Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
f30c2269 | 2 | * mm/page-writeback.c |
1da177e4 LT |
3 | * |
4 | * Copyright (C) 2002, Linus Torvalds. | |
5 | * | |
6 | * Contains functions related to writing back dirty pages at the | |
7 | * address_space level. | |
8 | * | |
9 | * 10Apr2002 akpm@zip.com.au | |
10 | * Initial version | |
11 | */ | |
12 | ||
13 | #include <linux/kernel.h> | |
14 | #include <linux/module.h> | |
15 | #include <linux/spinlock.h> | |
16 | #include <linux/fs.h> | |
17 | #include <linux/mm.h> | |
18 | #include <linux/swap.h> | |
19 | #include <linux/slab.h> | |
20 | #include <linux/pagemap.h> | |
21 | #include <linux/writeback.h> | |
22 | #include <linux/init.h> | |
23 | #include <linux/backing-dev.h> | |
55e829af | 24 | #include <linux/task_io_accounting_ops.h> |
1da177e4 LT |
25 | #include <linux/blkdev.h> |
26 | #include <linux/mpage.h> | |
d08b3851 | 27 | #include <linux/rmap.h> |
1da177e4 LT |
28 | #include <linux/percpu.h> |
29 | #include <linux/notifier.h> | |
30 | #include <linux/smp.h> | |
31 | #include <linux/sysctl.h> | |
32 | #include <linux/cpu.h> | |
33 | #include <linux/syscalls.h> | |
cf9a2ae8 | 34 | #include <linux/buffer_head.h> |
811d736f | 35 | #include <linux/pagevec.h> |
1da177e4 LT |
36 | |
37 | /* | |
38 | * The maximum number of pages to writeout in a single bdflush/kupdate | |
39 | * operation. We do this so we don't hold I_LOCK against an inode for | |
40 | * enormous amounts of time, which would block a userspace task which has | |
41 | * been forced to throttle against that inode. Also, the code reevaluates | |
42 | * the dirty each time it has written this many pages. | |
43 | */ | |
44 | #define MAX_WRITEBACK_PAGES 1024 | |
45 | ||
46 | /* | |
47 | * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited | |
48 | * will look to see if it needs to force writeback or throttling. | |
49 | */ | |
50 | static long ratelimit_pages = 32; | |
51 | ||
e236a166 | 52 | static int dirty_exceeded __cacheline_aligned_in_smp; /* Dirty mem may be over limit */ |
1da177e4 LT |
53 | |
54 | /* | |
55 | * When balance_dirty_pages decides that the caller needs to perform some | |
56 | * non-background writeback, this is how many pages it will attempt to write. | |
57 | * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably | |
58 | * large amounts of I/O are submitted. | |
59 | */ | |
60 | static inline long sync_writeback_pages(void) | |
61 | { | |
62 | return ratelimit_pages + ratelimit_pages / 2; | |
63 | } | |
64 | ||
65 | /* The following parameters are exported via /proc/sys/vm */ | |
66 | ||
67 | /* | |
68 | * Start background writeback (via pdflush) at this percentage | |
69 | */ | |
07db59bd | 70 | int dirty_background_ratio = 5; |
1da177e4 LT |
71 | |
72 | /* | |
73 | * The generator of dirty data starts writeback at this percentage | |
74 | */ | |
07db59bd | 75 | int vm_dirty_ratio = 10; |
1da177e4 LT |
76 | |
77 | /* | |
fd5403c7 | 78 | * The interval between `kupdate'-style writebacks, in jiffies |
1da177e4 | 79 | */ |
f6ef9438 | 80 | int dirty_writeback_interval = 5 * HZ; |
1da177e4 LT |
81 | |
82 | /* | |
fd5403c7 | 83 | * The longest number of jiffies for which data is allowed to remain dirty |
1da177e4 | 84 | */ |
f6ef9438 | 85 | int dirty_expire_interval = 30 * HZ; |
1da177e4 LT |
86 | |
87 | /* | |
88 | * Flag that makes the machine dump writes/reads and block dirtyings. | |
89 | */ | |
90 | int block_dump; | |
91 | ||
92 | /* | |
ed5b43f1 BS |
93 | * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: |
94 | * a full sync is triggered after this time elapses without any disk activity. | |
1da177e4 LT |
95 | */ |
96 | int laptop_mode; | |
97 | ||
98 | EXPORT_SYMBOL(laptop_mode); | |
99 | ||
100 | /* End of sysctl-exported parameters */ | |
101 | ||
102 | ||
103 | static void background_writeout(unsigned long _min_pages); | |
104 | ||
1da177e4 LT |
105 | /* |
106 | * Work out the current dirty-memory clamping and background writeout | |
107 | * thresholds. | |
108 | * | |
109 | * The main aim here is to lower them aggressively if there is a lot of mapped | |
110 | * memory around. To avoid stressing page reclaim with lots of unreclaimable | |
111 | * pages. It is better to clamp down on writers than to start swapping, and | |
112 | * performing lots of scanning. | |
113 | * | |
114 | * We only allow 1/2 of the currently-unmapped memory to be dirtied. | |
115 | * | |
116 | * We don't permit the clamping level to fall below 5% - that is getting rather | |
117 | * excessive. | |
118 | * | |
119 | * We make sure that the background writeout level is below the adjusted | |
120 | * clamping level. | |
121 | */ | |
1b424464 CL |
122 | |
123 | static unsigned long highmem_dirtyable_memory(unsigned long total) | |
124 | { | |
125 | #ifdef CONFIG_HIGHMEM | |
126 | int node; | |
