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> |
6fa3eb70 | 46 | #include <linux/debugfs.h> |
1da177e4 LT |
47 | |
48 | #include <asm/tlbflush.h> | |
49 | #include <asm/div64.h> | |
50 | ||
51 | #include <linux/swapops.h> | |
009dfd44 | 52 | #include <linux/balloon_compaction.h> |
1da177e4 | 53 | |
0f8053a5 NP |
54 | #include "internal.h" |
55 | ||
33906bc5 MG |
56 | #define CREATE_TRACE_POINTS |
57 | #include <trace/events/vmscan.h> | |
58 | ||
1da177e4 | 59 | struct scan_control { |
1da177e4 LT |
60 | /* Incremented by the number of inactive pages that were scanned */ |
61 | unsigned long nr_scanned; | |
62 | ||
a79311c1 RR |
63 | /* Number of pages freed so far during a call to shrink_zones() */ |
64 | unsigned long nr_reclaimed; | |
65 | ||
22fba335 KM |
66 | /* How many pages shrink_list() should reclaim */ |
67 | unsigned long nr_to_reclaim; | |
68 | ||
7b51755c KM |
69 | unsigned long hibernation_mode; |
70 | ||
1da177e4 | 71 | /* This context's GFP mask */ |
6daa0e28 | 72 | gfp_t gfp_mask; |
1da177e4 LT |
73 | |
74 | int may_writepage; | |
75 | ||
a6dc60f8 JW |
76 | /* Can mapped pages be reclaimed? */ |
77 | int may_unmap; | |
f1fd1067 | 78 | |
2e2e4259 KM |
79 | /* Can pages be swapped as part of reclaim? */ |
80 | int may_swap; | |
81 | ||
5ad333eb | 82 | int order; |
66e1707b | 83 | |
9e3b2f8c KK |
84 | /* Scan (total_size >> priority) pages at once */ |
85 | int priority; | |
86 | ||
f16015fb JW |
87 | /* |
88 | * The memory cgroup that hit its limit and as a result is the | |
89 | * primary target of this reclaim invocation. | |
90 | */ | |
91 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 92 | |
327c0e96 KH |
93 | /* |
94 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
95 | * are scanned. | |
96 | */ | |
97 | nodemask_t *nodemask; | |
1da177e4 LT |
98 | }; |
99 | ||
1da177e4 LT |
100 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
101 | ||
102 | #ifdef ARCH_HAS_PREFETCH | |
103 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
104 | do { \ | |
105 | if ((_page)->lru.prev != _base) { \ | |
106 | struct page *prev; \ | |
107 | \ | |
108 | prev = lru_to_page(&(_page->lru)); \ | |
109 | prefetch(&prev->_field); \ | |
110 | } \ | |
111 | } while (0) | |
112 | #else | |
113 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
114 | #endif | |
115 | ||
116 | #ifdef ARCH_HAS_PREFETCHW | |
117 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
118 | do { \ | |
119 | if ((_page)->lru.prev != _base) { \ | |
120 | struct page *prev; \ | |
121 | \ | |
122 | prev = lru_to_page(&(_page->lru)); \ | |
123 | prefetchw(&prev->_field); \ | |
124 | } \ | |
125 | } while (0) | |
126 | #else | |
127 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
128 | #endif | |
129 | ||
130 | /* | |
131 | * From 0 .. 100. Higher means more swappy. | |
132 | */ | |
133 | int vm_swappiness = 60; | |
b21e0b90 | 134 | unsigned long vm_total_pages; /* The total number of pages which the VM controls */ |
1da177e4 LT |
135 | |
136 | static LIST_HEAD(shrinker_list); | |
137 | static DECLARE_RWSEM(shrinker_rwsem); | |
138 | ||
c255a458 | 139 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
140 | static bool global_reclaim(struct scan_control *sc) |
141 | { | |
f16015fb | 142 | return !sc->target_mem_cgroup; |
89b5fae5 | 143 | } |
91a45470 | 144 | #else |
89b5fae5 JW |
145 | static bool global_reclaim(struct scan_control *sc) |
146 | { | |
147 | return true; | |
148 | } | |
91a45470 KH |
149 | #endif |
150 | ||
4d7dcca2 | 151 | static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) |
c9f299d9 | 152 | { |
c3c787e8 | 153 | if (!mem_cgroup_disabled()) |
4d7dcca2 | 154 | return mem_cgroup_get_lru_size(lruvec, lru); |
a3d8e054 | 155 | |
074291fe | 156 | return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); |
c9f299d9 KM |
157 | } |
158 | ||
6fa3eb70 S |
159 | struct dentry *debug_file; |
160 | ||
161 | static int debug_shrinker_show(struct seq_file *s, void *unused) | |
162 | { | |
163 | struct shrinker *shrinker; | |
164 | struct shrink_control sc; | |
165 | ||
166 | sc.gfp_mask = -1; | |
167 | sc.nr_to_scan = 0; | |
168 | ||
169 | down_read(&shrinker_rwsem); | |
170 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
171 | int num_objs; | |
172 | ||
173 | num_objs = shrinker->shrink(shrinker, &sc); | |
174 | seq_printf(s, "%pf %d\n", shrinker->shrink, num_objs); | |
175 | } | |
176 | up_read(&shrinker_rwsem); | |
177 | return 0; | |
178 | } | |
179 | ||
180 | static int debug_shrinker_open(struct inode *inode, struct file *file) | |
181 | { | |
182 | return single_open(file, debug_shrinker_show, inode->i_private); | |
183 | } | |
184 | ||
185 | static const struct file_operations debug_shrinker_fops = { | |
186 | .open = debug_shrinker_open, | |
187 | .read = seq_read, | |
188 | .llseek = seq_lseek, | |
189 | .release = single_release, | |
190 | }; | |
191 | ||
1da177e4 LT |
192 | /* |
193 | * Add a shrinker callback to be called from the vm | |
194 | */ | |
8e1f936b | 195 | void register_shrinker(struct shrinker *shrinker) |
1da177e4 | 196 | { |
83aeeada | 197 | atomic_long_set(&shrinker->nr_in_batch, 0); |
8e1f936b RR |
198 | down_write(&shrinker_rwsem); |
199 | list_add_tail(&shrinker->list, &shrinker_list); | |
200 | up_write(&shrinker_rwsem); | |
1da177e4 | 201 | } |
8e1f936b | 202 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 | 203 | |
6fa3eb70 S |
204 | static int __init add_shrinker_debug(void) |
205 | { | |
206 | debugfs_create_file("shrinker", 0644, NULL, NULL, | |
207 | &debug_shrinker_fops); | |
208 | return 0; | |
209 | } | |
210 | ||
211 | late_initcall(add_shrinker_debug); | |
212 | ||
1da177e4 LT |
213 | /* |
214 | * Remove one | |
215 | */ | |
8e1f936b | 216 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 LT |
217 | { |
218 | down_write(&shrinker_rwsem); | |
219 | list_del(&shrinker->list); | |
220 | up_write(&shrinker_rwsem); | |
1da177e4 | 221 | } |
8e1f936b | 222 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 | 223 | |
1495f230 YH |
224 | static inline int do_shrinker_shrink(struct shrinker *shrinker, |
225 | struct shrink_control *sc, | |
226 | unsigned long nr_to_scan) | |
227 | { | |
228 | sc->nr_to_scan = nr_to_scan; | |
229 | return (*shrinker->shrink)(shrinker, sc); | |
230 | } | |
231 | ||
1da177e4 LT |
232 | #define SHRINK_BATCH 128 |
233 | /* | |
234 | * Call the shrink functions to age shrinkable caches | |
235 | * | |
236 | * Here we assume it costs one seek to replace a lru page and that it also | |
237 | * takes a seek to recreate a cache object. With this in mind we age equal | |
238 | * percentages of the lru and ageable caches. This should balance the seeks | |
239 | * generated by these structures. | |
240 | * | |
183ff22b | 241 | * If the vm encountered mapped pages on the LRU it increase the pressure on |
1da177e4 LT |
242 | * slab to avoid swapping. |
243 | * | |
244 | * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. | |
245 | * | |
246 | * `lru_pages' represents the number of on-LRU pages in all the zones which | |
247 | * are eligible for the caller's allocation attempt. It is used for balancing | |
248 | * slab reclaim versus page reclaim. | |
b15e0905 AM |
249 | * |
250 | * Returns the number of slab objects which we shrunk. | |
1da177e4 | 251 | */ |
a09ed5e0 | 252 | unsigned long shrink_slab(struct shrink_control *shrink, |
1495f230 | 253 | unsigned long nr_pages_scanned, |
a09ed5e0 | 254 | unsigned long lru_pages) |
1da177e4 LT |
255 | { |
256 | struct shrinker *shrinker; | |
69e05944 | 257 | unsigned long ret = 0; |
1da177e4 | 258 | |
1495f230 YH |
259 | if (nr_pages_scanned == 0) |
260 | nr_pages_scanned = SWAP_CLUSTER_MAX; | |
1da177e4 | 261 | |
f06590bd MK |
262 | if (!down_read_trylock(&shrinker_rwsem)) { |
263 | /* Assume we'll be able to shrink next time */ | |
264 | ret = 1; | |
265 | goto out; | |
266 | } | |
1da177e4 LT |
267 | |
268 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
269 | unsigned long long delta; | |
635697c6 KK |
270 | long total_scan; |
271 | long max_pass; | |
09576073 | 272 | int shrink_ret = 0; |
acf92b48 DC |
273 | long nr; |
274 | long new_nr; | |
e9299f50 DC |
275 | long batch_size = shrinker->batch ? shrinker->batch |
276 | : SHRINK_BATCH; | |
1da177e4 | 277 | |
635697c6 KK |
278 | max_pass = do_shrinker_shrink(shrinker, shrink, 0); |
279 | if (max_pass <= 0) | |
280 | continue; | |
281 | ||
acf92b48 DC |
282 | /* |
283 | * copy the current shrinker scan count into a local variable | |
284 | * and zero it so that other concurrent shrinker invocations | |
285 | * don't also do this scanning work. | |
286 | */ | |
83aeeada | 287 | nr = atomic_long_xchg(&shrinker->nr_in_batch, 0); |
acf92b48 DC |
288 | |
289 | total_scan = nr; | |
1495f230 | 290 | delta = (4 * nr_pages_scanned) / shrinker->seeks; |
ea164d73 | 291 | delta *= max_pass; |
1da177e4 | 292 | do_div(delta, lru_pages + 1); |
acf92b48 DC |
293 | total_scan += delta; |
294 | if (total_scan < 0) { | |
88c3bd70 DR |
295 | printk(KERN_ERR "shrink_slab: %pF negative objects to " |
296 | "delete nr=%ld\n", | |
acf92b48 DC |
297 | shrinker->shrink, total_scan); |
298 | total_scan = max_pass; | |
ea164d73 AA |
299 | } |
300 | ||
3567b59a DC |
301 | /* |
302 | * We need to avoid excessive windup on filesystem shrinkers | |
303 | * due to large numbers of GFP_NOFS allocations causing the | |
304 | * shrinkers to return -1 all the time. This results in a large | |
305 | * nr being built up so when a shrink that can do some work | |
306 | * comes along it empties the entire cache due to nr >>> | |
307 | * max_pass. This is bad for sustaining a working set in | |
308 | * memory. | |
309 | * | |
310 | * Hence only allow the shrinker to scan the entire cache when | |
311 | * a large delta change is calculated directly. | |
312 | */ | |
313 | if (delta < max_pass / 4) | |
314 | total_scan = min(total_scan, max_pass / 2); | |
315 | ||
ea164d73 AA |
316 | /* |
317 | * Avoid risking looping forever due to too large nr value: | |
318 | * never try to free more than twice the estimate number of | |
319 | * freeable entries. | |
320 | */ | |
acf92b48 DC |
321 | if (total_scan > max_pass * 2) |
322 | total_scan = max_pass * 2; | |
1da177e4 | 323 | |
acf92b48 | 324 | trace_mm_shrink_slab_start(shrinker, shrink, nr, |
09576073 DC |
325 | nr_pages_scanned, lru_pages, |
326 | max_pass, delta, total_scan); | |
327 | ||
e9299f50 | 328 | while (total_scan >= batch_size) { |
b15e0905 | 329 | int nr_before; |
1da177e4 | 330 | |
1495f230 YH |
331 | nr_before = do_shrinker_shrink(shrinker, shrink, 0); |
332 | shrink_ret = do_shrinker_shrink(shrinker, shrink, | |
e9299f50 | 333 | batch_size); |
1da177e4 LT |
334 | if (shrink_ret == -1) |
335 | break; | |
b15e0905 AM |
336 | if (shrink_ret < nr_before) |
337 | ret += nr_before - shrink_ret; | |
e9299f50 DC |
338 | count_vm_events(SLABS_SCANNED, batch_size); |
339 | total_scan -= batch_size; | |
1da177e4 LT |
340 | |
341 | cond_resched(); | |
342 | } | |
343 | ||
acf92b48 DC |
344 | /* |
345 | * move the unused scan count back into the shrinker in a | |
346 | * manner that handles concurrent updates. If we exhausted the | |
347 | * scan, there is no need to do an update. | |
348 | */ | |
83aeeada KK |
349 | if (total_scan > 0) |
350 | new_nr = atomic_long_add_return(total_scan, | |
351 | &shrinker->nr_in_batch); | |
352 | else | |
353 | new_nr = atomic_long_read(&shrinker->nr_in_batch); | |
acf92b48 DC |
354 | |
355 | trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); | |
1da177e4 LT |
356 | } |
357 | up_read(&shrinker_rwsem); | |
f06590bd MK |
358 | out: |
359 | cond_resched(); | |
b15e0905 | 360 | return ret; |
1da177e4 LT |
361 | } |
362 | ||
1da177e4 LT |
363 | static inline int is_page_cache_freeable(struct page *page) |
364 | { | |
ceddc3a5 JW |
365 | /* |
366 | * A freeable page cache page is referenced only by the caller | |
367 | * that isolated the page, the page cache radix tree and | |
368 | * optional buffer heads at page->private. | |
369 | */ | |
edcf4748 | 370 | return page_count(page) - page_has_private(page) == 2; |
1da177e4 LT |
371 | } |
372 | ||
7d3579e8 KM |
373 | static int may_write_to_queue(struct backing_dev_info *bdi, |
374 | struct scan_control *sc) | |
1da177e4 | 375 | { |
930d9152 | 376 | if (current->flags & PF_SWAPWRITE) |
1da177e4 LT |
377 | return 1; |
378 | if (!bdi_write_congested(bdi)) | |
379 | return 1; | |
380 | if (bdi == current->backing_dev_info) | |
381 | return 1; | |
382 | return 0; | |
383 | } | |
384 | ||
385 | /* | |
386 | * We detected a synchronous write error writing a page out. Probably | |
387 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
388 | * fsync(), msync() or close(). | |
389 | * | |
390 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
391 | * prevents it from being freed up. But we have a ref on the page and once | |
392 | * that page is locked, the mapping is pinned. | |
393 | * | |
394 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
395 | * __GFP_FS. | |
396 | */ | |
397 | static void handle_write_error(struct address_space *mapping, | |
398 | struct page *page, int error) | |
399 | { | |
7eaceacc | 400 | lock_page(page); |
3e9f45bd GC |
401 | if (page_mapping(page) == mapping) |
402 | mapping_set_error(mapping, error); | |
1da177e4 LT |
403 | unlock_page(page); |
404 | } | |
405 | ||
04e62a29 CL |
406 | /* possible outcome of pageout() */ |
407 | typedef enum { | |
408 | /* failed to write page out, page is locked */ | |
409 | PAGE_KEEP, | |
410 | /* move page to the active list, page is locked */ | |
411 | PAGE_ACTIVATE, | |
412 | /* page has been sent to the disk successfully, page is unlocked */ | |
413 | PAGE_SUCCESS, | |
414 | /* page is clean and locked */ | |
415 | PAGE_CLEAN, | |
416 | } pageout_t; | |
417 | ||
1da177e4 | 418 | /* |
1742f19f AM |
419 | * pageout is called by shrink_page_list() for each dirty page. |
420 | * Calls ->writepage(). | |
1da177e4 | 421 | */ |
c661b078 | 422 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 423 | struct scan_control *sc) |
1da177e4 LT |
424 | { |
425 | /* | |
426 | * If the page is dirty, only perform writeback if that write | |
427 | * will be non-blocking. To prevent this allocation from being | |
428 | * stalled by pagecache activity. But note that there may be | |
429 | * stalls if we need to run get_block(). We could test | |
430 | * PagePrivate for that. | |
431 | * | |
6aceb53b | 432 | * If this process is currently in __generic_file_aio_write() against |
1da177e4 LT |
433 | * this page's queue, we can perform writeback even if that |
434 | * will block. | |
435 | * | |
436 | * If the page is swapcache, write it back even if that would | |
437 | * block, for some throttling. This happens by accident, because | |
438 | * swap_backing_dev_info is bust: it doesn't reflect the | |
439 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
440 | */ |
441 | if (!is_page_cache_freeable(page)) | |
442 | return PAGE_KEEP; | |
443 | if (!mapping) { | |
444 | /* | |
445 | * Some data journaling orphaned pages can have | |
446 | * page->mapping == NULL while being dirty with clean buffers. | |
447 | */ | |
266cf658 | 448 | if (page_has_private(page)) { |
1da177e4 LT |
449 | if (try_to_free_buffers(page)) { |
450 | ClearPageDirty(page); | |
d40cee24 | 451 | printk("%s: orphaned page\n", __func__); |
1da177e4 LT |
452 | return PAGE_CLEAN; |
453 | } | |
454 | } | |
455 | return PAGE_KEEP; | |
456 | } | |
457 | if (mapping->a_ops->writepage == NULL) | |
458 | return PAGE_ACTIVATE; | |
0e093d99 | 459 | if (!may_write_to_queue(mapping->backing_dev_info, sc)) |
1da177e4 LT |
460 | return PAGE_KEEP; |
461 | ||
462 | if (clear_page_dirty_for_io(page)) { | |
463 | int res; | |
464 | struct writeback_control wbc = { | |
465 | .sync_mode = WB_SYNC_NONE, | |
466 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
467 | .range_start = 0, |
468 | .range_end = LLONG_MAX, | |
1da177e4 LT |
469 | .for_reclaim = 1, |
470 | }; | |
471 | ||
472 | SetPageReclaim(page); | |
473 | res = mapping->a_ops->writepage(page, &wbc); | |
474 | if (res < 0) | |
475 | handle_write_error(mapping, page, res); | |
994fc28c | 476 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
477 | ClearPageReclaim(page); |
478 | return PAGE_ACTIVATE; | |
479 | } | |
c661b078 | 480 | |
1da177e4 LT |
481 | if (!PageWriteback(page)) { |
482 | /* synchronous write or broken a_ops? */ | |
483 | ClearPageReclaim(page); | |
484 | } | |
23b9da55 | 485 | trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); |
e129b5c2 | 486 | inc_zone_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
487 | return PAGE_SUCCESS; |
488 | } | |
489 | ||
490 | return PAGE_CLEAN; | |
491 | } | |
492 | ||
a649fd92 | 493 | /* |
e286781d NP |
494 | * Same as remove_mapping, but if the page is removed from the mapping, it |
495 | * gets returned with a refcount of 0. | |
a649fd92 | 496 | */ |
e286781d | 497 | static int __remove_mapping(struct address_space *mapping, struct page *page) |
49d2e9cc | 498 | { |
28e4d965 NP |
499 | BUG_ON(!PageLocked(page)); |
500 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 501 | |
19fd6231 | 502 | spin_lock_irq(&mapping->tree_lock); |
49d2e9cc | 503 | /* |
0fd0e6b0 NP |
504 | * The non racy check for a busy page. |
505 | * | |
506 | * Must be careful with the order of the tests. When someone has | |
507 | * a ref to the page, it may be possible that they dirty it then | |
508 | * drop the reference. So if PageDirty is tested before page_count | |
509 | * here, then the following race may occur: | |
510 | * | |
511 | * get_user_pages(&page); | |
512 | * [user mapping goes away] | |
513 | * write_to(page); | |
514 | * !PageDirty(page) [good] | |
515 | * SetPageDirty(page); | |
516 | * put_page(page); | |
517 | * !page_count(page) [good, discard it] | |
518 | * | |
519 | * [oops, our write_to data is lost] | |
520 | * | |
521 | * Reversing the order of the tests ensures such a situation cannot | |
522 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
523 | * load is not satisfied before that of page->_count. | |
524 | * | |
525 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
526 | * and thus under tree_lock, then this ordering is not required. | |
49d2e9cc | 527 | */ |
e286781d | 528 | if (!page_freeze_refs(page, 2)) |
49d2e9cc | 529 | goto cannot_free; |
e286781d NP |
530 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
531 | if (unlikely(PageDirty(page))) { | |
532 | page_unfreeze_refs(page, 2); | |
49d2e9cc | 533 | goto cannot_free; |
e286781d | 534 | } |
49d2e9cc CL |
535 | |
536 | if (PageSwapCache(page)) { | |
537 | swp_entry_t swap = { .val = page_private(page) }; | |
538 | __delete_from_swap_cache(page); | |
19fd6231 | 539 | spin_unlock_irq(&mapping->tree_lock); |
cb4b86ba | 540 | swapcache_free(swap, page); |
e286781d | 541 | } else { |
6072d13c LT |
542 | void (*freepage)(struct page *); |
543 | ||
544 | freepage = mapping->a_ops->freepage; | |
545 | ||
e64a782f | 546 | __delete_from_page_cache(page); |
19fd6231 | 547 | spin_unlock_irq(&mapping->tree_lock); |
e767e056 | 548 | mem_cgroup_uncharge_cache_page(page); |
6072d13c LT |
549 | |
550 | if (freepage != NULL) | |
551 | freepage(page); | |
49d2e9cc CL |
552 | } |
553 | ||
49d2e9cc CL |
554 | return 1; |
555 | ||
556 | cannot_free: | |
19fd6231 | 557 | spin_unlock_irq(&mapping->tree_lock); |
49d2e9cc CL |
558 | return 0; |
559 | } | |
560 | ||
e286781d NP |
561 | /* |
562 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
563 | * someone else has a ref on the page, abort and return 0. If it was | |
564 | * successfully detached, return 1. Assumes the caller has a single ref on | |
565 | * this page. | |
566 | */ | |
567 | int remove_mapping(struct address_space *mapping, struct page *page) | |
568 | { | |
569 | if (__remove_mapping(mapping, page)) { | |
570 | /* | |
571 | * Unfreezing the refcount with 1 rather than 2 effectively | |
572 | * drops the pagecache ref for us without requiring another | |
573 | * atomic operation. | |
574 | */ | |
575 | page_unfreeze_refs(page, 1); | |
