LS |
3733 | * |
3734 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3735 | * lists vs unevictable list. |
894bc310 LS |
3736 | * |
3737 | * Reasons page might not be evictable: | |
ba9ddf49 | 3738 | * (1) page's mapping marked unevictable |
b291f000 | 3739 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3740 | * |
894bc310 | 3741 | */ |
39b5f29a | 3742 | int page_evictable(struct page *page) |
894bc310 | 3743 | { |
39b5f29a | 3744 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
894bc310 | 3745 | } |
89e004ea | 3746 | |
85046579 | 3747 | #ifdef CONFIG_SHMEM |
89e004ea | 3748 | /** |
24513264 HD |
3749 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3750 | * @pages: array of pages to check | |
3751 | * @nr_pages: number of pages to check | |
89e004ea | 3752 | * |
24513264 | 3753 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3754 | * |
3755 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3756 | */ |
24513264 | 3757 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3758 | { |
925b7673 | 3759 | struct lruvec *lruvec; |
24513264 HD |
3760 | struct zone *zone = NULL; |
3761 | int pgscanned = 0; | |
3762 | int pgrescued = 0; | |
3763 | int i; | |
89e004ea | 3764 | |
24513264 HD |
3765 | for (i = 0; i < nr_pages; i++) { |
3766 | struct page *page = pages[i]; | |
3767 | struct zone *pagezone; | |
89e004ea | 3768 | |
24513264 HD |
3769 | pgscanned++; |
3770 | pagezone = page_zone(page); | |
3771 | if (pagezone != zone) { | |
3772 | if (zone) | |
3773 | spin_unlock_irq(&zone->lru_lock); | |
3774 | zone = pagezone; | |
3775 | spin_lock_irq(&zone->lru_lock); | |
3776 | } | |
fa9add64 | 3777 | lruvec = mem_cgroup_page_lruvec(page, zone); |
89e004ea | 3778 | |
24513264 HD |
3779 | if (!PageLRU(page) || !PageUnevictable(page)) |
3780 | continue; | |
89e004ea | 3781 | |
39b5f29a | 3782 | if (page_evictable(page)) { |
24513264 HD |
3783 | enum lru_list lru = page_lru_base_type(page); |
3784 | ||
309381fe | 3785 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 3786 | ClearPageUnevictable(page); |
fa9add64 HD |
3787 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
3788 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 3789 | pgrescued++; |
89e004ea | 3790 | } |
24513264 | 3791 | } |
89e004ea | 3792 | |
24513264 HD |
3793 | if (zone) { |
3794 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3795 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3796 | spin_unlock_irq(&zone->lru_lock); | |
89e004ea | 3797 | } |
89e004ea | 3798 | } |
85046579 | 3799 | #endif /* CONFIG_SHMEM */ |
af936a16 | 3800 | |
264e56d8 | 3801 | static void warn_scan_unevictable_pages(void) |
af936a16 | 3802 | { |
264e56d8 | 3803 | printk_once(KERN_WARNING |
25bd91bd | 3804 | "%s: The scan_unevictable_pages sysctl/node-interface has been " |
264e56d8 | 3805 | "disabled for lack of a legitimate use case. If you have " |
25bd91bd KM |
3806 | "one, please send an email to linux-mm@kvack.org.\n", |
3807 | current->comm); | |
af936a16 LS |
3808 | } |
3809 | ||
3810 | /* | |
3811 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3812 | * all nodes' unevictable lists for evictable pages | |
3813 | */ | |
3814 | unsigned long scan_unevictable_pages; | |
3815 | ||
3816 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
8d65af78 | 3817 | void __user *buffer, |
af936a16 LS |
3818 | size_t *length, loff_t *ppos) |
3819 | { | |
264e56d8 | 3820 | warn_scan_unevictable_pages(); |
8d65af78 | 3821 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
af936a16 LS |
3822 | scan_unevictable_pages = 0; |
3823 | return 0; | |
3824 | } | |
3825 | ||
e4455abb | 3826 | #ifdef CONFIG_NUMA |
af936a16 LS |
3827 | /* |
3828 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3829 | * a specified node's per zone unevictable lists for evictable pages. | |
3830 | */ | |
3831 | ||
10fbcf4c KS |
3832 | static ssize_t read_scan_unevictable_node(struct device *dev, |
3833 | struct device_attribute *attr, | |
af936a16 LS |
3834 | char *buf) |
3835 | { | |
264e56d8 | 3836 | warn_scan_unevictable_pages(); |
af936a16 LS |
3837 | return sprintf(buf, "0\n"); /* always zero; should fit... */ |
3838 | } | |
3839 | ||
10fbcf4c KS |
3840 | static ssize_t write_scan_unevictable_node(struct device *dev, |
3841 | struct device_attribute *attr, | |
af936a16 LS |
3842 | const char *buf, size_t count) |
3843 | { | |
264e56d8 | 3844 | warn_scan_unevictable_pages(); |
af936a16 LS |
3845 | return 1; |
3846 | } | |
3847 | ||
3848 | ||
10fbcf4c | 3849 | static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, |
af936a16 LS |
3850 | read_scan_unevictable_node, |
3851 | write_scan_unevictable_node); | |
3852 | ||
3853 | int scan_unevictable_register_node(struct node *node) | |
3854 | { | |
10fbcf4c | 3855 | return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 LS |
3856 | } |
3857 | ||
3858 | void scan_unevictable_unregister_node(struct node *node) | |
3859 | { | |
10fbcf4c | 3860 | device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 | 3861 | } |
e4455abb | 3862 | #endif |