127 | unsigned long x = 0; | |
128 | ||
129 | for_each_online_node(node) { | |
130 | struct zone *z = | |
131 | &NODE_DATA(node)->node_zones[ZONE_HIGHMEM]; | |
132 | ||
133 | x += zone_page_state(z, NR_FREE_PAGES) | |
134 | + zone_page_state(z, NR_INACTIVE) | |
135 | + zone_page_state(z, NR_ACTIVE); | |
136 | } | |
137 | /* | |
138 | * Make sure that the number of highmem pages is never larger | |
139 | * than the number of the total dirtyable memory. This can only | |
140 | * occur in very strange VM situations but we want to make sure | |
141 | * that this does not occur. | |
142 | */ | |
143 | return min(x, total); | |
144 | #else | |
145 | return 0; | |
146 | #endif | |
147 | } | |
148 | ||
149 | static unsigned long determine_dirtyable_memory(void) | |
150 | { | |
151 | unsigned long x; | |
152 | ||
153 | x = global_page_state(NR_FREE_PAGES) | |
154 | + global_page_state(NR_INACTIVE) | |
155 | + global_page_state(NR_ACTIVE); | |
156 | x -= highmem_dirtyable_memory(x); | |
157 | return x + 1; /* Ensure that we never return 0 */ | |
158 | } | |
159 | ||
1da177e4 | 160 | static void |
c24f21bd CL |
161 | get_dirty_limits(long *pbackground, long *pdirty, |
162 | struct address_space *mapping) | |
1da177e4 LT |
163 | { |
164 | int background_ratio; /* Percentages */ | |
165 | int dirty_ratio; | |
166 | int unmapped_ratio; | |
167 | long background; | |
168 | long dirty; | |
1b424464 | 169 | unsigned long available_memory = determine_dirtyable_memory(); |
1da177e4 LT |
170 | struct task_struct *tsk; |
171 | ||
c24f21bd CL |
172 | unmapped_ratio = 100 - ((global_page_state(NR_FILE_MAPPED) + |
173 | global_page_state(NR_ANON_PAGES)) * 100) / | |
1b424464 | 174 | available_memory; |
1da177e4 LT |
175 | |
176 | dirty_ratio = vm_dirty_ratio; | |
177 | if (dirty_ratio > unmapped_ratio / 2) | |
178 | dirty_ratio = unmapped_ratio / 2; | |
179 | ||
180 | if (dirty_ratio < 5) | |
181 | dirty_ratio = 5; | |
182 | ||
183 | background_ratio = dirty_background_ratio; | |
184 | if (background_ratio >= dirty_ratio) | |
185 | background_ratio = dirty_ratio / 2; | |
186 | ||
187 | background = (background_ratio * available_memory) / 100; | |
188 | dirty = (dirty_ratio * available_memory) / 100; | |
189 | tsk = current; | |
190 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) { | |
191 | background += background / 4; | |
192 | dirty += dirty / 4; | |
193 | } | |
194 | *pbackground = background; | |
195 | *pdirty = dirty; | |
196 | } | |
197 | ||
198 | /* | |
199 | * balance_dirty_pages() must be called by processes which are generating dirty | |
200 | * data. It looks at the number of dirty pages in the machine and will force | |
201 | * the caller to perform writeback if the system is over `vm_dirty_ratio'. | |
202 | * If we're over `background_thresh' then pdflush is woken to perform some | |
203 | * writeout. | |
204 | */ | |
205 | static void balance_dirty_pages(struct address_space *mapping) | |
206 | { | |
1da177e4 LT |
207 | long nr_reclaimable; |
208 | long background_thresh; | |
209 | long dirty_thresh; | |
210 | unsigned long pages_written = 0; | |
211 | unsigned long write_chunk = sync_writeback_pages(); | |
212 | ||
213 | struct backing_dev_info *bdi = mapping->backing_dev_info; | |
214 | ||
215 | for (;;) { | |
216 | struct writeback_control wbc = { | |
217 | .bdi = bdi, | |
218 | .sync_mode = WB_SYNC_NONE, | |
219 | .older_than_this = NULL, | |
220 | .nr_to_write = write_chunk, | |
111ebb6e | 221 | .range_cyclic = 1, |
1da177e4 LT |
222 | }; |
223 | ||
c24f21bd CL |
224 | get_dirty_limits(&background_thresh, &dirty_thresh, mapping); |
225 | nr_reclaimable = global_page_state(NR_FILE_DIRTY) + | |
226 | global_page_state(NR_UNSTABLE_NFS); | |
227 | if (nr_reclaimable + global_page_state(NR_WRITEBACK) <= | |
228 | dirty_thresh) | |
229 | break; | |
1da177e4 | 230 | |
e236a166 AM |
231 | if (!dirty_exceeded) |
232 | dirty_exceeded = 1; | |
1da177e4 LT |
233 | |
234 | /* Note: nr_reclaimable denotes nr_dirty + nr_unstable. | |
235 | * Unstable writes are a feature of certain networked | |
236 | * filesystems (i.e. NFS) in which data may have been | |
237 | * written to the server's write cache, but has not yet | |
238 | * been flushed to permanent storage. | |
239 | */ | |
240 | if (nr_reclaimable) { | |
241 | writeback_inodes(&wbc); | |
c24f21bd CL |
242 | get_dirty_limits(&background_thresh, |
243 | &dirty_thresh, mapping); | |
244 | nr_reclaimable = global_page_state(NR_FILE_DIRTY) + | |
245 | global_page_state(NR_UNSTABLE_NFS); | |
246 | if (nr_reclaimable + | |
247 | global_page_state(NR_WRITEBACK) | |
248 | <= dirty_thresh) | |
249 | break; | |
1da177e4 LT |
250 | pages_written += write_chunk - wbc.nr_to_write; |
251 | if (pages_written >= write_chunk) | |
252 | break; /* We've done our duty */ | |
253 | } | |