576 | return 1; | |
577 | } | |
578 | return 0; | |
579 | } | |
580 | ||
894bc310 LS |
581 | /** |
582 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
583 | * @page: page to be put back to appropriate lru list | |
584 | * | |
585 | * Add previously isolated @page to appropriate LRU list. | |
586 | * Page may still be unevictable for other reasons. | |
587 | * | |
588 | * lru_lock must not be held, interrupts must be enabled. | |
589 | */ | |
894bc310 LS |
590 | void putback_lru_page(struct page *page) |
591 | { | |
592 | int lru; | |
593 | int active = !!TestClearPageActive(page); | |
bbfd28ee | 594 | int was_unevictable = PageUnevictable(page); |
894bc310 LS |
595 | |
596 | VM_BUG_ON(PageLRU(page)); | |
597 | ||
598 | redo: | |
599 | ClearPageUnevictable(page); | |
600 | ||
39b5f29a | 601 | if (page_evictable(page)) { |
894bc310 LS |
602 | /* |
603 | * For evictable pages, we can use the cache. | |
604 | * In event of a race, worst case is we end up with an | |
605 | * unevictable page on [in]active list. | |
606 | * We know how to handle that. | |
607 | */ | |
401a8e1c | 608 | lru = active + page_lru_base_type(page); |
894bc310 LS |
609 | lru_cache_add_lru(page, lru); |
610 | } else { | |
611 | /* | |
612 | * Put unevictable pages directly on zone's unevictable | |
613 | * list. | |
614 | */ | |
615 | lru = LRU_UNEVICTABLE; | |
616 | add_page_to_unevictable_list(page); | |
6a7b9548 | 617 | /* |
21ee9f39 MK |
618 | * When racing with an mlock or AS_UNEVICTABLE clearing |
619 | * (page is unlocked) make sure that if the other thread | |
620 | * does not observe our setting of PG_lru and fails | |
24513264 | 621 | * isolation/check_move_unevictable_pages, |
21ee9f39 | 622 | * we see PG_mlocked/AS_UNEVICTABLE cleared below and move |
6a7b9548 JW |
623 | * the page back to the evictable list. |
624 | * | |
21ee9f39 | 625 | * The other side is TestClearPageMlocked() or shmem_lock(). |
6a7b9548 JW |
626 | */ |
627 | smp_mb(); | |
894bc310 | 628 | } |
894bc310 LS |
629 | |
630 | /* | |
631 | * page's status can change while we move it among lru. If an evictable | |
632 | * page is on unevictable list, it never be freed. To avoid that, | |
633 | * check after we added it to the list, again. | |
634 | */ | |
39b5f29a | 635 | if (lru == LRU_UNEVICTABLE && page_evictable(page)) { |
894bc310 LS |
636 | if (!isolate_lru_page(page)) { |
637 | put_page(page); | |
638 | goto redo; | |
639 | } | |
640 | /* This means someone else dropped this page from LRU | |
641 | * So, it will be freed or putback to LRU again. There is | |
642 | * nothing to do here. | |
643 | */ | |
644 | } | |
645 | ||
bbfd28ee LS |
646 | if (was_unevictable && lru != LRU_UNEVICTABLE) |
647 | count_vm_event(UNEVICTABLE_PGRESCUED); | |
648 | else if (!was_unevictable && lru == LRU_UNEVICTABLE) | |
649 | count_vm_event(UNEVICTABLE_PGCULLED); | |
650 | ||
894bc310 LS |
651 | put_page(page); /* drop ref from isolate */ |
652 | } | |
653 | ||
dfc8d636 JW |
654 | enum page_references { |
655 | PAGEREF_RECLAIM, | |
656 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 657 | PAGEREF_KEEP, |
dfc8d636 JW |
658 | PAGEREF_ACTIVATE, |
659 | }; | |
660 | ||
661 | static enum page_references page_check_references(struct page *page, | |
662 | struct scan_control *sc) | |
663 | { | |
64574746 | 664 | int referenced_ptes, referenced_page; |
dfc8d636 | 665 | unsigned long vm_flags; |
dfc8d636 | 666 | |
c3ac9a8a JW |
667 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
668 | &vm_flags); | |
64574746 | 669 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 670 | |
dfc8d636 JW |
671 | /* |
672 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
673 | * move the page to the unevictable list. | |
674 | */ | |
675 | if (vm_flags & VM_LOCKED) | |
676 | return PAGEREF_RECLAIM; | |
677 | ||
64574746 | 678 | if (referenced_ptes) { |
e4898273 | 679 | if (PageSwapBacked(page)) |
64574746 JW |
680 | return PAGEREF_ACTIVATE; |
681 | /* | |
682 | * All mapped pages start out with page table | |
683 | * references from the instantiating fault, so we need | |
684 | * to look twice if a mapped file page is used more | |
685 | * than once. | |
686 | * | |
687 | * Mark it and spare it for another trip around the | |
688 | * inactive list. Another page table reference will | |
689 | * lead to its activation. | |
690 | * | |
691 | * Note: the mark is set for activated pages as well | |
692 | * so that recently deactivated but used pages are | |
693 | * quickly recovered. | |
694 | */ | |
695 | SetPageReferenced(page); | |
696 | ||
34dbc67a | 697 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
698 | return PAGEREF_ACTIVATE; |
699 | ||
c909e993 KK |
700 | /* |
701 | * Activate file-backed executable pages after first usage. | |
702 | */ | |
703 | if (vm_flags & VM_EXEC) | |
704 | return PAGEREF_ACTIVATE; | |
705 | ||
64574746 JW |
706 | return PAGEREF_KEEP; |
707 | } | |
dfc8d636 JW |
708 | |
709 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 710 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
711 | return PAGEREF_RECLAIM_CLEAN; |
712 | ||
713 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
714 | } |
715 | ||
1da177e4 | 716 | /* |
1742f19f | 717 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 718 | */ |
1742f19f | 719 | static unsigned long shrink_page_list(struct list_head *page_list, |
6a18adb3 | 720 | struct zone *zone, |
f84f6e2b | 721 | struct scan_control *sc, |
02c6de8d | 722 | enum ttu_flags ttu_flags, |
92df3a72 | 723 | unsigned long *ret_nr_dirty, |
02c6de8d MK |
724 | unsigned long *ret_nr_writeback, |
725 | bool force_reclaim) | |
1da177e4 LT |
726 | { |
727 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 728 | LIST_HEAD(free_pages); |
1da177e4 | 729 | int pgactivate = 0; |
0e093d99 MG |
730 | unsigned long nr_dirty = 0; |
731 | unsigned long nr_congested = 0; | |
05ff5137 | 732 | unsigned long nr_reclaimed = 0; |
92df3a72 | 733 | unsigned long nr_writeback = 0; |
1da177e4 LT |
734 | |
735 | cond_resched(); | |
736 | ||
69980e31 | 737 | mem_cgroup_uncharge_start(); |
1da177e4 LT |
738 | while (!list_empty(page_list)) { |
739 | struct address_space *mapping; | |
740 | struct page *page; | |
741 | int may_enter_fs; | |
02c6de8d | 742 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
1da177e4 LT |
743 | |
744 | cond_resched(); | |
745 | ||
746 | page = lru_to_page(page_list); | |
747 | list_del(&page->lru); | |
748 | ||
529ae9aa | 749 | if (!trylock_page(page)) |
1da177e4 LT |
750 | goto keep; |
751 | ||
725d704e | 752 | VM_BUG_ON(PageActive(page)); |
6a18adb3 | 753 | VM_BUG_ON(page_zone(page) != zone); |
1da177e4 LT |
754 | |
755 | sc->nr_scanned++; | |
80e43426 | 756 | |
39b5f29a | 757 | if (unlikely(!page_evictable(page))) |
b291f000 | 758 | goto cull_mlocked; |
894bc310 | 759 | |
a6dc60f8 | 760 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
761 | goto keep_locked; |
762 | ||
1da177e4 LT |
763 | /* Double the slab pressure for mapped and swapcache pages */ |
764 | if (page_mapped(page) || PageSwapCache(page)) | |
765 | sc->nr_scanned++; | |
766 | ||
c661b078 AW |
767 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
768 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
769 | ||
770 | if (PageWriteback(page)) { | |
e62e384e MH |
771 | /* |
772 | * memcg doesn't have any dirty pages throttling so we | |
773 | * could easily OOM just because too many pages are in | |
c3b94f44 | 774 | * writeback and there is nothing else to reclaim. |
e62e384e | 775 | * |
c3b94f44 | 776 | * Check __GFP_IO, certainly because a loop driver |
e62e384e MH |
777 | * thread might enter reclaim, and deadlock if it waits |
778 | * on a page for which it is needed to do the write | |
779 | * (loop masks off __GFP_IO|__GFP_FS for this reason); | |
780 | * but more thought would probably show more reasons. | |
c3b94f44 HD |
781 | * |
782 | * Don't require __GFP_FS, since we're not going into | |
783 | * the FS, just waiting on its writeback completion. | |
784 | * Worryingly, ext4 gfs2 and xfs allocate pages with | |
785 | * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so | |
786 | * testing may_enter_fs here is liable to OOM on them. | |
e62e384e | 787 | */ |
c3b94f44 HD |
788 | if (global_reclaim(sc) || |
789 | !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) { | |
790 | /* | |
791 | * This is slightly racy - end_page_writeback() | |
792 | * might have just cleared PageReclaim, then | |
793 | * setting PageReclaim here end up interpreted | |
794 | * as PageReadahead - but that does not matter | |
795 | * enough to care. What we do want is for this | |
796 | * page to have PageReclaim set next time memcg | |
797 | * reclaim reaches the tests above, so it will | |
798 | * then wait_on_page_writeback() to avoid OOM; | |
799 | * and it's also appropriate in global reclaim. | |
800 | */ | |
801 | SetPageReclaim(page); | |
e62e384e | 802 | nr_writeback++; |
c3b94f44 | 803 | goto keep_locked; |
e62e384e | 804 | } |
c3b94f44 | 805 | wait_on_page_writeback(page); |
c661b078 | 806 | } |
1da177e4 | 807 | |
02c6de8d MK |
808 | if (!force_reclaim) |
809 | references = page_check_references(page, sc); | |
810 | ||
dfc8d636 JW |
811 | switch (references) { |
812 | case PAGEREF_ACTIVATE: | |
1da177e4 | 813 | goto activate_locked; |
64574746 JW |
814 | case PAGEREF_KEEP: |
815 | goto keep_locked; | |
dfc8d636 JW |
816 | case PAGEREF_RECLAIM: |
817 | case PAGEREF_RECLAIM_CLEAN: | |
818 | ; /* try to reclaim the page below */ | |
819 | } | |
1da177e4 | 820 | |
1da177e4 LT |
821 | /* |
822 | * Anonymous process memory has backing store? | |
823 | * Try to allocate it some swap space here. | |
824 | */ | |
b291f000 | 825 | if (PageAnon(page) && !PageSwapCache(page)) { |
63eb6b93 HD |
826 | if (!(sc->gfp_mask & __GFP_IO)) |
827 | goto keep_locked; | |
5bc7b8ac | 828 | if (!add_to_swap(page, page_list)) |
1da177e4 | 829 | goto activate_locked; |
63eb6b93 | 830 | may_enter_fs = 1; |
b291f000 | 831 | } |
1da177e4 LT |
832 | |
833 | mapping = page_mapping(page); | |
1da177e4 LT |
834 | |
835 | /* | |
836 | * The page is mapped into the page tables of one or more | |
837 | * processes. Try to unmap it here. | |
838 | */ | |
839 | if (page_mapped(page) && mapping) { | |
02c6de8d | 840 | switch (try_to_unmap(page, ttu_flags)) { |
1da177e4 LT |
841 | case SWAP_FAIL: |
842 | goto activate_locked; | |
843 | case SWAP_AGAIN: | |
844 | goto keep_locked; | |
b291f000 NP |
845 | case SWAP_MLOCK: |
846 | goto cull_mlocked; | |
1da177e4 LT |
847 | case SWAP_SUCCESS: |
848 | ; /* try to free the page below */ | |
849 | } | |
850 | } | |
851 | ||
852 | if (PageDirty(page)) { | |
0e093d99 MG |
853 | nr_dirty++; |
854 | ||
ee72886d MG |
855 | /* |
856 | * Only kswapd can writeback filesystem pages to | |
f84f6e2b MG |
857 | * avoid risk of stack overflow but do not writeback |
858 | * unless under significant pressure. | |
ee72886d | 859 | */ |
f84f6e2b | 860 | if (page_is_file_cache(page) && |
9e3b2f8c KK |
861 | (!current_is_kswapd() || |
862 | sc->priority >= DEF_PRIORITY - 2)) { | |
49ea7eb6 MG |
863 | /* |
864 | * Immediately reclaim when written back. | |
865 | * Similar in principal to deactivate_page() | |
866 | * except we already have the page isolated | |
867 | * and know it's dirty | |
868 | */ | |
869 | inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); | |
870 | SetPageReclaim(page); | |
871 | ||
ee72886d MG |
872 | goto keep_locked; |
873 | } | |
874 | ||
dfc8d636 | 875 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 876 | goto keep_locked; |
4dd4b920 | 877 | if (!may_enter_fs) |
1da177e4 | 878 | goto keep_locked; |
52a8363e | 879 | if (!sc->may_writepage) |
1da177e4 LT |
880 | goto keep_locked; |
881 | ||
882 | /* Page is dirty, try to write it out here */ | |
7d3579e8 | 883 | switch (pageout(page, mapping, sc)) { |
1da177e4 | 884 | case PAGE_KEEP: |
0e093d99 | 885 | nr_congested++; |
1da177e4 LT |
886 | goto keep_locked; |
887 | case PAGE_ACTIVATE: | |
888 | goto activate_locked; | |
889 | case PAGE_SUCCESS: | |
7d3579e8 | 890 | if (PageWriteback(page)) |
41ac1999 | 891 | goto keep; |
7d3579e8 | 892 | if (PageDirty(page)) |
1da177e4 | 893 | goto keep; |
7d3579e8 | 894 | |
1da177e4 LT |
895 | /* |
896 | * A synchronous write - probably a ramdisk. Go | |
897 | * ahead and try to reclaim the page. | |
898 | */ | |
529ae9aa | 899 | if (!trylock_page(page)) |
1da177e4 LT |
900 | goto keep; |
901 | if (PageDirty(page) || PageWriteback(page)) | |
902 | goto keep_locked; | |
903 | mapping = page_mapping(page); | |
904 | case PAGE_CLEAN: | |
905 | ; /* try to free the page below */ | |
906 | } | |
907 | } | |
908 | ||
909 | /* | |
910 | * If the page has buffers, try to free the buffer mappings | |
911 | * associated with this page. If we succeed we try to free | |
912 | * the page as well. | |
913 | * | |
914 | * We do this even if the page is PageDirty(). | |
915 | * try_to_release_page() does not perform I/O, but it is | |
916 | * possible for a page to have PageDirty set, but it is actually | |
917 | * clean (all its buffers are clean). This happens if the | |
918 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 919 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
920 | * try_to_release_page() will discover that cleanness and will |
921 | * drop the buffers and mark the page clean - it can be freed. | |
922 | * | |
923 | * Rarely, pages can have buffers and no ->mapping. These are | |
924 | * the pages which were not successfully invalidated in | |
925 | * truncate_complete_page(). We try to drop those buffers here | |
926 | * and if that worked, and the page is no longer mapped into | |
927 | * process address space (page_count == 1) it can be freed. | |
928 | * Otherwise, leave the page on the LRU so it is swappable. | |
929 | */ | |
266cf658 | 930 | if (page_has_private(page)) { |
1da177e4 LT |
931 | if (!try_to_release_page(page, sc->gfp_mask)) |
932 | goto activate_locked; | |
e286781d NP |
933 | if (!mapping && page_count(page) == 1) { |
934 | unlock_page(page); | |
935 | if (put_page_testzero(page)) | |
936 | goto free_it; | |
937 | else { | |
938 | /* | |
939 | * rare race with speculative reference. | |
940 | * the speculative reference will free | |
941 | * this page shortly, so we may | |
942 | * increment nr_reclaimed here (and | |
943 | * leave it off the LRU). | |
944 | */ | |
945 | nr_reclaimed++; | |
946 | continue; | |
947 | } | |
948 | } | |
1da177e4 LT |
949 | } |
950 | ||
e286781d | 951 | if (!mapping || !__remove_mapping(mapping, page)) |
49d2e9cc | 952 | goto keep_locked; |
1da177e4 | 953 | |
a978d6f5 NP |
954 | /* |
955 | * At this point, we have no other references and there is | |
956 | * no way to pick any more up (removed from LRU, removed | |
957 | * from pagecache). Can use non-atomic bitops now (and | |
958 | * we obviously don't have to worry about waking up a process | |
959 | * waiting on the page lock, because there are no references. | |
960 | */ | |
961 | __clear_page_locked(page); | |
e286781d | 962 | free_it: |
05ff5137 | 963 | nr_reclaimed++; |
abe4c3b5 MG |
964 | |
965 | /* | |
966 | * Is there need to periodically free_page_list? It would | |
967 | * appear not as the counts should be low | |
968 | */ | |
969 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
970 | continue; |
971 | ||
b291f000 | 972 | cull_mlocked: |
63d6c5ad HD |
973 | if (PageSwapCache(page)) |
974 | try_to_free_swap(page); | |
b291f000 NP |
975 | unlock_page(page); |
976 | putback_lru_page(page); | |
977 | continue; | |
978 | ||
1da177e4 | 979 | activate_locked: |
68a22394 RR |
980 | /* Not a candidate for swapping, so reclaim swap space. */ |
981 | if (PageSwapCache(page) && vm_swap_full()) | |
a2c43eed | 982 | try_to_free_swap(page); |
894bc310 | 983 | VM_BUG_ON(PageActive(page)); |
1da177e4 LT |
984 | SetPageActive(page); |
985 | pgactivate++; | |
986 | keep_locked: | |
987 | unlock_page(page); | |
988 | keep: | |
989 | list_add(&page->lru, &ret_pages); | |
b291f000 | 990 | VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); |
1da177e4 | 991 | } |
abe4c3b5 | 992 | |
0e093d99 MG |
993 | /* |
994 | * Tag a zone as congested if all the dirty pages encountered were | |
995 | * backed by a congested BDI. In this case, reclaimers should just | |
996 | * back off and wait for congestion to clear because further reclaim | |
997 | * will encounter the same problem | |
998 | */ | |
89b5fae5 | 999 | if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc)) |
6a18adb3 | 1000 | zone_set_flag(zone, ZONE_CONGESTED); |
0e093d99 | 1001 | |
cc59850e | 1002 | free_hot_cold_page_list(&free_pages, 1); |
abe4c3b5 | 1003 | |
1da177e4 | 1004 | list_splice(&ret_pages, page_list); |
f8891e5e | 1005 | count_vm_events(PGACTIVATE, pgactivate); |
69980e31 | 1006 | mem_cgroup_uncharge_end(); |
92df3a72 MG |
1007 | *ret_nr_dirty += nr_dirty; |
1008 | *ret_nr_writeback += nr_writeback; | |
05ff5137 | 1009 | return nr_reclaimed; |
1da177e4 LT |
1010 | } |
1011 | ||
02c6de8d MK |
1012 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1013 | struct list_head *page_list) | |
1014 | { | |
1015 | struct scan_control sc = { | |
1016 | .gfp_mask = GFP_KERNEL, | |
1017 | .priority = DEF_PRIORITY, | |
1018 | .may_unmap = 1, | |
1019 | }; | |
1020 | unsigned long ret, dummy1, dummy2; | |
1021 | struct page *page, *next; | |
1022 | LIST_HEAD(clean_pages); | |
1023 | ||
1024 | list_for_each_entry_safe(page, next, page_list, lru) { | |
009dfd44 RA |
1025 | if (page_is_file_cache(page) && !PageDirty(page) && |
1026 | !isolated_balloon_page(page)) { | |
02c6de8d MK |
1027 | ClearPageActive(page); |
1028 | list_move(&page->lru, &clean_pages); | |
1029 | } | |
1030 | } | |
1031 | ||
1032 | ret = shrink_page_list(&clean_pages, zone, &sc, | |
1033 | TTU_UNMAP|TTU_IGNORE_ACCESS, | |
1034 | &dummy1, &dummy2, true); | |
1035 | list_splice(&clean_pages, page_list); | |
1036 | __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); | |
1037 | return ret; | |
1038 | } | |
1039 | ||
5ad333eb AW |
1040 | /* |
1041 | * Attempt to remove the specified page from its LRU. Only take this page | |
1042 | * if it is of the appropriate PageActive status. Pages which are being | |
1043 | * freed elsewhere are also ignored. | |
1044 | * | |
1045 | * page: page to consider | |
1046 | * mode: one of the LRU isolation modes defined above | |
1047 | * | |
1048 | * returns 0 on success, -ve errno on failure. | |
1049 | */ | |
f3fd4a61 | 1050 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1051 | { |
1052 | int ret = -EINVAL; | |
1053 | ||
1054 | /* Only take pages on the LRU. */ | |
1055 | if (!PageLRU(page)) | |
1056 | return ret; | |
1057 | ||
e46a2879 MK |
1058 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1059 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1060 | return ret; |
1061 | ||
5ad333eb | 1062 | ret = -EBUSY; |
08e552c6 | 1063 | |
c8244935 MG |
1064 | /* |
1065 | * To minimise LRU disruption, the caller can indicate that it only | |
1066 | * wants to isolate pages it will be able to operate on without | |
1067 | * blocking - clean pages for the most part. | |
1068 | * | |
1069 | * ISOLATE_CLEAN means that only clean pages should be isolated. This | |
1070 | * is used by reclaim when it is cannot write to backing storage | |
1071 | * | |
1072 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages | |
1073 | * that it is possible to migrate without blocking | |
1074 | */ | |
1075 | if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { | |
1076 | /* All the caller can do on PageWriteback is block */ | |
1077 | if (PageWriteback(page)) | |
1078 | return ret; | |
1079 | ||
1080 | if (PageDirty(page)) { | |
1081 | struct address_space *mapping; | |
1082 | ||
1083 | /* ISOLATE_CLEAN means only clean pages */ | |
1084 | if (mode & ISOLATE_CLEAN) | |
1085 | return ret; | |
1086 | ||
1087 | /* | |
1088 | * Only pages without mappings or that have a | |
1089 | * ->migratepage callback are possible to migrate | |
1090 | * without blocking | |
1091 | */ | |
1092 | mapping = page_mapping(page); | |