3fcfab16 | 254 | congestion_wait(WRITE, HZ/10); |
1da177e4 LT |
255 | } |
256 | ||
c24f21bd CL |
257 | if (nr_reclaimable + global_page_state(NR_WRITEBACK) |
258 | <= dirty_thresh && dirty_exceeded) | |
259 | dirty_exceeded = 0; | |
1da177e4 LT |
260 | |
261 | if (writeback_in_progress(bdi)) | |
262 | return; /* pdflush is already working this queue */ | |
263 | ||
264 | /* | |
265 | * In laptop mode, we wait until hitting the higher threshold before | |
266 | * starting background writeout, and then write out all the way down | |
267 | * to the lower threshold. So slow writers cause minimal disk activity. | |
268 | * | |
269 | * In normal mode, we start background writeout at the lower | |
270 | * background_thresh, to keep the amount of dirty memory low. | |
271 | */ | |
272 | if ((laptop_mode && pages_written) || | |
273 | (!laptop_mode && (nr_reclaimable > background_thresh))) | |
274 | pdflush_operation(background_writeout, 0); | |
275 | } | |
276 | ||
edc79b2a PZ |
277 | void set_page_dirty_balance(struct page *page) |
278 | { | |
279 | if (set_page_dirty(page)) { | |
280 | struct address_space *mapping = page_mapping(page); | |
281 | ||
282 | if (mapping) | |
283 | balance_dirty_pages_ratelimited(mapping); | |
284 | } | |
285 | } | |
286 | ||
1da177e4 | 287 | /** |
fa5a734e | 288 | * balance_dirty_pages_ratelimited_nr - balance dirty memory state |
67be2dd1 | 289 | * @mapping: address_space which was dirtied |
a580290c | 290 | * @nr_pages_dirtied: number of pages which the caller has just dirtied |
1da177e4 LT |
291 | * |
292 | * Processes which are dirtying memory should call in here once for each page | |
293 | * which was newly dirtied. The function will periodically check the system's | |
294 | * dirty state and will initiate writeback if needed. | |
295 | * | |
296 | * On really big machines, get_writeback_state is expensive, so try to avoid | |
297 | * calling it too often (ratelimiting). But once we're over the dirty memory | |
298 | * limit we decrease the ratelimiting by a lot, to prevent individual processes | |
299 | * from overshooting the limit by (ratelimit_pages) each. | |
300 | */ | |
fa5a734e AM |
301 | void balance_dirty_pages_ratelimited_nr(struct address_space *mapping, |
302 | unsigned long nr_pages_dirtied) | |
1da177e4 | 303 | { |
fa5a734e AM |
304 | static DEFINE_PER_CPU(unsigned long, ratelimits) = 0; |
305 | unsigned long ratelimit; | |
306 | unsigned long *p; | |
1da177e4 LT |
307 | |
308 | ratelimit = ratelimit_pages; | |
309 | if (dirty_exceeded) | |
310 | ratelimit = 8; | |
311 | ||
312 | /* | |
313 | * Check the rate limiting. Also, we do not want to throttle real-time | |
314 | * tasks in balance_dirty_pages(). Period. | |
315 | */ | |
fa5a734e AM |
316 | preempt_disable(); |
317 | p = &__get_cpu_var(ratelimits); | |
318 | *p += nr_pages_dirtied; | |
319 | if (unlikely(*p >= ratelimit)) { | |
320 | *p = 0; | |
321 | preempt_enable(); | |
1da177e4 LT |
322 | balance_dirty_pages(mapping); |
323 | return; | |
324 | } | |
fa5a734e | 325 | preempt_enable(); |
1da177e4 | 326 | } |
fa5a734e | 327 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr); |
1da177e4 | 328 | |
232ea4d6 | 329 | void throttle_vm_writeout(gfp_t gfp_mask) |
1da177e4 | 330 | { |
1da177e4 LT |
331 | long background_thresh; |
332 | long dirty_thresh; | |
333 | ||
232ea4d6 AM |
334 | if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) { |
335 | /* | |
336 | * The caller might hold locks which can prevent IO completion | |
337 | * or progress in the filesystem. So we cannot just sit here | |
338 | * waiting for IO to complete. | |
339 | */ | |
340 | congestion_wait(WRITE, HZ/10); | |
341 | return; | |
342 | } | |
343 | ||
1da177e4 | 344 | for ( ; ; ) { |
c24f21bd | 345 | get_dirty_limits(&background_thresh, &dirty_thresh, NULL); |
1da177e4 LT |
346 | |
347 | /* | |
348 | * Boost the allowable dirty threshold a bit for page | |
349 | * allocators so they don't get DoS'ed by heavy writers | |
350 | */ | |
351 | dirty_thresh += dirty_thresh / 10; /* wheeee... */ | |
352 | ||
c24f21bd CL |
353 | if (global_page_state(NR_UNSTABLE_NFS) + |
354 | global_page_state(NR_WRITEBACK) <= dirty_thresh) | |
355 | break; | |
3fcfab16 | 356 | congestion_wait(WRITE, HZ/10); |
1da177e4 LT |
357 | } |
358 | } | |
359 | ||
1da177e4 LT |
360 | /* |
361 | * writeback at least _min_pages, and keep writing until the amount of dirty | |
362 | * memory is less than the background threshold, or until we're all clean. | |
363 | */ | |
364 | static void background_writeout(unsigned long _min_pages) | |
365 | { | |
366 | long min_pages = _min_pages; | |
367 | struct writeback_control wbc = { | |
368 | .bdi = NULL, | |
369 | .sync_mode = WB_SYNC_NONE, | |
370 | .older_than_this = NULL, | |
371 | .nr_to_write = 0, | |
372 | .nonblocking = 1, | |
111ebb6e | 373 | .range_cyclic = 1, |
1da177e4 LT |
374 | }; |
375 | ||
376 | for ( ; ; ) { | |
1da177e4 LT |
377 | long background_thresh; |
378 | long dirty_thresh; | |
379 | ||
c24f21bd CL |
380 | get_dirty_limits(&background_thresh, &dirty_thresh, NULL); |
381 | if (global_page_state(NR_FILE_DIRTY) + | |
382 | global_page_state(NR_UNSTABLE_NFS) < background_thresh | |
1da177e4 LT |
383 | && min_pages <= 0) |
384 | break; | |
385 | wbc.encountered_congestion = 0; | |
386 | wbc.nr_to_write = MAX_WRITEBACK_PAGES; | |
387 | wbc.pages_skipped = 0; | |
388 | writeback_inodes(&wbc); | |
389 | min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write; | |
390 | if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) { | |
391 | /* Wrote less than expected */ | |
3fcfab16 | 392 | congestion_wait(WRITE, HZ/10); |
1da177e4 LT |
393 | if (!wbc.encountered_congestion) |
394 | break; | |
395 | } | |
396 | } | |
397 | } | |
398 | ||
399 | /* | |
400 | * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back | |
401 | * the whole world. Returns 0 if a pdflush thread was dispatched. Returns | |
402 | * -1 if all pdflush threads were busy. | |
403 | */ | |
687a21ce | 404 | int wakeup_pdflush(long nr_pages) |
1da177e4 | 405 | { |
c24f21bd CL |
406 | if (nr_pages == 0) |
407 | nr_pages = global_page_state(NR_FILE_DIRTY) + | |
408 | global_page_state(NR_UNSTABLE_NFS); | |
1da177e4 LT |
409 | return pdflush_operation(background_writeout, nr_pages); |
410 | } | |
411 | ||
412 | static void wb_timer_fn(unsigned long unused); | |
413 | static void laptop_timer_fn(unsigned long unused); | |
414 | ||
8d06afab IM |
415 | static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0); |
416 | static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0); | |
1da177e4 LT |
417 | |
418 | /* | |
419 | * Periodic writeback of "old" data. | |
420 | * | |
421 | * Define "old": the first time one of an inode's pages is dirtied, we mark the | |
422 | * dirtying-time in the inode's address_space. So this periodic writeback code | |
423 | * just walks the superblock inode list, writing back any inodes which are | |
424 | * older than a specific point in time. | |
425 | * | |
f6ef9438 BS |
426 | * Try to run once per dirty_writeback_interval. But if a writeback event |
427 | * takes longer than a dirty_writeback_interval interval, then leave a | |
1da177e4 LT |
428 | * one-second gap. |
429 | * | |
430 | * older_than_this takes precedence over nr_to_write. So we'll only write back | |
431 | * all dirty pages if they are all attached to "old" mappings. | |
432 | */ | |
433 | static void wb_kupdate(unsigned long arg) | |
434 | { | |
435 | unsigned long oldest_jif; | |
436 | unsigned long start_jif; | |
437 | unsigned long next_jif; | |
438 | long nr_to_write; | |
1da177e4 LT |
439 | struct writeback_control wbc = { |
440 | .bdi = NULL, | |
441 | .sync_mode = WB_SYNC_NONE, | |
442 | .older_than_this = &oldest_jif, | |
443 | .nr_to_write = 0, | |
444 | .nonblocking = 1, | |
445 | .for_kupdate = 1, | |
111ebb6e | 446 | .range_cyclic = 1, |
1da177e4 LT |
447 | }; |
448 | ||
449 | sync_supers(); | |
450 | ||
f6ef9438 | 451 | oldest_jif = jiffies - dirty_expire_interval; |
1da177e4 | 452 | start_jif = jiffies; |
f6ef9438 | 453 | next_jif = start_jif + dirty_writeback_interval; |
c24f21bd CL |
454 | nr_to_write = global_page_state(NR_FILE_DIRTY) + |
455 | global_page_state(NR_UNSTABLE_NFS) + | |
1da177e4 LT |
456 | (inodes_stat.nr_inodes - inodes_stat.nr_unused); |
457 | while (nr_to_write > 0) { | |
458 | wbc.encountered_congestion = 0; | |
459 | wbc.nr_to_write = MAX_WRITEBACK_PAGES; | |
460 | writeback_inodes(&wbc); | |
461 | if (wbc.nr_to_write > 0) { | |
462 | if (wbc.encountered_congestion) | |
3fcfab16 | 463 | congestion_wait(WRITE, HZ/10); |
1da177e4 LT |
464 | else |
465 | break; /* All the old data is written */ | |
466 | } | |
467 | nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write; | |
468 | } | |
469 | if (time_before(next_jif, jiffies + HZ)) | |
470 | next_jif = jiffies + HZ; | |
f6ef9438 | 471 | if (dirty_writeback_interval) |
1da177e4 LT |
472 | mod_timer(&wb_timer, next_jif); |
473 | } | |
474 | ||
475 | /* | |
476 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
477 | */ | |
478 | int dirty_writeback_centisecs_handler(ctl_table *table, int write, | |
479 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) | |
480 | { | |
f6ef9438 BS |
481 | proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos); |
482 | if (dirty_writeback_interval) { | |
1da177e4 | 483 | mod_timer(&wb_timer, |
f6ef9438 BS |
484 | jiffies + dirty_writeback_interval); |
485 | } else { | |
1da177e4 LT |
486 | del_timer(&wb_timer); |
487 | } | |
488 | return 0; | |
489 | } | |
490 | ||
491 | static void wb_timer_fn(unsigned long unused) | |
492 | { | |