1093 | if (mapping && !mapping->a_ops->migratepage) | |
1094 | return ret; | |
1095 | } | |
1096 | } | |
39deaf85 | 1097 | |
f80c0673 MK |
1098 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1099 | return ret; | |
1100 | ||
5ad333eb AW |
1101 | if (likely(get_page_unless_zero(page))) { |
1102 | /* | |
1103 | * Be careful not to clear PageLRU until after we're | |
1104 | * sure the page is not being freed elsewhere -- the | |
1105 | * page release code relies on it. | |
1106 | */ | |
1107 | ClearPageLRU(page); | |
1108 | ret = 0; | |
1109 | } | |
1110 | ||
1111 | return ret; | |
1112 | } | |
1113 | ||
1da177e4 LT |
1114 | /* |
1115 | * zone->lru_lock is heavily contended. Some of the functions that | |
1116 | * shrink the lists perform better by taking out a batch of pages | |
1117 | * and working on them outside the LRU lock. | |
1118 | * | |
1119 | * For pagecache intensive workloads, this function is the hottest | |
1120 | * spot in the kernel (apart from copy_*_user functions). | |
1121 | * | |
1122 | * Appropriate locks must be held before calling this function. | |
1123 | * | |
1124 | * @nr_to_scan: The number of pages to look through on the list. | |
5dc35979 | 1125 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1126 | * @dst: The temp list to put pages on to. |
f626012d | 1127 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1128 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1129 | * @mode: One of the LRU isolation modes |
3cb99451 | 1130 | * @lru: LRU list id for isolating |
1da177e4 LT |
1131 | * |
1132 | * returns how many pages were moved onto *@dst. | |
1133 | */ | |
69e05944 | 1134 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1135 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1136 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1137 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1138 | { |
75b00af7 | 1139 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1140 | unsigned long nr_taken = 0; |
c9b02d97 | 1141 | unsigned long scan; |
1da177e4 | 1142 | |
c9b02d97 | 1143 | for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { |
5ad333eb | 1144 | struct page *page; |
fa9add64 | 1145 | int nr_pages; |
5ad333eb | 1146 | |
1da177e4 LT |
1147 | page = lru_to_page(src); |
1148 | prefetchw_prev_lru_page(page, src, flags); | |
1149 | ||
725d704e | 1150 | VM_BUG_ON(!PageLRU(page)); |
8d438f96 | 1151 | |
f3fd4a61 | 1152 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1153 | case 0: |
fa9add64 HD |
1154 | nr_pages = hpage_nr_pages(page); |
1155 | mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); | |
5ad333eb | 1156 | list_move(&page->lru, dst); |
fa9add64 | 1157 | nr_taken += nr_pages; |
5ad333eb AW |
1158 | break; |
1159 | ||
1160 | case -EBUSY: | |
1161 | /* else it is being freed elsewhere */ | |
1162 | list_move(&page->lru, src); | |
1163 | continue; | |
46453a6e | 1164 | |
5ad333eb AW |
1165 | default: |
1166 | BUG(); | |
1167 | } | |
1da177e4 LT |
1168 | } |
1169 | ||
f626012d | 1170 | *nr_scanned = scan; |
75b00af7 HD |
1171 | trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, |
1172 | nr_taken, mode, is_file_lru(lru)); | |
1da177e4 LT |
1173 | return nr_taken; |
1174 | } | |
1175 | ||
62695a84 NP |
1176 | /** |
1177 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1178 | * @page: page to isolate from its LRU list | |
1179 | * | |
1180 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1181 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1182 | * | |
1183 | * Returns 0 if the page was removed from an LRU list. | |
1184 | * Returns -EBUSY if the page was not on an LRU list. | |
1185 | * | |
1186 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1187 | * the active list, it will have PageActive set. If it was found on |
1188 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1189 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1190 | * |
1191 | * The vmstat statistic corresponding to the list on which the page was | |
1192 | * found will be decremented. | |
1193 | * | |
1194 | * Restrictions: | |
1195 | * (1) Must be called with an elevated refcount on the page. This is a | |
1196 | * fundamentnal difference from isolate_lru_pages (which is called | |
1197 | * without a stable reference). | |
1198 | * (2) the lru_lock must not be held. | |
1199 | * (3) interrupts must be enabled. | |
1200 | */ | |
1201 | int isolate_lru_page(struct page *page) | |
1202 | { | |
1203 | int ret = -EBUSY; | |
1204 | ||
0c917313 KK |
1205 | VM_BUG_ON(!page_count(page)); |
1206 | ||
62695a84 NP |
1207 | if (PageLRU(page)) { |
1208 | struct zone *zone = page_zone(page); | |
fa9add64 | 1209 | struct lruvec *lruvec; |
62695a84 NP |
1210 | |
1211 | spin_lock_irq(&zone->lru_lock); | |
fa9add64 | 1212 | lruvec = mem_cgroup_page_lruvec(page, zone); |
0c917313 | 1213 | if (PageLRU(page)) { |
894bc310 | 1214 | int lru = page_lru(page); |
0c917313 | 1215 | get_page(page); |
62695a84 | 1216 | ClearPageLRU(page); |
fa9add64 HD |
1217 | del_page_from_lru_list(page, lruvec, lru); |
1218 | ret = 0; | |
62695a84 NP |
1219 | } |
1220 | spin_unlock_irq(&zone->lru_lock); | |
1221 | } | |
1222 | return ret; | |
1223 | } | |
1224 | ||
35cd7815 | 1225 | /* |
d37dd5dc FW |
1226 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1227 | * then get resheduled. When there are massive number of tasks doing page | |
1228 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1229 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1230 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 RR |
1231 | */ |
1232 | static int too_many_isolated(struct zone *zone, int file, | |
1233 | struct scan_control *sc) | |
1234 | { | |
1235 | unsigned long inactive, isolated; | |
1236 | ||
6fa3eb70 | 1237 | if (current_is_kswapd() || sc->hibernation_mode) |
35cd7815 RR |
1238 | return 0; |
1239 | ||
89b5fae5 | 1240 | if (!global_reclaim(sc)) |
35cd7815 RR |
1241 | return 0; |
1242 | ||
1243 | if (file) { | |
1244 | inactive = zone_page_state(zone, NR_INACTIVE_FILE); | |
1245 | isolated = zone_page_state(zone, NR_ISOLATED_FILE); | |
1246 | } else { | |
1247 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1248 | isolated = zone_page_state(zone, NR_ISOLATED_ANON); | |
1249 | } | |
1250 | ||
3cf23841 FW |
1251 | /* |
1252 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1253 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1254 | * deadlock. | |
1255 | */ | |
1256 | if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS) | |
1257 | inactive >>= 3; | |
1258 | ||
35cd7815 RR |
1259 | return isolated > inactive; |
1260 | } | |
1261 | ||
66635629 | 1262 | static noinline_for_stack void |
75b00af7 | 1263 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1264 | { |
27ac81d8 KK |
1265 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1266 | struct zone *zone = lruvec_zone(lruvec); | |
3f79768f | 1267 | LIST_HEAD(pages_to_free); |
66635629 | 1268 | |
66635629 MG |
1269 | /* |
1270 | * Put back any unfreeable pages. | |
1271 | */ | |
66635629 | 1272 | while (!list_empty(page_list)) { |
3f79768f | 1273 | struct page *page = lru_to_page(page_list); |
66635629 | 1274 | int lru; |
3f79768f | 1275 | |
66635629 MG |
1276 | VM_BUG_ON(PageLRU(page)); |
1277 | list_del(&page->lru); | |
39b5f29a | 1278 | if (unlikely(!page_evictable(page))) { |
66635629 MG |
1279 | spin_unlock_irq(&zone->lru_lock); |
1280 | putback_lru_page(page); | |
1281 | spin_lock_irq(&zone->lru_lock); | |
1282 | continue; | |
1283 | } | |
fa9add64 HD |
1284 | |
1285 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
1286 | ||
7a608572 | 1287 | SetPageLRU(page); |
66635629 | 1288 | lru = page_lru(page); |
fa9add64 HD |
1289 | add_page_to_lru_list(page, lruvec, lru); |
1290 | ||
66635629 MG |
1291 | if (is_active_lru(lru)) { |
1292 | int file = is_file_lru(lru); | |
9992af10 RR |
1293 | int numpages = hpage_nr_pages(page); |
1294 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1295 | } |
2bcf8879 HD |
1296 | if (put_page_testzero(page)) { |
1297 | __ClearPageLRU(page); | |
1298 | __ClearPageActive(page); | |
fa9add64 | 1299 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1300 | |
1301 | if (unlikely(PageCompound(page))) { | |
1302 | spin_unlock_irq(&zone->lru_lock); | |
1303 | (*get_compound_page_dtor(page))(page); | |
1304 | spin_lock_irq(&zone->lru_lock); | |
1305 | } else | |
1306 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1307 | } |
1308 | } | |
66635629 | 1309 | |
3f79768f HD |
1310 | /* |
1311 | * To save our caller's stack, now use input list for pages to free. | |
1312 | */ | |
1313 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1314 | } |
1315 | ||
1da177e4 | 1316 | /* |
1742f19f AM |
1317 | * shrink_inactive_list() is a helper for shrink_zone(). It returns the number |
1318 | * of reclaimed pages | |
1da177e4 | 1319 | */ |
66635629 | 1320 | static noinline_for_stack unsigned long |
1a93be0e | 1321 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1322 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1323 | { |
1324 | LIST_HEAD(page_list); | |
e247dbce | 1325 | unsigned long nr_scanned; |
05ff5137 | 1326 | unsigned long nr_reclaimed = 0; |
e247dbce | 1327 | unsigned long nr_taken; |
92df3a72 MG |
1328 | unsigned long nr_dirty = 0; |
1329 | unsigned long nr_writeback = 0; | |
f3fd4a61 | 1330 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1331 | int file = is_file_lru(lru); |
1a93be0e KK |
1332 | struct zone *zone = lruvec_zone(lruvec); |
1333 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; | |
78dc583d | 1334 | |
35cd7815 | 1335 | while (unlikely(too_many_isolated(zone, file, sc))) { |
58355c78 | 1336 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
35cd7815 RR |
1337 | |
1338 | /* We are about to die and free our memory. Return now. */ | |
1339 | if (fatal_signal_pending(current)) | |
1340 | return SWAP_CLUSTER_MAX; | |
1341 | } | |
1342 | ||
1da177e4 | 1343 | lru_add_drain(); |
f80c0673 MK |
1344 | |
1345 | if (!sc->may_unmap) | |
61317289 | 1346 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1347 | if (!sc->may_writepage) |
61317289 | 1348 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1349 | |
1da177e4 | 1350 | spin_lock_irq(&zone->lru_lock); |
b35ea17b | 1351 | |
5dc35979 KK |
1352 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1353 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc KK |
1354 | |
1355 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); | |
1356 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); | |
1357 | ||
89b5fae5 | 1358 | if (global_reclaim(sc)) { |
e247dbce KM |
1359 | zone->pages_scanned += nr_scanned; |
1360 | if (current_is_kswapd()) | |
75b00af7 | 1361 | __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); |
e247dbce | 1362 | else |
75b00af7 | 1363 | __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); |
e247dbce | 1364 | } |
d563c050 | 1365 | spin_unlock_irq(&zone->lru_lock); |
b35ea17b | 1366 | |
d563c050 | 1367 | if (nr_taken == 0) |
66635629 | 1368 | return 0; |
5ad333eb | 1369 | |
02c6de8d MK |
1370 | nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, |
1371 | &nr_dirty, &nr_writeback, false); | |
c661b078 | 1372 | |
3f79768f HD |
1373 | spin_lock_irq(&zone->lru_lock); |
1374 | ||
95d918fc | 1375 | reclaim_stat->recent_scanned[file] += nr_taken; |
d563c050 | 1376 | |
904249aa YH |
1377 | if (global_reclaim(sc)) { |
1378 | if (current_is_kswapd()) | |
1379 | __count_zone_vm_events(PGSTEAL_KSWAPD, zone, | |
1380 | nr_reclaimed); | |
1381 | else | |
1382 | __count_zone_vm_events(PGSTEAL_DIRECT, zone, | |
1383 | nr_reclaimed); | |
1384 | } | |
a74609fa | 1385 | |
27ac81d8 | 1386 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1387 | |
95d918fc | 1388 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f HD |
1389 | |
1390 | spin_unlock_irq(&zone->lru_lock); | |
1391 | ||
1392 | free_hot_cold_page_list(&page_list, 1); | |
e11da5b4 | 1393 | |
92df3a72 MG |
1394 | /* |
1395 | * If reclaim is isolating dirty pages under writeback, it implies | |
1396 | * that the long-lived page allocation rate is exceeding the page | |
1397 | * laundering rate. Either the global limits are not being effective | |
1398 | * at throttling processes due to the page distribution throughout | |
1399 | * zones or there is heavy usage of a slow backing device. The | |
1400 | * only option is to throttle from reclaim context which is not ideal | |
1401 | * as there is no guarantee the dirtying process is throttled in the | |
1402 | * same way balance_dirty_pages() manages. | |
1403 | * | |
1404 | * This scales the number of dirty pages that must be under writeback | |
1405 | * before throttling depending on priority. It is a simple backoff | |
1406 | * function that has the most effect in the range DEF_PRIORITY to | |
1407 | * DEF_PRIORITY-2 which is the priority reclaim is considered to be | |
1408 | * in trouble and reclaim is considered to be in trouble. | |
1409 | * | |
1410 | * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle | |
1411 | * DEF_PRIORITY-1 50% must be PageWriteback | |
1412 | * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble | |
1413 | * ... | |
1414 | * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any | |
1415 | * isolated page is PageWriteback | |
1416 | */ | |
9e3b2f8c KK |
1417 | if (nr_writeback && nr_writeback >= |
1418 | (nr_taken >> (DEF_PRIORITY - sc->priority))) | |
92df3a72 MG |
1419 | wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); |
1420 | ||
e11da5b4 MG |
1421 | trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, |
1422 | zone_idx(zone), | |
1423 | nr_scanned, nr_reclaimed, | |
9e3b2f8c | 1424 | sc->priority, |
23b9da55 | 1425 | trace_shrink_flags(file)); |
05ff5137 | 1426 | return nr_reclaimed; |
1da177e4 LT |
1427 | } |
1428 | ||
1429 | /* | |
1430 | * This moves pages from the active list to the inactive list. | |
1431 | * | |
1432 | * We move them the other way if the page is referenced by one or more | |
1433 | * processes, from rmap. | |
1434 | * | |
1435 | * If the pages are mostly unmapped, the processing is fast and it is | |
1436 | * appropriate to hold zone->lru_lock across the whole operation. But if | |
1437 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1438 | * should drop zone->lru_lock around each page. It's impossible to balance | |
1439 | * this, so instead we remove the pages from the LRU while processing them. | |
1440 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1441 | * nobody will play with that bit on a non-LRU page. | |
1442 | * | |
1443 | * The downside is that we have to touch page->_count against each page. | |
1444 | * But we had to alter page->flags anyway. | |
1445 | */ | |
1cfb419b | 1446 | |
fa9add64 | 1447 | static void move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1448 | struct list_head *list, |
2bcf8879 | 1449 | struct list_head *pages_to_free, |
3eb4140f WF |
1450 | enum lru_list lru) |
1451 | { | |
fa9add64 | 1452 | struct zone *zone = lruvec_zone(lruvec); |
3eb4140f | 1453 | unsigned long pgmoved = 0; |
3eb4140f | 1454 | struct page *page; |
fa9add64 | 1455 | int nr_pages; |
3eb4140f | 1456 | |
3eb4140f WF |
1457 | while (!list_empty(list)) { |
1458 | page = lru_to_page(list); | |
fa9add64 | 1459 | lruvec = mem_cgroup_page_lruvec(page, zone); |
3eb4140f WF |
1460 | |
1461 | VM_BUG_ON(PageLRU(page)); | |
1462 | SetPageLRU(page); | |
1463 | ||
fa9add64 HD |
1464 | nr_pages = hpage_nr_pages(page); |
1465 | mem_cgroup_update_lru_size(lruvec, lru, nr_pages); | |
925b7673 | 1466 | list_move(&page->lru, &lruvec->lists[lru]); |
fa9add64 | 1467 | pgmoved += nr_pages; |
3eb4140f | 1468 | |
2bcf8879 HD |
1469 | if (put_page_testzero(page)) { |
1470 | __ClearPageLRU(page); | |
1471 | __ClearPageActive(page); | |
fa9add64 | 1472 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1473 | |
1474 | if (unlikely(PageCompound(page))) { | |
1475 | spin_unlock_irq(&zone->lru_lock); | |
1476 | (*get_compound_page_dtor(page))(page); | |
1477 | spin_lock_irq(&zone->lru_lock); | |
1478 | } else | |
1479 | list_add(&page->lru, pages_to_free); | |
3eb4140f WF |
1480 | } |
1481 | } | |
1482 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); | |
1483 | if (!is_active_lru(lru)) | |
1484 | __count_vm_events(PGDEACTIVATE, pgmoved); | |
1485 | } | |
1cfb419b | 1486 | |
f626012d | 1487 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1488 | struct lruvec *lruvec, |
f16015fb | 1489 | struct scan_control *sc, |
9e3b2f8c | 1490 | enum lru_list lru) |
1da177e4 | 1491 | { |
44c241f1 | 1492 | unsigned long nr_taken; |
f626012d | 1493 | unsigned long nr_scanned; |
6fe6b7e3 | 1494 | unsigned long vm_flags; |
1da177e4 | 1495 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 1496 | LIST_HEAD(l_active); |
b69408e8 | 1497 | LIST_HEAD(l_inactive); |
1da177e4 | 1498 | struct page *page; |
1a93be0e | 1499 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
44c241f1 | 1500 | unsigned long nr_rotated = 0; |
f3fd4a61 | 1501 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1502 | int file = is_file_lru(lru); |
1a93be0e | 1503 | struct zone *zone = lruvec_zone(lruvec); |
1da177e4 LT |
1504 | |
1505 | lru_add_drain(); | |
f80c0673 MK |
1506 | |
1507 | if (!sc->may_unmap) | |
61317289 | 1508 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1509 | if (!sc->may_writepage) |
61317289 | 1510 | isolate_mode |= ISOLATE_CLEAN; |
f80c0673 | 1511 | |
1da177e4 | 1512 | spin_lock_irq(&zone->lru_lock); |
925b7673 | 1513 | |
5dc35979 KK |
1514 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
1515 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 1516 | if (global_reclaim(sc)) |
f626012d | 1517 | zone->pages_scanned += nr_scanned; |
89b5fae5 | 1518 | |
b7c46d15 | 1519 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 1520 | |
f626012d | 1521 | __count_zone_vm_events(PGREFILL, zone, nr_scanned); |
3cb99451 | 1522 | __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); |
a731286d | 1523 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); |
1da177e4 LT |
1524 | spin_unlock_irq(&zone->lru_lock); |
1525 | ||
1da177e4 LT |
1526 | while (!list_empty(&l_hold)) { |
1527 | cond_resched(); | |
1528 | page = lru_to_page(&l_hold); | |
1529 | list_del(&page->lru); | |
7e9cd484 | 1530 | |
39b5f29a | 1531 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
1532 | putback_lru_page(page); |
1533 | continue; | |
1534 | } | |
1535 | ||
cc715d99 MG |
1536 | if (unlikely(buffer_heads_over_limit)) { |
1537 | if (page_has_private(page) && trylock_page(page)) { | |
1538 | if (page_has_private(page)) | |
1539 | try_to_release_page(page, 0); | |
1540 | unlock_page(page); | |
1541 | } | |
1542 | } | |
1543 | ||
c3ac9a8a JW |
1544 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
1545 | &vm_flags)) { | |
9992af10 | 1546 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
1547 | /* |
1548 | * Identify referenced, file-backed active pages and | |
1549 | * give them one more trip around the active list. So | |
1550 | * that executable code get better chances to stay in | |
1551 | * memory under moderate memory pressure. Anon pages | |
1552 | * are not likely to be evicted by use-once streaming | |
1553 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
1554 | * so we ignore them here. | |
1555 | */ | |
41e20983 | 1556 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
1557 | list_add(&page->lru, &l_active); |
1558 | continue; | |
1559 | } | |
1560 | } | |
7e9cd484 | 1561 | |
5205e56e | 1562 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
1563 | list_add(&page->lru, &l_inactive); |
1564 | } | |
1565 | ||
b555749a | 1566 | /* |
8cab4754 | 1567 | * Move pages back to the lru list. |
b555749a | 1568 | */ |
2a1dc509 | 1569 | spin_lock_irq(&zone->lru_lock); |
556adecb | 1570 | /* |
8cab4754 WF |
1571 | * Count referenced pages from currently used mappings as rotated, |
1572 | * even though only some of them are actually re-activated. This | |
1573 | * helps balance scan pressure between file and anonymous pages in | |
1574 | * get_scan_ratio. | |
7e9cd484 | 1575 | */ |