493 | if (pdflush_operation(wb_kupdate, 0) < 0) | |
494 | mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */ | |
495 | } | |
496 | ||
497 | static void laptop_flush(unsigned long unused) | |
498 | { | |
499 | sys_sync(); | |
500 | } | |
501 | ||
502 | static void laptop_timer_fn(unsigned long unused) | |
503 | { | |
504 | pdflush_operation(laptop_flush, 0); | |
505 | } | |
506 | ||
507 | /* | |
508 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
509 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
510 | * then push it back - the user is still using the disk. | |
511 | */ | |
512 | void laptop_io_completion(void) | |
513 | { | |
ed5b43f1 | 514 | mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
515 | } |
516 | ||
517 | /* | |
518 | * We're in laptop mode and we've just synced. The sync's writes will have | |
519 | * caused another writeback to be scheduled by laptop_io_completion. | |
520 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
521 | */ | |
522 | void laptop_sync_completion(void) | |
523 | { | |
524 | del_timer(&laptop_mode_wb_timer); | |
525 | } | |
526 | ||
527 | /* | |
528 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
529 | * if a large number of processes all perform writes at the same time. | |
530 | * If it is too low then SMP machines will call the (expensive) | |
531 | * get_writeback_state too often. | |
532 | * | |
533 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
534 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
535 | * thresholds before writeback cuts in. | |
536 | * | |
537 | * But the limit should not be set too high. Because it also controls the | |
538 | * amount of memory which the balance_dirty_pages() caller has to write back. | |
539 | * If this is too large then the caller will block on the IO queue all the | |
540 | * time. So limit it to four megabytes - the balance_dirty_pages() caller | |
541 | * will write six megabyte chunks, max. | |
542 | */ | |
543 | ||
2d1d43f6 | 544 | void writeback_set_ratelimit(void) |
1da177e4 | 545 | { |
40c99aae | 546 | ratelimit_pages = vm_total_pages / (num_online_cpus() * 32); |
1da177e4 LT |
547 | if (ratelimit_pages < 16) |
548 | ratelimit_pages = 16; | |
549 | if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024) | |
550 | ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE; | |
551 | } | |
552 | ||
26c2143b | 553 | static int __cpuinit |
1da177e4 LT |
554 | ratelimit_handler(struct notifier_block *self, unsigned long u, void *v) |
555 | { | |
2d1d43f6 | 556 | writeback_set_ratelimit(); |
aa0f0303 | 557 | return NOTIFY_DONE; |
1da177e4 LT |
558 | } |
559 | ||
74b85f37 | 560 | static struct notifier_block __cpuinitdata ratelimit_nb = { |
1da177e4 LT |
561 | .notifier_call = ratelimit_handler, |
562 | .next = NULL, | |
563 | }; | |
564 | ||
565 | /* | |
dc6e29da LT |
566 | * Called early on to tune the page writeback dirty limits. |
567 | * | |
568 | * We used to scale dirty pages according to how total memory | |
569 | * related to pages that could be allocated for buffers (by | |
570 | * comparing nr_free_buffer_pages() to vm_total_pages. | |
571 | * | |
572 | * However, that was when we used "dirty_ratio" to scale with | |
573 | * all memory, and we don't do that any more. "dirty_ratio" | |
574 | * is now applied to total non-HIGHPAGE memory (by subtracting | |
575 | * totalhigh_pages from vm_total_pages), and as such we can't | |
576 | * get into the old insane situation any more where we had | |
577 | * large amounts of dirty pages compared to a small amount of | |
578 | * non-HIGHMEM memory. | |
579 | * | |
580 | * But we might still want to scale the dirty_ratio by how | |
581 | * much memory the box has.. | |
1da177e4 LT |
582 | */ |
583 | void __init page_writeback_init(void) | |
584 | { | |
f6ef9438 | 585 | mod_timer(&wb_timer, jiffies + dirty_writeback_interval); |
2d1d43f6 | 586 | writeback_set_ratelimit(); |
1da177e4 LT |
587 | register_cpu_notifier(&ratelimit_nb); |
588 | } | |
589 | ||
811d736f | 590 | /** |
72fd4a35 | 591 | * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. |
811d736f DH |
592 | * @mapping: address space structure to write |
593 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
594 | * | |
595 | * This is a library function, which implements the writepages() | |
596 | * address_space_operation. | |
597 | * | |
598 | * If a page is already under I/O, generic_writepages() skips it, even | |
599 | * if it's dirty. This is desirable behaviour for memory-cleaning writeback, | |
600 | * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() | |
601 | * and msync() need to guarantee that all the data which was dirty at the time | |
602 | * the call was made get new I/O started against them. If wbc->sync_mode is | |
603 | * WB_SYNC_ALL then we were called for data integrity and we must wait for | |
604 | * existing IO to complete. | |
605 | * | |
606 | * Derived from mpage_writepages() - if you fix this you should check that | |
607 | * also! | |
608 | */ | |
609 | int generic_writepages(struct address_space *mapping, | |
610 | struct writeback_control *wbc) | |
611 | { | |
612 | struct backing_dev_info *bdi = mapping->backing_dev_info; | |
613 | int ret = 0; | |
614 | int done = 0; | |
615 | int (*writepage)(struct page *page, struct writeback_control *wbc); | |
616 | struct pagevec pvec; | |
617 | int nr_pages; | |
618 | pgoff_t index; | |
619 | pgoff_t end; /* Inclusive */ | |
620 | int scanned = 0; | |
621 | int range_whole = 0; | |
622 | ||
623 | if (wbc->nonblocking && bdi_write_congested(bdi)) { | |
624 | wbc->encountered_congestion = 1; | |
625 | return 0; | |
626 | } | |
627 | ||
628 | writepage = mapping->a_ops->writepage; | |
629 | ||
630 | /* deal with chardevs and other special file */ | |
631 | if (!writepage) | |
632 | return 0; | |
633 | ||
634 | pagevec_init(&pvec, 0); | |
635 | if (wbc->range_cyclic) { | |
636 | index = mapping->writeback_index; /* Start from prev offset */ | |
637 | end = -1; | |
638 | } else { | |
639 | index = wbc->range_start >> PAGE_CACHE_SHIFT; | |
640 | end = wbc->range_end >> PAGE_CACHE_SHIFT; | |
641 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) | |
642 | range_whole = 1; | |
643 | scanned = 1; | |
644 | } | |
645 | retry: | |
646 | while (!done && (index <= end) && | |
647 | (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, | |
648 | PAGECACHE_TAG_DIRTY, | |
649 | min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) { | |
650 | unsigned i; | |
651 | ||
652 | scanned = 1; | |
653 | for (i = 0; i < nr_pages; i++) { | |
654 | struct page *page = pvec.pages[i]; | |
655 | ||
656 | /* | |
657 | * At this point we hold neither mapping->tree_lock nor | |
658 | * lock on the page itself: the page may be truncated or | |
659 | * invalidated (changing page->mapping to NULL), or even | |
660 | * swizzled back from swapper_space to tmpfs file | |
661 | * mapping | |
662 | */ | |
663 | lock_page(page); | |
664 | ||
665 | if (unlikely(page->mapping != mapping)) { | |
666 | unlock_page(page); | |
667 | continue; | |
668 | } | |
669 | ||
670 | if (!wbc->range_cyclic && page->index > end) { | |
671 | done = 1; | |
672 | unlock_page(page); | |
673 | continue; | |
674 | } | |
675 | ||
676 | if (wbc->sync_mode != WB_SYNC_NONE) | |
677 | wait_on_page_writeback(page); | |
678 | ||
679 | if (PageWriteback(page) || | |
680 | !clear_page_dirty_for_io(page)) { | |
681 | unlock_page(page); | |
682 | continue; | |
683 | } | |
684 | ||
685 | ret = (*writepage)(page, wbc); | |
3e9f45bd | 686 | mapping_set_error(mapping, ret); |
811d736f DH |
687 | |
688 | if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) | |
689 | unlock_page(page); | |
690 | if (ret || (--(wbc->nr_to_write) <= 0)) | |
691 | done = 1; | |
692 | if (wbc->nonblocking && bdi_write_congested(bdi)) { | |
693 | wbc->encountered_congestion = 1; | |
694 | done = 1; | |
695 | } | |
696 | } | |
697 | pagevec_release(&pvec); | |
698 | cond_resched(); | |
699 | } | |
700 | if (!scanned && !done) { | |
701 | /* | |
702 | * We hit the last page and there is more work to be done: wrap | |
703 | * back to the start of the file | |
704 | */ | |
705 | scanned = 1; | |
706 | index = 0; | |
707 | goto retry; | |
708 | } | |
709 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) | |
710 | mapping->writeback_index = index; | |
711 | return ret; | |
712 | } | |
713 | ||
714 | EXPORT_SYMBOL(generic_writepages); | |
715 | ||
1da177e4 LT |
716 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
717 | { | |
22905f77 AM |
718 | int ret; |
719 | ||
1da177e4 LT |
720 | if (wbc->nr_to_write <= 0) |
721 | return 0; | |
22905f77 | 722 | wbc->for_writepages = 1; |
1da177e4 | 723 | if (mapping->a_ops->writepages) |
d08b3851 | 724 | ret = mapping->a_ops->writepages(mapping, wbc); |
22905f77 AM |
725 | else |
726 | ret = generic_writepages(mapping, wbc); | |
727 | wbc->for_writepages = 0; | |
728 | return ret; | |
1da177e4 LT |
729 | } |
730 | ||
731 | /** | |
732 | * write_one_page - write out a single page and optionally wait on I/O | |
67be2dd1 MW |
733 | * @page: the page to write |
734 | * @wait: if true, wait on writeout | |
1da177e4 LT |
735 | * |
736 | * The page must be locked by the caller and will be unlocked upon return. | |
737 | * | |
738 | * write_one_page() returns a negative error code if I/O failed. | |
739 | */ | |
740 | int write_one_page(struct page *page, int wait) | |
741 | { | |
742 | struct address_space *mapping = page->mapping; | |
743 | int ret = 0; | |
744 | struct writeback_control wbc = { | |
745 | .sync_mode = WB_SYNC_ALL, | |
746 | .nr_to_write = 1, | |
747 | }; | |
748 | ||
749 | BUG_ON(!PageLocked(page)); | |