b7c46d15 | 1576 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 1577 | |
fa9add64 HD |
1578 | move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
1579 | move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
a731286d | 1580 | __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); |
f8891e5e | 1581 | spin_unlock_irq(&zone->lru_lock); |
2bcf8879 HD |
1582 | |
1583 | free_hot_cold_page_list(&l_hold, 1); | |
1da177e4 LT |
1584 | } |
1585 | ||
74e3f3c3 | 1586 | #ifdef CONFIG_SWAP |
14797e23 | 1587 | static int inactive_anon_is_low_global(struct zone *zone) |
f89eb90e KM |
1588 | { |
1589 | unsigned long active, inactive; | |
1590 | ||
1591 | active = zone_page_state(zone, NR_ACTIVE_ANON); | |
1592 | inactive = zone_page_state(zone, NR_INACTIVE_ANON); | |
1593 | ||
1594 | if (inactive * zone->inactive_ratio < active) | |
1595 | return 1; | |
1596 | ||
1597 | return 0; | |
1598 | } | |
1599 | ||
14797e23 KM |
1600 | /** |
1601 | * inactive_anon_is_low - check if anonymous pages need to be deactivated | |
c56d5c7d | 1602 | * @lruvec: LRU vector to check |
14797e23 KM |
1603 | * |
1604 | * Returns true if the zone does not have enough inactive anon pages, | |
1605 | * meaning some active anon pages need to be deactivated. | |
1606 | */ | |
c56d5c7d | 1607 | static int inactive_anon_is_low(struct lruvec *lruvec) |
14797e23 | 1608 | { |
74e3f3c3 MK |
1609 | /* |
1610 | * If we don't have swap space, anonymous page deactivation | |
1611 | * is pointless. | |
1612 | */ | |
1613 | if (!total_swap_pages) | |
1614 | return 0; | |
1615 | ||
c3c787e8 | 1616 | if (!mem_cgroup_disabled()) |
c56d5c7d | 1617 | return mem_cgroup_inactive_anon_is_low(lruvec); |
f16015fb | 1618 | |
c56d5c7d | 1619 | return inactive_anon_is_low_global(lruvec_zone(lruvec)); |
14797e23 | 1620 | } |
74e3f3c3 | 1621 | #else |
c56d5c7d | 1622 | static inline int inactive_anon_is_low(struct lruvec *lruvec) |
74e3f3c3 MK |
1623 | { |
1624 | return 0; | |
1625 | } | |
1626 | #endif | |
14797e23 | 1627 | |
56e49d21 RR |
1628 | /** |
1629 | * inactive_file_is_low - check if file pages need to be deactivated | |
c56d5c7d | 1630 | * @lruvec: LRU vector to check |
56e49d21 RR |
1631 | * |
1632 | * When the system is doing streaming IO, memory pressure here | |
1633 | * ensures that active file pages get deactivated, until more | |
1634 | * than half of the file pages are on the inactive list. | |
1635 | * | |
1636 | * Once we get to that situation, protect the system's working | |
1637 | * set from being evicted by disabling active file page aging. | |
1638 | * | |
1639 | * This uses a different ratio than the anonymous pages, because | |
1640 | * the page cache uses a use-once replacement algorithm. | |
1641 | */ | |
c56d5c7d | 1642 | static int inactive_file_is_low(struct lruvec *lruvec) |
56e49d21 | 1643 | { |
e3790144 JW |
1644 | unsigned long inactive; |
1645 | unsigned long active; | |
1646 | ||
1647 | inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
1648 | active = get_lru_size(lruvec, LRU_ACTIVE_FILE); | |
56e49d21 | 1649 | |
e3790144 | 1650 | return active > inactive; |
56e49d21 RR |
1651 | } |
1652 | ||
75b00af7 | 1653 | static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) |
b39415b2 | 1654 | { |
75b00af7 | 1655 | if (is_file_lru(lru)) |
c56d5c7d | 1656 | return inactive_file_is_low(lruvec); |
b39415b2 | 1657 | else |
c56d5c7d | 1658 | return inactive_anon_is_low(lruvec); |
b39415b2 RR |
1659 | } |
1660 | ||
4f98a2fe | 1661 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
1a93be0e | 1662 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 1663 | { |
b39415b2 | 1664 | if (is_active_lru(lru)) { |
75b00af7 | 1665 | if (inactive_list_is_low(lruvec, lru)) |
1a93be0e | 1666 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
1667 | return 0; |
1668 | } | |
1669 | ||
1a93be0e | 1670 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
1671 | } |
1672 | ||
3d58ab5c | 1673 | static int vmscan_swappiness(struct scan_control *sc) |
1f4c025b | 1674 | { |
89b5fae5 | 1675 | if (global_reclaim(sc)) |
1f4c025b | 1676 | return vm_swappiness; |
3d58ab5c | 1677 | return mem_cgroup_swappiness(sc->target_mem_cgroup); |
1f4c025b KH |
1678 | } |
1679 | ||
9a265114 JW |
1680 | enum scan_balance { |
1681 | SCAN_EQUAL, | |
1682 | SCAN_FRACT, | |
1683 | SCAN_ANON, | |
1684 | SCAN_FILE, | |
1685 | }; | |
1686 | ||
6fa3eb70 S |
1687 | |
1688 | #ifdef CONFIG_ZRAM | |
1689 | static int vmscan_swap_file_ratio = 1; | |
1690 | module_param_named(swap_file_ratio, vmscan_swap_file_ratio, int, S_IRUGO | S_IWUSR); | |
1691 | ||
1692 | #if defined(CONFIG_ZRAM) && defined(CONFIG_MTK_LCA_RAM_OPTIMIZE) | |
1693 | ||
1694 | // vmscan debug | |
1695 | static int vmscan_swap_sum = 200; | |
1696 | module_param_named(swap_sum, vmscan_swap_sum, int, S_IRUGO | S_IWUSR); | |
1697 | ||
1698 | ||
1699 | static int vmscan_scan_file_sum = 0; | |
1700 | static int vmscan_scan_anon_sum = 0; | |
1701 | static int vmscan_recent_scanned_anon = 0; | |
1702 | static int vmscan_recent_scanned_file = 0; | |
1703 | static int vmscan_recent_rotated_anon = 0; | |
1704 | static int vmscan_recent_rotated_file = 0; | |
1705 | module_param_named(scan_file_sum, vmscan_scan_file_sum, int, S_IRUGO); | |
1706 | module_param_named(scan_anon_sum, vmscan_scan_anon_sum, int, S_IRUGO); | |
1707 | module_param_named(recent_scanned_anon, vmscan_recent_scanned_anon, int, S_IRUGO); | |
1708 | module_param_named(recent_scanned_file, vmscan_recent_scanned_file, int, S_IRUGO); | |
1709 | module_param_named(recent_rotated_anon, vmscan_recent_rotated_anon, int, S_IRUGO); | |
1710 | module_param_named(recent_rotated_file, vmscan_recent_rotated_file, int, S_IRUGO); | |
1711 | #endif // CONFIG_ZRAM | |
1712 | ||
1713 | ||
1714 | #if defined(CONFIG_ZRAM) && defined(CONFIG_MTK_LCA_RAM_OPTIMIZE) | |
1715 | //#define LOGTAG "VMSCAN" | |
1716 | static unsigned long t=0; | |
1717 | static unsigned long history[2] = {0}; | |
1718 | extern int lowmem_minfree[9]; | |
1719 | #endif | |
1720 | ||
1721 | #endif // CONFIG_ZRAM | |
1722 | ||
4f98a2fe RR |
1723 | /* |
1724 | * Determine how aggressively the anon and file LRU lists should be | |
1725 | * scanned. The relative value of each set of LRU lists is determined | |
1726 | * by looking at the fraction of the pages scanned we did rotate back | |
1727 | * onto the active list instead of evict. | |
1728 | * | |
be7bd59d WL |
1729 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
1730 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 1731 | */ |
90126375 | 1732 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
9e3b2f8c | 1733 | unsigned long *nr) |
4f98a2fe | 1734 | { |
9a265114 JW |
1735 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
1736 | u64 fraction[2]; | |
1737 | u64 denominator = 0; /* gcc */ | |
1738 | struct zone *zone = lruvec_zone(lruvec); | |
4f98a2fe | 1739 | unsigned long anon_prio, file_prio; |
9a265114 JW |
1740 | enum scan_balance scan_balance; |
1741 | unsigned long anon, file, free; | |
1742 | bool force_scan = false; | |
4f98a2fe | 1743 | unsigned long ap, fp; |
4111304d | 1744 | enum lru_list lru; |
6fa3eb70 S |
1745 | #if defined(CONFIG_ZRAM) && defined(CONFIG_MTK_LCA_RAM_OPTIMIZE) |
1746 | int cpu; | |
1747 | unsigned long SwapinCount, SwapoutCount, cached; | |
1748 | bool bThrashing = false; | |
1749 | #endif | |
246e87a9 | 1750 | |
f11c0ca5 JW |
1751 | /* |
1752 | * If the zone or memcg is small, nr[l] can be 0. This | |
1753 | * results in no scanning on this priority and a potential | |
1754 | * priority drop. Global direct reclaim can go to the next | |
1755 | * zone and tends to have no problems. Global kswapd is for | |
1756 | * zone balancing and it needs to scan a minimum amount. When | |
1757 | * reclaiming for a memcg, a priority drop can cause high | |
1758 | * latencies, so it's better to scan a minimum amount there as | |
1759 | * well. | |
1760 | */ | |
90126375 | 1761 | if (current_is_kswapd() && zone->all_unreclaimable) |
a4d3e9e7 | 1762 | force_scan = true; |
89b5fae5 | 1763 | if (!global_reclaim(sc)) |
a4d3e9e7 | 1764 | force_scan = true; |
76a33fc3 SL |
1765 | |
1766 | /* If we have no swap space, do not bother scanning anon pages. */ | |
ec8acf20 | 1767 | if (!sc->may_swap || (get_nr_swap_pages() <= 0)) { |
9a265114 | 1768 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
1769 | goto out; |
1770 | } | |
4f98a2fe | 1771 | |
10316b31 JW |
1772 | /* |
1773 | * Global reclaim will swap to prevent OOM even with no | |
1774 | * swappiness, but memcg users want to use this knob to | |
1775 | * disable swapping for individual groups completely when | |
1776 | * using the memory controller's swap limit feature would be | |
1777 | * too expensive. | |
1778 | */ | |
1779 | if (!global_reclaim(sc) && !vmscan_swappiness(sc)) { | |
9a265114 | 1780 | scan_balance = SCAN_FILE; |
10316b31 JW |
1781 | goto out; |
1782 | } | |
1783 | ||
1784 | /* | |
1785 | * Do not apply any pressure balancing cleverness when the | |
1786 | * system is close to OOM, scan both anon and file equally | |
1787 | * (unless the swappiness setting disagrees with swapping). | |
1788 | */ | |
1789 | if (!sc->priority && vmscan_swappiness(sc)) { | |
9a265114 | 1790 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
1791 | goto out; |
1792 | } | |
1793 | ||
4d7dcca2 HD |
1794 | anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + |
1795 | get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
1796 | file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + | |
1797 | get_lru_size(lruvec, LRU_INACTIVE_FILE); | |
a4d3e9e7 | 1798 | |
11d16c25 JW |
1799 | /* |
1800 | * If it's foreseeable that reclaiming the file cache won't be | |
1801 | * enough to get the zone back into a desirable shape, we have | |
1802 | * to swap. Better start now and leave the - probably heavily | |
1803 | * thrashing - remaining file pages alone. | |
1804 | */ | |
89b5fae5 | 1805 | if (global_reclaim(sc)) { |
11d16c25 | 1806 | free = zone_page_state(zone, NR_FREE_PAGES); |
90126375 | 1807 | if (unlikely(file + free <= high_wmark_pages(zone))) { |
9a265114 | 1808 | scan_balance = SCAN_ANON; |
76a33fc3 | 1809 | goto out; |
eeee9a8c | 1810 | } |
4f98a2fe RR |
1811 | } |
1812 | ||
7c5bd705 JW |
1813 | /* |
1814 | * There is enough inactive page cache, do not reclaim | |
1815 | * anything from the anonymous working set right now. | |
1816 | */ | |
1817 | if (!inactive_file_is_low(lruvec)) { | |
9a265114 | 1818 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
1819 | goto out; |
1820 | } | |
1821 | ||
9a265114 JW |
1822 | scan_balance = SCAN_FRACT; |
1823 | ||
58c37f6e KM |
1824 | /* |
1825 | * With swappiness at 100, anonymous and file have the same priority. | |
1826 | * This scanning priority is essentially the inverse of IO cost. | |
1827 | */ | |
3d58ab5c | 1828 | anon_prio = vmscan_swappiness(sc); |
75b00af7 | 1829 | file_prio = 200 - anon_prio; |
58c37f6e | 1830 | |
6fa3eb70 S |
1831 | /* |
1832 | * With swappiness at 100, anonymous and file have the same priority. | |
1833 | * This scanning priority is essentially the inverse of IO cost. | |
1834 | */ | |
1835 | #if defined(CONFIG_ZRAM) && defined(CONFIG_MTK_LCA_RAM_OPTIMIZE) | |
1836 | if (vmscan_swap_file_ratio) { | |
1837 | ||
1838 | if(t == 0) | |
1839 | t = jiffies; | |
1840 | ||
1841 | if (time_after(jiffies, t + 1 * HZ)) { | |
1842 | ||
1843 | for_each_online_cpu(cpu) { | |
1844 | struct vm_event_state *this = &per_cpu(vm_event_states, cpu); | |
1845 | SwapinCount += this->event[PSWPIN]; | |
1846 | SwapoutCount += this->event[PSWPOUT]; | |
1847 | } | |
1848 | ||
1849 | if( ((SwapinCount-history[0] + SwapoutCount - history[1]) / (jiffies-t) * HZ) > 3000){ | |
1850 | bThrashing = true; | |
1851 | //xlog_printk(ANDROID_LOG_ERROR, LOGTAG, "!!! thrashing !!!\n"); | |
1852 | }else{ | |
1853 | bThrashing = false; | |
1854 | //xlog_printk(ANDROID_LOG_WARN, LOGTAG, "!!! NO thrashing !!!\n"); | |
1855 | } | |
1856 | history[0] = SwapinCount; | |
1857 | history[1] = SwapoutCount; | |
1858 | ||
1859 | ||
1860 | t=jiffies; | |
1861 | } | |
1862 | ||
1863 | ||
1864 | if(!bThrashing){ | |
1865 | anon_prio = (vmscan_swappiness(sc) * anon) / (anon + file + 1); | |
1866 | file_prio = (vmscan_swap_sum - vmscan_swappiness(sc)) * file / (anon + file + 1); | |
1867 | //xlog_printk(ANDROID_LOG_DEBUG, LOGTAG, "1 anon_prio: %d, file_prio: %d \n", anon_prio, file_prio); | |
1868 | ||
1869 | } else { | |
1870 | cached = global_page_state(NR_FILE_PAGES) - global_page_state(NR_SHMEM) - total_swapcache_pages(); | |
1871 | if(cached > lowmem_minfree[2]) { | |
1872 | anon_prio = vmscan_swappiness(sc); | |
1873 | file_prio = vmscan_swap_sum - vmscan_swappiness(sc); | |
1874 | //xlog_printk(ANDROID_LOG_ERROR, LOGTAG, "2 anon_prio: %d, file_prio: %d \n", anon_prio, file_prio); | |
1875 | } else { | |
1876 | anon_prio = (vmscan_swappiness(sc) * anon) / (anon + file + 1); | |
1877 | file_prio = (vmscan_swap_sum - vmscan_swappiness(sc)) * file / (anon + file + 1); | |
1878 | //xlog_printk(ANDROID_LOG_ERROR, LOGTAG, "3 anon_prio: %d, file_prio: %d \n", anon_prio, file_prio); | |
1879 | } | |
1880 | } | |
1881 | ||
1882 | } else { | |
1883 | anon_prio = vmscan_swappiness(sc); | |
1884 | file_prio = vmscan_swap_sum - vmscan_swappiness(sc); | |
1885 | } | |
1886 | #elif defined(CONFIG_ZRAM) // CONFIG_ZRAM | |
1887 | if (vmscan_swap_file_ratio) { | |
1888 | anon_prio = anon_prio * anon / (anon + file + 1); | |
1889 | file_prio = file_prio * file / (anon + file + 1); | |
1890 | } | |
1891 | #endif // CONFIG_ZRAM | |
1892 | ||
1893 | ||
1894 | ||
4f98a2fe RR |
1895 | /* |
1896 | * OK, so we have swap space and a fair amount of page cache | |
1897 | * pages. We use the recently rotated / recently scanned | |
1898 | * ratios to determine how valuable each cache is. | |
1899 | * | |
1900 | * Because workloads change over time (and to avoid overflow) | |
1901 | * we keep these statistics as a floating average, which ends | |
1902 | * up weighing recent references more than old ones. | |
1903 | * | |
1904 | * anon in [0], file in [1] | |
1905 | */ | |
90126375 | 1906 | spin_lock_irq(&zone->lru_lock); |
6e901571 | 1907 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
1908 | reclaim_stat->recent_scanned[0] /= 2; |
1909 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
1910 | } |
1911 | ||
6e901571 | 1912 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
1913 | reclaim_stat->recent_scanned[1] /= 2; |
1914 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
1915 | } |
1916 | ||
4f98a2fe | 1917 | /* |
00d8089c RR |
1918 | * The amount of pressure on anon vs file pages is inversely |
1919 | * proportional to the fraction of recently scanned pages on | |
1920 | * each list that were recently referenced and in active use. | |
4f98a2fe | 1921 | */ |
fe35004f | 1922 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 1923 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 1924 | |
fe35004f | 1925 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 1926 | fp /= reclaim_stat->recent_rotated[1] + 1; |
90126375 | 1927 | spin_unlock_irq(&zone->lru_lock); |
4f98a2fe | 1928 | |
76a33fc3 SL |
1929 | fraction[0] = ap; |
1930 | fraction[1] = fp; | |
1931 | denominator = ap + fp + 1; | |
1932 | out: | |
4111304d HD |
1933 | for_each_evictable_lru(lru) { |
1934 | int file = is_file_lru(lru); | |
d778df51 | 1935 | unsigned long size; |
76a33fc3 | 1936 | unsigned long scan; |
6e08a369 | 1937 | |
d778df51 | 1938 | size = get_lru_size(lruvec, lru); |
10316b31 | 1939 | scan = size >> sc->priority; |
9a265114 | 1940 | |
10316b31 JW |
1941 | if (!scan && force_scan) |
1942 | scan = min(size, SWAP_CLUSTER_MAX); | |
9a265114 JW |
1943 | |
1944 | switch (scan_balance) { | |
1945 | case SCAN_EQUAL: | |
1946 | /* Scan lists relative to size */ | |
1947 | break; | |
1948 | case SCAN_FRACT: | |
1949 | /* | |
1950 | * Scan types proportional to swappiness and | |
1951 | * their relative recent reclaim efficiency. | |
1952 | */ | |
1953 | scan = div64_u64(scan * fraction[file], denominator); | |
1954 | break; | |
1955 | case SCAN_FILE: | |
1956 | case SCAN_ANON: | |
1957 | /* Scan one type exclusively */ | |
1958 | if ((scan_balance == SCAN_FILE) != file) | |
1959 | scan = 0; | |
1960 | break; | |
1961 | default: | |
1962 | /* Look ma, no brain */ | |
1963 | BUG(); | |
1964 | } | |
4111304d | 1965 | nr[lru] = scan; |
76a33fc3 | 1966 | } |
6e08a369 | 1967 | } |
4f98a2fe | 1968 | |
9b4f98cd JW |
1969 | /* |
1970 | * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. | |
1971 | */ | |
1972 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) | |
1973 | { | |
1974 | unsigned long nr[NR_LRU_LISTS]; | |
1975 | unsigned long nr_to_scan; | |
1976 | enum lru_list lru; | |
1977 | unsigned long nr_reclaimed = 0; | |
1978 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
1979 | struct blk_plug plug; | |
1980 | ||
1981 | get_scan_count(lruvec, sc, nr); | |
1982 | ||
1983 | blk_start_plug(&plug); | |
1984 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
1985 | nr[LRU_INACTIVE_FILE]) { | |
1986 | for_each_evictable_lru(lru) { | |
1987 | if (nr[lru]) { | |
1988 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
1989 | nr[lru] -= nr_to_scan; | |
1990 | ||
1991 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
1992 | lruvec, sc); | |
1993 | } | |
1994 | } | |
1995 | /* | |
1996 | * On large memory systems, scan >> priority can become | |
1997 | * really large. This is fine for the starting priority; | |
1998 | * we want to put equal scanning pressure on each zone. | |
1999 | * However, if the VM has a harder time of freeing pages, | |
2000 | * with multiple processes reclaiming pages, the total | |
2001 | * freeing target can get unreasonably large. | |
2002 | */ | |
2003 | if (nr_reclaimed >= nr_to_reclaim && | |
2004 | sc->priority < DEF_PRIORITY) | |
2005 | break; | |
2006 | } | |
2007 | blk_finish_plug(&plug); | |
2008 | sc->nr_reclaimed += nr_reclaimed; | |
2009 | ||
2010 | /* | |
2011 | * Even if we did not try to evict anon pages at all, we want to | |
2012 | * rebalance the anon lru active/inactive ratio. | |
2013 | */ | |
2014 | if (inactive_anon_is_low(lruvec)) | |
2015 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
2016 | sc, LRU_ACTIVE_ANON); | |
2017 | ||
2018 | throttle_vm_writeout(sc->gfp_mask); | |
2019 | } | |
2020 | ||
23b9da55 | 2021 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2022 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2023 | { |
d84da3f9 | 2024 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2025 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2026 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2027 | return true; |
2028 | ||
2029 | return false; | |
2030 | } | |
2031 | ||
3e7d3449 | 2032 | /* |
23b9da55 MG |
2033 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2034 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2035 | * true if more pages should be reclaimed such that when the page allocator | |
2036 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2037 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2038 | */ |
9b4f98cd | 2039 | static inline bool should_continue_reclaim(struct zone *zone, |
3e7d3449 MG |
2040 | unsigned long nr_reclaimed, |
2041 | unsigned long nr_scanned, | |
2042 | struct scan_control *sc) | |
2043 | { | |
2044 | unsigned long pages_for_compaction; | |
2045 | unsigned long inactive_lru_pages; | |
2046 | ||
2047 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2048 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2049 | return false; |
2050 | ||
2876592f MG |
2051 | /* Consider stopping depending on scan and reclaim activity */ |
2052 | if (sc->gfp_mask & __GFP_REPEAT) { | |
2053 | /* | |
2054 | * For __GFP_REPEAT allocations, stop reclaiming if the | |
2055 | * full LRU list has been scanned and we are still failing | |
2056 | * to reclaim pages. This full LRU scan is potentially | |
2057 | * expensive but a __GFP_REPEAT caller really wants to succeed | |
2058 | */ | |
2059 | if (!nr_reclaimed && !nr_scanned) | |
2060 | return false; | |