750 | ||
751 | if (wait) | |
752 | wait_on_page_writeback(page); | |
753 | ||
754 | if (clear_page_dirty_for_io(page)) { | |
755 | page_cache_get(page); | |
756 | ret = mapping->a_ops->writepage(page, &wbc); | |
757 | if (ret == 0 && wait) { | |
758 | wait_on_page_writeback(page); | |
759 | if (PageError(page)) | |
760 | ret = -EIO; | |
761 | } | |
762 | page_cache_release(page); | |
763 | } else { | |
764 | unlock_page(page); | |
765 | } | |
766 | return ret; | |
767 | } | |
768 | EXPORT_SYMBOL(write_one_page); | |
769 | ||
76719325 KC |
770 | /* |
771 | * For address_spaces which do not use buffers nor write back. | |
772 | */ | |
773 | int __set_page_dirty_no_writeback(struct page *page) | |
774 | { | |
775 | if (!PageDirty(page)) | |
776 | SetPageDirty(page); | |
777 | return 0; | |
778 | } | |
779 | ||
1da177e4 LT |
780 | /* |
781 | * For address_spaces which do not use buffers. Just tag the page as dirty in | |
782 | * its radix tree. | |
783 | * | |
784 | * This is also used when a single buffer is being dirtied: we want to set the | |
785 | * page dirty in that case, but not all the buffers. This is a "bottom-up" | |
786 | * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying. | |
787 | * | |
788 | * Most callers have locked the page, which pins the address_space in memory. | |
789 | * But zap_pte_range() does not lock the page, however in that case the | |
790 | * mapping is pinned by the vma's ->vm_file reference. | |
791 | * | |
792 | * We take care to handle the case where the page was truncated from the | |
793 | * mapping by re-checking page_mapping() insode tree_lock. | |
794 | */ | |
795 | int __set_page_dirty_nobuffers(struct page *page) | |
796 | { | |
1da177e4 LT |
797 | if (!TestSetPageDirty(page)) { |
798 | struct address_space *mapping = page_mapping(page); | |
799 | struct address_space *mapping2; | |
800 | ||
8c08540f AM |
801 | if (!mapping) |
802 | return 1; | |
803 | ||
804 | write_lock_irq(&mapping->tree_lock); | |
805 | mapping2 = page_mapping(page); | |
806 | if (mapping2) { /* Race with truncate? */ | |
807 | BUG_ON(mapping2 != mapping); | |
55e829af | 808 | if (mapping_cap_account_dirty(mapping)) { |
8c08540f | 809 | __inc_zone_page_state(page, NR_FILE_DIRTY); |
55e829af AM |
810 | task_io_account_write(PAGE_CACHE_SIZE); |
811 | } | |
8c08540f AM |
812 | radix_tree_tag_set(&mapping->page_tree, |
813 | page_index(page), PAGECACHE_TAG_DIRTY); | |
814 | } | |
815 | write_unlock_irq(&mapping->tree_lock); | |
816 | if (mapping->host) { | |
817 | /* !PageAnon && !swapper_space */ | |
818 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 819 | } |
4741c9fd | 820 | return 1; |
1da177e4 | 821 | } |
4741c9fd | 822 | return 0; |
1da177e4 LT |
823 | } |
824 | EXPORT_SYMBOL(__set_page_dirty_nobuffers); | |
825 | ||
826 | /* | |
827 | * When a writepage implementation decides that it doesn't want to write this | |
828 | * page for some reason, it should redirty the locked page via | |
829 | * redirty_page_for_writepage() and it should then unlock the page and return 0 | |
830 | */ | |
831 | int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page) | |
832 | { | |
833 | wbc->pages_skipped++; | |
834 | return __set_page_dirty_nobuffers(page); | |
835 | } | |
836 | EXPORT_SYMBOL(redirty_page_for_writepage); | |
837 | ||
838 | /* | |
839 | * If the mapping doesn't provide a set_page_dirty a_op, then | |
840 | * just fall through and assume that it wants buffer_heads. | |
841 | */ | |
842 | int fastcall set_page_dirty(struct page *page) | |
843 | { | |
844 | struct address_space *mapping = page_mapping(page); | |
845 | ||
846 | if (likely(mapping)) { | |
847 | int (*spd)(struct page *) = mapping->a_ops->set_page_dirty; | |
9361401e DH |
848 | #ifdef CONFIG_BLOCK |
849 | if (!spd) | |
850 | spd = __set_page_dirty_buffers; | |
851 | #endif | |
852 | return (*spd)(page); | |
1da177e4 | 853 | } |
4741c9fd AM |
854 | if (!PageDirty(page)) { |
855 | if (!TestSetPageDirty(page)) | |
856 | return 1; | |
857 | } | |
1da177e4 LT |
858 | return 0; |
859 | } | |
860 | EXPORT_SYMBOL(set_page_dirty); | |
861 | ||
862 | /* | |
863 | * set_page_dirty() is racy if the caller has no reference against | |
864 | * page->mapping->host, and if the page is unlocked. This is because another | |
865 | * CPU could truncate the page off the mapping and then free the mapping. | |
866 | * | |
867 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
868 | * holds a reference on the inode by having an open file. | |
869 | * | |
870 | * In other cases, the page should be locked before running set_page_dirty(). | |
871 | */ | |
872 | int set_page_dirty_lock(struct page *page) | |
873 | { | |
874 | int ret; | |
875 | ||
db37648c | 876 | lock_page_nosync(page); |
1da177e4 LT |
877 | ret = set_page_dirty(page); |
878 | unlock_page(page); | |