2061 | } else { | |
2062 | /* | |
2063 | * For non-__GFP_REPEAT allocations which can presumably | |
2064 | * fail without consequence, stop if we failed to reclaim | |
2065 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2066 | * pages that were scanned. This will return to the | |
2067 | * caller faster at the risk reclaim/compaction and | |
2068 | * the resulting allocation attempt fails | |
2069 | */ | |
2070 | if (!nr_reclaimed) | |
2071 | return false; | |
2072 | } | |
3e7d3449 MG |
2073 | |
2074 | /* | |
2075 | * If we have not reclaimed enough pages for compaction and the | |
2076 | * inactive lists are large enough, continue reclaiming | |
2077 | */ | |
2078 | pages_for_compaction = (2UL << sc->order); | |
9b4f98cd | 2079 | inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); |
ec8acf20 | 2080 | if (get_nr_swap_pages() > 0) |
9b4f98cd | 2081 | inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); |
3e7d3449 MG |
2082 | if (sc->nr_reclaimed < pages_for_compaction && |
2083 | inactive_lru_pages > pages_for_compaction) | |
2084 | return true; | |
2085 | ||
2086 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
9b4f98cd | 2087 | switch (compaction_suitable(zone, sc->order)) { |
3e7d3449 MG |
2088 | case COMPACT_PARTIAL: |
2089 | case COMPACT_CONTINUE: | |
2090 | return false; | |
2091 | default: | |
2092 | return true; | |
2093 | } | |
2094 | } | |
2095 | ||
9b4f98cd | 2096 | static void shrink_zone(struct zone *zone, struct scan_control *sc) |
1da177e4 | 2097 | { |
f0fdc5e8 | 2098 | unsigned long nr_reclaimed, nr_scanned; |
1da177e4 | 2099 | |
9b4f98cd JW |
2100 | do { |
2101 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2102 | struct mem_cgroup_reclaim_cookie reclaim = { | |
2103 | .zone = zone, | |
2104 | .priority = sc->priority, | |
2105 | }; | |
2106 | struct mem_cgroup *memcg; | |
3e7d3449 | 2107 | |
9b4f98cd JW |
2108 | nr_reclaimed = sc->nr_reclaimed; |
2109 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2110 | |
9b4f98cd JW |
2111 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2112 | do { | |
2113 | struct lruvec *lruvec; | |
5660048c | 2114 | |
9b4f98cd | 2115 | lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
f9be23d6 | 2116 | |
9b4f98cd | 2117 | shrink_lruvec(lruvec, sc); |
f16015fb | 2118 | |
9b4f98cd | 2119 | /* |
a394cb8e MH |
2120 | * Direct reclaim and kswapd have to scan all memory |
2121 | * cgroups to fulfill the overall scan target for the | |
9b4f98cd | 2122 | * zone. |
a394cb8e MH |
2123 | * |
2124 | * Limit reclaim, on the other hand, only cares about | |
2125 | * nr_to_reclaim pages to be reclaimed and it will | |
2126 | * retry with decreasing priority if one round over the | |
2127 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2128 | */ |
a394cb8e MH |
2129 | if (!global_reclaim(sc) && |
2130 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2131 | mem_cgroup_iter_break(root, memcg); |
2132 | break; | |
2133 | } | |
2134 | memcg = mem_cgroup_iter(root, memcg, &reclaim); | |
2135 | } while (memcg); | |
70ddf637 AV |
2136 | |
2137 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, | |
2138 | sc->nr_scanned - nr_scanned, | |
2139 | sc->nr_reclaimed - nr_reclaimed); | |
2140 | ||
9b4f98cd JW |
2141 | } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, |
2142 | sc->nr_scanned - nr_scanned, sc)); | |
f16015fb JW |
2143 | } |
2144 | ||
fe4b1b24 MG |
2145 | /* Returns true if compaction should go ahead for a high-order request */ |
2146 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) | |
2147 | { | |
2148 | unsigned long balance_gap, watermark; | |
2149 | bool watermark_ok; | |
2150 | ||
2151 | /* Do not consider compaction for orders reclaim is meant to satisfy */ | |
2152 | if (sc->order <= PAGE_ALLOC_COSTLY_ORDER) | |
2153 | return false; | |
2154 | ||
2155 | /* | |
2156 | * Compaction takes time to run and there are potentially other | |
2157 | * callers using the pages just freed. Continue reclaiming until | |
2158 | * there is a buffer of free pages available to give compaction | |
2159 | * a reasonable chance of completing and allocating the page | |
2160 | */ | |
2161 | balance_gap = min(low_wmark_pages(zone), | |
b40da049 | 2162 | (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / |
fe4b1b24 MG |
2163 | KSWAPD_ZONE_BALANCE_GAP_RATIO); |
2164 | watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order); | |
2165 | watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); | |
2166 | ||
2167 | /* | |
2168 | * If compaction is deferred, reclaim up to a point where | |
2169 | * compaction will have a chance of success when re-enabled | |
2170 | */ | |
aff62249 | 2171 | if (compaction_deferred(zone, sc->order)) |
fe4b1b24 MG |
2172 | return watermark_ok; |
2173 | ||
2174 | /* If compaction is not ready to start, keep reclaiming */ | |
2175 | if (!compaction_suitable(zone, sc->order)) | |
2176 | return false; | |
2177 | ||
2178 | return watermark_ok; | |
2179 | } | |
2180 | ||
1da177e4 LT |
2181 | /* |
2182 | * This is the direct reclaim path, for page-allocating processes. We only | |
2183 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2184 | * request. | |
2185 | * | |
41858966 MG |
2186 | * We reclaim from a zone even if that zone is over high_wmark_pages(zone). |
2187 | * Because: | |
1da177e4 LT |
2188 | * a) The caller may be trying to free *extra* pages to satisfy a higher-order |
2189 | * allocation or | |
41858966 MG |
2190 | * b) The target zone may be at high_wmark_pages(zone) but the lower zones |
2191 | * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' | |
2192 | * zone defense algorithm. | |
1da177e4 | 2193 | * |
1da177e4 LT |
2194 | * If a zone is deemed to be full of pinned pages then just give it a light |
2195 | * scan then give up on it. | |
e0c23279 MG |
2196 | * |
2197 | * This function returns true if a zone is being reclaimed for a costly | |
fe4b1b24 | 2198 | * high-order allocation and compaction is ready to begin. This indicates to |
0cee34fd MG |
2199 | * the caller that it should consider retrying the allocation instead of |
2200 | * further reclaim. | |
1da177e4 | 2201 | */ |
9e3b2f8c | 2202 | static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2203 | { |
dd1a239f | 2204 | struct zoneref *z; |
54a6eb5c | 2205 | struct zone *zone; |
d149e3b2 YH |
2206 | unsigned long nr_soft_reclaimed; |
2207 | unsigned long nr_soft_scanned; | |
0cee34fd | 2208 | bool aborted_reclaim = false; |
1cfb419b | 2209 | |
cc715d99 MG |
2210 | /* |
2211 | * If the number of buffer_heads in the machine exceeds the maximum | |
2212 | * allowed level, force direct reclaim to scan the highmem zone as | |
2213 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2214 | */ | |
2215 | if (buffer_heads_over_limit) | |
2216 | sc->gfp_mask |= __GFP_HIGHMEM; | |
2217 | ||
d4debc66 MG |
2218 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
2219 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
f3fe6512 | 2220 | if (!populated_zone(zone)) |
1da177e4 | 2221 | continue; |
1cfb419b KH |
2222 | /* |
2223 | * Take care memory controller reclaiming has small influence | |
2224 | * to global LRU. | |
2225 | */ | |
89b5fae5 | 2226 | if (global_reclaim(sc)) { |
1cfb419b KH |
2227 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2228 | continue; | |
9e3b2f8c KK |
2229 | if (zone->all_unreclaimable && |
2230 | sc->priority != DEF_PRIORITY) | |
1cfb419b | 2231 | continue; /* Let kswapd poll it */ |
6fa3eb70 | 2232 | if (IS_ENABLED(CONFIG_COMPACTION) && !sc->hibernation_mode) { |
e0887c19 | 2233 | /* |
e0c23279 MG |
2234 | * If we already have plenty of memory free for |
2235 | * compaction in this zone, don't free any more. | |
2236 | * Even though compaction is invoked for any | |
2237 | * non-zero order, only frequent costly order | |
2238 | * reclamation is disruptive enough to become a | |
c7cfa37b CA |
2239 | * noticeable problem, like transparent huge |
2240 | * page allocations. | |
e0887c19 | 2241 | */ |
fe4b1b24 | 2242 | if (compaction_ready(zone, sc)) { |
0cee34fd | 2243 | aborted_reclaim = true; |
e0887c19 | 2244 | continue; |
e0c23279 | 2245 | } |
e0887c19 | 2246 | } |
ac34a1a3 KH |
2247 | /* |
2248 | * This steals pages from memory cgroups over softlimit | |
2249 | * and returns the number of reclaimed pages and | |
2250 | * scanned pages. This works for global memory pressure | |
2251 | * and balancing, not for a memcg's limit. | |
2252 | */ | |
2253 | nr_soft_scanned = 0; | |
2254 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, | |
2255 | sc->order, sc->gfp_mask, | |
2256 | &nr_soft_scanned); | |
2257 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2258 | sc->nr_scanned += nr_soft_scanned; | |
2259 | /* need some check for avoid more shrink_zone() */ | |
1cfb419b | 2260 | } |
408d8544 | 2261 | |
9e3b2f8c | 2262 | shrink_zone(zone, sc); |
1da177e4 | 2263 | } |
e0c23279 | 2264 | |
0cee34fd | 2265 | return aborted_reclaim; |
d1908362 MK |
2266 | } |
2267 | ||
48526149 JW |
2268 | static unsigned long zone_reclaimable_pages(struct zone *zone) |
2269 | { | |
2270 | int nr; | |
2271 | ||
2272 | nr = zone_page_state(zone, NR_ACTIVE_FILE) + | |
2273 | zone_page_state(zone, NR_INACTIVE_FILE); | |
2274 | ||
2275 | if (get_nr_swap_pages() > 0) | |
2276 | nr += zone_page_state(zone, NR_ACTIVE_ANON) + | |
2277 | zone_page_state(zone, NR_INACTIVE_ANON); | |
2278 | ||
2279 | return nr; | |
2280 | } | |
2281 | ||
d1908362 MK |
2282 | static bool zone_reclaimable(struct zone *zone) |
2283 | { | |
2284 | return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; | |
2285 | } | |
2286 | ||
929bea7c | 2287 | /* All zones in zonelist are unreclaimable? */ |
d1908362 MK |
2288 | static bool all_unreclaimable(struct zonelist *zonelist, |
2289 | struct scan_control *sc) | |
2290 | { | |
2291 | struct zoneref *z; | |
2292 | struct zone *zone; | |
d1908362 MK |
2293 | |
2294 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2295 | gfp_zone(sc->gfp_mask), sc->nodemask) { | |
2296 | if (!populated_zone(zone)) | |
2297 | continue; | |
2298 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) | |
2299 | continue; | |
929bea7c KM |
2300 | if (!zone->all_unreclaimable) |
2301 | return false; | |
d1908362 MK |
2302 | } |
2303 | ||
929bea7c | 2304 | return true; |
1da177e4 | 2305 | } |
4f98a2fe | 2306 | |
1da177e4 LT |
2307 | /* |
2308 | * This is the main entry point to direct page reclaim. | |
2309 | * | |
2310 | * If a full scan of the inactive list fails to free enough memory then we | |
2311 | * are "out of memory" and something needs to be killed. | |
2312 | * | |
2313 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2314 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2315 | * caller can't do much about. We kick the writeback threads and take explicit |
2316 | * naps in the hope that some of these pages can be written. But if the | |
2317 | * allocating task holds filesystem locks which prevent writeout this might not | |
2318 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2319 | * |
2320 | * returns: 0, if no pages reclaimed | |
2321 | * else, the number of pages reclaimed | |
1da177e4 | 2322 | */ |
dac1d27b | 2323 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
a09ed5e0 YH |
2324 | struct scan_control *sc, |
2325 | struct shrink_control *shrink) | |
1da177e4 | 2326 | { |
69e05944 | 2327 | unsigned long total_scanned = 0; |
1da177e4 | 2328 | struct reclaim_state *reclaim_state = current->reclaim_state; |
dd1a239f | 2329 | struct zoneref *z; |
54a6eb5c | 2330 | struct zone *zone; |
22fba335 | 2331 | unsigned long writeback_threshold; |
0cee34fd | 2332 | bool aborted_reclaim; |
1da177e4 | 2333 | |
6fa3eb70 S |
2334 | #ifdef CONFIG_FREEZER |
2335 | if (unlikely(pm_freezing && !sc->hibernation_mode)) | |
2336 | return 0; | |
2337 | #endif | |
2338 | ||
873b4771 KK |
2339 | delayacct_freepages_start(); |
2340 | ||
89b5fae5 | 2341 | if (global_reclaim(sc)) |
1cfb419b | 2342 | count_vm_event(ALLOCSTALL); |
1da177e4 | 2343 | |
9e3b2f8c | 2344 | do { |
70ddf637 AV |
2345 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2346 | sc->priority); | |
66e1707b | 2347 | sc->nr_scanned = 0; |
9e3b2f8c | 2348 | aborted_reclaim = shrink_zones(zonelist, sc); |
e0c23279 | 2349 | |
66e1707b BS |
2350 | /* |
2351 | * Don't shrink slabs when reclaiming memory from | |
2352 | * over limit cgroups | |
2353 | */ | |
89b5fae5 | 2354 | if (global_reclaim(sc)) { |
c6a8a8c5 | 2355 | unsigned long lru_pages = 0; |
d4debc66 MG |
2356 | for_each_zone_zonelist(zone, z, zonelist, |
2357 | gfp_zone(sc->gfp_mask)) { | |
c6a8a8c5 KM |
2358 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
2359 | continue; | |
2360 | ||
2361 | lru_pages += zone_reclaimable_pages(zone); | |
2362 | } | |
2363 | ||
1495f230 | 2364 | shrink_slab(shrink, sc->nr_scanned, lru_pages); |
91a45470 | 2365 | if (reclaim_state) { |
a79311c1 | 2366 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; |
91a45470 KH |
2367 | reclaim_state->reclaimed_slab = 0; |
2368 | } | |
1da177e4 | 2369 | } |
66e1707b | 2370 | total_scanned += sc->nr_scanned; |
bb21c7ce | 2371 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
1da177e4 | 2372 | goto out; |
1da177e4 | 2373 | |
0e50ce3b MK |
2374 | /* |
2375 | * If we're getting trouble reclaiming, start doing | |
2376 | * writepage even in laptop mode. | |
2377 | */ | |
2378 | if (sc->priority < DEF_PRIORITY - 2) | |
2379 | sc->may_writepage = 1; | |
2380 | ||
1da177e4 LT |
2381 | /* |
2382 | * Try to write back as many pages as we just scanned. This | |
2383 | * tends to cause slow streaming writers to write data to the | |
2384 | * disk smoothly, at the dirtying rate, which is nice. But | |
2385 | * that's undesirable in laptop mode, where we *want* lumpy | |
2386 | * writeout. So in laptop mode, write out the whole world. | |
2387 | */ | |
22fba335 KM |
2388 | writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; |
2389 | if (total_scanned > writeback_threshold) { | |
0e175a18 CW |
2390 | wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, |
2391 | WB_REASON_TRY_TO_FREE_PAGES); | |
66e1707b | 2392 | sc->may_writepage = 1; |
1da177e4 LT |
2393 | } |
2394 | ||
2395 | /* Take a nap, wait for some writeback to complete */ | |
7b51755c | 2396 | if (!sc->hibernation_mode && sc->nr_scanned && |
9e3b2f8c | 2397 | sc->priority < DEF_PRIORITY - 2) { |
0e093d99 MG |
2398 | struct zone *preferred_zone; |
2399 | ||
2400 | first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), | |
f33261d7 DR |
2401 | &cpuset_current_mems_allowed, |
2402 | &preferred_zone); | |
0e093d99 MG |
2403 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); |
2404 | } | |
9e3b2f8c | 2405 | } while (--sc->priority >= 0); |
bb21c7ce | 2406 | |
1da177e4 | 2407 | out: |
873b4771 KK |
2408 | delayacct_freepages_end(); |
2409 | ||
bb21c7ce KM |
2410 | if (sc->nr_reclaimed) |
2411 | return sc->nr_reclaimed; | |
2412 | ||
929bea7c KM |
2413 | /* |
2414 | * As hibernation is going on, kswapd is freezed so that it can't mark | |
2415 | * the zone into all_unreclaimable. Thus bypassing all_unreclaimable | |
2416 | * check. | |
2417 | */ | |
2418 | if (oom_killer_disabled) | |
2419 | return 0; | |
2420 | ||
0cee34fd MG |
2421 | /* Aborted reclaim to try compaction? don't OOM, then */ |
2422 | if (aborted_reclaim) | |
7335084d MG |
2423 | return 1; |
2424 | ||
bb21c7ce | 2425 | /* top priority shrink_zones still had more to do? don't OOM, then */ |
89b5fae5 | 2426 | if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc)) |
bb21c7ce KM |
2427 | return 1; |
2428 | ||
2429 | return 0; | |
1da177e4 LT |
2430 | } |
2431 | ||
5515061d MG |
2432 | static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) |
2433 | { | |
2434 | struct zone *zone; | |
2435 | unsigned long pfmemalloc_reserve = 0; | |
2436 | unsigned long free_pages = 0; | |
2437 | int i; | |
2438 | bool wmark_ok; | |
2439 | ||
2440 | for (i = 0; i <= ZONE_NORMAL; i++) { | |
2441 | zone = &pgdat->node_zones[i]; | |
16838de6 MG |
2442 | if (!populated_zone(zone)) |
2443 | continue; | |
2444 | ||
5515061d MG |
2445 | pfmemalloc_reserve += min_wmark_pages(zone); |
2446 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
2447 | } | |
2448 | ||
16838de6 MG |
2449 | /* If there are no reserves (unexpected config) then do not throttle */ |
2450 | if (!pfmemalloc_reserve) | |
2451 | return true; | |
2452 | ||
5515061d MG |
2453 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
2454 | ||
2455 | /* kswapd must be awake if processes are being throttled */ | |
2456 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
2457 | pgdat->classzone_idx = min(pgdat->classzone_idx, | |
2458 | (enum zone_type)ZONE_NORMAL); | |
2459 | wake_up_interruptible(&pgdat->kswapd_wait); | |
2460 | } | |
2461 | ||
2462 | return wmark_ok; | |
2463 | } | |
2464 | ||
2465 | /* | |
2466 | * Throttle direct reclaimers if backing storage is backed by the network | |
2467 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
2468 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
2469 | * when the low watermark is reached. |
2470 | * | |
2471 | * Returns true if a fatal signal was delivered during throttling. If this | |
2472 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 2473 | */ |
50694c28 | 2474 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
2475 | nodemask_t *nodemask) |
2476 | { | |
16838de6 | 2477 | struct zoneref *z; |
5515061d | 2478 | struct zone *zone; |
16838de6 | 2479 | pg_data_t *pgdat = NULL; |
5515061d MG |
2480 | |
2481 | /* | |
2482 | * Kernel threads should not be throttled as they may be indirectly | |
2483 | * responsible for cleaning pages necessary for reclaim to make forward | |
2484 | * progress. kjournald for example may enter direct reclaim while | |
2485 | * committing a transaction where throttling it could forcing other | |
2486 | * processes to block on log_wait_commit(). | |
2487 | */ | |
2488 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
2489 | goto out; |
2490 | ||
2491 | /* | |
2492 | * If a fatal signal is pending, this process should not throttle. | |
2493 | * It should return quickly so it can exit and free its memory | |
2494 | */ | |
2495 | if (fatal_signal_pending(current)) | |
2496 | goto out; | |
5515061d | 2497 | |
16838de6 MG |
2498 | /* |
2499 | * Check if the pfmemalloc reserves are ok by finding the first node | |
2500 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
2501 | * GFP_KERNEL will be required for allocating network buffers when | |
2502 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
2503 | * | |
2504 | * Throttling is based on the first usable node and throttled processes | |
2505 | * wait on a queue until kswapd makes progress and wakes them. There | |
2506 | * is an affinity then between processes waking up and where reclaim | |
2507 | * progress has been made assuming the process wakes on the same node. | |
2508 | * More importantly, processes running on remote nodes will not compete | |
2509 | * for remote pfmemalloc reserves and processes on different nodes | |
2510 | * should make reasonable progress. | |
2511 | */ | |
2512 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
2513 | gfp_mask, nodemask) { | |
2514 | if (zone_idx(zone) > ZONE_NORMAL) | |
2515 | continue; | |
2516 | ||
2517 | /* Throttle based on the first usable node */ | |
2518 | pgdat = zone->zone_pgdat; | |
2519 | if (pfmemalloc_watermark_ok(pgdat)) | |
2520 | goto out; | |
2521 | break; | |
2522 | } | |
2523 | ||
2524 | /* If no zone was usable by the allocation flags then do not throttle */ | |
2525 | if (!pgdat) | |
50694c28 | 2526 | goto out; |
5515061d | 2527 | |
68243e76 MG |
2528 | /* Account for the throttling */ |
2529 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
2530 | ||
5515061d MG |
2531 | /* |
2532 | * If the caller cannot enter the filesystem, it's possible that it | |
2533 | * is due to the caller holding an FS lock or performing a journal | |
2534 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
2535 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
2536 | * blocked waiting on the same lock. Instead, throttle for up to a | |
2537 | * second before continuing. | |
2538 | */ | |
2539 | if (!(gfp_mask & __GFP_FS)) { | |
2540 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
2541 | pfmemalloc_watermark_ok(pgdat), HZ); | |
50694c28 MG |
2542 | |
2543 | goto check_pending; | |
5515061d MG |
2544 | } |
2545 | ||
2546 | /* Throttle until kswapd wakes the process */ | |
2547 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
2548 | pfmemalloc_watermark_ok(pgdat)); | |
50694c28 MG |
2549 | |
2550 | check_pending: | |
2551 | if (fatal_signal_pending(current)) | |
2552 | return true; | |
2553 | ||
2554 | out: | |
2555 | return false; | |
5515061d MG |
2556 | } |
2557 | ||
dac1d27b | 2558 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 2559 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 2560 | { |
33906bc5 | 2561 | unsigned long nr_reclaimed; |
66e1707b | 2562 | struct scan_control sc = { |
21caf2fc | 2563 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
66e1707b | 2564 | .may_writepage = !laptop_mode, |
22fba335 | 2565 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
a6dc60f8 | 2566 | .may_unmap = 1, |
2e2e4259 | 2567 | .may_swap = 1, |
66e1707b | 2568 | .order = order, |
9e3b2f8c | 2569 | .priority = DEF_PRIORITY, |
f16015fb | 2570 | .target_mem_cgroup = NULL, |
327c0e96 | 2571 | .nodemask = nodemask, |
66e1707b | 2572 | }; |
a09ed5e0 YH |
2573 | struct shrink_control shrink = { |
2574 | .gfp_mask = sc.gfp_mask, | |
2575 | }; | |
66e1707b | 2576 | |
5515061d | 2577 | /* |
50694c28 MG |
2578 | * Do not enter reclaim if fatal signal was delivered while throttled. |
2579 | * 1 is returned so that the page allocator does not OOM kill at this | |
2580 | * point. | |
5515061d | 2581 | */ |
50694c28 | 2582 | if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) |
5515061d MG |
2583 | return 1; |
2584 | ||
33906bc5 MG |
2585 | trace_mm_vmscan_direct_reclaim_begin(order, |
2586 | sc.may_writepage, | |
2587 | gfp_mask); | |
2588 | ||
a09ed5e0 | 2589 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
33906bc5 MG |
2590 | |
2591 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
2592 | ||
2593 | return nr_reclaimed; | |
66e1707b BS |
2594 | } |
2595 | ||
c255a458 | 2596 | #ifdef CONFIG_MEMCG |
66e1707b | 2597 | |
72835c86 | 2598 | unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, |
4e416953 | 2599 | gfp_t gfp_mask, bool noswap, |
0ae5e89c YH |
2600 | struct zone *zone, |
2601 | unsigned long *nr_scanned) | |
4e416953 BS |
2602 | { |
2603 | struct scan_control sc = { | |
0ae5e89c | 2604 | .nr_scanned = 0, |
b8f5c566 | 2605 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
4e416953 BS |
2606 | .may_writepage = !laptop_mode, |
2607 | .may_unmap = 1, | |
2608 | .may_swap = !noswap, | |
4e416953 | 2609 | .order = 0, |
9e3b2f8c | 2610 | .priority = 0, |
72835c86 | 2611 | .target_mem_cgroup = memcg, |
4e416953 | 2612 | }; |
f9be23d6 | 2613 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
0ae5e89c | 2614 | |
4e416953 BS |
2615 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2616 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 2617 | |
9e3b2f8c | 2618 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e KM |
2619 | sc.may_writepage, |
2620 | sc.gfp_mask); | |
2621 | ||
4e416953 BS |
2622 | /* |
2623 | * NOTE: Although we can get the priority field, using it | |
2624 | * here is not a good idea, since it limits the pages we can scan. | |
2625 | * if we don't reclaim here, the shrink_zone from balance_pgdat | |
2626 | * will pick up pages from other mem cgroup's as well. We hack | |
2627 | * the priority and make it zero. | |
2628 | */ | |
f9be23d6 | 2629 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
2630 | |
2631 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
2632 | ||
0ae5e89c | 2633 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
2634 | return sc.nr_reclaimed; |
2635 | } | |
2636 | ||
72835c86 | 2637 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
a7885eb8 | 2638 | gfp_t gfp_mask, |
185efc0f | 2639 | bool noswap) |
66e1707b | 2640 | { |
4e416953 | 2641 | struct zonelist *zonelist; |
bdce6d9e | 2642 | unsigned long nr_reclaimed; |
889976db | 2643 | int nid; |
66e1707b | 2644 | struct scan_control sc = { |
66e1707b | 2645 | .may_writepage = !laptop_mode, |
a6dc60f8 | 2646 | .may_unmap = 1, |
2e2e4259 | 2647 | .may_swap = !noswap, |
22fba335 | 2648 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
66e1707b | 2649 | .order = 0, |
9e3b2f8c | 2650 | .priority = DEF_PRIORITY, |
72835c86 | 2651 | .target_mem_cgroup = memcg, |
327c0e96 | 2652 | .nodemask = NULL, /* we don't care the placement */ |
a09ed5e0 YH |
2653 | .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
2654 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), | |
2655 | }; | |
2656 | struct shrink_control shrink = { | |
2657 | .gfp_mask = sc.gfp_mask, | |
66e1707b | 2658 | }; |
66e1707b | 2659 | |
889976db YH |
2660 | /* |
2661 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
2662 | * take care of from where we get pages. So the node where we start the | |
2663 | * scan does not need to be the current node. | |
2664 | */ | |
72835c86 | 2665 | nid = mem_cgroup_select_victim_node(memcg); |
889976db YH |
2666 | |
2667 | zonelist = NODE_DATA(nid)->node_zonelists; | |
bdce6d9e KM |
2668 | |
2669 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
2670 | sc.may_writepage, | |
2671 | sc.gfp_mask); | |
2672 | ||
a09ed5e0 | 2673 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
bdce6d9e KM |
2674 | |
2675 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
2676 | ||
2677 | return nr_reclaimed; | |
66e1707b BS |
2678 | } |
2679 | #endif | |
2680 | ||
9e3b2f8c | 2681 | static void age_active_anon(struct zone *zone, struct scan_control *sc) |
f16015fb | 2682 | { |
b95a2f2d | 2683 | struct mem_cgroup *memcg; |
f16015fb | 2684 | |
b95a2f2d JW |
2685 | if (!total_swap_pages) |
2686 | return; | |
2687 | ||
2688 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
2689 | do { | |
c56d5c7d | 2690 | struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); |
b95a2f2d | 2691 | |
c56d5c7d | 2692 | if (inactive_anon_is_low(lruvec)) |
1a93be0e | 2693 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 2694 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
2695 | |
2696 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
2697 | } while (memcg); | |
f16015fb JW |
2698 | } |
2699 | ||
60cefed4 JW |
2700 | static bool zone_balanced(struct zone *zone, int order, |
2701 | unsigned long balance_gap, int classzone_idx) | |
2702 | { | |
2703 | if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + | |
2704 | balance_gap, classzone_idx, 0)) | |
2705 | return false; | |
2706 | ||
d84da3f9 KS |
2707 | if (IS_ENABLED(CONFIG_COMPACTION) && order && |
2708 | !compaction_suitable(zone, order)) | |
60cefed4 JW |
2709 | return false; |
2710 | ||
2711 | return true; | |
2712 | } | |
2713 | ||
1741c877 | 2714 | /* |
4ae0a48b ZC |
2715 | * pgdat_balanced() is used when checking if a node is balanced. |
2716 | * | |
2717 | * For order-0, all zones must be balanced! | |
2718 | * | |
2719 | * For high-order allocations only zones that meet watermarks and are in a | |
2720 | * zone allowed by the callers classzone_idx are added to balanced_pages. The | |
2721 | * total of balanced pages must be at least 25% of the zones allowed by | |
2722 | * classzone_idx for the node to be considered balanced. Forcing all zones to | |
2723 | * be balanced for high orders can cause excessive reclaim when there are | |
2724 | * imbalanced zones. | |
1741c877 MG |
2725 | * The choice of 25% is due to |
2726 | * o a 16M DMA zone that is balanced will not balance a zone on any | |
2727 | * reasonable sized machine | |
2728 | * o On all other machines, the top zone must be at least a reasonable | |
25985edc | 2729 | * percentage of the middle zones. For example, on 32-bit x86, highmem |
1741c877 MG |
2730 | * would need to be at least 256M for it to be balance a whole node. |
2731 | * Similarly, on x86-64 the Normal zone would need to be at least 1G | |
2732 | * to balance a node on its own. These seemed like reasonable ratios. | |
2733 | */ | |
4ae0a48b | 2734 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) |
1741c877 | 2735 | { |
b40da049 | 2736 | unsigned long managed_pages = 0; |
4ae0a48b | 2737 | unsigned long balanced_pages = 0; |
1741c877 MG |
2738 | int i; |
2739 | ||
4ae0a48b ZC |
2740 | /* Check the watermark levels */ |
2741 | for (i = 0; i <= classzone_idx; i++) { | |
2742 | struct zone *zone = pgdat->node_zones + i; | |
1741c877 | 2743 | |
4ae0a48b ZC |
2744 | if (!populated_zone(zone)) |
2745 | continue; | |
2746 | ||
b40da049 | 2747 | managed_pages += zone->managed_pages; |
4ae0a48b ZC |
2748 | |
2749 | /* | |
2750 | * A special case here: | |
2751 | * | |
2752 | * balance_pgdat() skips over all_unreclaimable after | |
2753 | * DEF_PRIORITY. Effectively, it considers them balanced so | |
2754 | * they must be considered balanced here as well! | |
2755 | */ | |
2756 | if (zone->all_unreclaimable) { | |
b40da049 | 2757 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
2758 | continue; |
2759 | } | |
2760 | ||
2761 | if (zone_balanced(zone, order, 0, i)) | |
b40da049 | 2762 | balanced_pages += zone->managed_pages; |
4ae0a48b ZC |
2763 | else if (!order) |
2764 | return false; | |
2765 | } | |
2766 | ||
2767 | if (order) | |
b40da049 | 2768 | return balanced_pages >= (managed_pages >> 2); |
4ae0a48b ZC |
2769 | else |
2770 | return true; | |
1741c877 MG |
2771 | } |
2772 | ||
5515061d MG |
2773 | /* |
2774 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
2775 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
2776 | * | |
2777 | * Returns true if kswapd is ready to sleep | |
2778 | */ | |
2779 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, | |
dc83edd9 | 2780 | int classzone_idx) |
f50de2d3 | 2781 | { |
f50de2d3 MG |
2782 | /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ |
2783 | if (remaining) | |
5515061d MG |
2784 | return false; |
2785 | ||
2786 | /* | |
6bb148fb VB |
2787 | * The throttled processes are normally woken up in balance_pgdat() as |
2788 | * soon as pfmemalloc_watermark_ok() is true. But there is a potential | |
2789 | * race between when kswapd checks the watermarks and a process gets | |
2790 | * throttled. There is also a potential race if processes get | |
2791 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
2792 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
2793 | * the wake up checks. If kswapd is going to sleep, no process should | |
2794 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
2795 | * the wake up is premature, processes will wake kswapd and get | |
2796 | * throttled again. The difference from wake ups in balance_pgdat() is | |
2797 | * that here we are under prepare_to_wait(). | |
5515061d | 2798 | */ |
6bb148fb VB |
2799 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
2800 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 2801 | |
4ae0a48b | 2802 | return pgdat_balanced(pgdat, order, classzone_idx); |
f50de2d3 MG |
2803 | } |
2804 | ||
1da177e4 LT |
2805 | /* |
2806 | * For kswapd, balance_pgdat() will work across all this node's zones until | |
41858966 | 2807 | * they are all at high_wmark_pages(zone). |
1da177e4 | 2808 | * |
0abdee2b | 2809 | * Returns the final order kswapd was reclaiming at |
1da177e4 LT |
2810 | * |
2811 | * There is special handling here for zones which are full of pinned pages. | |
2812 | * This can happen if the pages are all mlocked, or if they are all used by | |
2813 | * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. | |
2814 | * What we do is to detect the case where all pages in the zone have been | |
2815 | * scanned twice and there has been zero successful reclaim. Mark the zone as | |
2816 | * dead and from now on, only perform a short scan. Basically we're polling | |
2817 | * the zone for when the problem goes away. | |
2818 | * | |
2819 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 MG |
2820 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
2821 | * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the | |
2822 | * lower zones regardless of the number of free pages in the lower zones. This | |
2823 | * interoperates with the page allocator fallback scheme to ensure that aging | |
2824 | * of pages is balanced across the zones. | |
1da177e4 | 2825 | */ |
99504748 | 2826 | static unsigned long balance_pgdat(pg_data_t *pgdat, int order, |
dc83edd9 | 2827 | int *classzone_idx) |
1da177e4 | 2828 | { |
dafcb73e | 2829 | bool pgdat_is_balanced = false; |
1da177e4 | 2830 | int i; |
99504748 | 2831 | int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ |
1da177e4 | 2832 | struct reclaim_state *reclaim_state = current->reclaim_state; |
0ae5e89c YH |
2833 | unsigned long nr_soft_reclaimed; |
2834 | unsigned long nr_soft_scanned; | |
179e9639 AM |
2835 | struct scan_control sc = { |
2836 | .gfp_mask = GFP_KERNEL, | |
a6dc60f8 | 2837 | .may_unmap = 1, |
2e2e4259 | 2838 | .may_swap = 1, |
22fba335 KM |
2839 | /* |
2840 | * kswapd doesn't want to be bailed out while reclaim. because | |
2841 | * we want to put equal scanning pressure on each zone. | |
2842 | */ | |
2843 | .nr_to_reclaim = ULONG_MAX, | |
5ad333eb | 2844 | .order = order, |
f16015fb | 2845 | .target_mem_cgroup = NULL, |
179e9639 | 2846 | }; |
a09ed5e0 YH |
2847 | struct shrink_control shrink = { |
2848 | .gfp_mask = sc.gfp_mask, | |
2849 | }; | |
1da177e4 | 2850 | loop_again: |
9e3b2f8c | 2851 | sc.priority = DEF_PRIORITY; |
a79311c1 | 2852 | sc.nr_reclaimed = 0; |
c0bbbc73 | 2853 | sc.may_writepage = !laptop_mode; |
f8891e5e | 2854 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 2855 | |
9e3b2f8c | 2856 | do { |
1da177e4 | 2857 | unsigned long lru_pages = 0; |
1da177e4 | 2858 | |
d6277db4 RW |
2859 | /* |
2860 | * Scan in the highmem->dma direction for the highest | |
2861 | * zone which needs scanning | |
2862 | */ | |
2863 | for (i = pgdat->nr_zones - 1; i >= 0; i--) { | |
2864 | struct zone *zone = pgdat->node_zones + i; | |
1da177e4 | 2865 | |
d6277db4 RW |
2866 | if (!populated_zone(zone)) |
2867 | continue; | |
1da177e4 | 2868 | |
9e3b2f8c KK |
2869 | if (zone->all_unreclaimable && |
2870 | sc.priority != DEF_PRIORITY) | |
d6277db4 | 2871 | continue; |
1da177e4 | 2872 | |
556adecb RR |
2873 | /* |
2874 | * Do some background aging of the anon list, to give | |
2875 | * pages a chance to be referenced before reclaiming. | |
2876 | */ | |
9e3b2f8c | 2877 | age_active_anon(zone, &sc); |
556adecb | 2878 | |
cc715d99 MG |
2879 | /* |
2880 | * If the number of buffer_heads in the machine | |
2881 | * exceeds the maximum allowed level and this node | |
2882 | * has a highmem zone, force kswapd to reclaim from | |
2883 | * it to relieve lowmem pressure. | |
2884 | */ | |
2885 | if (buffer_heads_over_limit && is_highmem_idx(i)) { | |
2886 | end_zone = i; | |
2887 | break; | |
2888 | } | |
2889 | ||
60cefed4 | 2890 | if (!zone_balanced(zone, order, 0, 0)) { |
d6277db4 | 2891 | end_zone = i; |
e1dbeda6 | 2892 | break; |
439423f6 SL |
2893 | } else { |
2894 | /* If balanced, clear the congested flag */ | |
2895 | zone_clear_flag(zone, ZONE_CONGESTED); | |
1da177e4 | 2896 | } |
1da177e4 | 2897 | } |
dafcb73e ZC |
2898 | |
2899 | if (i < 0) { | |
2900 | pgdat_is_balanced = true; | |
e1dbeda6 | 2901 | goto out; |
dafcb73e | 2902 | } |
e1dbeda6 | 2903 | |
1da177e4 LT |
2904 | for (i = 0; i <= end_zone; i++) { |
2905 | struct zone *zone = pgdat->node_zones + i; | |
2906 | ||
adea02a1 | 2907 | lru_pages += zone_reclaimable_pages(zone); |
1da177e4 LT |
2908 | } |
2909 | ||
2910 | /* | |
2911 | * Now scan the zone in the dma->highmem direction, stopping | |
2912 | * at the last zone which needs scanning. | |
2913 | * | |
2914 | * We do this because the page allocator works in the opposite | |
2915 | * direction. This prevents the page allocator from allocating | |
2916 | * pages behind kswapd's direction of progress, which would | |
2917 | * cause too much scanning of the lower zones. | |
2918 | */ | |
2919 | for (i = 0; i <= end_zone; i++) { | |
2920 | struct zone *zone = pgdat->node_zones + i; | |
fe2c2a10 | 2921 | int nr_slab, testorder; |
8afdcece | 2922 | unsigned long balance_gap; |
1da177e4 | 2923 | |
f3fe6512 | 2924 | if (!populated_zone(zone)) |
1da177e4 LT |
2925 | continue; |
2926 | ||
9e3b2f8c KK |
2927 | if (zone->all_unreclaimable && |
2928 | sc.priority != DEF_PRIORITY) | |
1da177e4 LT |
2929 | continue; |
2930 | ||
1da177e4 | 2931 | sc.nr_scanned = 0; |
4e416953 | 2932 | |
0ae5e89c | 2933 | nr_soft_scanned = 0; |
4e416953 BS |
2934 | /* |
2935 | * Call soft limit reclaim before calling shrink_zone. | |
4e416953 | 2936 | */ |
0ae5e89c YH |
2937 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, |
2938 | order, sc.gfp_mask, | |
2939 | &nr_soft_scanned); | |
2940 | sc.nr_reclaimed += nr_soft_reclaimed; | |
00918b6a | 2941 | |
32a4330d | 2942 | /* |
8afdcece MG |
2943 | * We put equal pressure on every zone, unless |
2944 | * one zone has way too many pages free | |
2945 | * already. The "too many pages" is defined | |
2946 | * as the high wmark plus a "gap" where the | |
2947 | * gap is either the low watermark or 1% | |
2948 | * of the zone, whichever is smaller. | |
32a4330d | 2949 | */ |
8afdcece | 2950 | balance_gap = min(low_wmark_pages(zone), |
b40da049 | 2951 | (zone->managed_pages + |
8afdcece MG |
2952 | KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / |
2953 | KSWAPD_ZONE_BALANCE_GAP_RATIO); | |
fe2c2a10 RR |
2954 | /* |
2955 | * Kswapd reclaims only single pages with compaction | |
2956 | * enabled. Trying too hard to reclaim until contiguous | |
2957 | * free pages have become available can hurt performance | |
2958 | * by evicting too much useful data from memory. | |
2959 | * Do not reclaim more than needed for compaction. | |
2960 | */ | |
2961 | testorder = order; | |
d84da3f9 | 2962 | if (IS_ENABLED(CONFIG_COMPACTION) && order && |
fe2c2a10 RR |
2963 | compaction_suitable(zone, order) != |
2964 | COMPACT_SKIPPED) | |
2965 | testorder = 0; | |
2966 | ||
cc715d99 | 2967 | if ((buffer_heads_over_limit && is_highmem_idx(i)) || |
60cefed4 JW |
2968 | !zone_balanced(zone, testorder, |
2969 | balance_gap, end_zone)) { | |
9e3b2f8c | 2970 | shrink_zone(zone, &sc); |
5a03b051 | 2971 | |
d7868dae MG |
2972 | reclaim_state->reclaimed_slab = 0; |
2973 | nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages); | |
2974 | sc.nr_reclaimed += reclaim_state->reclaimed_slab; | |
d7868dae MG |
2975 | |
2976 | if (nr_slab == 0 && !zone_reclaimable(zone)) | |
2977 | zone->all_unreclaimable = 1; | |
2978 | } | |
2979 | ||
1da177e4 | 2980 | /* |
0e50ce3b MK |
2981 | * If we're getting trouble reclaiming, start doing |
2982 | * writepage even in laptop mode. | |
1da177e4 | 2983 | */ |
0e50ce3b | 2984 | if (sc.priority < DEF_PRIORITY - 2) |
1da177e4 | 2985 | sc.may_writepage = 1; |
bb3ab596 | 2986 | |
215ddd66 MG |
2987 | if (zone->all_unreclaimable) { |
2988 | if (end_zone && end_zone == i) | |
2989 | end_zone--; | |
d7868dae | 2990 | continue; |
215ddd66 | 2991 | } |
d7868dae | 2992 | |
258401a6 | 2993 | if (zone_balanced(zone, testorder, 0, end_zone)) |
0e093d99 MG |
2994 | /* |
2995 | * If a zone reaches its high watermark, | |
2996 | * consider it to be no longer congested. It's | |
2997 | * possible there are dirty pages backed by | |
2998 | * congested BDIs but as pressure is relieved, | |
ab8704b8 | 2999 | * speculatively avoid congestion waits |
0e093d99 MG |
3000 | */ |
3001 | zone_clear_flag(zone, ZONE_CONGESTED); | |
1da177e4 | 3002 | } |
5515061d MG |
3003 | |
3004 | /* | |
3005 | * If the low watermark is met there is no need for processes | |
3006 | * to be throttled on pfmemalloc_wait as they should not be | |
3007 | * able to safely make forward progress. Wake them | |
3008 | */ | |
3009 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
3010 | pfmemalloc_watermark_ok(pgdat)) | |