879 | return ret; | |
880 | } | |
881 | EXPORT_SYMBOL(set_page_dirty_lock); | |
882 | ||
1da177e4 LT |
883 | /* |
884 | * Clear a page's dirty flag, while caring for dirty memory accounting. | |
885 | * Returns true if the page was previously dirty. | |
886 | * | |
887 | * This is for preparing to put the page under writeout. We leave the page | |
888 | * tagged as dirty in the radix tree so that a concurrent write-for-sync | |
889 | * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage | |
890 | * implementation will run either set_page_writeback() or set_page_dirty(), | |
891 | * at which stage we bring the page's dirty flag and radix-tree dirty tag | |
892 | * back into sync. | |
893 | * | |
894 | * This incoherency between the page's dirty flag and radix-tree tag is | |
895 | * unfortunate, but it only exists while the page is locked. | |
896 | */ | |
897 | int clear_page_dirty_for_io(struct page *page) | |
898 | { | |
899 | struct address_space *mapping = page_mapping(page); | |
900 | ||
7658cc28 LT |
901 | if (mapping && mapping_cap_account_dirty(mapping)) { |
902 | /* | |
903 | * Yes, Virginia, this is indeed insane. | |
904 | * | |
905 | * We use this sequence to make sure that | |
906 | * (a) we account for dirty stats properly | |
907 | * (b) we tell the low-level filesystem to | |
908 | * mark the whole page dirty if it was | |
909 | * dirty in a pagetable. Only to then | |
910 | * (c) clean the page again and return 1 to | |
911 | * cause the writeback. | |
912 | * | |
913 | * This way we avoid all nasty races with the | |
914 | * dirty bit in multiple places and clearing | |
915 | * them concurrently from different threads. | |
916 | * | |
917 | * Note! Normally the "set_page_dirty(page)" | |
918 | * has no effect on the actual dirty bit - since | |
919 | * that will already usually be set. But we | |
920 | * need the side effects, and it can help us | |
921 | * avoid races. | |
922 | * | |
923 | * We basically use the page "master dirty bit" | |
924 | * as a serialization point for all the different | |
925 | * threads doing their things. | |
926 | * | |
927 | * FIXME! We still have a race here: if somebody | |
928 | * adds the page back to the page tables in | |
929 | * between the "page_mkclean()" and the "TestClearPageDirty()", | |
930 | * we might have it mapped without the dirty bit set. | |
931 | */ | |
932 | if (page_mkclean(page)) | |
933 | set_page_dirty(page); | |
934 | if (TestClearPageDirty(page)) { | |
8c08540f | 935 | dec_zone_page_state(page, NR_FILE_DIRTY); |
7658cc28 | 936 | return 1; |
1da177e4 | 937 | } |
7658cc28 | 938 | return 0; |
1da177e4 | 939 | } |
7658cc28 | 940 | return TestClearPageDirty(page); |
1da177e4 | 941 | } |
58bb01a9 | 942 | EXPORT_SYMBOL(clear_page_dirty_for_io); |
1da177e4 LT |
943 | |
944 | int test_clear_page_writeback(struct page *page) | |
945 | { | |
946 | struct address_space *mapping = page_mapping(page); | |
947 | int ret; | |
948 | ||
949 | if (mapping) { | |
950 | unsigned long flags; | |
951 | ||
952 | write_lock_irqsave(&mapping->tree_lock, flags); | |
953 | ret = TestClearPageWriteback(page); | |
954 | if (ret) | |
955 | radix_tree_tag_clear(&mapping->page_tree, | |
956 | page_index(page), | |
957 | PAGECACHE_TAG_WRITEBACK); | |
958 | write_unlock_irqrestore(&mapping->tree_lock, flags); | |
959 | } else { | |
960 | ret = TestClearPageWriteback(page); | |
961 | } | |
962 | return ret; | |
963 | } | |
964 | ||
965 | int test_set_page_writeback(struct page *page) | |
966 | { | |
967 | struct address_space *mapping = page_mapping(page); | |
968 | int ret; | |
969 | ||
970 | if (mapping) { | |
971 | unsigned long flags; | |
972 | ||
973 | write_lock_irqsave(&mapping->tree_lock, flags); | |
974 | ret = TestSetPageWriteback(page); | |
975 | if (!ret) | |
976 | radix_tree_tag_set(&mapping->page_tree, | |
977 | page_index(page), | |
978 | PAGECACHE_TAG_WRITEBACK); | |
979 | if (!PageDirty(page)) | |
980 | radix_tree_tag_clear(&mapping->page_tree, | |
981 | page_index(page), | |
982 | PAGECACHE_TAG_DIRTY); | |
983 | write_unlock_irqrestore(&mapping->tree_lock, flags); | |
984 | } else { | |
985 | ret = TestSetPageWriteback(page); | |
986 | } | |
987 | return ret; | |
988 | ||
989 | } | |
990 | EXPORT_SYMBOL(test_set_page_writeback); | |
991 | ||
992 | /* | |
993 | * Return true if any of the pages in the mapping are marged with the | |
994 | * passed tag. | |
995 | */ | |
996 | int mapping_tagged(struct address_space *mapping, int tag) | |
997 | { | |
998 | unsigned long flags; | |
999 | int ret; | |
1000 | ||
1001 | read_lock_irqsave(&mapping->tree_lock, flags); | |
1002 | ret = radix_tree_tagged(&mapping->page_tree, tag); | |
1003 | read_unlock_irqrestore(&mapping->tree_lock, flags); | |
1004 | return ret; | |
1005 | } | |
1006 | EXPORT_SYMBOL(mapping_tagged); |