3011 | wake_up(&pgdat->pfmemalloc_wait); | |
3012 | ||
dafcb73e ZC |
3013 | if (pgdat_balanced(pgdat, order, *classzone_idx)) { |
3014 | pgdat_is_balanced = true; | |
1da177e4 | 3015 | break; /* kswapd: all done */ |
dafcb73e ZC |
3016 | } |
3017 | ||
1da177e4 LT |
3018 | /* |
3019 | * We do this so kswapd doesn't build up large priorities for | |
3020 | * example when it is freeing in parallel with allocators. It | |
3021 | * matches the direct reclaim path behaviour in terms of impact | |
3022 | * on zone->*_priority. | |
3023 | */ | |
a79311c1 | 3024 | if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) |
1da177e4 | 3025 | break; |
9e3b2f8c | 3026 | } while (--sc.priority >= 0); |
99504748 | 3027 | |
dafcb73e ZC |
3028 | out: |
3029 | if (!pgdat_is_balanced) { | |
1da177e4 | 3030 | cond_resched(); |
8357376d RW |
3031 | |
3032 | try_to_freeze(); | |
3033 | ||
73ce02e9 KM |
3034 | /* |
3035 | * Fragmentation may mean that the system cannot be | |
3036 | * rebalanced for high-order allocations in all zones. | |
3037 | * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, | |
3038 | * it means the zones have been fully scanned and are still | |
3039 | * not balanced. For high-order allocations, there is | |
3040 | * little point trying all over again as kswapd may | |
3041 | * infinite loop. | |
3042 | * | |
3043 | * Instead, recheck all watermarks at order-0 as they | |
3044 | * are the most important. If watermarks are ok, kswapd will go | |
3045 | * back to sleep. High-order users can still perform direct | |
3046 | * reclaim if they wish. | |
3047 | */ | |
3048 | if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) | |
3049 | order = sc.order = 0; | |
3050 | ||
1da177e4 LT |
3051 | goto loop_again; |
3052 | } | |
3053 | ||
99504748 MG |
3054 | /* |
3055 | * If kswapd was reclaiming at a higher order, it has the option of | |
3056 | * sleeping without all zones being balanced. Before it does, it must | |
3057 | * ensure that the watermarks for order-0 on *all* zones are met and | |
3058 | * that the congestion flags are cleared. The congestion flag must | |
3059 | * be cleared as kswapd is the only mechanism that clears the flag | |
3060 | * and it is potentially going to sleep here. | |
3061 | */ | |
3062 | if (order) { | |
7be62de9 RR |
3063 | int zones_need_compaction = 1; |
3064 | ||
99504748 MG |
3065 | for (i = 0; i <= end_zone; i++) { |
3066 | struct zone *zone = pgdat->node_zones + i; | |
3067 | ||
3068 | if (!populated_zone(zone)) | |
3069 | continue; | |
3070 | ||
7be62de9 RR |
3071 | /* Check if the memory needs to be defragmented. */ |
3072 | if (zone_watermark_ok(zone, order, | |
3073 | low_wmark_pages(zone), *classzone_idx, 0)) | |
3074 | zones_need_compaction = 0; | |
99504748 | 3075 | } |
7be62de9 RR |
3076 | |
3077 | if (zones_need_compaction) | |
3078 | compact_pgdat(pgdat, order); | |
99504748 MG |
3079 | } |
3080 | ||
0abdee2b | 3081 | /* |
5515061d | 3082 | * Return the order we were reclaiming at so prepare_kswapd_sleep() |
0abdee2b MG |
3083 | * makes a decision on the order we were last reclaiming at. However, |
3084 | * if another caller entered the allocator slow path while kswapd | |
3085 | * was awake, order will remain at the higher level | |
3086 | */ | |
dc83edd9 | 3087 | *classzone_idx = end_zone; |
0abdee2b | 3088 | return order; |
1da177e4 LT |
3089 | } |
3090 | ||
dc83edd9 | 3091 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f0bc0a60 KM |
3092 | { |
3093 | long remaining = 0; | |
3094 | DEFINE_WAIT(wait); | |
3095 | ||
3096 | if (freezing(current) || kthread_should_stop()) | |
3097 | return; | |
3098 | ||
3099 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3100 | ||
3101 | /* Try to sleep for a short interval */ | |
5515061d | 3102 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
3103 | remaining = schedule_timeout(HZ/10); |
3104 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3105 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3106 | } | |
3107 | ||
3108 | /* | |
3109 | * After a short sleep, check if it was a premature sleep. If not, then | |
3110 | * go fully to sleep until explicitly woken up. | |
3111 | */ | |
5515061d | 3112 | if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { |
f0bc0a60 KM |
3113 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3114 | ||
3115 | /* | |
3116 | * vmstat counters are not perfectly accurate and the estimated | |
3117 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3118 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3119 | * watermarks being breached while under pressure, we reduce the | |
3120 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3121 | * them before going back to sleep. | |
3122 | */ | |
3123 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c | 3124 | |
62997027 MG |
3125 | /* |
3126 | * Compaction records what page blocks it recently failed to | |
3127 | * isolate pages from and skips them in the future scanning. | |
3128 | * When kswapd is going to sleep, it is reasonable to assume | |
3129 | * that pages and compaction may succeed so reset the cache. | |
3130 | */ | |
3131 | reset_isolation_suitable(pgdat); | |
3132 | ||
1c7e7f6c AK |
3133 | if (!kthread_should_stop()) |
3134 | schedule(); | |
3135 | ||
f0bc0a60 KM |
3136 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3137 | } else { | |
3138 | if (remaining) | |
3139 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3140 | else | |
3141 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3142 | } | |
3143 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3144 | } | |
3145 | ||
1da177e4 LT |
3146 | /* |
3147 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3148 | * from the init process. |
1da177e4 LT |
3149 | * |
3150 | * This basically trickles out pages so that we have _some_ | |
3151 | * free memory available even if there is no other activity | |
3152 | * that frees anything up. This is needed for things like routing | |
3153 | * etc, where we otherwise might have all activity going on in | |
3154 | * asynchronous contexts that cannot page things out. | |
3155 | * | |
3156 | * If there are applications that are active memory-allocators | |
3157 | * (most normal use), this basically shouldn't matter. | |
3158 | */ | |
3159 | static int kswapd(void *p) | |
3160 | { | |
215ddd66 | 3161 | unsigned long order, new_order; |
d2ebd0f6 | 3162 | unsigned balanced_order; |
215ddd66 | 3163 | int classzone_idx, new_classzone_idx; |
d2ebd0f6 | 3164 | int balanced_classzone_idx; |
1da177e4 LT |
3165 | pg_data_t *pgdat = (pg_data_t*)p; |
3166 | struct task_struct *tsk = current; | |
f0bc0a60 | 3167 | |
1da177e4 LT |
3168 | struct reclaim_state reclaim_state = { |
3169 | .reclaimed_slab = 0, | |
3170 | }; | |
a70f7302 | 3171 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3172 | |
cf40bd16 NP |
3173 | lockdep_set_current_reclaim_state(GFP_KERNEL); |
3174 | ||
174596a0 | 3175 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3176 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3177 | current->reclaim_state = &reclaim_state; |
3178 | ||
3179 | /* | |
3180 | * Tell the memory management that we're a "memory allocator", | |
3181 | * and that if we need more memory we should get access to it | |
3182 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3183 | * never get caught in the normal page freeing logic. | |
3184 | * | |
3185 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3186 | * you need a small amount of memory in order to be able to | |
3187 | * page out something else, and this flag essentially protects | |
3188 | * us from recursively trying to free more memory as we're | |
3189 | * trying to free the first piece of memory in the first place). | |
3190 | */ | |
930d9152 | 3191 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3192 | set_freezable(); |
1da177e4 | 3193 | |
215ddd66 | 3194 | order = new_order = 0; |
d2ebd0f6 | 3195 | balanced_order = 0; |
215ddd66 | 3196 | classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; |
d2ebd0f6 | 3197 | balanced_classzone_idx = classzone_idx; |
1da177e4 | 3198 | for ( ; ; ) { |
6f6313d4 | 3199 | bool ret; |
3e1d1d28 | 3200 | |
215ddd66 MG |
3201 | /* |
3202 | * If the last balance_pgdat was unsuccessful it's unlikely a | |
3203 | * new request of a similar or harder type will succeed soon | |
3204 | * so consider going to sleep on the basis we reclaimed at | |
3205 | */ | |
d2ebd0f6 AS |
3206 | if (balanced_classzone_idx >= new_classzone_idx && |
3207 | balanced_order == new_order) { | |
215ddd66 MG |
3208 | new_order = pgdat->kswapd_max_order; |
3209 | new_classzone_idx = pgdat->classzone_idx; | |
3210 | pgdat->kswapd_max_order = 0; | |
3211 | pgdat->classzone_idx = pgdat->nr_zones - 1; | |
3212 | } | |
3213 | ||
99504748 | 3214 | if (order < new_order || classzone_idx > new_classzone_idx) { |
1da177e4 LT |
3215 | /* |
3216 | * Don't sleep if someone wants a larger 'order' | |
99504748 | 3217 | * allocation or has tigher zone constraints |
1da177e4 LT |
3218 | */ |
3219 | order = new_order; | |
99504748 | 3220 | classzone_idx = new_classzone_idx; |
1da177e4 | 3221 | } else { |
d2ebd0f6 AS |
3222 | kswapd_try_to_sleep(pgdat, balanced_order, |
3223 | balanced_classzone_idx); | |
1da177e4 | 3224 | order = pgdat->kswapd_max_order; |
99504748 | 3225 | classzone_idx = pgdat->classzone_idx; |
f0dfcde0 AS |
3226 | new_order = order; |
3227 | new_classzone_idx = classzone_idx; | |
4d40502e | 3228 | pgdat->kswapd_max_order = 0; |
215ddd66 | 3229 | pgdat->classzone_idx = pgdat->nr_zones - 1; |
1da177e4 | 3230 | } |
1da177e4 | 3231 | |
8fe23e05 DR |
3232 | ret = try_to_freeze(); |
3233 | if (kthread_should_stop()) | |
3234 | break; | |
3235 | ||
3236 | /* | |
3237 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3238 | * after returning from the refrigerator | |
3239 | */ | |
33906bc5 MG |
3240 | if (!ret) { |
3241 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); | |
d2ebd0f6 AS |
3242 | balanced_classzone_idx = classzone_idx; |
3243 | balanced_order = balance_pgdat(pgdat, order, | |
3244 | &balanced_classzone_idx); | |
33906bc5 | 3245 | } |
1da177e4 | 3246 | } |
b0a8cc58 | 3247 | |
f26bff27 | 3248 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
b0a8cc58 | 3249 | current->reclaim_state = NULL; |
f26bff27 JW |
3250 | lockdep_clear_current_reclaim_state(); |
3251 | ||
1da177e4 LT |
3252 | return 0; |
3253 | } | |
3254 | ||
3255 | /* | |
3256 | * A zone is low on free memory, so wake its kswapd task to service it. | |
3257 | */ | |
99504748 | 3258 | void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) |
1da177e4 LT |
3259 | { |
3260 | pg_data_t *pgdat; | |
3261 | ||
f3fe6512 | 3262 | if (!populated_zone(zone)) |
1da177e4 LT |
3263 | return; |
3264 | ||
6fa3eb70 S |
3265 | #ifdef CONFIG_FREEZER |
3266 | if (pm_freezing) | |
3267 | return; | |
3268 | #endif | |
3269 | ||
88f5acf8 | 3270 | if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) |
1da177e4 | 3271 | return; |
88f5acf8 | 3272 | pgdat = zone->zone_pgdat; |
99504748 | 3273 | if (pgdat->kswapd_max_order < order) { |
1da177e4 | 3274 | pgdat->kswapd_max_order = order; |
99504748 MG |
3275 | pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); |
3276 | } | |
8d0986e2 | 3277 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3278 | return; |
88f5acf8 MG |
3279 | if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) |
3280 | return; | |
3281 | ||
3282 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); | |
8d0986e2 | 3283 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3284 | } |
3285 | ||
c6f37f12 | 3286 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3287 | /* |
7b51755c | 3288 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3289 | * freed pages. |
3290 | * | |
3291 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3292 | * LRU order by reclaiming preferentially | |
3293 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3294 | */ |
6fa3eb70 | 3295 | unsigned long shrink_memory_mask(unsigned long nr_to_reclaim, gfp_t mask) |
1da177e4 | 3296 | { |
d6277db4 | 3297 | struct reclaim_state reclaim_state; |
d6277db4 | 3298 | struct scan_control sc = { |
6fa3eb70 | 3299 | .gfp_mask = mask, |
7b51755c KM |
3300 | .may_swap = 1, |
3301 | .may_unmap = 1, | |
d6277db4 | 3302 | .may_writepage = 1, |
7b51755c KM |
3303 | .nr_to_reclaim = nr_to_reclaim, |
3304 | .hibernation_mode = 1, | |
7b51755c | 3305 | .order = 0, |
9e3b2f8c | 3306 | .priority = DEF_PRIORITY, |
1da177e4 | 3307 | }; |
a09ed5e0 YH |
3308 | struct shrink_control shrink = { |
3309 | .gfp_mask = sc.gfp_mask, | |
3310 | }; | |
3311 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); | |
7b51755c KM |
3312 | struct task_struct *p = current; |
3313 | unsigned long nr_reclaimed; | |
1da177e4 | 3314 | |
7b51755c KM |
3315 | p->flags |= PF_MEMALLOC; |
3316 | lockdep_set_current_reclaim_state(sc.gfp_mask); | |
3317 | reclaim_state.reclaimed_slab = 0; | |
3318 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3319 | |
a09ed5e0 | 3320 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); |
d979677c | 3321 | |
7b51755c KM |
3322 | p->reclaim_state = NULL; |
3323 | lockdep_clear_current_reclaim_state(); | |
3324 | p->flags &= ~PF_MEMALLOC; | |
d6277db4 | 3325 | |
7b51755c | 3326 | return nr_reclaimed; |
1da177e4 | 3327 | } |
6fa3eb70 S |
3328 | EXPORT_SYMBOL_GPL(shrink_memory_mask); |
3329 | ||
3330 | #ifdef CONFIG_MTKPASR | |
3331 | extern void shrink_mtkpasr_all(void); | |
3332 | #else | |
3333 | #define shrink_mtkpasr_all() do {} while (0) | |
3334 | #endif | |
3335 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) | |
3336 | { | |
3337 | shrink_mtkpasr_all(); | |
3338 | return shrink_memory_mask(nr_to_reclaim, GFP_HIGHUSER_MOVABLE); | |
3339 | } | |
3340 | EXPORT_SYMBOL_GPL(shrink_all_memory); | |
c6f37f12 | 3341 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3342 | |
1da177e4 LT |
3343 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3344 | not required for correctness. So if the last cpu in a node goes | |
3345 | away, we get changed to run anywhere: as the first one comes back, | |
3346 | restore their cpu bindings. */ | |
fcb35a9b GKH |
3347 | static int cpu_callback(struct notifier_block *nfb, unsigned long action, |
3348 | void *hcpu) | |
1da177e4 | 3349 | { |
58c0a4a7 | 3350 | int nid; |
1da177e4 | 3351 | |
8bb78442 | 3352 | if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { |
48fb2e24 | 3353 | for_each_node_state(nid, N_MEMORY) { |
c5f59f08 | 3354 | pg_data_t *pgdat = NODE_DATA(nid); |
a70f7302 RR |
3355 | const struct cpumask *mask; |
3356 | ||
3357 | mask = cpumask_of_node(pgdat->node_id); | |
c5f59f08 | 3358 | |
3e597945 | 3359 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
1da177e4 | 3360 | /* One of our CPUs online: restore mask */ |
c5f59f08 | 3361 | set_cpus_allowed_ptr(pgdat->kswapd, mask); |
1da177e4 LT |
3362 | } |
3363 | } | |
3364 | return NOTIFY_OK; | |
3365 | } | |
1da177e4 | 3366 | |
3218ae14 YG |
3367 | /* |
3368 | * This kswapd start function will be called by init and node-hot-add. | |
3369 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3370 | */ | |
3371 | int kswapd_run(int nid) | |
3372 | { | |
3373 | pg_data_t *pgdat = NODE_DATA(nid); | |
3374 | int ret = 0; | |
3375 | ||
3376 | if (pgdat->kswapd) | |
3377 | return 0; | |
3378 | ||
3379 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3380 | if (IS_ERR(pgdat->kswapd)) { | |
3381 | /* failure at boot is fatal */ | |
3382 | BUG_ON(system_state == SYSTEM_BOOTING); | |
d5dc0ad9 GS |
3383 | pr_err("Failed to start kswapd on node %d\n", nid); |
3384 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3385 | pgdat->kswapd = NULL; |
3218ae14 YG |
3386 | } |
3387 | return ret; | |
3388 | } | |
3389 | ||
8fe23e05 | 3390 | /* |
d8adde17 JL |
3391 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
3392 | * hold lock_memory_hotplug(). | |
8fe23e05 DR |
3393 | */ |
3394 | void kswapd_stop(int nid) | |
3395 | { | |
3396 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3397 | ||
d8adde17 | 3398 | if (kswapd) { |
8fe23e05 | 3399 | kthread_stop(kswapd); |
d8adde17 JL |
3400 | NODE_DATA(nid)->kswapd = NULL; |
3401 | } | |
8fe23e05 DR |
3402 | } |
3403 | ||
1da177e4 LT |
3404 | static int __init kswapd_init(void) |
3405 | { | |
3218ae14 | 3406 | int nid; |
69e05944 | 3407 | |
1da177e4 | 3408 | swap_setup(); |
48fb2e24 | 3409 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3410 | kswapd_run(nid); |
1da177e4 LT |
3411 | hotcpu_notifier(cpu_callback, 0); |
3412 | return 0; | |
3413 | } | |
3414 | ||
3415 | module_init(kswapd_init) | |
9eeff239 CL |
3416 | |
3417 | #ifdef CONFIG_NUMA | |
3418 | /* | |
3419 | * Zone reclaim mode | |
3420 | * | |
3421 | * If non-zero call zone_reclaim when the number of free pages falls below | |
3422 | * the watermarks. | |
9eeff239 CL |
3423 | */ |
3424 | int zone_reclaim_mode __read_mostly; | |
3425 | ||
1b2ffb78 | 3426 | #define RECLAIM_OFF 0 |
7d03431c | 3427 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 CL |
3428 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
3429 | #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ | |
3430 | ||
a92f7126 CL |
3431 | /* |
3432 | * Priority for ZONE_RECLAIM. This determines the fraction of pages | |
3433 | * of a node considered for each zone_reclaim. 4 scans 1/16th of | |
3434 | * a zone. | |
3435 | */ | |
3436 | #define ZONE_RECLAIM_PRIORITY 4 | |
3437 | ||
9614634f CL |
3438 | /* |
3439 | * Percentage of pages in a zone that must be unmapped for zone_reclaim to | |
3440 | * occur. | |
3441 | */ | |
3442 | int sysctl_min_unmapped_ratio = 1; | |
3443 | ||
0ff38490 CL |
3444 | /* |
3445 | * If the number of slab pages in a zone grows beyond this percentage then | |
3446 | * slab reclaim needs to occur. | |
3447 | */ | |
3448 | int sysctl_min_slab_ratio = 5; | |
3449 | ||
90afa5de MG |
3450 | static inline unsigned long zone_unmapped_file_pages(struct zone *zone) |
3451 | { | |
3452 | unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); | |
3453 | unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + | |
3454 | zone_page_state(zone, NR_ACTIVE_FILE); | |
3455 | ||
3456 | /* | |
3457 | * It's possible for there to be more file mapped pages than | |
3458 | * accounted for by the pages on the file LRU lists because | |
3459 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3460 | */ | |
3461 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3462 | } | |
3463 | ||
3464 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
3465 | static long zone_pagecache_reclaimable(struct zone *zone) | |
3466 | { | |
3467 | long nr_pagecache_reclaimable; | |
3468 | long delta = 0; | |
3469 | ||
3470 | /* | |
3471 | * If RECLAIM_SWAP is set, then all file pages are considered | |
3472 | * potentially reclaimable. Otherwise, we have to worry about | |
3473 | * pages like swapcache and zone_unmapped_file_pages() provides | |
3474 | * a better estimate | |
3475 | */ | |
3476 | if (zone_reclaim_mode & RECLAIM_SWAP) | |
3477 | nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); | |
3478 | else | |
3479 | nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); | |
3480 | ||
3481 | /* If we can't clean pages, remove dirty pages from consideration */ | |
3482 | if (!(zone_reclaim_mode & RECLAIM_WRITE)) | |
3483 | delta += zone_page_state(zone, NR_FILE_DIRTY); | |
3484 | ||
3485 | /* Watch for any possible underflows due to delta */ | |
3486 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
3487 | delta = nr_pagecache_reclaimable; | |
3488 | ||
3489 | return nr_pagecache_reclaimable - delta; | |
3490 | } | |
3491 | ||
9eeff239 CL |
3492 | /* |
3493 | * Try to free up some pages from this zone through reclaim. | |
3494 | */ | |
179e9639 | 3495 | static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 3496 | { |
7fb2d46d | 3497 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 3498 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
3499 | struct task_struct *p = current; |
3500 | struct reclaim_state reclaim_state; | |
179e9639 AM |
3501 | struct scan_control sc = { |
3502 | .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), | |
a6dc60f8 | 3503 | .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), |
2e2e4259 | 3504 | .may_swap = 1, |
62b726c1 | 3505 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
21caf2fc | 3506 | .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), |
bd2f6199 | 3507 | .order = order, |
9e3b2f8c | 3508 | .priority = ZONE_RECLAIM_PRIORITY, |
179e9639 | 3509 | }; |
a09ed5e0 YH |
3510 | struct shrink_control shrink = { |
3511 | .gfp_mask = sc.gfp_mask, | |
3512 | }; | |
15748048 | 3513 | unsigned long nr_slab_pages0, nr_slab_pages1; |
9eeff239 | 3514 | |
9eeff239 | 3515 | cond_resched(); |
d4f7796e CL |
3516 | /* |
3517 | * We need to be able to allocate from the reserves for RECLAIM_SWAP | |
3518 | * and we also need to be able to write out pages for RECLAIM_WRITE | |
3519 | * and RECLAIM_SWAP. | |
3520 | */ | |
3521 | p->flags |= PF_MEMALLOC | PF_SWAPWRITE; | |
76ca542d | 3522 | lockdep_set_current_reclaim_state(gfp_mask); |
9eeff239 CL |
3523 | reclaim_state.reclaimed_slab = 0; |
3524 | p->reclaim_state = &reclaim_state; | |
c84db23c | 3525 | |
90afa5de | 3526 | if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { |
0ff38490 CL |
3527 | /* |
3528 | * Free memory by calling shrink zone with increasing | |
3529 | * priorities until we have enough memory freed. | |
3530 | */ | |
0ff38490 | 3531 | do { |
9e3b2f8c KK |
3532 | shrink_zone(zone, &sc); |
3533 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); | |
0ff38490 | 3534 | } |
c84db23c | 3535 | |
15748048 KM |
3536 | nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3537 | if (nr_slab_pages0 > zone->min_slab_pages) { | |
2a16e3f4 | 3538 | /* |
7fb2d46d | 3539 | * shrink_slab() does not currently allow us to determine how |
0ff38490 CL |
3540 | * many pages were freed in this zone. So we take the current |
3541 | * number of slab pages and shake the slab until it is reduced | |
3542 | * by the same nr_pages that we used for reclaiming unmapped | |
3543 | * pages. | |
2a16e3f4 | 3544 | * |
0ff38490 CL |
3545 | * Note that shrink_slab will free memory on all zones and may |
3546 | * take a long time. | |
2a16e3f4 | 3547 | */ |
4dc4b3d9 KM |
3548 | for (;;) { |
3549 | unsigned long lru_pages = zone_reclaimable_pages(zone); | |
3550 | ||
3551 | /* No reclaimable slab or very low memory pressure */ | |
1495f230 | 3552 | if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) |
4dc4b3d9 KM |
3553 | break; |
3554 | ||
3555 | /* Freed enough memory */ | |
3556 | nr_slab_pages1 = zone_page_state(zone, | |
3557 | NR_SLAB_RECLAIMABLE); | |
3558 | if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) | |
3559 | break; | |
3560 | } | |
83e33a47 CL |
3561 | |
3562 | /* | |
3563 | * Update nr_reclaimed by the number of slab pages we | |
3564 | * reclaimed from this zone. | |
3565 | */ | |
15748048 KM |
3566 | nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); |
3567 | if (nr_slab_pages1 < nr_slab_pages0) | |
3568 | sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; | |
2a16e3f4 CL |
3569 | } |
3570 | ||
9eeff239 | 3571 | p->reclaim_state = NULL; |
d4f7796e | 3572 | current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); |
76ca542d | 3573 | lockdep_clear_current_reclaim_state(); |
a79311c1 | 3574 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 3575 | } |
179e9639 AM |
3576 | |
3577 | int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) | |
3578 | { | |
179e9639 | 3579 | int node_id; |
d773ed6b | 3580 | int ret; |
179e9639 AM |
3581 | |
3582 | /* | |
0ff38490 CL |
3583 | * Zone reclaim reclaims unmapped file backed pages and |
3584 | * slab pages if we are over the defined limits. | |
34aa1330 | 3585 | * |
9614634f CL |
3586 | * A small portion of unmapped file backed pages is needed for |
3587 | * file I/O otherwise pages read by file I/O will be immediately | |
3588 | * thrown out if the zone is overallocated. So we do not reclaim | |
3589 | * if less than a specified percentage of the zone is used by | |
3590 | * unmapped file backed pages. | |
179e9639 | 3591 | */ |
90afa5de MG |
3592 | if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && |
3593 | zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) | |
fa5e084e | 3594 | return ZONE_RECLAIM_FULL; |
179e9639 | 3595 | |
93e4a89a | 3596 | if (zone->all_unreclaimable) |
fa5e084e | 3597 | return ZONE_RECLAIM_FULL; |
d773ed6b | 3598 | |
179e9639 | 3599 | /* |
d773ed6b | 3600 | * Do not scan if the allocation should not be delayed. |
179e9639 | 3601 | */ |
d773ed6b | 3602 | if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) |
fa5e084e | 3603 | return ZONE_RECLAIM_NOSCAN; |
179e9639 AM |
3604 | |
3605 | /* | |
3606 | * Only run zone reclaim on the local zone or on zones that do not | |
3607 | * have associated processors. This will favor the local processor | |
3608 | * over remote processors and spread off node memory allocations | |
3609 | * as wide as possible. | |
3610 | */ | |
89fa3024 | 3611 | node_id = zone_to_nid(zone); |
37c0708d | 3612 | if (node_state(node_id, N_CPU) && node_id != numa_node_id()) |
fa5e084e | 3613 | return ZONE_RECLAIM_NOSCAN; |
d773ed6b DR |
3614 | |
3615 | if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) | |
fa5e084e MG |
3616 | return ZONE_RECLAIM_NOSCAN; |
3617 | ||
d773ed6b DR |
3618 | ret = __zone_reclaim(zone, gfp_mask, order); |
3619 | zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); | |
3620 | ||
24cf7251 MG |
3621 | if (!ret) |
3622 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
3623 | ||
d773ed6b | 3624 | return ret; |
179e9639 | 3625 | } |
9eeff239 | 3626 | #endif |
894bc310 | 3627 | |
894bc310 LS |
3628 | /* |
3629 | * page_evictable - test whether a page is evictable | |
3630 | * @page: the page to test | |
894bc310 LS |
3631 | * |
3632 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 3633 | * lists vs unevictable list. |
894bc310 LS |
3634 | * |
3635 | * Reasons page might not be evictable: | |
ba9ddf49 | 3636 | * (1) page's mapping marked unevictable |
b291f000 | 3637 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 3638 | * |
894bc310 | 3639 | */ |
39b5f29a | 3640 | int page_evictable(struct page *page) |
894bc310 | 3641 | { |
39b5f29a | 3642 | return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); |
894bc310 | 3643 | } |
89e004ea | 3644 | |
85046579 | 3645 | #ifdef CONFIG_SHMEM |
89e004ea | 3646 | /** |
24513264 HD |
3647 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
3648 | * @pages: array of pages to check | |
3649 | * @nr_pages: number of pages to check | |
89e004ea | 3650 | * |
24513264 | 3651 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
3652 | * |
3653 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 3654 | */ |
24513264 | 3655 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 3656 | { |
925b7673 | 3657 | struct lruvec *lruvec; |
24513264 HD |
3658 | struct zone *zone = NULL; |
3659 | int pgscanned = 0; | |
3660 | int pgrescued = 0; | |
3661 | int i; | |
89e004ea | 3662 | |
24513264 HD |
3663 | for (i = 0; i < nr_pages; i++) { |
3664 | struct page *page = pages[i]; | |
3665 | struct zone *pagezone; | |
89e004ea | 3666 | |
24513264 HD |
3667 | pgscanned++; |
3668 | pagezone = page_zone(page); | |
3669 | if (pagezone != zone) { | |
3670 | if (zone) | |
3671 | spin_unlock_irq(&zone->lru_lock); | |
3672 | zone = pagezone; | |
3673 | spin_lock_irq(&zone->lru_lock); | |
3674 | } | |
fa9add64 | 3675 | lruvec = mem_cgroup_page_lruvec(page, zone); |
89e004ea | 3676 | |
24513264 HD |
3677 | if (!PageLRU(page) || !PageUnevictable(page)) |
3678 | continue; | |
89e004ea | 3679 | |
39b5f29a | 3680 | if (page_evictable(page)) { |
24513264 HD |
3681 | enum lru_list lru = page_lru_base_type(page); |
3682 | ||
3683 | VM_BUG_ON(PageActive(page)); | |
3684 | ClearPageUnevictable(page); | |
fa9add64 HD |
3685 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
3686 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 3687 | pgrescued++; |
89e004ea | 3688 | } |
24513264 | 3689 | } |
89e004ea | 3690 | |
24513264 HD |
3691 | if (zone) { |
3692 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); | |
3693 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
3694 | spin_unlock_irq(&zone->lru_lock); | |
89e004ea | 3695 | } |
89e004ea | 3696 | } |
85046579 | 3697 | #endif /* CONFIG_SHMEM */ |
af936a16 | 3698 | |
264e56d8 | 3699 | static void warn_scan_unevictable_pages(void) |
af936a16 | 3700 | { |
264e56d8 | 3701 | printk_once(KERN_WARNING |
25bd91bd | 3702 | "%s: The scan_unevictable_pages sysctl/node-interface has been " |
264e56d8 | 3703 | "disabled for lack of a legitimate use case. If you have " |
25bd91bd KM |
3704 | "one, please send an email to linux-mm@kvack.org.\n", |
3705 | current->comm); | |
af936a16 LS |
3706 | } |
3707 | ||
3708 | /* | |
3709 | * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of | |
3710 | * all nodes' unevictable lists for evictable pages | |
3711 | */ | |
3712 | unsigned long scan_unevictable_pages; | |
3713 | ||
3714 | int scan_unevictable_handler(struct ctl_table *table, int write, | |
8d65af78 | 3715 | void __user *buffer, |
af936a16 LS |
3716 | size_t *length, loff_t *ppos) |
3717 | { | |
264e56d8 | 3718 | warn_scan_unevictable_pages(); |
8d65af78 | 3719 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
af936a16 LS |
3720 | scan_unevictable_pages = 0; |
3721 | return 0; | |
3722 | } | |
3723 | ||
e4455abb | 3724 | #ifdef CONFIG_NUMA |
af936a16 LS |
3725 | /* |
3726 | * per node 'scan_unevictable_pages' attribute. On demand re-scan of | |
3727 | * a specified node's per zone unevictable lists for evictable pages. | |
3728 | */ | |
3729 | ||
10fbcf4c KS |
3730 | static ssize_t read_scan_unevictable_node(struct device *dev, |
3731 | struct device_attribute *attr, | |
af936a16 LS |
3732 | char *buf) |
3733 | { | |
264e56d8 | 3734 | warn_scan_unevictable_pages(); |
af936a16 LS |
3735 | return sprintf(buf, "0\n"); /* always zero; should fit... */ |
3736 | } | |
3737 | ||
10fbcf4c KS |
3738 | static ssize_t write_scan_unevictable_node(struct device *dev, |
3739 | struct device_attribute *attr, | |
af936a16 LS |
3740 | const char *buf, size_t count) |
3741 | { | |
264e56d8 | 3742 | warn_scan_unevictable_pages(); |
af936a16 LS |
3743 | return 1; |
3744 | } | |
3745 | ||
3746 | ||
10fbcf4c | 3747 | static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, |
af936a16 LS |
3748 | read_scan_unevictable_node, |
3749 | write_scan_unevictable_node); | |
3750 | ||
3751 | int scan_unevictable_register_node(struct node *node) | |
3752 | { | |
10fbcf4c | 3753 | return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 LS |
3754 | } |
3755 | ||
3756 | void scan_unevictable_unregister_node(struct node *node) | |
3757 | { | |
10fbcf4c | 3758 | device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages); |
af936a16 | 3759 | } |
e4455abb | 3760 | #endif |
6fa3eb70 S |
3761 | |
3762 | #ifdef CONFIG_MTKPASR | |
3763 | void try_to_shrink_slab(void) | |
3764 | { | |
3765 | struct shrinker *shrinker; | |
3766 | struct shrink_control shrink = { | |
3767 | .gfp_mask = GFP_KERNEL|__GFP_HIGHMEM, | |
3768 | }; | |
3769 | ||
3770 | if (!down_read_trylock(&shrinker_rwsem)) { | |
3771 | return; | |
3772 | } | |
3773 | ||
3774 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
3775 | int num_objs; | |
3776 | int shrink_ret = 0; | |
3777 | int retry = 2; | |
3778 | ||
3779 | num_objs = do_shrinker_shrink(shrinker, &shrink, 0); | |
3780 | if (num_objs <= 0) | |
3781 | continue; | |
3782 | ||
3783 | do { | |
3784 | /* To shrink */ | |
3785 | shrink_ret = do_shrinker_shrink(shrinker, &shrink, num_objs); | |
3786 | if (shrink_ret == -1) | |
3787 | break; | |
3788 | /* Check empty */ | |
3789 | num_objs = do_shrinker_shrink(shrinker, &shrink, 0); | |
3790 | if (num_objs <= 0) | |
3791 | break; | |
3792 | } while (--retry); | |
3793 | } | |
3794 | ||
3795 | up_read(&shrinker_rwsem); | |
3796 | } | |
3797 | ||
3798 | extern void free_hot_cold_page(struct page *page, int cold); | |
3799 | /* Isolate pages for PASR */ | |
3800 | #ifdef CONFIG_MTKPASR_ALLEXTCOMP | |
3801 | int mtkpasr_isolate_page(struct page *page, int check_swap) | |
3802 | #else | |
3803 | int mtkpasr_isolate_page(struct page *page) | |
3804 | #endif | |
3805 | { | |
3806 | struct zone *zone = page_zone(page); | |
3807 | struct lruvec *lruvec; | |
3808 | unsigned long flags; | |
3809 | isolate_mode_t mode = ISOLATE_ASYNC_MIGRATE; | |
3810 | ||
3811 | /* Lock this zone - USE trylock version! */ | |
3812 | if (!spin_trylock_irqsave(&zone->lru_lock, flags)) { | |
3813 | printk(KERN_ALERT"\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n"); | |
3814 | printk(KERN_ALERT"[%s][%d] Failed to lock this zone!\n",__FUNCTION__,__LINE__); | |
3815 | printk(KERN_ALERT"\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n"); | |
3816 | return -EAGAIN; | |
3817 | } | |
3818 | ||
3819 | #ifdef CONFIG_MTKPASR_ALLEXTCOMP | |
3820 | /* Check whether we should handle SwapBacked, SwapCache pages */ | |
3821 | if (check_swap) { | |
3822 | if (PageSwapBacked(page) || PageSwapCache(page)) { | |
3823 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
3824 | return -EACCES; | |
3825 | } | |
3826 | } | |
3827 | #endif | |
3828 | ||
3829 | /* Try to isolate this page */ | |
3830 | if (__isolate_lru_page(page, mode) != 0) { | |
3831 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
3832 | return -EACCES; | |
3833 | } | |
3834 | ||
3835 | /* Successfully isolated */ | |
3836 | lruvec = mem_cgroup_page_lruvec(page, zone); | |
3837 | del_page_from_lru_list(page, lruvec, page_lru(page)); | |
3838 | ||
3839 | /* Unlock this zone */ | |
3840 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
3841 | ||
3842 | return 0; | |
3843 | } | |
3844 | ||
3845 | /* Drop page (in File/Anon LRUs) (Imitate the behavior of shrink_page_list) */ | |
3846 | /* If returns error, caller needs to putback page by itself. */ | |
3847 | int mtkpasr_drop_page(struct page *page) | |
3848 | { | |
3849 | int ret; | |
3850 | unsigned long vm_flags = 0x0; | |
3851 | bool active = false; | |
3852 | struct address_space *mapping; | |
3853 | enum ttu_flags unmap_flags = TTU_UNMAP; | |
3854 | ||
3855 | /* Suitable scan control */ | |
3856 | struct scan_control sc = { | |
3857 | .gfp_mask = GFP_KERNEL, | |
3858 | .order = PAGE_ALLOC_COSTLY_ORDER + 1, | |
3859 | //.reclaim_mode = RECLAIM_MODE_SINGLE|RECLAIM_MODE_SYNC, // We only handle "SwapBacked" pages in this reclaim_mode! | |
3860 | }; | |
3861 | ||
3862 | /* Try to isolate this page */ | |
3863 | #ifdef CONFIG_MTKPASR_ALLEXTCOMP | |
3864 | ret = mtkpasr_isolate_page(page, 0x1); | |
3865 | #else | |
3866 | ret = mtkpasr_isolate_page(page); | |
3867 | #endif | |
3868 | if (ret) { | |
3869 | return ret; | |
3870 | } | |
3871 | ||
3872 | /* Check whether it is evictable! */ | |
3873 | if (unlikely(!page_evictable(page))) { | |
3874 | putback_lru_page(page); | |
3875 | return -EACCES; | |
3876 | } | |
3877 | ||
3878 | /* If it is Active, reference and deactivate it */ | |
3879 | if (PageActive(page)) { | |
3880 | active = TestClearPageActive(page); | |
3881 | } | |
3882 | ||
3883 | /* If we fail to lock this page, ignore it */ | |
3884 | if (!trylock_page(page)) { | |
3885 | goto putback; | |
3886 | } | |
3887 | ||
3888 | /* If page is in writeback, we don't handle it here! */ | |
3889 | if (PageWriteback(page)) { | |
3890 | goto unlock; | |
3891 | } | |
3892 | ||
3893 | /* | |
3894 | * Anonymous process memory has backing store? | |
3895 | * Try to allocate it some swap space here. | |
3896 | */ | |
3897 | if (PageAnon(page) && !PageSwapCache(page)) { | |
3898 | /* Check whether we have enough free memory */ | |
3899 | if (vm_swap_full()) { | |
3900 | goto unlock; | |
3901 | } | |
3902 | ||
3903 | /* Ok! It is safe to add this page to swap. */ | |
3904 | if (!add_to_swap(page, NULL)){ | |
3905 | goto unlock; | |
3906 | } | |
3907 | } | |
3908 | ||
3909 | /* We don't handle dirty file cache here (Related devices may be suspended) */ | |
3910 | if (page_is_file_cache(page)) { | |
3911 | /* How do we handle pages in VM_EXEC vmas? */ | |
3912 | if ((vm_flags & VM_EXEC)) { | |
3913 | goto unlock; | |
3914 | } | |
3915 | /* We don't handle dirty file pages! */ | |
3916 | if (PageDirty(page)) { | |
3917 | #ifdef CONFIG_MTKPASR_DEBUG | |
3918 | printk(KERN_ALERT "\n\n\n\n\n\n [%s][%d]\n\n\n\n\n\n",__FUNCTION__,__LINE__); | |
3919 | #endif | |
3920 | goto unlock; | |
3921 | } | |
3922 | } | |
3923 | ||
3924 | /* | |
3925 | * The page is mapped into the page tables of one or more | |
3926 | * processes. Try to unmap it here. | |
3927 | */ | |
3928 | mapping = page_mapping(page); | |
3929 | if (page_mapped(page) && mapping) { | |
3930 | #if 0 | |
3931 | /* Indicate unmap action for SwapBacked pages */ | |
3932 | if (PageSwapBacked(page)) { | |
3933 | unmap_flags |= TTU_IGNORE_ACCESS; | |
3934 | } | |
3935 | #endif | |
3936 | /* To unmap */ | |
3937 | switch (try_to_unmap(page, unmap_flags)) { | |
3938 | case SWAP_SUCCESS: | |
3939 | /* try to free the page below */ | |
3940 | break; | |
3941 | case SWAP_FAIL: | |
3942 | goto restore_swap; | |
3943 | case SWAP_AGAIN: | |
3944 | goto restore_swap; | |
3945 | case SWAP_MLOCK: | |
3946 | goto restore_swap; | |
3947 | ||
3948 | } | |
3949 | } | |
3950 | ||
3951 | /* Check whether it is dirtied. | |
3952 | * We have filtered out dirty file pages above. (IMPORTANT!) | |
3953 | * "VM_BUG_ON(!PageSwapBacked(page))" | |
3954 | * */ | |
3955 | if (PageDirty(page)) { | |
3956 | /* Page is dirty, try to write it out here */ | |
3957 | /* It's ok for zram swap! */ | |
3958 | /* Should we need to apply GFP_IOFS? */ | |
3959 | switch (pageout(page, mapping, &sc)) { | |
3960 | case PAGE_SUCCESS: | |
3961 | if (PageWriteback(page)) { | |
3962 | goto putback; | |
3963 | } | |
3964 | if (PageDirty(page)) { | |
3965 | goto putback; | |
3966 | } | |
3967 | ||
3968 | /* | |
3969 | * A synchronous write - probably a ramdisk. Go | |
3970 | * ahead and try to reclaim the page. | |
3971 | */ | |
3972 | if (!trylock_page(page)) { | |
3973 | goto putback; | |
3974 | } | |
3975 | if (PageDirty(page) || PageWriteback(page)) { | |
3976 | goto unlock; | |
3977 | } | |
3978 | mapping = page_mapping(page); | |
3979 | case PAGE_CLEAN: | |
3980 | /* try to free the page below */ | |
3981 | break; | |
3982 | default: | |
3983 | #ifdef CONFIG_MTKPASR_DEBUG | |
3984 | /*printk(KERN_ALERT "\n\n\n\n\n\n [%s][%d]\n\n\n\n\n\n",__FUNCTION__,__LINE__);*/ | |
3985 | #endif | |
3986 | goto restore_unmap; | |
3987 | } | |
3988 | } | |
3989 | ||
3990 | /* Release buffer */ | |
3991 | if (page_has_private(page)) { | |
3992 | if (!try_to_release_page(page, sc.gfp_mask)) { | |
3993 | goto unlock; | |
3994 | } | |
3995 | if (!mapping && page_count(page) == 1) { | |
3996 | unlock_page(page); | |
3997 | if (put_page_testzero(page)) { | |
3998 | goto freeit; | |
3999 | } else { | |
4000 | /* Race! TOCHECK */ | |
4001 | printk(KERN_ALERT "\n\n\n\n\n\n [%s][%d] RACE!!\n\n\n\n\n\n",__FUNCTION__,__LINE__); | |
4002 | goto notask; | |
4003 | } | |
4004 | } | |
4005 | } | |
4006 | if (!mapping || !__remove_mapping(mapping, page)) { | |
4007 | goto unlock; | |
4008 | } | |
4009 | ||
4010 | __clear_page_locked(page); | |
4011 | ||
4012 | freeit: | |
4013 | free_hot_cold_page(page, 0); | |
4014 | return 0; | |
4015 | ||
4016 | restore_unmap: | |
4017 | /* Do something */ | |
4018 | ||
4019 | restore_swap: | |
4020 | if (PageSwapCache(page)) | |
4021 | try_to_free_swap(page); | |
4022 | ||
4023 | unlock: | |
4024 | unlock_page(page); | |
4025 | ||
4026 | putback: | |
4027 | /* Activate it again if needed! */ | |
4028 | if (active) | |
4029 | SetPageActive(page); | |
4030 | ||
4031 | /* We don't putback them to corresponding LRUs, because we want to do more tasks outside this function! | |
4032 | putback_lru_page(page); */ | |
4033 | ||
4034 | /* Failedly dropped pages. Do migration! */ | |
4035 | return -EBUSY; | |
4036 | ||
4037 | notask: | |
4038 | return 0; | |
4039 | } | |
4040 | #endif |